├── .gitignore ├── .pylintrc ├── LICENSE ├── README.rst ├── pysemseg ├── __init__.py ├── datasets │ ├── __init__.py │ ├── base.py │ ├── camvid │ │ ├── __init__.py │ │ └── camvid.py │ ├── pascal_voc │ │ ├── __init__.py │ │ ├── pascal.py │ │ └── prepare_dataset.py │ └── transformer.py ├── evaluate.py ├── loggers │ ├── __init__.py │ ├── console_logger.py │ ├── tensorboard_logger.py │ └── visdom_logger.py ├── losses │ ├── __init__.py │ ├── binary.py │ └── focal.py ├── lr_schedulers │ ├── __init__.py │ ├── constant.py │ └── poly.py ├── metrics.py ├── models │ ├── __init__.py │ ├── deeplab.py │ ├── densenet.py │ ├── fcn.py │ ├── resnet.py │ └── unet.py ├── train.py ├── transforms │ ├── __init__.py │ ├── convert.py │ ├── loaders.py │ └── transforms.py └── utils.py ├── requirements.txt └── setup.py /.gitignore: -------------------------------------------------------------------------------- 1 | *.swp 2 | *.pyc 3 | pysemseg.egg-info/ 4 | -------------------------------------------------------------------------------- /.pylintrc: -------------------------------------------------------------------------------- 1 | [MASTER] 2 | 3 | # A comma-separated list of package or module names from where C extensions may 4 | # be loaded. Extensions are loading into the active Python interpreter and may 5 | # run arbitrary code. 6 | extension-pkg-whitelist=cv2,scipy.io 7 | 8 | # Add files or directories to the blacklist. They should be base names, not 9 | # paths. 10 | ignore=CVS 11 | 12 | # Add files or directories matching the regex patterns to the blacklist. The 13 | # regex matches against base names, not paths. 14 | ignore-patterns= 15 | 16 | # Python code to execute, usually for sys.path manipulation such as 17 | # pygtk.require(). 18 | #init-hook= 19 | 20 | # Use multiple processes to speed up Pylint. Specifying 0 will auto-detect the 21 | # number of processors available to use. 22 | jobs=1 23 | 24 | # Control the amount of potential inferred values when inferring a single 25 | # object. This can help the performance when dealing with large functions or 26 | # complex, nested conditions. 27 | limit-inference-results=100 28 | 29 | # List of plugins (as comma separated values of python modules names) to load, 30 | # usually to register additional checkers. 31 | load-plugins= 32 | 33 | # Pickle collected data for later comparisons. 34 | persistent=yes 35 | 36 | # Specify a configuration file. 37 | #rcfile= 38 | 39 | # When enabled, pylint would attempt to guess common misconfiguration and emit 40 | # user-friendly hints instead of false-positive error messages. 41 | suggestion-mode=yes 42 | 43 | # Allow loading of arbitrary C extensions. Extensions are imported into the 44 | # active Python interpreter and may run arbitrary code. 45 | unsafe-load-any-extension=no 46 | 47 | 48 | [MESSAGES CONTROL] 49 | 50 | # Only show warnings with the listed confidence levels. Leave empty to show 51 | # all. Valid levels: HIGH, INFERENCE, INFERENCE_FAILURE, UNDEFINED. 52 | confidence= 53 | 54 | # Disable the message, report, category or checker with the given id(s). You 55 | # can either give multiple identifiers separated by comma (,) or put this 56 | # option multiple times (only on the command line, not in the configuration 57 | # file where it should appear only once). You can also use "--disable=all" to 58 | # disable everything first and then reenable specific checks. For example, if 59 | # you want to run only the similarities checker, you can use "--disable=all 60 | # --enable=similarities". If you want to run only the classes checker, but have 61 | # no Warning level messages displayed, use "--disable=all --enable=classes 62 | # --disable=W". 63 | disable=print-statement, 64 | parameter-unpacking, 65 | unpacking-in-except, 66 | old-raise-syntax, 67 | backtick, 68 | long-suffix, 69 | old-ne-operator, 70 | old-octal-literal, 71 | import-star-module-level, 72 | non-ascii-bytes-literal, 73 | raw-checker-failed, 74 | bad-inline-option, 75 | locally-disabled, 76 | locally-enabled, 77 | file-ignored, 78 | suppressed-message, 79 | useless-suppression, 80 | deprecated-pragma, 81 | use-symbolic-message-instead, 82 | apply-builtin, 83 | basestring-builtin, 84 | buffer-builtin, 85 | cmp-builtin, 86 | coerce-builtin, 87 | execfile-builtin, 88 | file-builtin, 89 | long-builtin, 90 | raw_input-builtin, 91 | reduce-builtin, 92 | standarderror-builtin, 93 | unicode-builtin, 94 | xrange-builtin, 95 | coerce-method, 96 | delslice-method, 97 | getslice-method, 98 | setslice-method, 99 | no-absolute-import, 100 | old-division, 101 | dict-iter-method, 102 | dict-view-method, 103 | next-method-called, 104 | metaclass-assignment, 105 | indexing-exception, 106 | raising-string, 107 | reload-builtin, 108 | oct-method, 109 | hex-method, 110 | nonzero-method, 111 | cmp-method, 112 | input-builtin, 113 | round-builtin, 114 | intern-builtin, 115 | unichr-builtin, 116 | map-builtin-not-iterating, 117 | zip-builtin-not-iterating, 118 | range-builtin-not-iterating, 119 | filter-builtin-not-iterating, 120 | using-cmp-argument, 121 | eq-without-hash, 122 | div-method, 123 | idiv-method, 124 | rdiv-method, 125 | exception-message-attribute, 126 | invalid-str-codec, 127 | sys-max-int, 128 | bad-python3-import, 129 | deprecated-string-function, 130 | deprecated-str-translate-call, 131 | deprecated-itertools-function, 132 | deprecated-types-field, 133 | next-method-defined, 134 | dict-items-not-iterating, 135 | dict-keys-not-iterating, 136 | dict-values-not-iterating, 137 | deprecated-operator-function, 138 | deprecated-urllib-function, 139 | xreadlines-attribute, 140 | deprecated-sys-function, 141 | exception-escape, 142 | comprehension-escape, 143 | C0111, # missing docstring 144 | C0414, # Import alias does not rename original package (useless-import-alias) 145 | too-few-public-methods, 146 | 147 | # Enable the message, report, category or checker with the given id(s). You can 148 | # either give multiple identifier separated by comma (,) or put this option 149 | # multiple time (only on the command line, not in the configuration file where 150 | # it should appear only once). See also the "--disable" option for examples. 151 | enable=c-extension-no-member 152 | 153 | 154 | [REPORTS] 155 | 156 | # Python expression which should return a note less than 10 (10 is the highest 157 | # note). You have access to the variables errors warning, statement which 158 | # respectively contain the number of errors / warnings messages and the total 159 | # number of statements analyzed. This is used by the global evaluation report 160 | # (RP0004). 161 | evaluation=10.0 - ((float(5 * error + warning + refactor + convention) / statement) * 10) 162 | 163 | # Template used to display messages. This is a python new-style format string 164 | # used to format the message information. See doc for all details. 165 | #msg-template= 166 | 167 | # Set the output format. Available formats are text, parseable, colorized, json 168 | # and msvs (visual studio). You can also give a reporter class, e.g. 169 | # mypackage.mymodule.MyReporterClass. 170 | output-format=text 171 | 172 | # Tells whether to display a full report or only the messages. 173 | reports=no 174 | 175 | # Activate the evaluation score. 176 | score=yes 177 | 178 | 179 | [REFACTORING] 180 | 181 | # Maximum number of nested blocks for function / method body 182 | max-nested-blocks=5 183 | 184 | # Complete name of functions that never returns. When checking for 185 | # inconsistent-return-statements if a never returning function is called then 186 | # it will be considered as an explicit return statement and no message will be 187 | # printed. 188 | never-returning-functions=sys.exit 189 | 190 | 191 | [FORMAT] 192 | 193 | # Expected format of line ending, e.g. empty (any line ending), LF or CRLF. 194 | expected-line-ending-format= 195 | 196 | # Regexp for a line that is allowed to be longer than the limit. 197 | ignore-long-lines=^\s*(# )??$ 198 | 199 | # Number of spaces of indent required inside a hanging or continued line. 200 | indent-after-paren=4 201 | 202 | # String used as indentation unit. This is usually " " (4 spaces) or "\t" (1 203 | # tab). 204 | indent-string=' ' 205 | 206 | # Maximum number of characters on a single line. 207 | max-line-length=100 208 | 209 | # Maximum number of lines in a module. 210 | max-module-lines=1000 211 | 212 | # List of optional constructs for which whitespace checking is disabled. `dict- 213 | # separator` is used to allow tabulation in dicts, etc.: {1 : 1,\n222: 2}. 214 | # `trailing-comma` allows a space between comma and closing bracket: (a, ). 215 | # `empty-line` allows space-only lines. 216 | no-space-check=trailing-comma, 217 | dict-separator 218 | 219 | # Allow the body of a class to be on the same line as the declaration if body 220 | # contains single statement. 221 | single-line-class-stmt=no 222 | 223 | # Allow the body of an if to be on the same line as the test if there is no 224 | # else. 225 | single-line-if-stmt=no 226 | 227 | 228 | [TYPECHECK] 229 | 230 | # List of decorators that produce context managers, such as 231 | # contextlib.contextmanager. Add to this list to register other decorators that 232 | # produce valid context managers. 233 | contextmanager-decorators=contextlib.contextmanager 234 | 235 | # List of members which are set dynamically and missed by pylint inference 236 | # system, and so shouldn't trigger E1101 when accessed. Python regular 237 | # expressions are accepted. 238 | generated-members= 239 | 240 | # Tells whether missing members accessed in mixin class should be ignored. A 241 | # mixin class is detected if its name ends with "mixin" (case insensitive). 242 | ignore-mixin-members=yes 243 | 244 | # Tells whether to warn about missing members when the owner of the attribute 245 | # is inferred to be None. 246 | ignore-none=yes 247 | 248 | # This flag controls whether pylint should warn about no-member and similar 249 | # checks whenever an opaque object is returned when inferring. The inference 250 | # can return multiple potential results while evaluating a Python object, but 251 | # some branches might not be evaluated, which results in partial inference. In 252 | # that case, it might be useful to still emit no-member and other checks for 253 | # the rest of the inferred objects. 254 | ignore-on-opaque-inference=yes 255 | 256 | # List of class names for which member attributes should not be checked (useful 257 | # for classes with dynamically set attributes). This supports the use of 258 | # qualified names. 259 | ignored-classes=optparse.Values,thread._local,_thread._local 260 | 261 | # List of module names for which member attributes should not be checked 262 | # (useful for modules/projects where namespaces are manipulated during runtime 263 | # and thus existing member attributes cannot be deduced by static analysis. It 264 | # supports qualified module names, as well as Unix pattern matching. 265 | ignored-modules= 266 | 267 | # Show a hint with possible names when a member name was not found. The aspect 268 | # of finding the hint is based on edit distance. 269 | missing-member-hint=yes 270 | 271 | # The minimum edit distance a name should have in order to be considered a 272 | # similar match for a missing member name. 273 | missing-member-hint-distance=1 274 | 275 | # The total number of similar names that should be taken in consideration when 276 | # showing a hint for a missing member. 277 | missing-member-max-choices=1 278 | 279 | 280 | [BASIC] 281 | 282 | # Naming style matching correct argument names. 283 | argument-naming-style=snake_case 284 | 285 | # Regular expression matching correct argument names. Overrides argument- 286 | # naming-style. 287 | #argument-rgx= 288 | 289 | # Naming style matching correct attribute names. 290 | attr-naming-style=snake_case 291 | 292 | # Regular expression matching correct attribute names. Overrides attr-naming- 293 | # style. 294 | #attr-rgx= 295 | 296 | # Bad variable names which should always be refused, separated by a comma. 297 | bad-names=foo, 298 | bar, 299 | baz, 300 | toto, 301 | tutu, 302 | tata 303 | 304 | # Naming style matching correct class attribute names. 305 | class-attribute-naming-style=any 306 | 307 | # Regular expression matching correct class attribute names. Overrides class- 308 | # attribute-naming-style. 309 | #class-attribute-rgx= 310 | 311 | # Naming style matching correct class names. 312 | class-naming-style=PascalCase 313 | 314 | # Regular expression matching correct class names. Overrides class-naming- 315 | # style. 316 | #class-rgx= 317 | 318 | # Naming style matching correct constant names. 319 | const-naming-style=UPPER_CASE 320 | 321 | # Regular expression matching correct constant names. Overrides const-naming- 322 | # style. 323 | #const-rgx= 324 | 325 | # Minimum line length for functions/classes that require docstrings, shorter 326 | # ones are exempt. 327 | docstring-min-length=-1 328 | 329 | # Naming style matching correct function names. 330 | function-naming-style=snake_case 331 | 332 | # Regular expression matching correct function names. Overrides function- 333 | # naming-style. 334 | #function-rgx= 335 | 336 | # Good variable names which should always be accepted, separated by a comma. 337 | good-names=i, 338 | j, 339 | k, 340 | ex, 341 | Run, 342 | _ 343 | 344 | # Include a hint for the correct naming format with invalid-name. 345 | include-naming-hint=no 346 | 347 | # Naming style matching correct inline iteration names. 348 | inlinevar-naming-style=any 349 | 350 | # Regular expression matching correct inline iteration names. Overrides 351 | # inlinevar-naming-style. 352 | #inlinevar-rgx= 353 | 354 | # Naming style matching correct method names. 355 | method-naming-style=snake_case 356 | 357 | # Regular expression matching correct method names. Overrides method-naming- 358 | # style. 359 | #method-rgx= 360 | 361 | # Naming style matching correct module names. 362 | module-naming-style=snake_case 363 | 364 | # Regular expression matching correct module names. Overrides module-naming- 365 | # style. 366 | #module-rgx= 367 | 368 | # Colon-delimited sets of names that determine each other's naming style when 369 | # the name regexes allow several styles. 370 | name-group= 371 | 372 | # Regular expression which should only match function or class names that do 373 | # not require a docstring. 374 | no-docstring-rgx=^_ 375 | 376 | # List of decorators that produce properties, such as abc.abstractproperty. Add 377 | # to this list to register other decorators that produce valid properties. 378 | # These decorators are taken in consideration only for invalid-name. 379 | property-classes=abc.abstractproperty 380 | 381 | # Naming style matching correct variable names. 382 | variable-naming-style=snake_case 383 | 384 | # Regular expression matching correct variable names. Overrides variable- 385 | # naming-style. 386 | #variable-rgx= 387 | 388 | 389 | [VARIABLES] 390 | 391 | # List of additional names supposed to be defined in builtins. Remember that 392 | # you should avoid to define new builtins when possible. 393 | additional-builtins= 394 | 395 | # Tells whether unused global variables should be treated as a violation. 396 | allow-global-unused-variables=yes 397 | 398 | # List of strings which can identify a callback function by name. A callback 399 | # name must start or end with one of those strings. 400 | callbacks=cb_, 401 | _cb 402 | 403 | # A regular expression matching the name of dummy variables (i.e. expected to 404 | # not be used). 405 | dummy-variables-rgx=_+$|(_[a-zA-Z0-9_]*[a-zA-Z0-9]+?$)|dummy|^ignored_|^unused_ 406 | 407 | # Argument names that match this expression will be ignored. Default to name 408 | # with leading underscore. 409 | ignored-argument-names=_.*|^ignored_|^unused_ 410 | 411 | # Tells whether we should check for unused import in __init__ files. 412 | init-import=no 413 | 414 | # List of qualified module names which can have objects that can redefine 415 | # builtins. 416 | redefining-builtins-modules=six.moves,past.builtins,future.builtins,builtins,io 417 | 418 | 419 | [SPELLING] 420 | 421 | # Limits count of emitted suggestions for spelling mistakes. 422 | max-spelling-suggestions=4 423 | 424 | # Spelling dictionary name. Available dictionaries: none. To make it working 425 | # install python-enchant package.. 426 | spelling-dict= 427 | 428 | # List of comma separated words that should not be checked. 429 | spelling-ignore-words= 430 | 431 | # A path to a file that contains private dictionary; one word per line. 432 | spelling-private-dict-file= 433 | 434 | # Tells whether to store unknown words to indicated private dictionary in 435 | # --spelling-private-dict-file option instead of raising a message. 436 | spelling-store-unknown-words=no 437 | 438 | 439 | [LOGGING] 440 | 441 | # Logging modules to check that the string format arguments are in logging 442 | # function parameter format. 443 | logging-modules=logging 444 | 445 | 446 | [SIMILARITIES] 447 | 448 | # Ignore comments when computing similarities. 449 | ignore-comments=yes 450 | 451 | # Ignore docstrings when computing similarities. 452 | ignore-docstrings=yes 453 | 454 | # Ignore imports when computing similarities. 455 | ignore-imports=no 456 | 457 | # Minimum lines number of a similarity. 458 | min-similarity-lines=4 459 | 460 | 461 | [MISCELLANEOUS] 462 | 463 | # List of note tags to take in consideration, separated by a comma. 464 | notes=FIXME, 465 | XXX, 466 | TODO 467 | 468 | 469 | [IMPORTS] 470 | 471 | # Allow wildcard imports from modules that define __all__. 472 | allow-wildcard-with-all=no 473 | 474 | # Analyse import fallback blocks. This can be used to support both Python 2 and 475 | # 3 compatible code, which means that the block might have code that exists 476 | # only in one or another interpreter, leading to false positives when analysed. 477 | analyse-fallback-blocks=no 478 | 479 | # Deprecated modules which should not be used, separated by a comma. 480 | deprecated-modules=optparse,tkinter.tix 481 | 482 | # Create a graph of external dependencies in the given file (report RP0402 must 483 | # not be disabled). 484 | ext-import-graph= 485 | 486 | # Create a graph of every (i.e. internal and external) dependencies in the 487 | # given file (report RP0402 must not be disabled). 488 | import-graph= 489 | 490 | # Create a graph of internal dependencies in the given file (report RP0402 must 491 | # not be disabled). 492 | int-import-graph= 493 | 494 | # Force import order to recognize a module as part of the standard 495 | # compatibility libraries. 496 | known-standard-library= 497 | 498 | # Force import order to recognize a module as part of a third party library. 499 | known-third-party=enchant 500 | 501 | 502 | [DESIGN] 503 | 504 | # Maximum number of arguments for function / method. 505 | max-args=5 506 | 507 | # Maximum number of attributes for a class (see R0902). 508 | max-attributes=7 509 | 510 | # Maximum number of boolean expressions in an if statement. 511 | max-bool-expr=5 512 | 513 | # Maximum number of branch for function / method body. 514 | max-branches=12 515 | 516 | # Maximum number of locals for function / method body. 517 | max-locals=15 518 | 519 | # Maximum number of parents for a class (see R0901). 520 | max-parents=7 521 | 522 | # Maximum number of public methods for a class (see R0904). 523 | max-public-methods=20 524 | 525 | # Maximum number of return / yield for function / method body. 526 | max-returns=6 527 | 528 | # Maximum number of statements in function / method body. 529 | max-statements=50 530 | 531 | # Minimum number of public methods for a class (see R0903). 532 | min-public-methods=2 533 | 534 | 535 | [CLASSES] 536 | 537 | # List of method names used to declare (i.e. assign) instance attributes. 538 | defining-attr-methods=__init__, 539 | __new__, 540 | setUp 541 | 542 | # List of member names, which should be excluded from the protected access 543 | # warning. 544 | exclude-protected=_asdict, 545 | _fields, 546 | _replace, 547 | _source, 548 | _make 549 | 550 | # List of valid names for the first argument in a class method. 551 | valid-classmethod-first-arg=cls 552 | 553 | # List of valid names for the first argument in a metaclass class method. 554 | valid-metaclass-classmethod-first-arg=cls 555 | 556 | 557 | [EXCEPTIONS] 558 | 559 | # Exceptions that will emit a warning when being caught. Defaults to 560 | # "Exception". 561 | overgeneral-exceptions=Exception 562 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.rst: -------------------------------------------------------------------------------- 1 | PySemSeg 2 | ======== 3 | 4 | PySemSeg is a library for training Deep Learning Models for Semantic Segmentation in Pytorch. 5 | The goal of the library is to provide implementations of SOTA segmentation models, with pretrained versions 6 | on popular datasets, as well as an easy-to-use training loop for new models and datasets. Most Semantic Segmentation datasets 7 | with fine-grained annotations are small, so Transfer Learning is crucial for success and is a core capability of the library. PySemSeg can use visdom or tensorboardX for training summary visualialization. 8 | 9 | 10 | Installation 11 | ============= 12 | 13 | Using pip: 14 | 15 | .. code:: bash 16 | 17 | pip install git+https://github.com/petko-nikolov/pysemseg 18 | 19 | 20 | Models 21 | ====== 22 | 23 | - FCN [`paper `_] - FCN32, FCN16, FCN8 with pre-trained VGG16 24 | - UNet [`paper `_] 25 | - Tiramisu (FC DenseNets)[`paper `_] - FC DenseNet 56, FC DenseNet 67, FC DensetNet 103 with efficient checkpointing 26 | - DeepLab V3 [`paper `_] - Multi-grid, ASPP and BatchNorm fine-tuning with pre-trained resnets backbone 27 | - DeepLab V3+ [`paper `_] 28 | - RefineNet [`paper `_] - [Upcoming ...] 29 | - PSPNet [`paper `_] - [Upcoming ...] 30 | 31 | 32 | Datasets 33 | ======== 34 | - `Pascal VOC `_ 35 | - `CamVid `_ 36 | - Cityscapes [Upcoming ...] 37 | - ADE20K [Upcoming ...] 38 | 39 | 40 | Train a model from command line 41 | =============================== 42 | 43 | The following is an example command to train a VGGFCN8 model on the Pascal VOC 2012 dataset. In addition to the dataset and the model, a transformer class should be passed (PascalVOCTransform in this case) - a callable where all input image and mask augmentations and tensor transforms are implemented. Run :code:`pysemseg-train -h` for a full list of options. 44 | 45 | .. code:: bash 46 | 47 | pysemseg-train \ 48 | --model VGGFCN8 \ 49 | --model-dir ~/models/vgg8_pascal_model/ \ 50 | --dataset PascalVOCSegmentation \ 51 | --data-dir ~/datasets/PascalVOC/ \ 52 | --batch-size 4 \ 53 | --test-batch-size 1 \ 54 | --epochs 40 \ 55 | --lr 0.001 \ 56 | -- optimizer SGD \ 57 | -- optimizer-args '{"weight_decay": 0.0005, "momentum": 0.9}' \ 58 | --transformer PascalVOCTransform \ 59 | --lr-scheduler PolyLR \ 60 | --lr-scheduler_args '{"max_epochs": 40, "gamma": 0.8}' 61 | 62 | 63 | or pass a YAML config 64 | 65 | 66 | 67 | .. code:: bash 68 | 69 | pysemseg-train --config config.yaml 70 | 71 | 72 | .. code:: YAML 73 | 74 | model: VGGFCN32 75 | model-dir: models/vgg8_pascal_model/ 76 | dataset: PascalVOCSegmentation 77 | data-dir: datasets/PascalVOC/ 78 | batch-size: 4 79 | test-batch-size: 1 80 | epochs: 40 81 | lr: 0.001 82 | optimizer: SGD 83 | optimizer-args: 84 | weight_decay: 0.0005 85 | momentum: 0.9 86 | transformer: PascalVOCTransform 87 | no-cuda: true 88 | lr-scheduler: PolyLR 89 | lr-scheduler-args: 90 | max_epochs: 40 91 | gamma: 0.8 92 | 93 | Load and predict with a trained model 94 | ===================================== 95 | 96 | To use a checkpoint for inference you have to call :code:`load_model` with a checkpoint, the model class and the transformer class used during training. 97 | 98 | .. code:: python 99 | 100 | import torch.nn.functional as F 101 | from pysemseg.transforms import CV2ImageLoader 102 | from pysemseg.utils import load_model 103 | from pysemseg.models import VGGFCN32 104 | from pysemseg.datasets import PascalVOCTransform 105 | 106 | model = load_model( 107 | './checkpoint_path', 108 | VGGFCN32, 109 | PascalVOCTransform 110 | ) 111 | 112 | image = CV2ImageLoader()('./image_path') 113 | logits = model(image) 114 | probabilities = F.softmax(logits, dim=1) 115 | predictions = torch.argmax(logits, dim=1) 116 | -------------------------------------------------------------------------------- /pysemseg/__init__.py: -------------------------------------------------------------------------------- 1 | import pysemseg.datasets 2 | import pysemseg.models 3 | import pysemseg.lr_schedulers 4 | import pysemseg.transforms 5 | import pysemseg.loggers 6 | import pysemseg.losses 7 | -------------------------------------------------------------------------------- /pysemseg/datasets/__init__.py: -------------------------------------------------------------------------------- 1 | from .base import SegmentationDataset 2 | from .pascal_voc.pascal import PascalVOCSegmentation, PascalVOCTransform 3 | from .camvid import CamVid, CamVidTransform 4 | from .transformer import DatasetTransformer 5 | 6 | 7 | def create_dataset(data_dir, dataset_cls, dataset_args, 8 | transformer_cls, transformer_args, mode): 9 | dataset = dataset_cls(data_dir, mode, **dataset_args) 10 | transformer = transformer_cls(mode=mode, **transformer_args) 11 | return DatasetTransformer(dataset, transformer, mode) 12 | -------------------------------------------------------------------------------- /pysemseg/datasets/base.py: -------------------------------------------------------------------------------- 1 | from torch.utils.data import Dataset 2 | from abc import ABCMeta, abstractmethod 3 | from pysemseg.utils import ColorPalette 4 | 5 | 6 | class SegmentationDataset(Dataset, metaclass=ABCMeta): 7 | def __init__(self): 8 | self.color_palette_ = ColorPalette(256) 9 | 10 | @property 11 | @abstractmethod 12 | def number_of_classes(self): 13 | pass 14 | 15 | @property 16 | def labels(self): 17 | return None 18 | 19 | @property 20 | def color_palette(self): 21 | return self.color_palette_ 22 | 23 | @property 24 | def ignore_index(self): 25 | return -1 26 | 27 | @property 28 | def in_channels(self): 29 | return 3 30 | -------------------------------------------------------------------------------- /pysemseg/datasets/camvid/__init__.py: -------------------------------------------------------------------------------- 1 | from .camvid import CamVid, CamVidTransform 2 | -------------------------------------------------------------------------------- /pysemseg/datasets/camvid/camvid.py: -------------------------------------------------------------------------------- 1 | import os 2 | import glob 3 | from torchvision.transforms import Normalize 4 | import cv2 5 | 6 | from pysemseg import transforms 7 | from pysemseg.datasets.base import SegmentationDataset 8 | from pysemseg.utils import ColorPalette 9 | 10 | 11 | CAMVID_CLASSES = [ 12 | "Sky", 13 | "Building", 14 | "Column_pole", 15 | "Road", 16 | "Sidewalk", 17 | "Tree", 18 | "SignSymbol", 19 | "Fence", 20 | "Car", 21 | "Pedestrian", 22 | "Bicyclist" 23 | ] 24 | 25 | CAMVID_COLORS = [ 26 | (128, 128, 128), 27 | (128, 0, 0), 28 | (192, 192, 128), 29 | (128, 64, 128), 30 | (0, 0, 192), 31 | (128, 128, 0), 32 | (192, 128, 128), 33 | (64, 64, 128), 34 | (64, 0, 128), 35 | (64, 64, 0), 36 | (0, 128, 192), 37 | (0, 0, 0) 38 | ] 39 | 40 | 41 | def _parse_image_paths(images_dir, annotations_dir): 42 | image_data = [] 43 | for image_filepath in glob.glob(images_dir + '/*.png'): 44 | image_filename = os.path.basename(image_filepath) 45 | annotation_filepath = os.path.join(annotations_dir, image_filename) 46 | assert os.path.exists(annotation_filepath) 47 | image_data.append({ 48 | 'id': image_filename, 49 | 'image_filepath': image_filepath, 50 | 'gt_filepath': ( 51 | annotation_filepath 52 | if os.path.exists(annotation_filepath) else None) 53 | }) 54 | return image_data 55 | 56 | 57 | class CamVid(SegmentationDataset): 58 | def __init__(self, root_dir, split): 59 | super().__init__() 60 | assert split in ['train', 'val', 'test'] 61 | self.color_palette_ = ColorPalette(CAMVID_COLORS) 62 | self.image_loader = transforms.CV2ImageLoader() 63 | self.target_loader = transforms.CV2ImageLoader(grayscale=True) 64 | self.root_dir = root_dir 65 | self.split = split 66 | self.image_data = _parse_image_paths( 67 | os.path.join(self.root_dir, split), 68 | os.path.join(self.root_dir, split + 'annot') 69 | ) 70 | 71 | @property 72 | def number_of_classes(self): 73 | return 12 74 | 75 | @property 76 | def labels(self): 77 | return CAMVID_CLASSES 78 | 79 | @property 80 | def ignore_index(self): 81 | return 11 82 | 83 | def __getitem__(self, index): 84 | item = self.image_data[index] 85 | return ( 86 | item['id'], 87 | self.image_loader(item['image_filepath']), 88 | self.target_loader(item['gt_filepath']) 89 | ) 90 | 91 | def __len__(self): 92 | return len(self.image_data) 93 | 94 | 95 | 96 | class CamVidTransform: 97 | def __init__(self, mode): 98 | self.mode = mode 99 | self.image_loader = transforms.Compose([ 100 | transforms.ToFloatImage() 101 | ]) 102 | 103 | self.image_augmentations = transforms.Compose([ 104 | transforms.RandomHueSaturation( 105 | hue_delta=0.05, saturation_scale_range=(0.7, 1.3)), 106 | transforms.RandomContrast(0.5, 1.5), 107 | transforms.RandomBrightness(-32.0 / 255, 32. / 255) 108 | ]) 109 | 110 | self.joint_augmentations = transforms.Compose([ 111 | transforms.RandomCropFixedSize((224, 224)), 112 | transforms.RandomHorizontalFlip() 113 | ]) 114 | 115 | self.tensor_transforms = transforms.Compose([ 116 | transforms.ToTensor(), 117 | Normalize( 118 | mean=[0.39068785, 0.40521392, 0.41434407], 119 | std=[0.29652068, 0.30514979, 0.30080369]) 120 | ]) 121 | 122 | def __call__(self, image, target): 123 | image = self.image_loader(image) 124 | if self.mode == 'train': 125 | image, target = self.joint_augmentations(image, target) 126 | image = self.image_augmentations(image) 127 | image = self.tensor_transforms(image) 128 | target = transforms.ToCategoryTensor()(target) 129 | return image, target 130 | -------------------------------------------------------------------------------- /pysemseg/datasets/pascal_voc/__init__.py: -------------------------------------------------------------------------------- 1 | from .pascal import PascalVOCSegmentation, PascalVOCTransform 2 | -------------------------------------------------------------------------------- /pysemseg/datasets/pascal_voc/pascal.py: -------------------------------------------------------------------------------- 1 | import os 2 | from torchvision.transforms import Normalize 3 | import cv2 4 | 5 | from pysemseg.datasets.base import SegmentationDataset 6 | from pysemseg import transforms 7 | 8 | 9 | PASCAL_CLASSES = [ 10 | 'background', 11 | 'aeroplane', 12 | 'bicycle', 13 | 'bird', 14 | 'boat', 15 | 'bottle', 16 | 'bus', 17 | 'car', 18 | 'cat', 19 | 'chair', 20 | 'cow', 21 | 'diningtable', 22 | 'dog', 23 | 'horse', 24 | 'motorbike', 25 | 'person', 26 | 'potted-plant', 27 | 'sheep', 28 | 'sofa', 29 | 'train', 30 | 'tv/monitor' 31 | ] 32 | 33 | 34 | def _read_image_ids(split_filepath): 35 | with open(split_filepath, 'r') as split_file: 36 | return [s.strip() for s in split_file] 37 | 38 | 39 | def _parse_image_paths(image_dir, ground_truth_dir, image_ids): 40 | image_data = [] 41 | 42 | for image_id in image_ids: 43 | img_path = os.path.join(image_dir, image_id + '.jpg') 44 | mask_path = os.path.join(ground_truth_dir, image_id + '.png') 45 | if os.path.exists(img_path): 46 | image_data.append({ 47 | 'id': image_id, 48 | 'image_filepath': img_path, 49 | 'gt_filepath': ( 50 | mask_path if os.path.exists(mask_path) else None) 51 | }) 52 | return image_data 53 | 54 | 55 | class PascalVOCSegmentation(SegmentationDataset): 56 | def __init__(self, root, split='train'): 57 | super().__init__() 58 | 59 | assert split in ['train', 'test', 'val'] 60 | self.root = os.path.expanduser(root) 61 | self.split = split 62 | self.image_loader = transforms.CV2ImageLoader() 63 | self.target_loader = transforms.CV2ImageLoader(grayscale=True) 64 | 65 | benchmark_train_ids = set( 66 | _read_image_ids( 67 | os.path.join(root, 'benchmark_RELEASE/dataset/train.txt') 68 | ) 69 | ) 70 | 71 | benchmark_val_ids = set( 72 | _read_image_ids( 73 | os.path.join(root, 'benchmark_RELEASE/dataset/val.txt') 74 | ) 75 | ) 76 | 77 | voc2012_train_ids = set( 78 | _read_image_ids( 79 | os.path.join( 80 | root, 81 | 'VOCdevkit/VOC2012/ImageSets/Segmentation/train.txt') 82 | ) 83 | ) 84 | 85 | voc2012_val_ids = set( 86 | _read_image_ids( 87 | os.path.join( 88 | root, 89 | 'VOCdevkit/VOC2012/ImageSets/Segmentation/val.txt') 90 | ) 91 | ) 92 | 93 | self.train_image_data = _parse_image_paths( 94 | os.path.join(root, 'VOCdevkit/VOC2012/JPEGImages'), 95 | os.path.join(root, 'VOCdevkit/VOC2012/SegmentationClassLabels'), 96 | voc2012_train_ids 97 | ) 98 | 99 | self.train_image_data.extend(_parse_image_paths( 100 | os.path.join(root, 'benchmark_RELEASE/dataset/img'), 101 | os.path.join(root, 'benchmark_RELEASE/dataset/cls_labels'), 102 | benchmark_train_ids | benchmark_val_ids 103 | )) 104 | 105 | self.train_ids = ( 106 | voc2012_train_ids | benchmark_val_ids | benchmark_train_ids 107 | ) 108 | 109 | self.val_ids = voc2012_val_ids - self.train_ids 110 | 111 | self.val_image_data = _parse_image_paths( 112 | os.path.join(root, 'VOCdevkit/VOC2012/JPEGImages'), 113 | os.path.join(root, 'VOCdevkit/VOC2012/SegmentationClassLabels'), 114 | self.val_ids 115 | ) 116 | 117 | self.image_data = { 118 | 'train': self.train_image_data, 119 | 'val': self.val_image_data 120 | }[self.split] 121 | 122 | @property 123 | def number_of_classes(self): 124 | return 21 125 | 126 | @property 127 | def labels(self): 128 | return PASCAL_CLASSES 129 | 130 | @property 131 | def ignore_index(self): 132 | return 255 133 | 134 | def __getitem__(self, index): 135 | item = self.image_data[index] 136 | return ( 137 | item['id'], 138 | self.image_loader(item['image_filepath']), 139 | self.target_loader(item['gt_filepath']) 140 | ) 141 | 142 | def __len__(self): 143 | return len(self.image_data) 144 | 145 | 146 | class PascalVOCTransform: 147 | def __init__(self, mode): 148 | self.ignore_index = 255 149 | self.mode = mode 150 | self.image_loader = transforms.Compose([ 151 | transforms.ToFloatImage() 152 | ]) 153 | 154 | self.image_augmentations = transforms.Compose([ 155 | transforms.RandomHueSaturation( 156 | hue_delta=0.05, saturation_scale_range=(0.7, 1.3)), 157 | transforms.RandomContrast(0.8, 1.2), 158 | transforms.RandomBrightness(-32.0 / 255, 32. / 255) 159 | ]) 160 | 161 | self.joint_augmentations = transforms.Compose([ 162 | transforms.RandomCrop(), 163 | transforms.ScaleTo((513, 513)), 164 | transforms.Concat([ 165 | transforms.PadTo((513, 513)), 166 | transforms.PadTo((513, 513), 255) 167 | ]), 168 | transforms.RandomHorizontalFlip()] 169 | ) 170 | 171 | self.tensor_transforms = transforms.Compose([ 172 | transforms.ToTensor(), 173 | Normalize( 174 | mean=[0.485, 0.456, 0.406], 175 | std=[0.229, 0.224, 0.225]) 176 | ]) 177 | 178 | def __call__(self, image, target): 179 | image = self.image_loader(image) 180 | if self.mode == 'train': 181 | image, target = self.joint_augmentations(image, target) 182 | image = self.image_augmentations(image) 183 | image = self.tensor_transforms(image) 184 | target = transforms.ToCategoryTensor()(target) 185 | return image, target 186 | -------------------------------------------------------------------------------- /pysemseg/datasets/pascal_voc/prepare_dataset.py: -------------------------------------------------------------------------------- 1 | import os 2 | import argparse 3 | import glob 4 | import shutil 5 | from multiprocessing import Pool, cpu_count 6 | import scipy.io 7 | import cv2 8 | from tqdm import tqdm 9 | import numpy as np 10 | from transforms.loaders import CV2ImageLoader 11 | 12 | from utils import prompt_delete_dir, ColorPalette 13 | 14 | 15 | OUTPUT_DIR = None 16 | color_palette = ColorPalette(256) 17 | 18 | 19 | def process_mask(filename): 20 | basename = os.path.basename(filename) 21 | palette_mask = CV2ImageLoader()(filename) 22 | label_mask = color_palette.decode_color(palette_mask) 23 | cv2.imwrite(os.path.join(OUTPUT_DIR, basename), label_mask) 24 | 25 | 26 | def process_berkley_gt(filename): 27 | basename = os.path.basename(filename) 28 | image_id, _ = basename.split('.') 29 | mat = scipy.io.loadmat( 30 | filename, mat_dtype=True, squeeze_me=True, struct_as_record=False 31 | ) 32 | segmentation = mat['GTcls'].Segmentation 33 | cv2.imwrite( 34 | os.path.join(OUTPUT_DIR, image_id + '.png'), mat['GTcls'].Segmentation 35 | ) 36 | 37 | 38 | def init_process(output_dir): 39 | global OUTPUT_DIR 40 | OUTPUT_DIR = output_dir 41 | 42 | 43 | def convert_voc2012_labels(input_dir, output_dir, ncpus=None, overwrite=False): 44 | if overwrite: 45 | shutil.rmtree(output_dir, ignore_errors=True) 46 | else: 47 | prompt_delete_dir(output_dir) 48 | 49 | os.makedirs(output_dir) 50 | 51 | files = glob.glob(os.path.join(input_dir, '*.png')) 52 | 53 | with Pool(processes=ncpus or cpu_count(), 54 | initializer=init_process, initargs=(output_dir,)) as pool: 55 | with tqdm(total=len(files), desc='Prepare VOC2012') as pbar: 56 | for _ in enumerate(pool.imap_unordered(process_mask, files)): 57 | pbar.update() 58 | 59 | 60 | 61 | def convert_berkley_labels(input_dir, output_dir, ncpus=None, overwrite=False): 62 | if overwrite: 63 | shutil.rmtree(output_dir, ignore_errors=True) 64 | else: 65 | prompt_delete_dir(output_dir) 66 | os.makedirs(output_dir) 67 | 68 | files = glob.glob(os.path.join(input_dir, '*.mat')) 69 | 70 | with Pool(processes=ncpus or cpu_count(), 71 | initializer=init_process, initargs=(output_dir,)) as pool: 72 | with tqdm(total=len(files), desc='Prepare Berkley') as pbar: 73 | for _ in enumerate(pool.imap_unordered( 74 | process_berkley_gt, files)): 75 | pbar.update() 76 | 77 | 78 | def prepare_dataset(rootdir, ncpus=None, overwrite=False): 79 | voc2012_root = os.path.join(rootdir, 'VOCdevkit/VOC2012') 80 | berkley_root = os.path.join(rootdir, 'benchmark_RELEASE') 81 | assert os.path.exists(voc2012_root) 82 | convert_voc2012_labels( 83 | os.path.join(voc2012_root, 'SegmentationClass'), 84 | os.path.join(voc2012_root, 'SegmentationClassLabels'), 85 | ncpus, 86 | overwrite) 87 | convert_berkley_labels( 88 | os.path.join(berkley_root, 'dataset/cls/'), 89 | os.path.join(berkley_root, 'dataset/cls_labels'), 90 | ncpus, 91 | overwrite) 92 | 93 | 94 | def main(): 95 | parser = argparse.ArgumentParser("Convert palette to labels") 96 | parser.add_argument("--dataset_dir", required=True, type=str) 97 | parser.add_argument("--overwrite", action='store_true', default=False, 98 | help="Overwrite existing data.") 99 | parser.add_argument("--ncpus", required=False, type=int) 100 | args = parser.parse_args() 101 | prepare_dataset(args.dataset_dir, args.ncpus, args.overwrite) 102 | 103 | 104 | if __name__ == '__main__': 105 | main() 106 | -------------------------------------------------------------------------------- /pysemseg/datasets/transformer.py: -------------------------------------------------------------------------------- 1 | from torch.utils.data import Dataset 2 | 3 | 4 | class DatasetTransformer(Dataset): 5 | def __init__( 6 | self, dataset, transform, mode='train'): 7 | self.dataset = dataset 8 | self.transform = transform 9 | self.mode = mode 10 | 11 | def __getitem__(self, index): 12 | example_id, image, mask = self.dataset[index] 13 | image, mask = self.transform(image, mask) 14 | if self.mode == 'test': 15 | return example_id, image 16 | return example_id, image, mask 17 | 18 | def __len__(self): 19 | return len(self.dataset) 20 | 21 | @property 22 | def labels(self): 23 | return self.dataset.labels 24 | 25 | @property 26 | def number_of_classes(self): 27 | return self.dataset.number_of_classes 28 | 29 | @property 30 | def color_palette(self): 31 | return self.dataset.color_palette 32 | 33 | @property 34 | def ignore_index(self): 35 | return self.dataset.ignore_index 36 | 37 | @property 38 | def in_channels(self): 39 | return self.dataset.in_channels 40 | -------------------------------------------------------------------------------- /pysemseg/evaluate.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from torch.autograd import Variable 3 | import torch 4 | from pysemseg.metrics import TorchSegmentationMetrics 5 | from pysemseg.utils import tensor_to_numpy, flatten_dict 6 | 7 | 8 | def evaluate( 9 | model, loader, criterion, console_logger, epoch, 10 | visual_logger, device, log_images_interval): 11 | model.eval() 12 | 13 | metrics = TorchSegmentationMetrics( 14 | loader.dataset.number_of_classes, 15 | loader.dataset.labels, 16 | ignore_index=loader.dataset.ignore_index, 17 | device=device 18 | ) 19 | 20 | with torch.no_grad(): 21 | for step, (_, data, target) in enumerate(loader): 22 | data, target = data.to(device), target.to(device) 23 | 24 | data, target = Variable(data), Variable(target) 25 | output = model(data) 26 | loss = criterion(output, target) 27 | 28 | predictions = torch.argmax(output, dim=1) 29 | 30 | loss = loss / torch.sum(target != loader.dataset.ignore_index).float() 31 | 32 | metrics.add(predictions, target, loss) 33 | 34 | if step % log_images_interval == 0: 35 | visual_logger.log_prediction_images( 36 | step, 37 | tensor_to_numpy(data.data), 38 | tensor_to_numpy(target.data), 39 | tensor_to_numpy(predictions), 40 | name='images', 41 | prefix='Validation' 42 | ) 43 | 44 | metrics_dict = metrics.metrics() 45 | 46 | console_logger.log( 47 | len(loader), epoch, loader, data, 48 | metrics_dict, mode='Validation') 49 | 50 | if visual_logger is not None: 51 | visual_logger.log_metrics(epoch, metrics_dict, prefix='Validation') 52 | 53 | return predictions 54 | -------------------------------------------------------------------------------- /pysemseg/loggers/__init__.py: -------------------------------------------------------------------------------- 1 | from .tensorboard_logger import TensorboardLogger 2 | from .visdom_logger import VisdomLogger 3 | from .console_logger import ConsoleLogger 4 | -------------------------------------------------------------------------------- /pysemseg/loggers/console_logger.py: -------------------------------------------------------------------------------- 1 | import json 2 | from pysemseg.utils import flatten_dict 3 | 4 | 5 | class ConsoleLogger(): 6 | 7 | def __init__(self, filename=None, continue_training=False): 8 | self.filename = filename 9 | self.log_file = None 10 | self.continue_training = continue_training 11 | 12 | def __enter__(self): 13 | mode = 'w' 14 | if self.continue_training: 15 | mode += 'a' 16 | self.log_file = open(self.filename, mode) 17 | return self 18 | 19 | def __exit__(self, *args, **kwargs): 20 | self.log_file.close() 21 | 22 | def log(self, index, epoch, loader, data, metrics, mode='Train'): 23 | metric_str = ", ".join([ 24 | '{}:{:.6f}'.format(k, v) for k, v in flatten_dict(metrics).items() 25 | ]) 26 | print('{} Epoch: {} [{}/{} ({:.0f}%)] [{}]'.format( 27 | mode, epoch, index * len(data), len(loader.dataset), 28 | 100. * index / len(loader), metric_str)) 29 | if self.log_file is not None: 30 | log_data = {**metrics, 'epoch': epoch, 'step': index, "mode": mode} 31 | self.log_file.write(json.dumps(log_data) + '\n') 32 | -------------------------------------------------------------------------------- /pysemseg/loggers/tensorboard_logger.py: -------------------------------------------------------------------------------- 1 | from tensorboardX import SummaryWriter 2 | import numpy as np 3 | from pysemseg.utils import flatten_dict, ColorPalette 4 | 5 | 6 | class TensorboardLogger: 7 | def __init__(self, log_directory, color_palette): 8 | self.log_directory = log_directory 9 | self.summary_writer = SummaryWriter(self.log_directory) 10 | self.color_palette = color_palette 11 | 12 | def log_args(self, args_dict): 13 | args_text = "\n".join("{}: {}".format(k, v) for k, v in args_dict.items()) 14 | self.summary_writer.add_text( 15 | 'Args', 16 | text=args_text, 17 | ) 18 | 19 | def log_metrics(self, iteration, metrics, prefix): 20 | for key, value in flatten_dict(metrics).items(): 21 | self.summary_writer.add_scalar( 22 | "{}/{}".format(prefix, key), value, iteration) 23 | 24 | def log_prediction_images(self, iteration, image, gt, prediction, name, prefix): 25 | gt = self.color_palette.encode_color(gt).transpose([0, 3, 1, 2]) 26 | prediction = self.color_palette.encode_color(prediction).transpose([0, 3, 1, 2]) 27 | combined_images = np.concatenate((image, gt, prediction), axis=-1) 28 | for i in range(image.shape[0]): 29 | self.summary_writer.add_image( 30 | "{}/{}/{}".format(prefix, name, i), combined_images[i], iteration) 31 | 32 | def log_learning_rate(self, iteration, lr): 33 | self.summary_writer.add_scalar('learning_rate', value, iteration) 34 | 35 | -------------------------------------------------------------------------------- /pysemseg/loggers/visdom_logger.py: -------------------------------------------------------------------------------- 1 | import os 2 | import numpy as np 3 | import cv2 4 | import visdom 5 | 6 | from pysemseg.utils import flatten_dict 7 | from pysemseg.transforms import ResizeBatch 8 | 9 | 10 | IMAGES_WIDTH = 128 11 | 12 | 13 | class VisdomLogger: 14 | def __init__(self, log_directory, color_palette, continue_logging=False): 15 | self.log_directory = log_directory 16 | self.color_palette = color_palette 17 | self.coninue_logging = continue_logging 18 | visdom_env = os.path.basename(log_directory.rstrip('/')) 19 | self.visdom = visdom.Visdom( 20 | env=visdom_env, 21 | log_to_filename=os.path.join(log_directory, 'viz.log') 22 | ) 23 | if not continue_logging: 24 | self.visdom.delete_env(visdom_env) 25 | 26 | def log_args(self, args_dict): 27 | args_text = "
".join("{}: {}".format(k, v) for k, v in args_dict.items()) 28 | self.visdom.text( 29 | text=args_text, 30 | win='Args', 31 | ) 32 | 33 | def _update_metric_plots(self, iteration, metrics, prefix): 34 | for key, value in flatten_dict(metrics).items(): 35 | name = "{}/{}".format(prefix, key) 36 | self.visdom.line( 37 | np.array([value]), 38 | np.array([iteration]), 39 | win=name, 40 | update='append' if iteration > 0 else None, 41 | opts={'title': name} 42 | ) 43 | 44 | def _log_current_class_metrics(self, metrics, prefix): 45 | class_metric_names = list(next(iter(metrics['class'].values())).keys()) 46 | for metric_name in class_metric_names: 47 | name = '{}/Current/{}'.format(prefix, metric_name) 48 | class_names = list(metrics['class'].keys()) 49 | values = np.array([v[metric_name] for v in metrics['class'].values()]) 50 | self.visdom.bar( 51 | values, 52 | win=name, 53 | opts={'rownames': class_names, 'title': name} 54 | ) 55 | 56 | def log_metrics(self, iteration, metrics, prefix): 57 | self._update_metric_plots(iteration, metrics, prefix) 58 | self._log_current_class_metrics(metrics, prefix) 59 | 60 | def log_prediction_images( 61 | self, iteration, image, ground_truth, prediction, name, prefix): 62 | title = '{}/{}'.format(prefix, name) 63 | ground_truth = self.color_palette.encode_color(ground_truth) 64 | prediction = self.color_palette.encode_color(prediction) 65 | height = int(IMAGES_WIDTH / image.shape[3] * image.shape[2]) 66 | image = ResizeBatch((height, IMAGES_WIDTH))( 67 | image.transpose([0, 2, 3, 1])) 68 | prediction = ResizeBatch( 69 | (height, IMAGES_WIDTH), interpolation=cv2.INTER_NEAREST)( 70 | prediction) 71 | ground_truth = ResizeBatch( 72 | (height, IMAGES_WIDTH), interpolation=cv2.INTER_NEAREST)( 73 | ground_truth) 74 | combined_images = np.concatenate( 75 | [image[:, :, :, :3], ground_truth, prediction], axis=2) 76 | self.visdom.images( 77 | combined_images.transpose([0, 3, 1, 2]), 78 | nrow=1, 79 | win=title, 80 | opts={'caption': title} 81 | ) 82 | 83 | def log_learning_rate(self, iteration, lr): 84 | self.visdom.line( 85 | np.array([lr]), 86 | np.array([iteration]), 87 | win='learning_rate', 88 | update='append' if iteration > 0 else None, 89 | opts={'title': 'Learning rate'} 90 | ) 91 | 92 | -------------------------------------------------------------------------------- /pysemseg/losses/__init__.py: -------------------------------------------------------------------------------- 1 | from .binary import DiceLoss, LogDiceLoss, JaccardLoss, LogJaccardLoss 2 | from .focal import FocalLoss 3 | -------------------------------------------------------------------------------- /pysemseg/losses/binary.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | 6 | def _compute_dice(inputs, targets, ignore_index, pos_index=1): 7 | smooth = 1. 8 | inputs = F.softmax(inputs, dim=1) 9 | foreground_probs = inputs[:, pos_index, :, :] 10 | targets = (targets == pos_index).float() 11 | intersection = (foreground_probs * targets).sum() 12 | denom = (foreground_probs ** 2 + targets ** 2).sum() 13 | return (2 * intersection + smooth)/ (denom + smooth) 14 | 15 | 16 | def _compute_jaccard(inputs, targets, ignore_index, pos_index=1): 17 | smooth = 1. 18 | inputs = F.softmax(inputs, dim=1) 19 | foreground_probs = inputs[:, pos_index, :, :] 20 | targets = (targets == pos_index).float() 21 | intersection = (foreground_probs * targets).sum() 22 | union = (foreground_probs + targets).sum() 23 | return (intersection + smooth) / (union - intersection + smooth) 24 | 25 | 26 | class DiceLoss(nn.Module): 27 | def __init__(self, ignore_index=-1, pos_index=1, rescale=True): 28 | super().__init__() 29 | self.ignore_index = ignore_index 30 | self.rescale = rescale 31 | self.pos_index = pos_index 32 | 33 | def forward(self, inputs, targets): 34 | loss = 1.0 - _compute_dice( 35 | inputs, targets, self.ignore_index, self.pos_index) 36 | if self.rescale: 37 | num_targets = torch.sum(targets != self.ignore_index).float() 38 | loss *= num_targets 39 | return loss 40 | 41 | 42 | class LogDiceLoss(nn.Module): 43 | def __init__(self, ignore_index=-1, pos_index=1, rescale=True): 44 | super().__init__() 45 | self.ignore_index = ignore_index 46 | self.rescale = rescale 47 | self.pos_index = pos_index 48 | 49 | def forward(self, inputs, targets): 50 | loss = -torch.log(_compute_dice( 51 | inputs, targets, self.ignore_index, self.pos_index)) 52 | if self.rescale: 53 | num_targets = torch.sum(targets != self.ignore_index).float() 54 | loss *= num_targets 55 | return loss 56 | 57 | 58 | class JaccardLoss(nn.Module): 59 | def __init__(self, ignore_index=-1, pos_index=1, rescale=True): 60 | super().__init__() 61 | self.ignore_index = ignore_index 62 | self.rescale = rescale 63 | self.pos_index = pos_index 64 | 65 | def forward(self, inputs, targets): 66 | loss = 1.0 - _compute_jaccard( 67 | inputs, targets, self.ignore_index, self.pos_index) 68 | if self.rescale: 69 | num_targets = torch.sum(targets != self.ignore_index).float() 70 | loss *= num_targets 71 | return loss 72 | 73 | 74 | class LogJaccardLoss(nn.Module): 75 | def __init__(self, ignore_index=-1, pos_index=1, rescale=True): 76 | super().__init__() 77 | self.ignore_index = ignore_index 78 | self.rescale = rescale 79 | self.pos_index = pos_index 80 | 81 | def forward(self, inputs, targets): 82 | loss = -torch.log(_compute_jaccard( 83 | inputs, targets, self.ignore_index, self.pos_index)) 84 | if self.rescale: 85 | num_targets = torch.sum(targets != self.ignore_index).float() 86 | loss *= num_targets 87 | return loss 88 | -------------------------------------------------------------------------------- /pysemseg/losses/focal.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | def one_hot_encode(targets, N): 6 | targets = targets.unsqueeze(1) 7 | one_hot = torch.zeros( 8 | (targets.size(0), N, targets.size(2), targets.size(3)), 9 | device=targets.device 10 | ) 11 | one_hot = one_hot.scatter_(1, targets.data, 1) 12 | one_hot = torch.autograd.Variable(one_hot) 13 | return one_hot 14 | 15 | 16 | class FocalLoss(nn.Module): 17 | def __init__(self, gamma=1.0, weights=None, ignore_index=-1): 18 | super().__init__() 19 | self.gamma = gamma 20 | self.weights = weights 21 | self.ignore_index = ignore_index 22 | self.ce = nn.CrossEntropyLoss( 23 | weight=self.weights, ignore_index=ignore_index, reduction='none' 24 | ) 25 | 26 | def forward(self, inputs, targets): 27 | ce_loss = self.ce(inputs, targets) 28 | one_hot = one_hot_encode(targets, inputs.size(1)) 29 | inv_probs = 1.0 - F.softmax(inputs, dim=1) 30 | focal = (inv_probs * one_hot).sum(dim=1) ** self.gamma 31 | mask = (targets != self.ignore_index).float() 32 | return torch.sum(focal * ce_loss * mask) 33 | -------------------------------------------------------------------------------- /pysemseg/lr_schedulers/__init__.py: -------------------------------------------------------------------------------- 1 | from .poly import PolyLR 2 | from .constant import ConstantLR 3 | -------------------------------------------------------------------------------- /pysemseg/lr_schedulers/constant.py: -------------------------------------------------------------------------------- 1 | from torch.optim.lr_scheduler import _LRScheduler 2 | 3 | 4 | class ConstantLR(_LRScheduler): 5 | def __init__(self, optimizer, last_epoch=-1): 6 | super().__init__(optimizer, last_epoch) 7 | 8 | def get_lr(self): 9 | return self.base_lrs 10 | -------------------------------------------------------------------------------- /pysemseg/lr_schedulers/poly.py: -------------------------------------------------------------------------------- 1 | from torch.optim.lr_scheduler import _LRScheduler 2 | 3 | 4 | class PolyLR(_LRScheduler): 5 | def __init__(self, optimizer, max_epochs, gamma=0.9, last_epoch=-1): 6 | self.max_epochs = max_epochs 7 | self.gamma = gamma 8 | super().__init__(optimizer, last_epoch) 9 | 10 | def get_lr(self): 11 | return [ 12 | base_lr * (1.0 - self.last_epoch / self.max_epochs) ** self.gamma 13 | for base_lr in self.base_lrs 14 | ] 15 | -------------------------------------------------------------------------------- /pysemseg/metrics.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | from pysemseg.utils import tensor_to_numpy 4 | 5 | def _apply(fcn, acc, counter): 6 | result = {} 7 | for key, value in acc.items(): 8 | if isinstance(value, dict): 9 | result[key] = _apply(fcn, value, counter.get(key, {})) 10 | else: 11 | result[key] = fcn(acc[key], counter.get(key, 0.0)) 12 | return result 13 | 14 | 15 | class _Accumulator(): 16 | def __init__(self): 17 | self.numerator = {} 18 | self.denominator = {} 19 | 20 | def mean(self): 21 | return _apply( 22 | lambda n, d: n / d if d > 0 else 0.0, 23 | self.numerator, self.denominator) 24 | 25 | def update(self, data): 26 | self.numerator = _apply( 27 | lambda a, d: a[0] + d, data, self.numerator) 28 | self.denominator = _apply( 29 | lambda a, d: a[1] + d, 30 | data, self.denominator) 31 | 32 | 33 | class SegmentationMetrics: 34 | def __init__(self, num_classes, labels=None, ignore_index=-1): 35 | labels = labels or list(range(num_classes)) 36 | self.labels = dict(enumerate(labels)) 37 | self.accumulator = _Accumulator() 38 | self.num_classes = num_classes 39 | self.cm = np.zeros( 40 | (self.num_classes, self.num_classes), dtype=np.int32) 41 | self.ignore_index = ignore_index 42 | 43 | 44 | def metrics(self): 45 | metrics = {} 46 | metrics['accuracy'] = self._accuracy() 47 | metrics['class'] = {} 48 | metrics['class']['iou'] = dict(zip(self.labels, self._iou())) 49 | metrics['mIOU'] = np.mean(list(metrics['class']['iou'].values())) 50 | return metrics 51 | 52 | def _accuracy(self): 53 | return self.cm.diagonal().sum() / self.cm.sum() 54 | 55 | def _iou(self): 56 | colsum = self.cm.sum(axis=0) 57 | rowsum = self.cm.sum(axis=1) 58 | diag = self.cm.diagonal() 59 | return diag / (colsum + rowsum - diag) 60 | 61 | def _confusion_matrix(self, outputs, targets): 62 | outputs = outputs.reshape(-1,) 63 | targets = targets.reshape(-1,) 64 | mask = targets != self.ignore_index 65 | comb = self.num_classes * outputs[mask] + targets[mask] 66 | comb = np.bincount(comb, minlength=self.num_classes ** 2) 67 | return comb.reshape(self.num_classes, self.num_classes) 68 | 69 | 70 | def add(self, outputs, targets, loss): 71 | self.accumulator.update({'loss': (loss, 1)}) 72 | self.cm += self._confusion_matrix(outputs, targets) 73 | 74 | 75 | class TorchSegmentationMetrics: 76 | def __init__(self, num_classes, labels=None, ignore_index=-1, device=None): 77 | labels = labels or list(range(num_classes)) 78 | self.labels = dict(enumerate(labels)) 79 | self.accumulator = _Accumulator() 80 | self.num_classes = num_classes 81 | self.cm = torch.zeros( 82 | (self.num_classes, self.num_classes), dtype=torch.long, 83 | requires_grad=False, device=device) 84 | self.ignore_index = ignore_index 85 | 86 | 87 | def metrics(self): 88 | metrics = {} 89 | metrics['accuracy'] = self._accuracy() 90 | metrics['class'] = { 91 | k: {'iou': v} 92 | for k, v in zip(self.labels, self._iou()) 93 | } 94 | metrics['mIOU'] = np.mean([c['iou'] for c in metrics['class'].values()]) 95 | metrics.update(self.accumulator.mean()) 96 | metrics = _apply( 97 | lambda x, y: float(x), 98 | metrics, 99 | {} 100 | ) 101 | return metrics 102 | 103 | def _accuracy(self): 104 | accuracy = self.cm.diagonal().sum().float() / self.cm.sum().float() 105 | return tensor_to_numpy(accuracy) 106 | 107 | def _iou(self): 108 | colsum = self.cm.sum(dim=0) 109 | rowsum = self.cm.sum(dim=1) 110 | diag = self.cm.diagonal() 111 | iou = diag.float() / (colsum + rowsum - diag).float() 112 | return tensor_to_numpy(iou) 113 | 114 | def _confusion_matrix(self, outputs, targets): 115 | outputs = outputs.view(-1).contiguous() 116 | targets = targets.view(-1).contiguous() 117 | mask = (targets != self.ignore_index) 118 | comb = self.num_classes * outputs[mask] + targets[mask] 119 | comb = torch.bincount(comb, minlength=self.num_classes ** 2) 120 | return comb.reshape(self.num_classes, self.num_classes).long() 121 | 122 | 123 | def add(self, outputs, targets, loss): 124 | self.accumulator.update({'loss': (loss, 1)}) 125 | self.cm += self._confusion_matrix(outputs, targets) 126 | 127 | 128 | def compute_example_segmentation_metrics( 129 | n_classes, outputs, targets, loss, ignore_index=-1): 130 | metrics = TorchSegmentationMetrics(n_classes, ignore_index) 131 | metrics.add(outputs, targets, loss) 132 | return metrics.metrics() 133 | -------------------------------------------------------------------------------- /pysemseg/models/__init__.py: -------------------------------------------------------------------------------- 1 | from .unet import UNet, unet_resnet101, unet_resnet152, unet_big_resnet101 2 | from .fcn import VGGFCN32, VGGFCN16, VGGFCN8 3 | from .densenet import fcdensenet56, fcdensenet67, fcdensenet103 4 | from .deeplab import ( 5 | deeplabv3_resnet50, deeplabv3_resnet101, deeplabv3_resnet152, 6 | deeplabv3plus_resnet101, deeplabv3plus_resnet50 7 | ) 8 | -------------------------------------------------------------------------------- /pysemseg/models/deeplab.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from .resnet import resnet50, resnet101, resnet152 5 | 6 | class ASPPModule(nn.Module): 7 | def __init__(self, in_channels, out_channels, rates): 8 | super().__init__() 9 | assert len(rates) == 3 10 | self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=1, 11 | bias=False) 12 | self.conv2 = nn.Conv2d( 13 | in_channels, out_channels, kernel_size=3, dilation=rates[0], 14 | padding=rates[0], bias=False) 15 | self.conv3 = nn.Conv2d( 16 | in_channels, out_channels, kernel_size=3, dilation=rates[1], 17 | padding=rates[1], bias=False) 18 | self.conv4 = nn.Conv2d( 19 | in_channels, out_channels, kernel_size=3, dilation=rates[2], 20 | padding=rates[2], bias=False) 21 | self.global_avg_pool = nn.Sequential( 22 | nn.AdaptiveAvgPool2d((1, 1)), 23 | nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) 24 | ) 25 | self.bn = nn.BatchNorm2d(5 * out_channels) 26 | self.relu = nn.ReLU(inplace=True) 27 | self.final_conv = nn.Conv2d( 28 | 5 * out_channels, out_channels, kernel_size=1, bias=False) 29 | 30 | def forward(self, x): 31 | x1 = self.conv1(x) 32 | x2 = self.conv2(x) 33 | x3 = self.conv3(x) 34 | x4 = self.conv4(x) 35 | x5 = self.global_avg_pool(x) 36 | x5 = F.interpolate( 37 | x5, size=x1.size()[2:], mode='bilinear', align_corners=True 38 | ) 39 | x = torch.cat([x1, x2, x3, x4, x5], dim=1) 40 | x = self.relu(self.bn(x)) 41 | x = self.final_conv(x) 42 | return x 43 | 44 | 45 | class Decoder(nn.Module): 46 | def __init__(self, low_level_channels, encoder_channels, 47 | low_level_reduced_channels=48): 48 | super().__init__() 49 | self.reduce_low_level = nn.Conv2d( 50 | low_level_channels, low_level_reduced_channels, kernel_size=1) 51 | self.fuse = nn.Sequential( 52 | nn.Conv2d( 53 | encoder_channels + low_level_reduced_channels, 256, 54 | kernel_size=3, padding=1), 55 | nn.ReLU(inplace=True), 56 | nn.BatchNorm2d(256), 57 | nn.Conv2d(256, 256, kernel_size=3, padding=1), 58 | nn.BatchNorm2d(256), 59 | nn.ReLU(inplace=True) 60 | ) 61 | 62 | def forward(self, x, low_level_features): 63 | low_level_features = self.reduce_low_level(low_level_features) 64 | x = F.interpolate( 65 | x, size=low_level_features.shape[2:], mode='bilinear', 66 | align_corners=True 67 | ) 68 | x = torch.cat([x, low_level_features], dim=1) 69 | x = self.fuse(x) 70 | return x 71 | 72 | 73 | class Deeplab(nn.Module): 74 | def __init__( 75 | self, in_channels, n_classes, backbone_cls, decoder=None, 76 | score_channels=256, aspp_rates=[6, 12, 18], pretrained=True, 77 | output_stride=16, multi_grid=[1, 2, 4], finetune_bn=True): 78 | super().__init__() 79 | self.backbone = backbone_cls( 80 | pretrained=pretrained, 81 | in_channels=in_channels, 82 | output_stride=output_stride, 83 | multi_grid=multi_grid 84 | ) 85 | rate = 2 if output_stride == 8 else 1 86 | self.aspp = ASPPModule(2048, 256, [r * rate for r in aspp_rates]) 87 | self.finetune_bn = finetune_bn 88 | self.decoder = decoder 89 | self.score = nn.Sequential( 90 | nn.Conv2d(score_channels, score_channels, kernel_size=3, padding=1), 91 | nn.BatchNorm2d(score_channels), 92 | nn.ReLU(inplace=True), 93 | nn.Conv2d(score_channels, n_classes, kernel_size=1) 94 | ) 95 | 96 | def forward(self, x): 97 | input_size = x.shape[2:] 98 | x, block_outputs = self.backbone(x) 99 | x = self.aspp(x) 100 | if self.decoder: 101 | x = self.decoder(x, block_outputs[0]) 102 | x = F.interpolate( 103 | x, size=input_size, mode='bilinear', align_corners=True 104 | ) 105 | x = self.score(x) 106 | return x 107 | 108 | def train(self, mode=True): 109 | super().train(mode) 110 | if not self.finetune_bn: 111 | for module in self.modules(): 112 | if isinstance(module, nn.BatchNorm2d): 113 | module.train(False) 114 | 115 | 116 | def deeplabv3_resnet50(in_channels, n_classes, **kwargs): 117 | return Deeplab(in_channels, n_classes, resnet50, **kwargs) 118 | 119 | 120 | def deeplabv3_resnet101(in_channels, n_classes, **kwargs): 121 | return Deeplab(in_channels, n_classes, resnet101, **kwargs) 122 | 123 | 124 | def deeplabv3_resnet152(in_channels, n_classes, **kwargs): 125 | return Deeplab(in_channels, n_classes, resnet152, **kwargs) 126 | 127 | 128 | def deeplabv3plus_resnet101(in_channels, n_classes, **kwargs): 129 | return Deeplab( 130 | in_channels, n_classes, resnet101, 131 | decoder=Decoder(256, 256, 48), **kwargs 132 | ) 133 | 134 | 135 | def deeplabv3plus_resnet50(in_channels, n_classes, **kwargs): 136 | return Deeplab( 137 | in_channels, n_classes, resnet50, 138 | decoder=Decoder(256, 256, 48), **kwargs 139 | ) 140 | -------------------------------------------------------------------------------- /pysemseg/models/densenet.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from torch.utils import checkpoint 5 | 6 | 7 | class DenseLayer(nn.Sequential): 8 | def __init__(self, n_input_features, growth_rate, drop_rate): 9 | super().__init__() 10 | self.add_module('norm', nn.BatchNorm2d(n_input_features)) 11 | self.add_module('relu', nn.ReLU(inplace=True)) 12 | self.add_module( 13 | 'conv', 14 | nn.Conv2d( 15 | n_input_features, growth_rate, kernel_size=3, stride=1, 16 | padding=1)) 17 | self.add_module('drop', nn.Dropout(drop_rate)) 18 | self.reset_parameters() 19 | 20 | def reset_parameters(self): 21 | torch.nn.init.kaiming_uniform_(self.conv.weight, nonlinearity='relu') 22 | 23 | 24 | def _dense_layer_cp_factory(layer): 25 | def _dense_layer_fn(*features): 26 | concat_features = torch.cat(features, 1) 27 | return layer(concat_features) 28 | return _dense_layer_fn 29 | 30 | 31 | class DenseBlock(nn.Module): 32 | def __init__( 33 | self, num_layers, num_input_features, growth_rate, drop_rate, 34 | efficient=False): 35 | super(DenseBlock, self).__init__() 36 | self.efficient = efficient 37 | self.layers = nn.ModuleList([ 38 | DenseLayer(num_input_features + i * growth_rate, growth_rate, drop_rate) 39 | for i in range(num_layers) 40 | ]) 41 | 42 | def forward(self, x): 43 | layer_outputs = [] 44 | for layer in self.layers: 45 | if self.efficient: 46 | output = checkpoint.checkpoint( 47 | _dense_layer_cp_factory(layer), x, *layer_outputs) 48 | else: 49 | output = layer(x) 50 | x = torch.cat([x, output], 1) 51 | layer_outputs.append(output) 52 | return torch.cat(layer_outputs, 1) 53 | 54 | 55 | class TransitionDown(nn.Sequential): 56 | def __init__(self, n_input_features, n_output_features, drop_rate=0.2): 57 | super().__init__() 58 | self.add_module('norm', nn.BatchNorm2d(n_input_features)) 59 | self.add_module('relu', nn.ReLU(inplace=True)) 60 | self.add_module('conv', nn.Conv2d( 61 | n_input_features, n_output_features, kernel_size=1, stride=1)) 62 | self.add_module('drop', nn.Dropout(drop_rate)) 63 | self.add_module('pool', nn.MaxPool2d(kernel_size=2, stride=2)) 64 | self.reset_parameters() 65 | 66 | def reset_parameters(self): 67 | torch.nn.init.kaiming_uniform_(self.conv.weight, nonlinearity='relu') 68 | 69 | 70 | class TransitionUp(nn.Module): 71 | def __init__(self, n_input_features, n_output_features): 72 | super().__init__() 73 | self.conv = nn.ConvTranspose2d( 74 | n_input_features, n_output_features, kernel_size=3, stride=2, 75 | padding=1, output_padding=1) 76 | 77 | def reset_parameters(self): 78 | torch.nn.init.kaiming_uniform_(self.conv.weight, nonlinearity='relu') 79 | torch.nn.init.kaiming_uniform_(self.conv.bias, nonlinearity='relu') 80 | 81 | def forward(self, x): 82 | return self.conv(x) 83 | 84 | 85 | class FCDenseNet(nn.Module): 86 | def __init__( 87 | self, in_channels, n_classes, growth_rate=32, 88 | n_init_features=48, drop_rate=0.2, blocks=(4, 5, 7, 10, 12, 15), 89 | efficient=False): 90 | super().__init__() 91 | 92 | self.efficient = efficient 93 | self.initial_conv = nn.Conv2d( 94 | in_channels, n_init_features, kernel_size=3, stride=1, padding=1) 95 | 96 | self.blocks_down = nn.ModuleList() 97 | self.transitions_down = nn.ModuleList() 98 | 99 | n_features = n_init_features 100 | self.skip_features = [] 101 | 102 | for i, n_layers in enumerate(blocks[:-1]): 103 | self.blocks_down.append( 104 | DenseBlock( 105 | n_layers, n_features, growth_rate, drop_rate, 106 | efficient=self.efficient) 107 | ) 108 | n_features = n_features + n_layers * growth_rate 109 | self.skip_features.append(n_features) 110 | self.transitions_down.append( 111 | TransitionDown(n_features, n_features, drop_rate) 112 | ) 113 | 114 | self.blocks_down.append( 115 | DenseBlock( 116 | blocks[-1], n_features, growth_rate, drop_rate, 117 | efficient=self.efficient)) 118 | n_features = blocks[-1] * growth_rate 119 | 120 | self.blocks_up = nn.ModuleList() 121 | self.transitions_up = nn.ModuleList() 122 | 123 | for i in range(len(blocks) - 2, -1, -1): 124 | self.transitions_up.append(TransitionUp(n_features, n_features)) 125 | n_features = n_features + self.skip_features[i] 126 | self.blocks_up.append( 127 | DenseBlock(blocks[i], n_features, growth_rate, drop_rate) 128 | ) 129 | n_features = growth_rate * blocks[i] 130 | 131 | self.score = nn.Conv2d(n_features, n_classes, kernel_size=1, stride=1) 132 | 133 | self.reset_parameters() 134 | 135 | def reset_parameters(self): 136 | torch.nn.init.kaiming_uniform_( 137 | self.initial_conv.weight, nonlinearity='relu' 138 | ) 139 | 140 | def _maybe_pad(self, x, size): 141 | hpad = max(size[0] - x.shape[2], 0) 142 | wpad = max(size[1] - x.shape[3], 0) 143 | if hpad + wpad > 0: 144 | lhpad = hpad // 2 145 | rhpad = hpad // 2 + hpad % 2 146 | lwpad = wpad // 2 147 | rwpad = wpad // 2 + wpad % 2 148 | 149 | x = F.pad(x, (lwpad, rwpad, lhpad, rhpad, 0, 0, 0, 0 )) 150 | return x 151 | 152 | 153 | 154 | def forward(self, x): 155 | x = self.initial_conv(x) 156 | skips = [] 157 | for block, trans in zip(self.blocks_down, self.transitions_down): 158 | c = block(x) 159 | x = torch.cat([c, x], 1) 160 | skips.append(x) 161 | x = trans(x) 162 | 163 | x = self.blocks_down[-1](x) 164 | skips = skips[::-1] 165 | 166 | for block, trans, skip in zip( 167 | self.blocks_up, self.transitions_up, skips): 168 | x = trans(x) 169 | x = self._maybe_pad(x, skip.shape[2:]) 170 | x = torch.cat([x, skip], 1) 171 | x = block(x) 172 | 173 | x = self.score(x) 174 | return x 175 | 176 | 177 | def fcdensenet56(in_channels, n_classes): 178 | return FCDenseNet( 179 | in_channels, n_classes, growth_rate=12, n_init_features=48, 180 | drop_rate=0.2, blocks=(4,) * 5, efficient=True 181 | ) 182 | 183 | 184 | def fcdensenet67(in_channels, n_classes): 185 | return FCDenseNet( 186 | in_channels, n_classes, growth_rate=16, n_init_features=48, 187 | drop_rate=0.2, blocks=(5,) * 5, efficient=True 188 | ) 189 | 190 | 191 | def fcdensenet103(in_channels, n_classes): 192 | return FCDenseNet( 193 | in_channels, n_classes, growth_rate=16, n_init_features=48, 194 | drop_rate=0.2, blocks=(4, 5, 7, 10, 12, 15), efficient=True 195 | ) 196 | -------------------------------------------------------------------------------- /pysemseg/models/fcn.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.nn.functional as F 3 | from abc import ABCMeta 4 | import torchvision.models as models 5 | 6 | 7 | def _maybe_pad(x, size): 8 | hpad = size[0] - x.shape[2] 9 | wpad = size[1] - x.shape[3] 10 | if hpad + wpad > 0: 11 | x = F.pad(x, (0, wpad, 0, hpad, 0, 0, 0, 0 )) 12 | return x 13 | 14 | 15 | class VGGFCN(nn.Module, metaclass=ABCMeta): 16 | def __init__(self, in_channels, n_classes): 17 | super().__init__() 18 | assert in_channels == 3 19 | self.n_classes = n_classes 20 | self.vgg16 = models.vgg16(pretrained=True) 21 | self.classifier = nn.Sequential( 22 | nn.Conv2d(512, 4096, kernel_size=7, padding=3), 23 | nn.ReLU(True), 24 | nn.Dropout(), 25 | nn.Conv2d(4096, 4096, kernel_size=1), 26 | nn.ReLU(True), 27 | nn.Dropout(), 28 | nn.Conv2d(4096, n_classes, kernel_size=1), 29 | ) 30 | 31 | self._initialize_weights() 32 | 33 | def _initialize_weights(self): 34 | self.classifier[0].weight.data = ( 35 | self.vgg16.classifier[0].weight.data.view( 36 | self.classifier[0].weight.size()) 37 | ) 38 | self.classifier[3].weight.data = ( 39 | self.vgg16.classifier[3].weight.data.view( 40 | self.classifier[3].weight.size()) 41 | ) 42 | 43 | 44 | class VGGFCN32(VGGFCN): 45 | def forward(self, x): 46 | input_height, input_width = x.shape[2], x.shape[3] 47 | x = self.vgg16.features(x) 48 | x = self.classifier(x) 49 | x = F.interpolate(x, size=(input_height, input_width), 50 | mode='bilinear', align_corners=True) 51 | return x 52 | 53 | 54 | class VGGFCN16(VGGFCN): 55 | def __init__(self, in_channels, n_classes): 56 | super().__init__(in_channels, n_classes) 57 | self.score4 = nn.Conv2d(512, n_classes, kernel_size=1) 58 | self.upscale5 = nn.ConvTranspose2d( 59 | n_classes, n_classes, kernel_size=2, stride=2) 60 | 61 | def forward(self, x): 62 | input_height, input_width = x.shape[2], x.shape[3] 63 | pool4 = self.vgg16.features[:-7](x) 64 | pool5 = self.vgg16.features[-7:](pool4) 65 | pool5_upscaled = self.upscale5(self.classifier(pool5)) 66 | pool4 = self.score4(pool4) 67 | x = pool4 + pool5_upscaled 68 | x = F.interpolate(x, size=(input_height, input_width), 69 | mode='bilinear', align_corners=True) 70 | return x 71 | 72 | 73 | class VGGFCN8(VGGFCN): 74 | def __init__(self, in_channels, n_classes): 75 | super().__init__(in_channels, n_classes) 76 | self.upscale4 = nn.ConvTranspose2d( 77 | n_classes, n_classes, kernel_size=2, stride=2) 78 | self.score4 = nn.Conv2d( 79 | 512, n_classes, kernel_size=1, stride=1) 80 | self.score3 = nn.Conv2d( 81 | 256, n_classes, kernel_size=1, stride=1) 82 | self.upscale5 = nn.ConvTranspose2d( 83 | n_classes, n_classes, kernel_size=2, stride=2) 84 | 85 | def forward(self, x): 86 | input_height, input_width = x.shape[2], x.shape[3] 87 | pool3 = self.vgg16.features[:-14](x) 88 | pool4 = self.vgg16.features[-14:-7](pool3) 89 | pool5 = self.vgg16.features[-7:](pool4) 90 | pool5_upscaled = self.upscale5(self.classifier(pool5)) 91 | pool5_upscaled = _maybe_pad(pool5_upscaled, pool4.shape[2:]) 92 | pool4_scores = self.score4(pool4) 93 | pool4_fused = pool4_scores + pool5_upscaled 94 | pool4_upscaled = self.upscale4(pool4_fused) 95 | pool4_upscaled = _maybe_pad(pool4_upscaled, pool3.shape[2:]) 96 | x = self.score3(pool3) + pool4_upscaled 97 | x = F.interpolate(x, size=(input_height, input_width), 98 | mode='bilinear', align_corners=True) 99 | return x 100 | -------------------------------------------------------------------------------- /pysemseg/models/resnet.py: -------------------------------------------------------------------------------- 1 | import math 2 | import re 3 | from collections import OrderedDict 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | import torch.utils.model_zoo as model_zoo 7 | 8 | 9 | RESNET_CKPT = { 10 | 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 11 | 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 12 | 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth' 13 | } 14 | 15 | 16 | class Bottleneck(nn.Module): 17 | def __init__(self, in_channels, bottleneck_channels, out_channels, 18 | stride=1, dilation=1, downsample=None): 19 | super().__init__() 20 | self.conv1 = nn.Conv2d( 21 | in_channels, bottleneck_channels, kernel_size=1, bias=False) 22 | self.bn1 = nn.BatchNorm2d(bottleneck_channels) 23 | self.conv2 = nn.Conv2d( 24 | bottleneck_channels, bottleneck_channels, kernel_size=3, 25 | stride=stride, padding=(dilation, dilation), 26 | dilation=dilation, bias=False) 27 | self.bn2 = nn.BatchNorm2d(bottleneck_channels) 28 | self.conv3 = nn.Conv2d( 29 | bottleneck_channels, out_channels, kernel_size=1, 30 | bias=False) 31 | self.bn3 = nn.BatchNorm2d(out_channels) 32 | self.relu = nn.ReLU(inplace=True) 33 | self.downsample = downsample 34 | self.stride = stride 35 | 36 | def forward(self, x): 37 | residual = x 38 | 39 | out = self.conv1(x) 40 | out = self.bn1(out) 41 | out = self.relu(out) 42 | 43 | out = self.conv2(out) 44 | out = self.bn2(out) 45 | out = self.relu(out) 46 | 47 | out = self.conv3(out) 48 | out = self.bn3(out) 49 | 50 | if self.downsample is not None: 51 | residual = self.downsample(x) 52 | 53 | out += residual 54 | out = self.relu(out) 55 | 56 | return out 57 | 58 | 59 | class ResBlock(nn.Sequential): 60 | def __init__(self, in_channels, channels, n_layers, dilations=None, 61 | stride=1, expansion=4): 62 | super().__init__() 63 | if dilations is None: 64 | dilations = [1] * n_layers 65 | downsample = None 66 | if stride != 1 or in_channels != channels * expansion: 67 | downsample = nn.Sequential( 68 | nn.Conv2d(in_channels, channels * expansion, 69 | kernel_size=1, stride=stride, bias=False), 70 | nn.BatchNorm2d(channels * expansion), 71 | ) 72 | 73 | layers = [] 74 | self.add_module( 75 | '0', Bottleneck( 76 | in_channels, channels, channels * expansion, stride, 77 | dilations[0], downsample=downsample)) 78 | in_channels = channels * expansion 79 | for i in range(1, n_layers): 80 | self.add_module( 81 | '{}'.format(i), 82 | Bottleneck(in_channels, channels, in_channels, 83 | dilation=dilations[i])) 84 | 85 | 86 | 87 | class ResNet(nn.Module): 88 | def __init__(self, in_channels, layers_config): 89 | super().__init__() 90 | self.expansion = 4 91 | channels = layers_config[0]['channels'] 92 | self.conv1 = nn.Conv2d( 93 | in_channels, channels, kernel_size=7, stride=2, padding=3, 94 | bias=False) 95 | self.bn1 = nn.BatchNorm2d(channels) 96 | self.relu = nn.ReLU(inplace=True) 97 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) 98 | self.layers = nn.ModuleList() 99 | for i, layer_config in enumerate(layers_config): 100 | self.layers.add_module( 101 | 'layer{}'.format(i + 1), 102 | ResBlock(channels, **layer_config, expansion=self.expansion) 103 | ) 104 | channels = layer_config['channels'] * self.expansion 105 | 106 | self.reset_parameters() 107 | 108 | def load_pretrained_model(self, url): 109 | def replace_key(key): 110 | return re.sub(r'layer(\d)', r'layers.layer\g<1>', key) 111 | state_dict = model_zoo.load_url(url) 112 | state_dict = OrderedDict([ 113 | (replace_key(key), value) for key, value in state_dict.items() 114 | ]) 115 | self.load_state_dict(state_dict, strict=False) 116 | 117 | def reset_parameters(self): 118 | for m in self.modules(): 119 | if isinstance(m, nn.Conv2d): 120 | n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels 121 | m.weight.data.normal_(0, math.sqrt(2. / n)) 122 | elif isinstance(m, nn.BatchNorm2d): 123 | m.weight.data.fill_(1) 124 | m.bias.data.zero_() 125 | 126 | def forward(self, x): 127 | x = self.conv1(x) 128 | x = self.bn1(x) 129 | x = self.relu(x) 130 | x = self.maxpool(x) 131 | block_outputs = [] 132 | for layer in self.layers: 133 | x = layer(x) 134 | block_outputs.append(x) 135 | return x, block_outputs 136 | 137 | 138 | def resnet(layers, pretrained_model=None, in_channels=3, output_stride=16, 139 | multi_grid=(1, 2, 4), add_blocks=0): 140 | 141 | assert output_stride in [8, 16, 32] 142 | if output_stride == 8: 143 | rate3 = [2] * layers[2] 144 | stride3 = 1 145 | rate4 = [2 * 2 * mg for mg in multi_grid] 146 | stride4 = 1 147 | elif output_stride == 16: 148 | rate3 = [1] * layers[2] 149 | stride3 = 2 150 | rate4 = [2 * mg for mg in multi_grid] 151 | stride4 = 1 152 | elif output_stride == 32: 153 | rate3 = [1] * layers[2] 154 | stride3 = 2 155 | rate4 = [1] * layers[3] 156 | stride4 = 2 157 | 158 | blocks = [ 159 | { 160 | 'stride': 1, 161 | 'n_layers': layers[0], 162 | 'channels': 64 163 | }, 164 | { 165 | 'stride': 2, 166 | 'n_layers': layers[1], 167 | 'channels': 128 168 | }, 169 | { 170 | 'stride': stride3, 171 | 'n_layers': layers[2], 172 | 'dilations': rate3, 173 | 'channels': 256 174 | }, 175 | { 176 | 'stride': stride4, 177 | 'n_layers': layers[3], 178 | 'dilations': rate4, 179 | 'channels': 512 180 | } 181 | ] 182 | rate = rate3[0] * 4 183 | for _ in range(add_blocks): 184 | dilations = [rate * mg for mg in multi_grid] 185 | blocks.append( 186 | { 187 | 'stride': 1, 188 | 'n_layers': 3, 189 | 'dilations': dilations, 190 | 'channels': 512 191 | } 192 | ) 193 | resnet = ResNet(in_channels, blocks) 194 | if pretrained_model is not None: 195 | resnet.load_pretrained_model(pretrained_model) 196 | 197 | return resnet 198 | 199 | 200 | def resnet50(pretrained=True, **kwargs): 201 | return resnet( 202 | [3, 4, 6, 3], 203 | pretrained_model=RESNET_CKPT['resnet50'] if pretrained else None, 204 | **kwargs 205 | ) 206 | 207 | 208 | def resnet101(pretrained=True, **kwargs): 209 | return resnet( 210 | [3, 4, 23, 3], 211 | pretrained_model=RESNET_CKPT['resnet101'] if pretrained else None, 212 | **kwargs 213 | ) 214 | 215 | 216 | def resnet152(pretrained=True, **kwargs): 217 | return resnet( 218 | [3, 8, 36, 3], 219 | pretrained_model=RESNET_CKPT['resnet152'] if pretrained else None 220 | **kwargs 221 | ) 222 | -------------------------------------------------------------------------------- /pysemseg/models/unet.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from pysemseg.models.resnet import resnet50, resnet101, resnet152, ResBlock 5 | 6 | 7 | def _maybe_pad(x, size): 8 | hpad = max(size[0] - x.shape[2], 0) 9 | wpad = max(size[1] - x.shape[3], 0) 10 | if hpad + wpad > 0: 11 | lhpad = hpad // 2 12 | rhpad = hpad // 2 + hpad % 2 13 | lwpad = wpad // 2 14 | rwpad = wpad // 2 + wpad % 2 15 | 16 | x = F.pad(x, (lwpad, rwpad, lhpad, rhpad, 0, 0, 0, 0 )) 17 | return x 18 | 19 | 20 | class DownLayer(nn.Module): 21 | def __init__(self, in_units, out_units, batch_norm=True): 22 | super().__init__() 23 | self.batch_norm = batch_norm 24 | self.conv1 = nn.Conv2d(in_units, out_units, kernel_size=3, padding=1) 25 | if batch_norm: 26 | self.bn1 = nn.BatchNorm2d(out_units) 27 | self.conv2 = nn.Conv2d(out_units, out_units, kernel_size=3, padding=1) 28 | if batch_norm: 29 | self.bn2 = nn.BatchNorm2d(out_units) 30 | self.dropout = nn.Dropout(p=0.0) 31 | 32 | def forward(self, x): 33 | x = F.relu(self.conv1(x), inplace=True) 34 | if self.batch_norm: 35 | x = self.bn1(x) 36 | x = F.relu(self.conv2(x), inplace=True) 37 | if self.batch_norm: 38 | x = self.bn2(x) 39 | if self.training: 40 | x = self.dropout(x) 41 | return x 42 | 43 | 44 | class UpResBlock(nn.Module): 45 | def __init__(self, in_units, out_units, upsample=True, batch_norm=True): 46 | super().__init__() 47 | self.upsample = upsample 48 | self.batch_norm = batch_norm 49 | self.block = ResBlock( 50 | in_units, out_units, 3, dilations=[1, 2, 1], expansion=1) 51 | if self.upsample: 52 | self.conv3 = nn.ConvTranspose2d( 53 | out_units, out_units // 2, kernel_size=2, stride=2, 54 | padding=0, output_padding=0) 55 | if self.batch_norm: 56 | self.bn3 = nn.BatchNorm2d(out_units // 2) 57 | 58 | 59 | def forward(self, x): 60 | x = self.block(x) 61 | if self.upsample: 62 | x = F.relu(self.conv3(x), inplace=True) 63 | if self.batch_norm: 64 | x = self.bn3(x) 65 | return x 66 | 67 | 68 | class UpLayer(nn.Module): 69 | def __init__(self, in_units, out_units, upsample=True, batch_norm=True): 70 | super().__init__() 71 | self.batch_norm = batch_norm 72 | self.conv1 = nn.Conv2d(in_units, out_units, kernel_size=3, padding=1) 73 | if batch_norm: 74 | self.bn1 = nn.BatchNorm2d(out_units) 75 | self.conv2 = nn.Conv2d(out_units, out_units, kernel_size=3, padding=1) 76 | if batch_norm: 77 | self.bn2 = nn.BatchNorm2d(out_units) 78 | self.upsample = upsample 79 | self.dropout = nn.Dropout(p=0.0) 80 | if self.upsample: 81 | self.conv3 = nn.ConvTranspose2d( 82 | out_units, out_units // 2, kernel_size=2, stride=2, 83 | padding=0, output_padding=0) 84 | if self.batch_norm: 85 | self.bn3 = nn.BatchNorm2d(out_units // 2) 86 | 87 | 88 | def forward(self, x): 89 | x = F.relu(self.conv1(x), inplace=True) 90 | if self.batch_norm: 91 | x = self.bn1(x) 92 | x = F.relu(self.conv2(x), inplace=True) 93 | if self.batch_norm: 94 | x = self.bn2(x) 95 | if self.training: 96 | x = self.dropout(x) 97 | if self.upsample: 98 | x = F.relu(self.conv3(x), inplace=True) 99 | if self.batch_norm: 100 | x = self.bn3(x) 101 | return x 102 | 103 | 104 | class UNet(nn.Module): 105 | def __init__(self, n_classes, in_channels, batch_norm=True): 106 | super().__init__() 107 | self.down_layers = [ 108 | DownLayer(in_channels, 64, batch_norm), 109 | DownLayer(64, 128, batch_norm), 110 | DownLayer(128, 256, batch_norm), 111 | DownLayer(256, 512, batch_norm) 112 | ] 113 | self.interface1 = nn.Conv2d(512, 1024, kernel_size=3, padding=1) 114 | self.interface2 = nn.Conv2d(1024, 512, kernel_size=3, padding=1) 115 | self.interface_up = nn.ConvTranspose2d( 116 | 512, 512, kernel_size=2, stride=2) 117 | self.up_layers = [ 118 | UpLayer(1024, 512, batch_norm=batch_norm), 119 | UpLayer(512, 256, batch_norm=batch_norm), 120 | UpLayer(256, 128, batch_norm=batch_norm), 121 | UpLayer(128, 64, upsample=False, batch_norm=batch_norm) 122 | ] 123 | self.down_layers = nn.ModuleList(self.down_layers) 124 | self.up_layers = nn.ModuleList(self.up_layers) 125 | self.conv_classes = nn.Conv2d(64, n_classes, kernel_size=1) 126 | 127 | def forward(self, x): 128 | input_height, input_width = x.shape[2:] 129 | down_outputs = [] 130 | for layer in self.down_layers: 131 | x = layer(x) 132 | down_outputs.append(x) 133 | x = F.max_pool2d(x, kernel_size=2) 134 | x = F.relu(self.interface1(x), inplace=True) 135 | x = F.relu(self.interface2(x), inplace=True) 136 | x = F.relu(self.interface_up(x), inplace=True) 137 | for i, layer in enumerate(self.up_layers): 138 | skip = down_outputs[-(i+1)] 139 | x = _maybe_pad(x, skip.shape[2:]) 140 | skip = _maybe_pad(skip, x.shape[2:]) 141 | x = torch.cat([x, skip], dim=1) 142 | x = layer(x) 143 | x = F.interpolate( 144 | x, (input_height, input_width), mode='bilinear', align_corners=True) 145 | x = self.conv_classes(x) 146 | return x 147 | 148 | 149 | class UNetResNetV1(nn.Module): 150 | def __init__( 151 | self, n_classes, network, skip_channels, interface_channels, 152 | up_channels): 153 | super().__init__() 154 | self.network = network 155 | self.bottleneck_channels = interface_channels // self.network.expansion 156 | self.interface = UpLayer(interface_channels, self.bottleneck_channels) 157 | self.up_layers = nn.ModuleList([ 158 | UpLayer(skip_channels[-1] + self.bottleneck_channels // 2, up_channels[0]), 159 | UpLayer(skip_channels[-2] + up_channels[0] // 2, up_channels[1]), 160 | UpLayer(skip_channels[-3] + up_channels[1] // 2, up_channels[2]), 161 | UpLayer(skip_channels[-4] + up_channels[2] // 2, up_channels[3]) 162 | ]) 163 | self.conv_classes = nn.Conv2d(up_channels[3] // 2, n_classes, kernel_size=1) 164 | 165 | def _resnet_forward(self, x): 166 | skips = [] 167 | x = self.network.conv1(x) 168 | skips.append(x) 169 | x = self.network.bn1(x) 170 | x = self.network.relu(x) 171 | x = self.network.maxpool(x) 172 | skips.append(x) 173 | for i, layer in enumerate(self.network.layers): 174 | height, width = x.shape[2:] 175 | x = layer(x) 176 | if (x.shape[2] < height and x.shape[3] < width and 177 | i < len(self.network.layers) - 1): 178 | skips.append(x) 179 | return x, skips 180 | 181 | def forward(self, x): 182 | input_tensor = x 183 | x, skips = self._resnet_forward(x) 184 | x = self.interface(x) 185 | for i in range(0, len(self.up_layers)): 186 | skip = _maybe_pad(skips[-i - 1], x.shape[2:]) 187 | x = torch.cat([x, skip], dim=1) 188 | x = self.up_layers[i](x) 189 | x = _maybe_pad(x, input_tensor.shape[2:]) 190 | x = F.interpolate( 191 | x, size=input_tensor.shape[2:], mode='bilinear', align_corners=True) 192 | scores = self.conv_classes(x) 193 | return scores 194 | 195 | 196 | def unet_resnet( 197 | resnet_model_fn, in_channels, n_classes, pretrained=True, 198 | up_channels=[512, 256, 128, 64], **kwargs): 199 | net = resnet_model_fn( 200 | in_channels=in_channels, output_stride=32, pretrained=pretrained, 201 | **kwargs) 202 | return UNetResNetV1( 203 | n_classes=n_classes, network=net, 204 | skip_channels=[64, 64, 512, 1024], 205 | interface_channels=2048, 206 | up_channels=up_channels 207 | ) 208 | 209 | 210 | def unet_resnet101(*args, **kwargs): 211 | return unet_resnet(resnet101, *args, **kwargs) 212 | 213 | 214 | def unet_resnet152(*args, **kwargs): 215 | return unet_resnet(resnet152, *args, **kwargs) 216 | 217 | def unet_big_resnet101(*args, **kwargs): 218 | return unet_resnet( 219 | resnet101, *args, up_channels=[1024, 256, 128, 64], **kwargs) 220 | 221 | -------------------------------------------------------------------------------- /pysemseg/train.py: -------------------------------------------------------------------------------- 1 | import cProfile 2 | import sys 3 | import os 4 | import ast 5 | import time 6 | import yaml 7 | import numpy as np 8 | import configargparse 9 | 10 | import torch 11 | import torch.nn.functional as F 12 | from torch.autograd import Variable 13 | 14 | from pysemseg import datasets 15 | from pysemseg.metrics import SegmentationMetrics, TorchSegmentationMetrics 16 | from pysemseg.loggers import TensorboardLogger, VisdomLogger, ConsoleLogger 17 | from pysemseg.evaluate import evaluate 18 | from pysemseg.utils import ( 19 | prompt_delete_dir, restore, tensor_to_numpy, import_type, 20 | flatten_dict, get_latest_checkpoint, save 21 | ) 22 | 23 | 24 | def define_args(): 25 | parser = configargparse.ArgParser(description='PyTorch Segmentation Framework', 26 | config_file_parser_class=configargparse.YAMLConfigFileParser, 27 | ) 28 | parser.add_argument( 29 | '--config', is_config_file=True, required=False, help='Config file') 30 | parser.add_argument('--model', type=str, required=True, 31 | help=('A path to the model including the module. ' 32 | 'Should be resolvable')) 33 | parser.add_argument('--model-args', type=ast.literal_eval, required=False, default={}, 34 | help=('Args passed to the model constructor')) 35 | parser.add_argument('--data-dir', type=str, required=True, 36 | help='Path to the dataset root dir.') 37 | parser.add_argument('--model-dir', type=str, required=True, 38 | help='Path to store output data.') 39 | parser.add_argument('--dataset', type=str, required=True, 40 | help=('Path to the dataset class including the module')) 41 | parser.add_argument('--dataset-args', type=ast.literal_eval, default={}, 42 | required=False, 43 | help='Dataset args.') 44 | parser.add_argument('--criterion', type=str, required=False, 45 | default='torch.nn.CrossEntropyLoss', 46 | help=('Path to the loss class including the module')) 47 | parser.add_argument('--criterion-args', type=ast.literal_eval, 48 | default={'reduction': 'sum'}, required=False, 49 | help='Criterion args.') 50 | parser.add_argument('--batch-size', type=int, default=64, metavar='N', 51 | help='input batch size for training (default: 64)') 52 | parser.add_argument('--max-gpu-batch-size', type=int, default=None, 53 | help='Effective GPU batch size. Gradients will be' 54 | 'accumulated to the batch-size before update.') 55 | parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N', 56 | help='input batch size for testing (default: 1000)') 57 | parser.add_argument('--epochs', type=int, default=10, metavar='N', 58 | help='number of epochs to train (default: 10)') 59 | parser.add_argument('--optimizer', type=str, default='RMSprop', 60 | required=False, 61 | help='Optimizer type.') 62 | parser.add_argument('--optimizer-args', type=ast.literal_eval, default={}, 63 | required=False, 64 | help='Optimizer args.') 65 | parser.add_argument('--lr-scheduler', type=str, required=False, 66 | default='lr_schedulers.ConstantLR', 67 | help='Learning rate scheduler type.') 68 | parser.add_argument('--lr-scheduler-args', type=ast.literal_eval, default={}, 69 | required=False, 70 | help='Learning rate scheduler args.') 71 | parser.add_argument('--transformer', type=str, required=False, 72 | help='Transformer type') 73 | parser.add_argument('--transformer-args', type=ast.literal_eval, default={}, 74 | required=False, 75 | help='Transformer args.') 76 | parser.add_argument('--lr', type=float, default=0.001, metavar='LR', 77 | help='learning rate (default: 0.001)') 78 | parser.add_argument('--no-cuda', action='store_true', default=False, 79 | help='disables CUDA training') 80 | parser.add_argument('--seed', type=int, default=8861, metavar='S', 81 | help='random seed (default: 8861)') 82 | parser.add_argument('--log-interval', type=int, default=20, metavar='N', 83 | help='logging training status frequency') 84 | parser.add_argument('--log-images-interval', type=int, default=200, metavar='N', 85 | help='Frequency of logging images and larger plots') 86 | parser.add_argument('--ngpu', type=int, default=1, help='number of GPUs to use') 87 | parser.add_argument('--num-workers', type=int, default=1, 88 | help='Number of CPU data workers') 89 | parser.add_argument('--checkpoint', type=str, 90 | required=False, 91 | help='Load model on checkpoint.') 92 | parser.add_argument('--save-model-frequency', type=int, 93 | required=False, default=5, 94 | help='Save model checkpoint every nth epoch.') 95 | parser.add_argument('--profile', type=str, required=False, 96 | help='Runs a profiler and saves the results in a file.') 97 | group = parser.add_mutually_exclusive_group() 98 | group.add_argument('--allow-missing-keys', action='store_true', default=False, 99 | help='Whether to allow module keys to differ from checkpoint keys' 100 | ' when loading a checkpoint') 101 | group.add_argument('--continue-training', action='store_true', default=False, 102 | help='Continue experiment from the last checkpoint in the model dir') 103 | return parser 104 | 105 | 106 | def train_epoch( 107 | model, loader, criterion, optimizer, lr_scheduler, 108 | epoch, console_logger, visual_logger, device, log_interval, 109 | log_images_interval, accumulate_steps): 110 | model.train() 111 | 112 | ignore_index = loader.dataset.ignore_index 113 | 114 | metrics = TorchSegmentationMetrics( 115 | loader.dataset.number_of_classes, 116 | loader.dataset.labels, 117 | ignore_index=ignore_index, 118 | device=device 119 | ) 120 | epoch_metrics = TorchSegmentationMetrics( 121 | loader.dataset.number_of_classes, 122 | loader.dataset.labels, 123 | ignore_index=ignore_index, 124 | device=device 125 | ) 126 | 127 | start_time = time.time() 128 | for step, (ids, data, target) in enumerate(loader): 129 | data , target = Variable(data.to(device)), Variable(target.to(device, non_blocking=True)) 130 | output = model(data) 131 | loss = criterion(output, target) 132 | num_targets = torch.sum(target != ignore_index).float() 133 | loss = loss / (num_targets * accumulate_steps) 134 | loss.backward() 135 | 136 | if step % accumulate_steps == 0: 137 | optimizer.step() 138 | optimizer.zero_grad() 139 | 140 | output = F.softmax(output.detach(), dim=1) 141 | predictions = torch.argmax(output, dim=1) 142 | 143 | predictions, target, loss = ( 144 | predictions.detach(), target.detach(), 145 | (loss * accumulate_steps).detach() 146 | ) 147 | 148 | metrics.add(predictions, target, loss) 149 | epoch_metrics.add(predictions, target, loss) 150 | 151 | if step % (accumulate_steps * log_interval) == 0 and step > 0: 152 | avg_time = ( 153 | time.time() - start_time) / log_interval 154 | start_time = time.time() 155 | metrics_dict = metrics.metrics() 156 | metrics_dict['time'] = avg_time 157 | metrics_dict.pop('class') 158 | console_logger.log(step, epoch, loader, tensor_to_numpy(data), metrics_dict) 159 | 160 | metrics = TorchSegmentationMetrics( 161 | loader.dataset.number_of_classes, 162 | loader.dataset.labels, 163 | ignore_index=ignore_index, 164 | device=device 165 | ) 166 | 167 | if step % (accumulate_steps * log_images_interval) == 0 and step > 0: 168 | visual_logger.log_prediction_images( 169 | step, 170 | tensor_to_numpy(data.data), 171 | tensor_to_numpy(target.data), 172 | tensor_to_numpy(predictions.data), 173 | name='images', 174 | prefix='Train' 175 | ) 176 | 177 | visual_logger.log_metrics(epoch, epoch_metrics.metrics(), 'Train') 178 | visual_logger.log_learning_rate(epoch, optimizer.param_groups[0]['lr']) 179 | 180 | 181 | def _create_data_loaders( 182 | data_dir, dataset_cls, dataset_args, transformer_cls, transformer_args, 183 | train_batch_size, val_batch_size, num_workers, pin_memory=False): 184 | train_dataset = datasets.create_dataset( 185 | data_dir, dataset_cls, dataset_args, 186 | transformer_cls, transformer_args, mode='train') 187 | 188 | train_loader = torch.utils.data.DataLoader( 189 | train_dataset, batch_size=train_batch_size, 190 | shuffle=True, num_workers=num_workers, pin_memory=pin_memory) 191 | 192 | validate_dataset = datasets.create_dataset( 193 | data_dir, dataset_cls, dataset_args, 194 | transformer_cls, transformer_args, mode='val') 195 | 196 | validate_loader = torch.utils.data.DataLoader( 197 | validate_dataset, batch_size=val_batch_size, 198 | shuffle=False, num_workers=num_workers, pin_memory=pin_memory) 199 | 200 | return train_loader, validate_loader 201 | 202 | 203 | def _store_args(args, model_dir): 204 | with open(os.path.join(model_dir, 'args.yaml'), 'w') as args_file: 205 | yaml.dump( 206 | {**args.__dict__, 'command': " ".join(sys.argv)}, 207 | args_file 208 | ) 209 | 210 | 211 | def _set_seed(seed, cuda): 212 | torch.manual_seed(seed) 213 | np.random.seed(seed) 214 | if cuda: 215 | torch.cuda.manual_seed(seed) 216 | 217 | 218 | def train(args): 219 | if not args.continue_training: 220 | prompt_delete_dir(args.model_dir) 221 | os.makedirs(args.model_dir) 222 | 223 | _store_args(args, args.model_dir) 224 | 225 | # seed torch and cuda 226 | args.cuda = not args.no_cuda and torch.cuda.is_available() 227 | _set_seed(args.seed, args.cuda) 228 | 229 | device = torch.device('cuda:0' if args.cuda else 'cpu:0') 230 | 231 | dataset_cls = import_type(args.dataset, ['pysemseg.datasets']) 232 | transformer_cls = import_type(args.transformer, ['pysemseg.datasets']) 233 | 234 | args.max_gpu_batch_size = args.max_gpu_batch_size or args.batch_size 235 | 236 | train_loader, validate_loader = _create_data_loaders( 237 | args.data_dir, dataset_cls, args.dataset_args, transformer_cls, 238 | args.transformer_args, args.max_gpu_batch_size, 239 | args.test_batch_size, args.num_workers, pin_memory=args.cuda 240 | ) 241 | 242 | visual_logger = VisdomLogger( 243 | log_directory=args.model_dir, 244 | color_palette=train_loader.dataset.color_palette, 245 | continue_logging=args.continue_training 246 | ) 247 | 248 | visual_logger.log_args(args.__dict__) 249 | 250 | model_class = import_type(args.model, ['pysemseg.models']) 251 | model = model_class( 252 | in_channels=train_loader.dataset.in_channels, 253 | n_classes=train_loader.dataset.number_of_classes, 254 | **args.model_args 255 | ) 256 | 257 | if torch.cuda.device_count() > 1: 258 | model = torch.nn.DataParallel(model) 259 | 260 | model = model.to(device) 261 | 262 | criterion_cls = import_type(args.criterion, ['pysemseg.losses']) 263 | criterion = criterion_cls( 264 | ignore_index=train_loader.dataset.ignore_index, 265 | **args.criterion_args 266 | ) 267 | 268 | criterion = criterion.to(device) 269 | 270 | optimizer_class = import_type(args.optimizer, ['torch.optim']) 271 | optimizer = optimizer_class( 272 | model.parameters(), lr=args.lr, **args.optimizer_args 273 | ) 274 | 275 | start_epoch = 0 276 | 277 | if args.continue_training: 278 | args.checkpoint = get_latest_checkpoint(args.model_dir) 279 | assert args.checkpoint is not None 280 | 281 | lr_scheduler_cls = import_type( 282 | args.lr_scheduler, ['pysemseg.lr_schedulers', 'torch.optim.lr_scheduler'] 283 | ) 284 | lr_scheduler = lr_scheduler_cls(optimizer, **args.lr_scheduler_args) 285 | 286 | if args.checkpoint: 287 | start_epoch = restore( 288 | args.checkpoint, model, optimizer, lr_scheduler, 289 | strict=not args.allow_missing_keys) + 1 290 | 291 | log_filepath = os.path.join(args.model_dir, 'train.log') 292 | 293 | with ConsoleLogger(filename=log_filepath) as logger: 294 | for epoch in range(start_epoch, start_epoch + args.epochs): 295 | train_epoch( 296 | model, train_loader, criterion, optimizer, lr_scheduler, 297 | epoch, logger, visual_logger, device, args.log_interval, 298 | args.log_images_interval, 299 | args.batch_size // args.max_gpu_batch_size) 300 | evaluate( 301 | model, validate_loader, criterion, logger, epoch, 302 | visual_logger, device, args.log_images_interval) 303 | if epoch % args.save_model_frequency == 0: 304 | save(model, optimizer, lr_scheduler, args.model_dir, 305 | train_loader.dataset.in_channels, 306 | train_loader.dataset.number_of_classes, epoch, args) 307 | lr_scheduler.step() 308 | 309 | 310 | def main(): 311 | sys.path.append('./') 312 | parser = define_args() 313 | args = parser.parse_args() 314 | if args.profile is not None: 315 | cProfile.runctx('train(args)', globals(), locals(), args.profile) 316 | else: 317 | train(args) 318 | 319 | 320 | if __name__ == '__main__': 321 | main() 322 | -------------------------------------------------------------------------------- /pysemseg/transforms/__init__.py: -------------------------------------------------------------------------------- 1 | from .loaders import PILImageLoader, CV2ImageLoader 2 | from .convert import Grayscale, ToCategoryTensor, ToTensor, ToFloatImage 3 | from .transforms import ( 4 | Compose, Resize, Binarize, RandomContrast, RandomBrightness, 5 | RandomHueSaturation, RandomHorizontalFlip, RandomRotate, 6 | RandomTranslate, RandomGammaCorrection, ResizeBatch, 7 | RandomCrop, PadTo, ScaleTo, RandomCropFixedSize, Concat, 8 | RandomScale, Choice, RandomPerspective, RandomShear, 9 | RandomElasticTransform, RandomGaussianBlur 10 | ) 11 | -------------------------------------------------------------------------------- /pysemseg/transforms/convert.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import cv2 3 | import torch 4 | 5 | 6 | class Grayscale: 7 | def __call__(self, image): 8 | return cv2.cvtColor(image, cv2.COLOR_RGB2GRAY) 9 | 10 | 11 | class ToCategoryTensor: 12 | def __init__(self, remap=None): 13 | self.remap = remap 14 | 15 | def __call__(self, image): 16 | if self.remap: 17 | for k, v in self.remap.items(): 18 | image[image == k] = v 19 | return torch.LongTensor(image) 20 | 21 | 22 | class ToTensor: 23 | def __call__(self, image): 24 | if len(image.shape) == 2: 25 | image = np.expand_dims(image, axis=-1) 26 | image = np.transpose(image, [2, 0, 1]) 27 | return torch.FloatTensor(image) 28 | 29 | 30 | class ToFloatImage: 31 | def __call__(self, image): 32 | if image.dtype == np.uint8: 33 | return image / 255. 34 | elif image.dtype == np.uint16: 35 | return image / np.iinfo(np.uint16).max 36 | elif image.dtype == np.float: 37 | return image 38 | raise ValueError("Unsupported image type") 39 | -------------------------------------------------------------------------------- /pysemseg/transforms/loaders.py: -------------------------------------------------------------------------------- 1 | from PIL import Image 2 | import cv2 3 | 4 | 5 | class PILImageLoader(): 6 | def __call__(self, path): 7 | with open(path, 'rb') as f: 8 | with Image.open(f) as img: 9 | return img.convert('RGB') 10 | 11 | 12 | class CV2ImageLoader(): 13 | def __init__(self, grayscale=False): 14 | self.grayscale = grayscale 15 | 16 | def __call__(self, path): 17 | if self.grayscale: 18 | return cv2.imread(path, cv2.IMREAD_GRAYSCALE) 19 | img = cv2.imread(path) 20 | return cv2.cvtColor(img, cv2.COLOR_BGR2RGB) 21 | -------------------------------------------------------------------------------- /pysemseg/transforms/transforms.py: -------------------------------------------------------------------------------- 1 | import cv2 2 | import numpy as np 3 | from scipy.ndimage.filters import gaussian_filter 4 | 5 | 6 | 7 | class Compose: 8 | def __init__(self, transforms): 9 | self.transforms = transforms 10 | 11 | def __call__(self, image, *args): 12 | for transform in self.transforms: 13 | if args: 14 | image, *args = transform(image, *args) 15 | else: 16 | image = transform(image) 17 | if args: 18 | return (image,) + tuple(args) 19 | else: 20 | return image 21 | 22 | class Resize: 23 | """ 24 | params: 25 | size: (height, width) 26 | returns: 27 | resized image 28 | """ 29 | def __init__(self, size, interpolation=cv2.INTER_LANCZOS4): 30 | self.size = size 31 | self.interpolation = interpolation 32 | 33 | def __call__(self, image): 34 | dims = len(image.shape) 35 | image = cv2.resize( 36 | image, self.size[::-1], interpolation=self.interpolation) 37 | if len(image.shape) == dims -1: 38 | image = np.expand_dims(image, -1) 39 | return image 40 | 41 | 42 | 43 | class ResizeBatch: 44 | """ 45 | params: 46 | size: (height, width) 47 | returns: 48 | resized image 49 | """ 50 | def __init__(self, size, interpolation=cv2.INTER_LANCZOS4): 51 | self.size = size 52 | self.interpolation = interpolation 53 | self.resize_op = Resize(size, interpolation=interpolation) 54 | 55 | def __call__(self, images): 56 | resized_images = [] 57 | for i in range(images.shape[0]): 58 | resized_images.append(self.resize_op(images[i])) 59 | return np.stack(resized_images) 60 | 61 | 62 | class Binarize: 63 | """ 64 | params: 65 | threshold: binary threshold 66 | returns: 67 | binarized image 68 | """ 69 | def __init__(self, threshold): 70 | self.threshold = threshold 71 | 72 | def __call__(self, image): 73 | _, binarized_image = cv2.threshold( 74 | image, self.threshold, 255, cv2.THRESH_BINARY) 75 | return binarized_image 76 | 77 | 78 | class RandomContrast: 79 | """ 80 | params: 81 | contrast level: 82 | returns: 83 | image with changed contrast 84 | """ 85 | def __init__(self, low, high): 86 | self.low = low 87 | self.high = high 88 | 89 | def __call__(self, image): 90 | assert np.issubdtype(image.dtype, np.floating) 91 | contrast = np.random.uniform(self.low, self.high) 92 | image = image * contrast 93 | return np.clip(image, 0.0, 1.0) 94 | 95 | 96 | class RandomBrightness: 97 | """ 98 | params: 99 | brightness level 100 | returns: 101 | image with changed birghtness 102 | """ 103 | def __init__(self, low, high): 104 | self.low = low 105 | self.high = high 106 | 107 | def __call__(self, image): 108 | assert np.issubdtype(image.dtype, np.floating) 109 | brightness = np.random.uniform(self.low, self.high) 110 | image = image + brightness 111 | return np.clip(image, 0.0, 1.0) 112 | 113 | 114 | class RandomHueSaturation: 115 | """ 116 | params: 117 | saturation adjustment 118 | returns: 119 | saturated image 120 | """ 121 | def __init__(self, hue_delta=0.1, saturation_scale_range=(0.7, 1.3)): 122 | self.hue_delta = hue_delta 123 | self.saturation_scale_range = saturation_scale_range 124 | 125 | def _adjust_hue(self, hue, delta): 126 | hue += delta * 360 127 | hue[hue < 0] += 360 128 | hue[hue >= 360] -= 360 129 | return hue 130 | 131 | def __call__(self, image): 132 | assert np.issubdtype(image.dtype, np.floating) 133 | image = image.astype(np.float32) 134 | saturation = np.random.uniform(*self.saturation_scale_range) 135 | hue = np.random.uniform(-self.hue_delta, self.hue_delta) 136 | hsvimage = cv2.cvtColor(image, cv2.COLOR_RGB2HSV) 137 | (h, s, v) = cv2.split(hsvimage) 138 | s = np.clip(s * saturation, 0.0, 1.0) 139 | h = self._adjust_hue(h, hue) 140 | hsvimage = cv2.merge([h, s, v]) 141 | image = cv2.cvtColor(hsvimage, cv2.COLOR_HSV2RGB) 142 | image = np.clip(image, 0.0, 1.0) 143 | return image 144 | 145 | 146 | class RandomGammaCorrection: 147 | def __init__(self, min_gamma=0.75, max_gamma=1.25): 148 | self.min_gamma = min_gamma 149 | self.max_gamma = max_gamma 150 | 151 | def __call__(self, image): 152 | if np.issubdtype(image.dtype, np.floating): 153 | image = (image * 255).astype(np.uint8) 154 | 155 | gamma = np.random.uniform(self.min_gamma, self.max_gamma) 156 | inv_gamma = 1.0 / gamma 157 | table = np.array([((i / 255.0) ** inv_gamma) * 255 158 | for i in np.arange(0, 256)]).astype("uint8") 159 | 160 | image = cv2.LUT(image, table) 161 | image = image / 255. 162 | image = np.clip(image, 0.0, 1.0) 163 | return image 164 | 165 | class RandomHorizontalFlip: 166 | """ 167 | params: 168 | Probability to flip the image horizontally 169 | returns: 170 | Maybe flipped image 171 | """ 172 | def __init__(self, flip_probability=0.5): 173 | self.flip_probability = flip_probability 174 | 175 | def __call__(self, image, mask): 176 | if np.random.random() < self.flip_probability: 177 | image = cv2.flip(image, 1) 178 | mask = cv2.flip(mask, 1) 179 | return image, mask 180 | 181 | 182 | class RandomRotate: 183 | def __init__(self, max_delta=5.0): 184 | self.max_delta = max_delta 185 | 186 | def __call__(self, image, mask): 187 | angle = np.random.uniform(-self.max_delta, self.max_delta) 188 | image = self._rotate(image, angle, interpolation=cv2.INTER_LINEAR) 189 | mask = self._rotate(mask, angle, interpolation=cv2.INTER_NEAREST) 190 | return image, mask 191 | 192 | def _rotate(self, image, angle, interpolation=cv2.INTER_LANCZOS4): 193 | image_center = (image.shape[1] / 2, image.shape[0] / 2) 194 | rot_mat = cv2.getRotationMatrix2D(image_center, angle, 1.0) 195 | result = cv2.warpAffine( 196 | image, rot_mat, (image.shape[1], image.shape[0]), flags=interpolation, 197 | borderMode=cv2.BORDER_REPLICATE) 198 | return result 199 | 200 | 201 | class RandomTranslate: 202 | def __init__(self, max_percent_delta=0.02): 203 | self.max_percent_delta = max_percent_delta 204 | 205 | def __call__(self, image, mask): 206 | delta_x = np.random.uniform(-self.max_percent_delta, self.max_percent_delta) 207 | delta_y = np.random.uniform(-self.max_percent_delta, self.max_percent_delta) 208 | image = self._translate(image, delta_x, delta_y) 209 | mask = self._translate(mask, delta_x, delta_y) 210 | return image, mask 211 | 212 | def _translate(self, image, delta_x, delta_y): 213 | height, width = image.shape[:2] 214 | translation_matrix = np.float32([ 215 | [1, 0, int(width * delta_x)], 216 | [0, 1, int(height * delta_y)] 217 | ]) 218 | translated_image = cv2.warpAffine( 219 | image, translation_matrix, (width, height), 220 | borderMode=cv2.BORDER_REPLICATE 221 | ) 222 | return translated_image 223 | 224 | 225 | class RandomCropFixedSize: 226 | def __init__(self, size): 227 | self.height, self.width = size 228 | 229 | def __call__(self, image, mask): 230 | assert image.shape[:2] == mask.shape[:2] 231 | assert image.shape[0] >= self.height and image.shape[1] >= self.width 232 | scol = np.random.randint(0, image.shape[0] - self.height + 1) 233 | srow = np.random.randint(0, image.shape[1] - self.width + 1) 234 | image_crop = image[scol:scol+self.height, srow:srow+self.width] 235 | mask_crop = mask[scol:scol+self.height, srow:srow+self.width] 236 | return image_crop, mask_crop 237 | 238 | 239 | class RandomCrop: 240 | def __init__(self, scale_height=(0.8, 1.0), scale_width=(0.8, 1.0)): 241 | self.scale_height = scale_height 242 | self.scale_width = scale_width 243 | 244 | def __call__(self, image, mask): 245 | assert image.shape[:2] == mask.shape[:2] 246 | sh = np.random.uniform(*self.scale_height) 247 | sw = np.random.uniform(*self.scale_width) 248 | crop_height = int(sh * image.shape[0]) 249 | crop_width = int(sw * image.shape[1]) 250 | return RandomCropFixedSize((crop_height, crop_width))(image, mask) 251 | 252 | 253 | class RandomScale: 254 | def __init__(self, scale_height=(0.8, 1.2), scale_width=(0.8, 1.2), aspect_range=[0.8, 1.2]): 255 | self.scale_height = scale_height 256 | self.scale_width = scale_width 257 | self.aspect_range = aspect_range 258 | 259 | def __call__(self, image, mask): 260 | hs = np.random.uniform(*self.scale_height) 261 | height = int(image.shape[0] * hs) 262 | lw = max(hs * self.aspect_range[0], self.scale_width[0]) 263 | hw = min(hs * self.aspect_range[1], self.scale_width[1]) 264 | ws = np.random.uniform(lw, hw) 265 | width = int(image.shape[1] * ws) 266 | return ( 267 | Resize((height, width))(image), 268 | Resize((height, width), cv2.INTER_NEAREST)(mask) 269 | ) 270 | 271 | 272 | class ScaleTo: 273 | def __init__(self, size, interpolation=cv2.INTER_LANCZOS4): 274 | self.height, self.width = size 275 | self.interpolation = interpolation 276 | 277 | def __call__(self, image, mask): 278 | scale = min(self.height / image.shape[0], self.width / image.shape[1]) 279 | height = int(scale * image.shape[0]) 280 | width = int(scale * image.shape[1]) 281 | return ( 282 | cv2.resize(image, (width, height), 283 | interpolation=self.interpolation), 284 | cv2.resize(mask, (width, height), 285 | interpolation=cv2.INTER_NEAREST) 286 | ) 287 | 288 | 289 | class Concat: 290 | def __init__(self, transforms): 291 | self.transforms = transforms 292 | 293 | def __call__(self, *args): 294 | assert len(args) == len(self.transforms) 295 | return [ 296 | tr(x) for tr, x in zip(self.transforms, args) 297 | ] 298 | 299 | class PadTo: 300 | def __init__(self, size, pad_value=0): 301 | self.height, self.width = size 302 | self.pad_value = pad_value 303 | 304 | def __call__(self, image): 305 | hpad = max(0, self.height - image.shape[0]) 306 | wpad = max(0, self.width - image.shape[1]) 307 | top_pad = hpad // 2 308 | bottom_pad = hpad // 2 + hpad % 2 309 | left_pad = wpad // 2 310 | right_pad = wpad // 2 + wpad % 2 311 | image = cv2.copyMakeBorder( 312 | image, top_pad, bottom_pad, left_pad, 313 | right_pad, cv2.BORDER_CONSTANT, value=self.pad_value 314 | ) 315 | return image 316 | 317 | 318 | class Choice: 319 | def __init__(self, choices, p=None): 320 | self.choices = choices 321 | self.p = p 322 | if not self.p: 323 | self.p = np.ones(len(self.choices)) / len(self.choices) 324 | 325 | def __call__(self, *args): 326 | c = np.random.choice(self.choices, p=self.p) 327 | return c(*args) 328 | 329 | 330 | class RandomPerspective: 331 | def __init__(self, corner_offset=0.15): 332 | self.corner_offset = corner_offset 333 | 334 | def __call__(self, image, mask): 335 | height, width = image.shape[:2] 336 | lh = int(self.corner_offset * height) 337 | hh = int(height - self.corner_offset * height) 338 | lw = int(self.corner_offset * width) 339 | hw = int(width - self.corner_offset * width) 340 | x1, x3 = np.random.randint(0, lw, [2]) 341 | y1, y2 = np.random.randint(0, lh, [2]) 342 | x2, x4 = np.random.randint(hw, width, [2]) 343 | y3, y4 = np.random.randint(hh, height, [2]) 344 | mapped = np.float32([ 345 | [x1, y1], 346 | [x2, y2], 347 | [x3, y3], 348 | [x4, y4] 349 | ]) 350 | corners = np.float32([ 351 | [0, 0], 352 | [width, 0], 353 | [0, height], 354 | [width, height] 355 | ]) 356 | M = cv2.getPerspectiveTransform(mapped, corners) 357 | image = cv2.warpPerspective( 358 | image, M, (width, height), borderMode=cv2.BORDER_REFLECT) 359 | mask = cv2.warpPerspective( 360 | mask, M, (width, height), flags=cv2.INTER_NEAREST, 361 | borderMode=cv2.BORDER_REFLECT) 362 | 363 | return image, mask 364 | 365 | class RandomShear: 366 | def __init__(self, max_range=0.01): 367 | self.max_range = max_range 368 | 369 | def __call__(self, image, mask): 370 | height, width = image.shape[:2] 371 | cx = width / 2 372 | cy = height / 2 373 | pts = np.float32([ 374 | [cx, cy], 375 | [cx + 0.1 * width, cy], 376 | [cx, cy + 0.1 * height] 377 | ]) 378 | ru = lambda x: x * np.random.uniform( 379 | -self.max_range, self.max_range) 380 | 381 | mapped = np.float32([ 382 | [pts[0][0] + ru(width), pts[0][1]], 383 | [pts[1][0] + ru(width), pts[1][1] + ru(height)], 384 | [pts[2][0], pts[2][1] + ru(height)] 385 | ]) 386 | 387 | M = cv2.getAffineTransform(pts, mapped) 388 | 389 | image = cv2.warpAffine( 390 | image, M, (width, height), borderMode=cv2.BORDER_REFLECT) 391 | mask = cv2.warpAffine( 392 | mask, M, (width, height), flags=cv2.INTER_NEAREST, borderMode=cv2.BORDER_REFLECT) 393 | 394 | return image, mask 395 | 396 | 397 | class RandomElasticTransform: 398 | """ 399 | Adapted from 400 | https://www.kaggle.com/bguberfain/elastic-transform-for-data-augmentation 401 | 402 | """ 403 | def __init__(self, sigma=(0.07, 0.15), alpha=(3, 3.5)): 404 | self.sigma = sigma 405 | self.alpha = alpha 406 | 407 | def __call__(self, image, mask): 408 | height, width = image.shape[:2] 409 | sigma = np.random.uniform(*self.sigma) 410 | x_sigma, y_sigma = sigma * width, sigma * height 411 | alpha = np.random.uniform(*self.alpha) 412 | x_alpha, y_alpha = alpha * width, alpha * height 413 | x_, y_ = np.meshgrid(np.arange(0, width), np.arange(0, height), indexing='xy') 414 | dx = gaussian_filter((np.random.rand(height, width) * 2 - 1), x_sigma) * x_alpha 415 | dy = gaussian_filter((np.random.rand(height, width) * 2 - 1), y_sigma) * y_alpha 416 | x = (x_ + dx).astype(np.float32) 417 | y = (y_ + dy).astype(np.float32) 418 | image = cv2.remap( 419 | image, x, y, interpolation=cv2.INTER_AREA, 420 | borderMode=cv2.BORDER_REFLECT) 421 | mask = cv2.remap( 422 | mask, x, y, interpolation=cv2.INTER_NEAREST, 423 | borderMode=cv2.BORDER_REFLECT) 424 | return image, mask 425 | 426 | 427 | class RandomGaussianBlur: 428 | def __init__(self, kernel_sizes=(3, 5, 7), apply_prob=0.2): 429 | self.kernel_sizes = kernel_sizes 430 | self.apply_prob = apply_prob 431 | 432 | def __call__(self, image): 433 | if np.random.rand() < self.apply_prob: 434 | ksize = np.random.choice(self.kernel_sizes) 435 | image = cv2.GaussianBlur(image, (ksize,) * 2, 0) 436 | return image 437 | -------------------------------------------------------------------------------- /pysemseg/utils.py: -------------------------------------------------------------------------------- 1 | import shutil 2 | import sys 3 | import os 4 | import glob 5 | import re 6 | import numpy as np 7 | import torch 8 | import importlib 9 | 10 | def save(model, optimizer, lr_scheduler, model_dir, 11 | in_channels, n_classes, epoch, train_args): 12 | save_dict = { 13 | 'model': model.state_dict(), 14 | 'model_args': { 15 | 'in_channels': in_channels, 16 | 'n_classes': n_classes, 17 | **train_args.model_args, 18 | }, 19 | 'transformer_args': train_args.transformer_args, 20 | 'epoch': epoch, 21 | 'optimizer': optimizer.state_dict(), 22 | 'lr_scheduler': lr_scheduler.state_dict(), 23 | 'train_args': train_args.__dict__ 24 | } 25 | torch.save( 26 | save_dict, 27 | os.path.join(model_dir, 'checkpoint-{}'.format(epoch))) 28 | 29 | 30 | def prompt_delete_dir(directory): 31 | if os.path.exists(directory): 32 | answer = input( 33 | "{} exists. Do you want to delete it?[y/n]".format(directory)) 34 | if answer == 'y': 35 | shutil.rmtree(directory) 36 | elif answer != 'n': 37 | sys.exit(1) 38 | 39 | 40 | def restore(checkpoint_path, model, optimizer=None, lr_scheduler=None, 41 | restore_cpu=False, strict=True, continue_training=False): 42 | checkpoint = torch.load( 43 | checkpoint_path, 44 | map_location=lambda storage, location: storage if restore_cpu else None) 45 | checkpoint['model'] = { 46 | re.sub('^module[.]', '', key): value 47 | for key, value in checkpoint['model'].items() 48 | } 49 | if not strict: 50 | msd = model.state_dict() 51 | checkpoint['model'] = { 52 | k: v for k, v in checkpoint['model'].items() 53 | if k in msd and msd[k].shape == v.shape 54 | } 55 | print("Loading {} parameters".format(len(checkpoint['model']))) 56 | 57 | model.load_state_dict(checkpoint['model'], strict=strict) 58 | if optimizer is not None and continue_training: 59 | optimizer.load_state_dict(checkpoint['optimizer']) 60 | if (lr_scheduler is not None and 'lr_scheduler' in checkpoint and 61 | continue_training): 62 | lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) 63 | return checkpoint['epoch'] 64 | 65 | 66 | def get_latest_checkpoint(model_dir): 67 | def key_fn(path): 68 | matches = re.match(r'checkpoint-(\d+)', os.path.basename(path)) 69 | return int(matches.groups()[0]) 70 | return sorted( 71 | glob.glob(os.path.join(model_dir, 'checkpoint-*')), 72 | key=key_fn 73 | )[-1] 74 | 75 | 76 | def tensor_to_numpy(tensor): 77 | if tensor.is_cuda: 78 | tensor = tensor.cpu() 79 | if isinstance(tensor, torch.autograd.Variable): 80 | tensor = tensor.data 81 | return tensor.numpy() 82 | 83 | 84 | def import_type(name, modules=[]): 85 | type_paths = [name] + [m + '.' + name for m in modules] 86 | for tp in type_paths: 87 | try: 88 | components = tp.split('.') 89 | mod = importlib.import_module(components[0]) 90 | for comp in components[1:]: 91 | mod = getattr(mod, comp) 92 | return mod 93 | except: 94 | pass 95 | raise ImportError(name + ' not found') 96 | 97 | 98 | def flatten_dict(dict_obj): 99 | result = {} 100 | for key, value in dict_obj.items(): 101 | if isinstance(value, dict): 102 | flattened = flatten_dict(value) 103 | result.update({ 104 | "{}/{}".format(key, kk): fv 105 | for kk, fv in flattened.items()}) 106 | else: 107 | result[key] = value 108 | return result 109 | 110 | 111 | def _get_palette_map(n_classes): 112 | def bitget(byteval, idx): 113 | return (byteval & (1 << idx)) != 0 114 | color_to_label = {} 115 | for k in range(0, n_classes): 116 | red = green = blue = 0 117 | cls = k 118 | for j in range(8): 119 | red = red | (bitget(cls, 0) << 7 - j) 120 | green = green | (bitget(cls, 1) << 7 - j) 121 | blue = blue | (bitget(cls, 2) << 7 - j) 122 | cls = cls >> 3 123 | color_to_label[(red, green, blue)] = k 124 | return color_to_label 125 | 126 | 127 | class ColorPalette: 128 | def __init__(self, colors): 129 | if isinstance(colors, int): 130 | self.color_to_label = _get_palette_map(colors) 131 | else: 132 | self.color_to_label = {v: k for k, v in enumerate(colors)} 133 | self.label_to_color = {v: k for k, v in self.color_to_label.items()} 134 | 135 | def encode_color(self, label_mask): 136 | """ 137 | Encodes a label mask with its RGB color representation 138 | Arguments: 139 | label_mask: A numpy array with dimensions either 140 | (height, width), (height, width, 1), (batch_size, height, width) 141 | or (batch_size, height, width, 1) 142 | 143 | Returns: Color encoded representation 144 | """ 145 | input_shape = label_mask.shape 146 | if len(input_shape) == 2: 147 | label_mask = np.reshape(label_mask, [1, *input_shape, 1]) 148 | output_shape = input_shape + (3,) 149 | elif len(input_shape) == 3 and input_shape[-1] == 1: 150 | label_mask = np.expand_dims(label_mask, axis=0) 151 | output_shape = input_shape[:2] + (3,) 152 | elif len(input_shape) == 3: 153 | label_mask = np.expand_dims(label_mask, axis=-1) 154 | output_shape = input_shape + (3,) 155 | elif len(input_shape) == 4: 156 | output_shape = input_shape[:3] + (3,) 157 | 158 | palette_mask = np.zeros(label_mask.shape[:3] + (3,), dtype=np.uint8) 159 | for i, mask in self.label_to_color.items(): 160 | palette_mask[np.where(np.all(label_mask == i, axis=-1))[:3]] = mask 161 | 162 | palette_mask = palette_mask.reshape(output_shape) 163 | 164 | return palette_mask 165 | 166 | 167 | def decode_color(self, palette_mask): 168 | label_mask = np.zeros(palette_mask.shape[:2], dtype=np.uint8) 169 | for mask, i in self.color_to_label.items(): 170 | label_mask[np.where(np.all(palette_mask == mask, axis=-1))[:2]] = i 171 | return label_mask 172 | 173 | 174 | def load_model( 175 | checkpoint_path, model_cls, transformer_cls, 176 | model_args={}, transformer_args={}, device=None): 177 | checkpoint = torch.load( 178 | checkpoint_path, 179 | map_location=lambda storage, location: storage) 180 | model = model_cls(**{**checkpoint.get('model_args', {}), **model_args}) 181 | if device is not None: 182 | model = model.to(device) 183 | transformer = transformer_cls( 184 | 'test', 185 | **{**checkpoint.get('transformer_args', {}), **transformer_args} 186 | ) 187 | checkpoint['model'] = { 188 | re.sub('^module[.]', '', key): value 189 | for key, value in checkpoint['model'].items() 190 | } 191 | model.load_state_dict(checkpoint['model'], strict=True) 192 | model.eval() 193 | def predict(input_tensor): 194 | with torch.no_grad(): 195 | transformed_image = transformer( 196 | input_tensor, 197 | torch.zeros(input_tensor.shape[1:], dtype=torch.long) 198 | )[0].unsqueeze(0) 199 | if device is not None: 200 | transformed_image = transformed_image.to(device) 201 | return model(transformed_image) 202 | return predict 203 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | pytorh>=0.4.4 2 | tqdm>=4.19.5 3 | tensorboardX>=1.1 4 | ipython>=6.2.1 5 | opencv-python>=3.4.0.12 6 | Pillow>=5.0.0 7 | numpy>=1.14.0 8 | scikit-image>=0.13.1 9 | visdom>=0.1.8.5 10 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | import os 2 | import io 3 | import re 4 | from setuptools import setup, find_packages 5 | 6 | 7 | def read(*names, **kwargs): 8 | with io.open(os.path.join(os.path.dirname(__file__), *names), 9 | encoding=kwargs.get("encoding", "utf8")) as fp: 10 | return fp.read() 11 | 12 | 13 | readme = read('README.rst') 14 | 15 | VERSION = '0.1.1' 16 | 17 | requirements = [ 18 | "torch==1.0.0", 19 | "torchvision==0.2.1", 20 | "tqdm>=4.19.5", 21 | "tensorboardX>=1.1", 22 | "ipython>=6.2.1", 23 | "opencv-python>=3.4.0.12", 24 | "Pillow>=5.0.0", 25 | "numpy>=1.14.0", 26 | "scikit-image>=0.13.1", 27 | "visdom>=0.1.8.5" 28 | "cython==0.29.1", 29 | "pycocotools==2.0.0", 30 | "ConfigArgParse==0.13.0" 31 | ] 32 | 33 | setup( 34 | # Metadata 35 | name='pysemseg', 36 | version=VERSION, 37 | author='Petko Nikolov', 38 | author_email='py.nikolov@gmail.com', 39 | url='https://github.com/petko-nikolov/pysemseg', 40 | description='Pytorch library for training Deep Learning models for Semantic Segmentation', 41 | long_description=readme, 42 | license='BSD', 43 | 44 | # Package info 45 | packages=find_packages(), 46 | 47 | zip_safe=True, 48 | install_requires=requirements, 49 | entry_points={ 50 | 'console_scripts': [ 51 | 'pysemseg-train=pysemseg.train:main' 52 | ] 53 | } 54 | ) 55 | --------------------------------------------------------------------------------