├── .gitignore
├── .idea
    └── workspace.xml
├── README.md
├── config.py
├── data.py
├── emb_build
    ├── read_csv.py
    └── tran_emb.py
├── fine_tune.py
├── gen_result.py
├── main.py
├── main_boost.py
├── models
    ├── AttLSTM.py
    ├── BasicModule.py
    ├── FastText.py
    ├── GRU.py
    ├── Inception.py
    ├── LSTM.py
    ├── RCNN.py
    ├── RCNN1.py
    ├── TextCNN.py
    ├── __init__.py
    ├── bigru_att.py
    └── bilstm_conv.py
├── result
    └── ensemble.py
├── script
    ├── boost.sh
    ├── run.sh
    ├── tune.sh
    └── val_gen.sh
├── test_ensemble.py
├── util
    └── split_val.py
└── val_result
    ├── avg_ens.py
    └── ensemble.py


/.gitignore:
--------------------------------------------------------------------------------
  1 | # add by myself
  2 | .vector_cache/
  3 | .data/
  4 | emb-100.txt
  5 | val_set.csv
  6 | snapshot/
  7 | .idea/
  8 | .idea/workspace.xml
  9 | emb/
 10 | dataset/
 11 | *.txt
 12 | 
 13 | # Byte-compiled / optimized / DLL files
 14 | __pycache__/
 15 | *.py[cod]
 16 | *$py.class
 17 | 
 18 | # C extensions
 19 | *.so
 20 | 
 21 | # Distribution / packaging
 22 | .Python
 23 | build/
 24 | develop-eggs/
 25 | dist/
 26 | downloads/
 27 | eggs/
 28 | .eggs/
 29 | lib/
 30 | lib64/
 31 | parts/
 32 | sdist/
 33 | var/
 34 | wheels/
 35 | *.egg-info/
 36 | .installed.cfg
 37 | *.egg
 38 | MANIFEST
 39 | 
 40 | # PyInstaller
 41 | #  Usually these files are written by a python script from a template
 42 | #  before PyInstaller builds the exe, so as to inject date/other infos into it.
 43 | *.manifest
 44 | *.spec
 45 | 
 46 | # Installer logs
 47 | pip-log.txt
 48 | pip-delete-this-directory.txt
 49 | 
 50 | # Unit test / coverage reports
 51 | htmlcov/
 52 | .tox/
 53 | .coverage
 54 | .coverage.*
 55 | .cache
 56 | nosetests.xml
 57 | coverage.xml
 58 | *.cover
 59 | .hypothesis/
 60 | .pytest_cache/
 61 | 
 62 | # Translations
 63 | *.mo
 64 | *.pot
 65 | 
 66 | # Django stuff:
 67 | *.log
 68 | local_settings.py
 69 | db.sqlite3
 70 | 
 71 | # Flask stuff:
 72 | instance/
 73 | .webassets-cache
 74 | 
 75 | # Scrapy stuff:
 76 | .scrapy
 77 | 
 78 | # Sphinx documentation
 79 | docs/_build/
 80 | 
 81 | # PyBuilder
 82 | target/
 83 | 
 84 | # Jupyter Notebook
 85 | .ipynb_checkpoints
 86 | 
 87 | # pyenv
 88 | .python-version
 89 | 
 90 | # celery beat schedule file
 91 | celerybeat-schedule
 92 | 
 93 | # SageMath parsed files
 94 | *.sage.py
 95 | 
 96 | # Environments
 97 | .env
 98 | .venv
 99 | env/
100 | venv/
101 | ENV/
102 | env.bak/
103 | venv.bak/
104 | 
105 | # Spyder project settings
106 | .spyderproject
107 | .spyproject
108 | 
109 | # Rope project settings
110 | .ropeproject
111 | 
112 | # mkdocs documentation
113 | /site
114 | 
115 | # mypy
116 | .mypy_cache/
117 | 


--------------------------------------------------------------------------------
/.idea/workspace.xml:
--------------------------------------------------------------------------------
  1 | <?xml version="1.0" encoding="UTF-8"?>
  2 | <project version="4">
  3 |   <component name="ChangeListManager">
  4 |     <list default="true" id="fff4d958-5015-4c87-83d4-c1dc1da633cf" name="Default" comment="">
  5 |       <change beforePath="$PROJECT_DIR$/.idea/data_grand.iml" beforeDir="false" />
  6 |       <change beforePath="$PROJECT_DIR$/.idea/inspectionProfiles/Project_Default.xml" beforeDir="false" />
  7 |       <change beforePath="$PROJECT_DIR$/.idea/inspectionProfiles/profiles_settings.xml" beforeDir="false" />
  8 |       <change beforePath="$PROJECT_DIR$/.idea/markdown-navigator.xml" beforeDir="false" />
  9 |       <change beforePath="$PROJECT_DIR$/.idea/markdown-navigator/profiles_settings.xml" beforeDir="false" />
 10 |       <change beforePath="$PROJECT_DIR$/.idea/misc.xml" beforeDir="false" />
 11 |       <change beforePath="$PROJECT_DIR$/.idea/modules.xml" beforeDir="false" />
 12 |       <change beforePath="$PROJECT_DIR$/.idea/vcs.xml" beforeDir="false" />
 13 |       <change beforePath="$PROJECT_DIR$/.idea/workspace.xml" beforeDir="false" />
 14 |     </list>
 15 |     <option name="EXCLUDED_CONVERTED_TO_IGNORED" value="true" />
 16 |     <option name="SHOW_DIALOG" value="false" />
 17 |     <option name="HIGHLIGHT_CONFLICTS" value="true" />
 18 |     <option name="HIGHLIGHT_NON_ACTIVE_CHANGELIST" value="false" />
 19 |     <option name="LAST_RESOLUTION" value="IGNORE" />
 20 |   </component>
 21 |   <component name="CreatePatchCommitExecutor">
 22 |     <option name="PATCH_PATH" value="" />
 23 |   </component>
 24 |   <component name="FUSProjectUsageTrigger">
 25 |     <session id="-1872809769">
 26 |       <usages-collector id="statistics.lifecycle.project">
 27 |         <counts>
 28 |           <entry key="project.closed" value="17" />
 29 |           <entry key="project.open.time.0" value="3" />
 30 |           <entry key="project.open.time.1" value="4" />
 31 |           <entry key="project.open.time.10" value="1" />
 32 |           <entry key="project.open.time.11" value="7" />
 33 |           <entry key="project.open.time.12" value="1" />
 34 |           <entry key="project.open.time.13" value="1" />
 35 |           <entry key="project.open.time.15" value="1" />
 36 |           <entry key="project.open.time.6" value="1" />
 37 |           <entry key="project.opened" value="19" />
 38 |         </counts>
 39 |       </usages-collector>
 40 |       <usages-collector id="statistics.file.extensions.edit">
 41 |         <counts>
 42 |           <entry key="gitignore" value="8" />
 43 |           <entry key="py" value="14829" />
 44 |           <entry key="sh" value="86" />
 45 |         </counts>
 46 |       </usages-collector>
 47 |       <usages-collector id="statistics.file.types.edit">
 48 |         <counts>
 49 |           <entry key="PLAIN_TEXT" value="94" />
 50 |           <entry key="Python" value="14829" />
 51 |         </counts>
 52 |       </usages-collector>
 53 |       <usages-collector id="statistics.file.extensions.open">
 54 |         <counts>
 55 |           <entry key="gitignore" value="2" />
 56 |           <entry key="py" value="173" />
 57 |           <entry key="pyi" value="1" />
 58 |           <entry key="sh" value="2" />
 59 |         </counts>
 60 |       </usages-collector>
 61 |       <usages-collector id="statistics.file.types.open">
 62 |         <counts>
 63 |           <entry key="PLAIN_TEXT" value="4" />
 64 |           <entry key="Python" value="173" />
 65 |           <entry key="PythonStub" value="1" />
 66 |         </counts>
 67 |       </usages-collector>
 68 |     </session>
 69 |   </component>
 70 |   <component name="FavoritesManager">
 71 |     <favorites_list name="data_grand" />
 72 |   </component>
 73 |   <component name="FileEditorManager">
 74 |     <leaf SIDE_TABS_SIZE_LIMIT_KEY="300">
 75 |       <file pinned="false" current-in-tab="false">
 76 |         <entry file="file://$PROJECT_DIR$/models/BasicModule.py">
 77 |           <provider selected="true" editor-type-id="text-editor">
 78 |             <state relative-caret-position="1170">
 79 |               <caret line="46" selection-start-line="46" selection-end-line="46" />
 80 |             </state>
 81 |           </provider>
 82 |         </entry>
 83 |       </file>
 84 |       <file pinned="false" current-in-tab="false">
 85 |         <entry file="file://$PROJECT_DIR$/util/split_val.py">
 86 |           <provider selected="true" editor-type-id="text-editor">
 87 |             <state relative-caret-position="416">
 88 |               <caret line="18" selection-start-line="18" selection-end-line="18" />
 89 |               <folding>
 90 |                 <element signature="e#0#19#0" expanded="true" />
 91 |               </folding>
 92 |             </state>
 93 |           </provider>
 94 |         </entry>
 95 |       </file>
 96 |       <file pinned="false" current-in-tab="false">
 97 |         <entry file="file://$PROJECT_DIR$/data.py">
 98 |           <provider selected="true" editor-type-id="text-editor">
 99 |             <state relative-caret-position="754">
100 |               <caret line="36" column="24" selection-start-line="36" selection-start-column="24" selection-end-line="36" selection-end-column="24" />
101 |               <folding>
102 |                 <element signature="e#105#131#0" expanded="true" />
103 |               </folding>
104 |             </state>
105 |           </provider>
106 |         </entry>
107 |       </file>
108 |       <file pinned="false" current-in-tab="false">
109 |         <entry file="file://$PROJECT_DIR$/main.py">
110 |           <provider selected="true" editor-type-id="text-editor">
111 |             <state relative-caret-position="390">
112 |               <caret line="25" column="16" selection-start-line="25" selection-start-column="16" selection-end-line="25" selection-end-column="16" />
113 |               <folding>
114 |                 <element signature="e#98#110#0" expanded="true" />
115 |               </folding>
116 |             </state>
117 |           </provider>
118 |         </entry>
119 |       </file>
120 |       <file pinned="false" current-in-tab="false">
121 |         <entry file="file://$PROJECT_DIR$/main_boost.py">
122 |           <provider selected="true" editor-type-id="text-editor">
123 |             <state relative-caret-position="234">
124 |               <caret line="9" column="11" selection-start-line="9" selection-start-column="11" selection-end-line="9" selection-end-column="11" />
125 |             </state>
126 |           </provider>
127 |         </entry>
128 |       </file>
129 |       <file pinned="false" current-in-tab="true">
130 |         <entry file="file://$PROJECT_DIR$/.gitignore">
131 |           <provider selected="true" editor-type-id="text-editor">
132 |             <state relative-caret-position="260">
133 |               <caret line="10" column="5" selection-start-line="10" selection-start-column="5" selection-end-line="10" selection-end-column="5" />
134 |             </state>
135 |           </provider>
136 |         </entry>
137 |       </file>
138 |       <file pinned="false" current-in-tab="false">
139 |         <entry file="file://$PROJECT_DIR$/config.py">
140 |           <provider selected="true" editor-type-id="text-editor">
141 |             <state relative-caret-position="520">
142 |               <caret line="20" column="18" selection-start-line="20" selection-start-column="18" selection-end-line="20" selection-end-column="18" />
143 |             </state>
144 |           </provider>
145 |         </entry>
146 |       </file>
147 |       <file pinned="false" current-in-tab="false">
148 |         <entry file="file://$PROJECT_DIR$/test_ensemble.py">
149 |           <provider selected="true" editor-type-id="text-editor">
150 |             <state relative-caret-position="260">
151 |               <caret line="10" column="11" selection-start-line="10" selection-start-column="11" selection-end-line="10" selection-end-column="11" />
152 |             </state>
153 |           </provider>
154 |         </entry>
155 |       </file>
156 |       <file pinned="false" current-in-tab="false">
157 |         <entry file="file://$PROJECT_DIR$/fine_tune.py">
158 |           <provider selected="true" editor-type-id="text-editor">
159 |             <state relative-caret-position="312">
160 |               <caret line="12" column="11" selection-start-line="12" selection-start-column="11" selection-end-line="12" selection-end-column="11" />
161 |             </state>
162 |           </provider>
163 |         </entry>
164 |       </file>
165 |       <file pinned="false" current-in-tab="false">
166 |         <entry file="file://$PROJECT_DIR$/gen_result.py">
167 |           <provider selected="true" editor-type-id="text-editor">
168 |             <state relative-caret-position="208">
169 |               <caret line="15" column="19" selection-start-line="15" selection-start-column="19" selection-end-line="15" selection-end-column="19" />
170 |               <folding>
171 |                 <element signature="e#105#117#0" expanded="true" />
172 |               </folding>
173 |             </state>
174 |           </provider>
175 |         </entry>
176 |       </file>
177 |     </leaf>
178 |   </component>
179 |   <component name="FileTemplateManagerImpl">
180 |     <option name="RECENT_TEMPLATES">
181 |       <list>
182 |         <option value="Python Script" />
183 |       </list>
184 |     </option>
185 |   </component>
186 |   <component name="FindInProjectRecents">
187 |     <findStrings>
188 |       <find>conv</find>
189 |       <find>fix_length</find>
190 |       <find>save_dir</find>
191 |       <find>use_gp</find>
192 |       <find>vola</find>
193 |       <find>self</find>
194 |       <find>device</find>
195 |       <find>sort_key</find>
196 |       <find>VOcab</find>
197 |       <find>cuda</find>
198 |       <find>Vocab</find>
199 |       <find>self.analyzer</find>
200 |       <find>resul</find>
201 |       <find>pth</find>
202 |       <find>args</find>
203 |       <find>opt</find>
204 |       <find>yie</find>
205 |       <find>repeat</find>
206 |       <find>save</find>
207 |       <find>boost</find>
208 |       <find>weight</find>
209 |       <find>.idea</find>
210 |     </findStrings>
211 |   </component>
212 |   <component name="Git.Settings">
213 |     <option name="RECENT_GIT_ROOT_PATH" value="$PROJECT_DIR$" />
214 |   </component>
215 |   <component name="IdeDocumentHistory">
216 |     <option name="CHANGED_PATHS">
217 |       <list>
218 |         <option value="$PROJECT_DIR$/baseline.py" />
219 |         <option value="$PROJECT_DIR$/models/bilstm_conv.py" />
220 |         <option value="$PROJECT_DIR$/MLP/MLP.py" />
221 |         <option value="$PROJECT_DIR$/MLP/main.py" />
222 |         <option value="$PROJECT_DIR$/models/BiLSTM.py" />
223 |         <option value="$PROJECT_DIR$/models/AttLSTM.py" />
224 |         <option value="$PROJECT_DIR$/models/BasicModule.py" />
225 |         <option value="$PROJECT_DIR$/run.sh" />
226 |         <option value="$PROJECT_DIR$/models/LSTM.py" />
227 |         <option value="$PROJECT_DIR$/util_new.py" />
228 |         <option value="$PROJECT_DIR$/data_process.py" />
229 |         <option value="$PROJECT_DIR$/main_new.py" />
230 |         <option value="$PROJECT_DIR$/ensemble.py" />
231 |         <option value="$PROJECT_DIR$/models/GRU.py" />
232 |         <option value="$PROJECT_DIR$/models/TextCNN.py" />
233 |         <option value="$PROJECT_DIR$/models/RCNN.py" />
234 |         <option value="$PROJECT_DIR$/models/FastText.py" />
235 |         <option value="$PROJECT_DIR$/models/RCNN1.py" />
236 |         <option value="$PROJECT_DIR$/main_nllloss.py" />
237 |         <option value="$PROJECT_DIR$/util.py" />
238 |         <option value="$PROJECT_DIR$/models/Inception.py" />
239 |         <option value="$PROJECT_DIR$/main_boost (2).py" />
240 |         <option value="$PROJECT_DIR$/config.py" />
241 |         <option value="$PROJECT_DIR$/models/__init__.py" />
242 |         <option value="$PROJECT_DIR$/emb_build/tran_emb.py" />
243 |         <option value="$PROJECT_DIR$/main.py" />
244 |         <option value="$PROJECT_DIR$/main_boost.py" />
245 |         <option value="$PROJECT_DIR$/test_ensemble.py" />
246 |         <option value="$PROJECT_DIR$/fine_tune.py" />
247 |         <option value="$PROJECT_DIR$/gen_result.py" />
248 |         <option value="$PROJECT_DIR$/.gitignore" />
249 |       </list>
250 |     </option>
251 |   </component>
252 |   <component name="ProjectFrameBounds" extendedState="6">
253 |     <option name="x" value="100" />
254 |     <option name="y" value="107" />
255 |     <option name="width" value="1863" />
256 |     <option name="height" value="1093" />
257 |   </component>
258 |   <component name="ProjectLevelVcsManager">
259 |     <ConfirmationsSetting value="1" id="Add" />
260 |   </component>
261 |   <component name="ProjectView">
262 |     <navigator proportions="" version="1">
263 |       <foldersAlwaysOnTop value="true" />
264 |     </navigator>
265 |     <panes>
266 |       <pane id="ProjectPane">
267 |         <subPane>
268 |           <expand>
269 |             <path>
270 |               <item name="data_grand" type="b2602c69:ProjectViewProjectNode" />
271 |               <item name="data_grand" type="462c0819:PsiDirectoryNode" />
272 |             </path>
273 |           </expand>
274 |           <select />
275 |         </subPane>
276 |       </pane>
277 |       <pane id="Scope" />
278 |     </panes>
279 |   </component>
280 |   <component name="PropertiesComponent">
281 |     <property name="last_opened_file_path" value="$PROJECT_DIR$/../data_grand_new" />
282 |     <property name="settings.editor.selected.configurable" value="com.jetbrains.python.configuration.PyIntegratedToolsModulesConfigurable" />
283 |   </component>
284 |   <component name="RecentsManager">
285 |     <key name="MoveFile.RECENT_KEYS">
286 |       <recent name="D:\git\data_grand\script" />
287 |     </key>
288 |     <key name="CopyFile.RECENT_KEYS">
289 |       <recent name="$PROJECT_DIR$/models" />
290 |     </key>
291 |   </component>
292 |   <component name="RunDashboard">
293 |     <option name="ruleStates">
294 |       <list>
295 |         <RuleState>
296 |           <option name="name" value="ConfigurationTypeDashboardGroupingRule" />
297 |         </RuleState>
298 |         <RuleState>
299 |           <option name="name" value="StatusDashboardGroupingRule" />
300 |         </RuleState>
301 |       </list>
302 |     </option>
303 |   </component>
304 |   <component name="RunManager">
305 |     <configuration default="true" type="BashConfigurationType" factoryName="Bash">
306 |       <option name="INTERPRETER_OPTIONS" value="" />
307 |       <option name="INTERPRETER_PATH" value="/bin/bash" />
308 |       <option name="WORKING_DIRECTORY" value="" />
309 |       <option name="PARENT_ENVS" value="true" />
310 |       <option name="SCRIPT_NAME" value="" />
311 |       <option name="PARAMETERS" value="" />
312 |       <module name="" />
313 |       <envs />
314 |       <method />
315 |     </configuration>
316 |     <configuration default="true" type="tests" factoryName="Attests">
317 |       <option name="INTERPRETER_OPTIONS" value="" />
318 |       <option name="PARENT_ENVS" value="true" />
319 |       <envs />
320 |       <option name="SDK_HOME" value="" />
321 |       <option name="WORKING_DIRECTORY" value="" />
322 |       <option name="IS_MODULE_SDK" value="false" />
323 |       <option name="ADD_CONTENT_ROOTS" value="true" />
324 |       <option name="ADD_SOURCE_ROOTS" value="true" />
325 |       <module name="data_grand" />
326 |       <option name="SCRIPT_NAME" value="" />
327 |       <option name="CLASS_NAME" value="" />
328 |       <option name="METHOD_NAME" value="" />
329 |       <option name="FOLDER_NAME" value="" />
330 |       <option name="TEST_TYPE" value="TEST_SCRIPT" />
331 |       <option name="PATTERN" value="" />
332 |       <option name="USE_PATTERN" value="false" />
333 |       <method />
334 |     </configuration>
335 |     <configuration default="true" type="PythonConfigurationType" factoryName="Python">
336 |       <option name="INTERPRETER_OPTIONS" value="" />
337 |       <option name="PARENT_ENVS" value="true" />
338 |       <envs>
339 |         <env name="PYTHONUNBUFFERED" value="1" />
340 |       </envs>
341 |       <option name="SDK_HOME" value="" />
342 |       <option name="WORKING_DIRECTORY" value="" />
343 |       <option name="IS_MODULE_SDK" value="false" />
344 |       <option name="ADD_CONTENT_ROOTS" value="true" />
345 |       <option name="ADD_SOURCE_ROOTS" value="true" />
346 |       <module name="" />
347 |       <option name="SCRIPT_NAME" value="" />
348 |       <option name="PARAMETERS" value="" />
349 |       <option name="SHOW_COMMAND_LINE" value="false" />
350 |       <option name="EMULATE_TERMINAL" value="false" />
351 |       <option name="MODULE_MODE" value="false" />
352 |       <option name="REDIRECT_INPUT" value="false" />
353 |       <option name="INPUT_FILE" value="" />
354 |       <method v="2" />
355 |     </configuration>
356 |     <configuration name="main" type="PythonConfigurationType" factoryName="Python">
357 |       <module name="data_grand" />
358 |       <option name="INTERPRETER_OPTIONS" value="" />
359 |       <option name="PARENT_ENVS" value="true" />
360 |       <envs>
361 |         <env name="PYTHONUNBUFFERED" value="1" />
362 |       </envs>
363 |       <option name="SDK_HOME" value="" />
364 |       <option name="WORKING_DIRECTORY" value="$PROJECT_DIR$" />
365 |       <option name="IS_MODULE_SDK" value="true" />
366 |       <option name="ADD_CONTENT_ROOTS" value="true" />
367 |       <option name="ADD_SOURCE_ROOTS" value="true" />
368 |       <option name="SCRIPT_NAME" value="$PROJECT_DIR$/main.py" />
369 |       <option name="PARAMETERS" value="main" />
370 |       <option name="SHOW_COMMAND_LINE" value="false" />
371 |       <option name="EMULATE_TERMINAL" value="false" />
372 |       <option name="MODULE_MODE" value="false" />
373 |       <option name="REDIRECT_INPUT" value="false" />
374 |       <option name="INPUT_FILE" value="" />
375 |       <method v="2" />
376 |     </configuration>
377 |   </component>
378 |   <component name="SvnConfiguration">
379 |     <configuration />
380 |   </component>
381 |   <component name="TaskManager">
382 |     <task active="true" id="Default" summary="Default task">
383 |       <changelist id="fff4d958-5015-4c87-83d4-c1dc1da633cf" name="Default" comment="" />
384 |       <created>1531118794879</created>
385 |       <option name="number" value="Default" />
386 |       <option name="presentableId" value="Default" />
387 |       <updated>1531118794879</updated>
388 |     </task>
389 |     <servers />
390 |   </component>
391 |   <component name="ToolWindowManager">
392 |     <frame x="62" y="-7" width="1684" height="993" extended-state="6" />
393 |     <editor active="true" />
394 |     <layout>
395 |       <window_info active="true" content_ui="combo" id="Project" order="0" visible="true" weight="0.13038205" />
396 |       <window_info id="Structure" order="1" side_tool="true" weight="0.25" />
397 |       <window_info id="Favorites" order="2" side_tool="true" />
398 |       <window_info anchor="bottom" id="Message" order="0" />
399 |       <window_info anchor="bottom" id="Find" order="1" weight="0.3299285" />
400 |       <window_info anchor="bottom" id="Run" order="2" weight="0.49212122" />
401 |       <window_info anchor="bottom" id="Debug" order="3" sideWeight="0.49941793" weight="0.5636574" />
402 |       <window_info anchor="bottom" id="Cvs" order="4" weight="0.25" />
403 |       <window_info anchor="bottom" id="Inspection" order="5" weight="0.4" />
404 |       <window_info anchor="bottom" id="TODO" order="6" />
405 |       <window_info anchor="bottom" id="Terminal" order="7" weight="0.3298611" />
406 |       <window_info anchor="bottom" id="Event Log" order="8" sideWeight="0.5005821" side_tool="true" weight="0.2630303" />
407 |       <window_info anchor="bottom" id="Version Control" order="9" />
408 |       <window_info anchor="bottom" id="Python Console" order="10" />
409 |       <window_info anchor="right" id="Commander" order="0" weight="0.4" />
410 |       <window_info anchor="right" id="Ant Build" order="1" weight="0.25" />
411 |       <window_info anchor="right" content_ui="combo" id="Hierarchy" order="2" weight="0.25" />
412 |     </layout>
413 |   </component>
414 |   <component name="Vcs.Log.UiProperties">
415 |     <option name="RECENTLY_FILTERED_USER_GROUPS">
416 |       <collection />
417 |     </option>
418 |     <option name="RECENTLY_FILTERED_BRANCH_GROUPS">
419 |       <collection />
420 |     </option>
421 |   </component>
422 |   <component name="VcsContentAnnotationSettings">
423 |     <option name="myLimit" value="2678400000" />
424 |   </component>
425 |   <component name="XDebuggerManager">
426 |     <watches-manager>
427 |       <configuration name="PythonConfigurationType">
428 |         <watch expression="train.examples" />
429 |         <watch expression="len(text)" language="Python" />
430 |         <watch expression="__py_debug_temp_var_1900047073" />
431 |       </configuration>
432 |     </watches-manager>
433 |   </component>
434 |   <component name="debuggerHistoryManager">
435 |     <expressions id="watch">
436 |       <expression>
437 |         <expression-string>len(text)</expression-string>
438 |         <language-id>Python</language-id>
439 |         <evaluation-mode>EXPRESSION</evaluation-mode>
440 |       </expression>
441 |     </expressions>
442 |     <expressions id="evaluateExpression">
443 |       <expression>
444 |         <expression-string>len(text.split())</expression-string>
445 |         <language-id>Python</language-id>
446 |         <evaluation-mode>EXPRESSION</evaluation-mode>
447 |       </expression>
448 |       <expression>
449 |         <expression-string>len(set(indexs))</expression-string>
450 |         <language-id>Python</language-id>
451 |         <evaluation-mode>EXPRESSION</evaluation-mode>
452 |       </expression>
453 |       <expression>
454 |         <expression-string>len(indexs)</expression-string>
455 |         <language-id>Python</language-id>
456 |         <evaluation-mode>EXPRESSION</evaluation-mode>
457 |       </expression>
458 |       <expression>
459 |         <expression-string>text</expression-string>
460 |         <language-id>Python</language-id>
461 |         <evaluation-mode>EXPRESSION</evaluation-mode>
462 |       </expression>
463 |       <expression>
464 |         <expression-string>textr</expression-string>
465 |         <language-id>Python</language-id>
466 |         <evaluation-mode>EXPRESSION</evaluation-mode>
467 |       </expression>
468 |       <expression>
469 |         <expression-string>optimizer</expression-string>
470 |         <language-id>Python</language-id>
471 |         <evaluation-mode>EXPRESSION</evaluation-mode>
472 |       </expression>
473 |     </expressions>
474 |   </component>
475 |   <component name="editorHistoryManager">
476 |     <entry file="file://$PROJECT_DIR$/data_process.py" />
477 |     <entry file="file://$PROJECT_DIR$/main_new.py" />
478 |     <entry file="file://E:/anaconda/Lib/site-packages/torchtext/data/__init__.py">
479 |       <provider selected="true" editor-type-id="text-editor">
480 |         <state relative-caret-position="130">
481 |           <caret line="12" column="23" selection-start-line="12" selection-start-column="23" selection-end-line="12" selection-end-column="23" />
482 |         </state>
483 |       </provider>
484 |     </entry>
485 |     <entry file="file://E:/anaconda/Lib/site-packages/torchtext/data/dataset.py">
486 |       <provider selected="true" editor-type-id="text-editor">
487 |         <state relative-caret-position="3484">
488 |           <caret line="144" column="31" selection-start-line="144" selection-start-column="31" selection-end-line="144" selection-end-column="31" />
489 |         </state>
490 |       </provider>
491 |     </entry>
492 |     <entry file="file://E:/anaconda/Lib/site-packages/torch/cuda/random.py">
493 |       <provider selected="true" editor-type-id="text-editor">
494 |         <state relative-caret-position="280">
495 |           <caret line="66" column="37" lean-forward="true" selection-start-line="66" selection-start-column="37" selection-end-line="66" selection-end-column="37" />
496 |         </state>
497 |       </provider>
498 |     </entry>
499 |     <entry file="file://$PROJECT_DIR$/ensemble.py" />
500 |     <entry file="file://E:/anaconda/Lib/site-packages/fire/core.py">
501 |       <provider selected="true" editor-type-id="text-editor">
502 |         <state relative-caret-position="481">
503 |           <caret line="544" lean-forward="true" selection-start-line="544" selection-end-line="544" />
504 |         </state>
505 |       </provider>
506 |     </entry>
507 |     <entry file="file://$USER_HOME$/.PyCharmCE2018.2/system/python_stubs/-1980697110/_random.py">
508 |       <provider selected="true" editor-type-id="text-editor">
509 |         <state relative-caret-position="358">
510 |           <caret line="23" lean-forward="true" selection-start-line="23" selection-end-line="23" />
511 |         </state>
512 |       </provider>
513 |     </entry>
514 |     <entry file="file://E:/anaconda/Lib/random.py">
515 |       <provider selected="true" editor-type-id="text-editor">
516 |         <state relative-caret-position="592">
517 |           <caret line="756" column="39" lean-forward="true" selection-start-line="756" selection-start-column="39" selection-end-line="756" selection-end-column="39" />
518 |         </state>
519 |       </provider>
520 |     </entry>
521 |     <entry file="file://$USER_HOME$/.PyCharmCE2018.2/system/python_stubs/-1980697110/numpy/random/mtrand.py">
522 |       <provider selected="true" editor-type-id="text-editor">
523 |         <state relative-caret-position="280">
524 |           <caret line="228" column="32" selection-start-line="228" selection-start-column="25" selection-end-line="228" selection-end-column="32" />
525 |         </state>
526 |       </provider>
527 |     </entry>
528 |     <entry file="file://$USER_HOME$/.PyCharmCE2018.2/system/python_stubs/-1980697110/builtins.py">
529 |       <provider selected="true" editor-type-id="text-editor">
530 |         <state relative-caret-position="280">
531 |           <caret line="5009" column="8" selection-start-line="5009" selection-start-column="8" selection-end-line="5009" selection-end-column="8" />
532 |         </state>
533 |       </provider>
534 |     </entry>
535 |     <entry file="file://$APPLICATION_HOME_DIR$/helpers/typeshed/third_party/3/six/__init__.pyi">
536 |       <provider selected="true" editor-type-id="text-editor">
537 |         <state relative-caret-position="280">
538 |           <caret line="43" selection-start-line="43" selection-end-line="43" />
539 |         </state>
540 |       </provider>
541 |     </entry>
542 |     <entry file="file://E:/anaconda/Lib/site-packages/torchtext/data/example.py">
543 |       <provider selected="true" editor-type-id="text-editor">
544 |         <state relative-caret-position="272">
545 |           <caret line="42" column="19" selection-start-line="42" selection-start-column="19" selection-end-line="42" selection-end-column="19" />
546 |         </state>
547 |       </provider>
548 |     </entry>
549 |     <entry file="file://E:/anaconda/Lib/site-packages/nbconvert/preprocessors/csshtmlheader.py">
550 |       <provider selected="true" editor-type-id="text-editor">
551 |         <state relative-caret-position="514">
552 |           <caret line="44" column="25" lean-forward="true" selection-start-line="44" selection-start-column="25" selection-end-line="44" selection-end-column="25" />
553 |         </state>
554 |       </provider>
555 |     </entry>
556 |     <entry file="file://E:/anaconda/Lib/site-packages/torchtext/data/utils.py">
557 |       <provider selected="true" editor-type-id="text-editor">
558 |         <state relative-caret-position="78">
559 |           <caret line="5" column="4" selection-start-line="5" selection-start-column="4" selection-end-line="5" selection-end-column="4" />
560 |         </state>
561 |       </provider>
562 |     </entry>
563 |     <entry file="file://E:/anaconda/Lib/site-packages/torchtext/data/field.py">
564 |       <provider selected="true" editor-type-id="text-editor">
565 |         <state relative-caret-position="220">
566 |           <caret line="147" column="21" selection-start-line="147" selection-start-column="21" selection-end-line="147" selection-end-column="21" />
567 |         </state>
568 |       </provider>
569 |     </entry>
570 |     <entry file="file://E:/anaconda/Lib/site-packages/torchtext/data/iterator.py">
571 |       <provider selected="true" editor-type-id="text-editor">
572 |         <state relative-caret-position="844">
573 |           <caret line="130" column="29" lean-forward="true" selection-start-line="130" selection-start-column="29" selection-end-line="130" selection-end-column="29" />
574 |         </state>
575 |       </provider>
576 |     </entry>
577 |     <entry file="file://$PROJECT_DIR$/models/bilstm_conv.py">
578 |       <provider selected="true" editor-type-id="text-editor">
579 |         <state relative-caret-position="234">
580 |           <caret line="12" lean-forward="true" selection-start-line="12" selection-end-line="12" />
581 |         </state>
582 |       </provider>
583 |     </entry>
584 |     <entry file="file://E:/anaconda/Lib/site-packages/torch/nn/modules/linear.py">
585 |       <provider selected="true" editor-type-id="text-editor">
586 |         <state relative-caret-position="644">
587 |           <caret line="38" column="38" lean-forward="true" selection-start-line="38" selection-start-column="38" selection-end-line="38" selection-end-column="38" />
588 |         </state>
589 |       </provider>
590 |     </entry>
591 |     <entry file="file://E:/anaconda/Lib/site-packages/torch/nn/modules/batchnorm.py">
592 |       <provider selected="true" editor-type-id="text-editor">
593 |         <state relative-caret-position="-1106">
594 |           <caret line="88" column="22" selection-start-line="88" selection-start-column="22" selection-end-line="88" selection-end-column="22" />
595 |         </state>
596 |       </provider>
597 |     </entry>
598 |     <entry file="file://E:/anaconda/Lib/site-packages/torch/nn/modules/activation.py">
599 |       <provider selected="true" editor-type-id="text-editor">
600 |         <state relative-caret-position="350">
601 |           <caret line="72" column="29" lean-forward="true" selection-start-line="72" selection-start-column="29" selection-end-line="72" selection-end-column="29" />
602 |         </state>
603 |       </provider>
604 |     </entry>
605 |     <entry file="file://$PROJECT_DIR$/models/FastText.py">
606 |       <provider selected="true" editor-type-id="text-editor">
607 |         <state relative-caret-position="260">
608 |           <caret line="10" column="12" lean-forward="true" selection-start-line="10" selection-start-column="12" selection-end-line="10" selection-end-column="12" />
609 |         </state>
610 |       </provider>
611 |     </entry>
612 |     <entry file="file://E:/anaconda/Lib/site-packages/torch/nn/modules/rnn.py">
613 |       <provider selected="true" editor-type-id="text-editor">
614 |         <state relative-caret-position="1528">
615 |           <caret line="507" column="36" selection-start-line="507" selection-start-column="36" selection-end-line="507" selection-end-column="36" />
616 |         </state>
617 |       </provider>
618 |     </entry>
619 |     <entry file="file://$PROJECT_DIR$/models/RCNN1.py">
620 |       <provider selected="true" editor-type-id="text-editor">
621 |         <state relative-caret-position="272">
622 |           <caret line="14" selection-start-line="14" selection-end-line="14" />
623 |         </state>
624 |       </provider>
625 |     </entry>
626 |     <entry file="file://$PROJECT_DIR$/models/RCNN.py">
627 |       <provider selected="true" editor-type-id="text-editor">
628 |         <state relative-caret-position="168">
629 |           <caret line="25" lean-forward="true" selection-start-line="25" selection-end-line="25" />
630 |         </state>
631 |       </provider>
632 |     </entry>
633 |     <entry file="file://$PROJECT_DIR$/models/AttLSTM.py">
634 |       <provider selected="true" editor-type-id="text-editor">
635 |         <state relative-caret-position="272">
636 |           <caret line="36" column="33" selection-start-line="36" selection-start-column="33" selection-end-line="36" selection-end-column="33" />
637 |         </state>
638 |       </provider>
639 |     </entry>
640 |     <entry file="file://$PROJECT_DIR$/models/GRU.py">
641 |       <provider selected="true" editor-type-id="text-editor">
642 |         <state relative-caret-position="171">
643 |           <caret line="42" column="91" selection-start-line="42" selection-start-column="67" selection-end-line="42" selection-end-column="92" />
644 |         </state>
645 |       </provider>
646 |     </entry>
647 |     <entry file="file://E:/anaconda/Lib/site-packages/torch/nn/_functions/rnn.py">
648 |       <provider selected="true" editor-type-id="text-editor">
649 |         <state relative-caret-position="754">
650 |           <caret line="36" lean-forward="true" selection-start-line="36" selection-end-line="36" />
651 |         </state>
652 |       </provider>
653 |     </entry>
654 |     <entry file="file://$PROJECT_DIR$/models/TextCNN.py">
655 |       <provider selected="true" editor-type-id="text-editor">
656 |         <state relative-caret-position="263">
657 |           <caret line="46" lean-forward="true" selection-start-line="46" selection-end-line="46" />
658 |         </state>
659 |       </provider>
660 |     </entry>
661 |     <entry file="file://$PROJECT_DIR$/main_boost (2).py" />
662 |     <entry file="file://E:/anaconda/Lib/site-packages/torch/nn/modules/conv.py">
663 |       <provider selected="true" editor-type-id="text-editor">
664 |         <state relative-caret-position="728">
665 |           <caret line="33" column="16" selection-start-line="33" selection-start-column="16" selection-end-line="33" selection-end-column="16" />
666 |         </state>
667 |       </provider>
668 |     </entry>
669 |     <entry file="file://$PROJECT_DIR$/main_nllloss.py" />
670 |     <entry file="file://$PROJECT_DIR$/script/run.sh">
671 |       <provider selected="true" editor-type-id="text-editor">
672 |         <state relative-caret-position="26">
673 |           <caret line="1" column="40" selection-start-line="1" selection-start-column="40" selection-end-line="1" selection-end-column="40" />
674 |         </state>
675 |       </provider>
676 |     </entry>
677 |     <entry file="file://$PROJECT_DIR$/models/Inception.py">
678 |       <provider selected="true" editor-type-id="text-editor">
679 |         <state relative-caret-position="1664">
680 |           <caret line="69" column="32" selection-start-line="69" selection-start-column="32" selection-end-line="69" selection-end-column="32" />
681 |         </state>
682 |       </provider>
683 |     </entry>
684 |     <entry file="file://E:/anaconda/Lib/site-packages/torch/nn/modules/module.py">
685 |       <provider selected="true" editor-type-id="text-editor">
686 |         <state relative-caret-position="12922">
687 |           <caret line="504" selection-start-line="504" selection-end-line="504" />
688 |         </state>
689 |       </provider>
690 |     </entry>
691 |     <entry file="file://E:/anaconda/Lib/site-packages/torch/nn/functional.py">
692 |       <provider selected="true" editor-type-id="text-editor">
693 |         <state relative-caret-position="26260">
694 |           <caret line="1023" selection-start-line="1023" selection-end-line="1023" />
695 |         </state>
696 |       </provider>
697 |     </entry>
698 |     <entry file="file://$PROJECT_DIR$/models/LSTM.py">
699 |       <provider selected="true" editor-type-id="text-editor">
700 |         <state relative-caret-position="520">
701 |           <caret line="24" column="10" selection-start-line="24" selection-start-column="6" selection-end-line="24" selection-end-column="10" />
702 |         </state>
703 |       </provider>
704 |     </entry>
705 |     <entry file="file://$PROJECT_DIR$/models/__init__.py">
706 |       <provider selected="true" editor-type-id="text-editor">
707 |         <state relative-caret-position="26">
708 |           <caret line="1" selection-start-line="1" selection-end-line="1" />
709 |         </state>
710 |       </provider>
711 |     </entry>
712 |     <entry file="file://$PROJECT_DIR$/util.py" />
713 |     <entry file="file://$PROJECT_DIR$/emb_build/read_csv.py">
714 |       <provider selected="true" editor-type-id="text-editor">
715 |         <state relative-caret-position="130">
716 |           <caret line="5" selection-start-line="5" selection-end-line="5" />
717 |         </state>
718 |       </provider>
719 |     </entry>
720 |     <entry file="file://$PROJECT_DIR$/emb_build/tran_emb.py">
721 |       <provider selected="true" editor-type-id="text-editor">
722 |         <state relative-caret-position="260">
723 |           <caret line="10" column="15" selection-start-line="10" selection-start-column="15" selection-end-line="10" selection-end-column="15" />
724 |         </state>
725 |       </provider>
726 |     </entry>
727 |     <entry file="file://$PROJECT_DIR$/models/BasicModule.py">
728 |       <provider selected="true" editor-type-id="text-editor">
729 |         <state relative-caret-position="1170">
730 |           <caret line="46" selection-start-line="46" selection-end-line="46" />
731 |         </state>
732 |       </provider>
733 |     </entry>
734 |     <entry file="file://$PROJECT_DIR$/util/split_val.py">
735 |       <provider selected="true" editor-type-id="text-editor">
736 |         <state relative-caret-position="416">
737 |           <caret line="18" selection-start-line="18" selection-end-line="18" />
738 |           <folding>
739 |             <element signature="e#0#19#0" expanded="true" />
740 |           </folding>
741 |         </state>
742 |       </provider>
743 |     </entry>
744 |     <entry file="file://$PROJECT_DIR$/data.py">
745 |       <provider selected="true" editor-type-id="text-editor">
746 |         <state relative-caret-position="754">
747 |           <caret line="36" column="24" selection-start-line="36" selection-start-column="24" selection-end-line="36" selection-end-column="24" />
748 |           <folding>
749 |             <element signature="e#105#131#0" expanded="true" />
750 |           </folding>
751 |         </state>
752 |       </provider>
753 |     </entry>
754 |     <entry file="file://$PROJECT_DIR$/main.py">
755 |       <provider selected="true" editor-type-id="text-editor">
756 |         <state relative-caret-position="390">
757 |           <caret line="25" column="16" selection-start-line="25" selection-start-column="16" selection-end-line="25" selection-end-column="16" />
758 |           <folding>
759 |             <element signature="e#98#110#0" expanded="true" />
760 |           </folding>
761 |         </state>
762 |       </provider>
763 |     </entry>
764 |     <entry file="file://$PROJECT_DIR$/main_boost.py">
765 |       <provider selected="true" editor-type-id="text-editor">
766 |         <state relative-caret-position="234">
767 |           <caret line="9" column="11" selection-start-line="9" selection-start-column="11" selection-end-line="9" selection-end-column="11" />
768 |         </state>
769 |       </provider>
770 |     </entry>
771 |     <entry file="file://$PROJECT_DIR$/config.py">
772 |       <provider selected="true" editor-type-id="text-editor">
773 |         <state relative-caret-position="520">
774 |           <caret line="20" column="18" selection-start-line="20" selection-start-column="18" selection-end-line="20" selection-end-column="18" />
775 |         </state>
776 |       </provider>
777 |     </entry>
778 |     <entry file="file://$PROJECT_DIR$/test_ensemble.py">
779 |       <provider selected="true" editor-type-id="text-editor">
780 |         <state relative-caret-position="260">
781 |           <caret line="10" column="11" selection-start-line="10" selection-start-column="11" selection-end-line="10" selection-end-column="11" />
782 |         </state>
783 |       </provider>
784 |     </entry>
785 |     <entry file="file://$PROJECT_DIR$/fine_tune.py">
786 |       <provider selected="true" editor-type-id="text-editor">
787 |         <state relative-caret-position="312">
788 |           <caret line="12" column="11" selection-start-line="12" selection-start-column="11" selection-end-line="12" selection-end-column="11" />
789 |         </state>
790 |       </provider>
791 |     </entry>
792 |     <entry file="file://$PROJECT_DIR$/gen_result.py">
793 |       <provider selected="true" editor-type-id="text-editor">
794 |         <state relative-caret-position="208">
795 |           <caret line="15" column="19" selection-start-line="15" selection-start-column="19" selection-end-line="15" selection-end-column="19" />
796 |           <folding>
797 |             <element signature="e#105#117#0" expanded="true" />
798 |           </folding>
799 |         </state>
800 |       </provider>
801 |     </entry>
802 |     <entry file="file://$PROJECT_DIR$/.gitignore">
803 |       <provider selected="true" editor-type-id="text-editor">
804 |         <state relative-caret-position="260">
805 |           <caret line="10" column="5" selection-start-line="10" selection-start-column="5" selection-end-line="10" selection-end-column="5" />
806 |         </state>
807 |       </provider>
808 |     </entry>
809 |   </component>
810 | </project>


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # 2018“达观杯”文本智能处理挑战赛
 2 | ## Top 10 “万里阳光号”解决方案
 3 | 更详细的比赛经验分享见我的知乎专栏文章https://zhuanlan.zhihu.com/p/45391378  
 4 | 
 5 | 比赛详情见 [达观杯文本智能处理挑战赛](http://www.dcjingsai.com/common/cmpt/%E2%80%9C%E8%BE%BE%E8%A7%82%E6%9D%AF%E2%80%9D%E6%96%87%E6%9C%AC%E6%99%BA%E8%83%BD%E5%A4%84%E7%90%86%E6%8C%91%E6%88%98%E8%B5%9B_%E7%AB%9E%E8%B5%9B%E4%BF%A1%E6%81%AF.html)  
 6 | 最终排名见 [排行榜](
 7 | http://www.dcjingsai.com/common/cmpt/%E2%80%9C%E8%BE%BE%E8%A7%82%E6%9D%AF%E2%80%9D%E6%96%87%E6%9C%AC%E6%99%BA%E8%83%BD%E5%A4%84%E7%90%86%E6%8C%91%E6%88%98%E8%B5%9B_%E6%8E%92%E8%A1%8C%E6%A6%9C.html)
 8 | ### 环境配置
 9 | 代码基于Pytorch，版本为0.4.1，Python版本为3.6。需安装：
10 | - pytorch
11 | - torchtext
12 | - word2vec
13 | - pandas
14 | - sklearn
15 | - numpy
16 | - fire
17 | ### 文件说明
18 | ```
19 | emb_build/: word2vec训练词／字向量
20 | models/: 深度学习模型
21 | result/: 生成测试集提交csv
22 | val_result/: 模型融合
23 | script/: 脚本文件
24 | config.py: 模型配置
25 | fine_tune.py: 模型fine tune
26 | gen_result.py: 生成模型在测试集上的预测概率结果
27 | test_ensemble.py: 生成模型在验证集上的预测概率结果
28 | data.py: 数据预处理
29 | main.py: 模型训练
30 | ```
31 | 
32 | ### 词／字向量训练
33 | 词／字向量训练使用word2vec包，见[word2vec](https://github.com/danielfrg/word2vec)。分别使用训练集和测试集中所有词文本和字文本训练词向量和字向量，向量维度设置为300维。  
34 | 代码见emb_build文件夹下read_csv.py和tran_emb.py，注意修改代码中训练集和测试集的文件路径，依次运行即可得到词／字向量文件word_300.txt和article_300.txt。
35 | ```
36 | python read_csv.py
37 | python tran_emb.py
38 | ```
39 | ### 文本预处理
40 | 将比赛提供的训练数据按9:1的比例，划分为训练集和验证集。
41 | ```
42 | # util/
43 | python split_val.py
44 | ```
45 | word文本平均长度为717，按照覆盖95%样本的标准，取截断长度为2000；article文本平均长度为1177，按同样的标准取截断长度为3200。  
46 | 从csv文件中提取文本数据，使用torchtext进行文本预处理，并进一步构造batch，这部分代码见data.py的类GrandDataset和方法load_data()。
47 | ### 训练模型
48 | 主要用到了五个模型
49 | 
50 | - TextCNN: models/TextCNN.py
51 | - GRU: models/GRU.py
52 | - RCNN: models/RCNN.py
53 | - FastText: models/FastText.py
54 | - Attention: models/bigru_att.py  
55 | 
56 | 分别训练两个对应的word模型和article模型，注意文本数据和词/字向量的存放路径。 
57 | 注意模型配置位于 config.py，模型训练代码位于main.py中的main方法，命令示例如下（也可见script/run.sh）:  
58 | ```
59 | python main.py main --model='LSTM' --device=5 --id='word4'
60 | python main.py main --model='GRU' --device=6 --id='word4'
61 | python main.py main --model='RCNN1' --device=4 --id='word4'
62 | python main.py main --model='GRU' --device=8 --id='word41'
63 | python main.py main --model='TextCNN' --device=10 --id='rev4'
64 | ```
65 | ### 训练策略
66 | - 优化器选用torch.optim.Adam，初始学习率设置为1e-3（注意embedding层的学习率另外设置）。
67 | - 先固定embedding层的参数，每个epoch后计算模型在验证集上的f1值，如果开始下降（一般为2-3个epoch之后），将embedding层的学习率设为2e-4。
68 | - 每个epoch后计算验证集上的f1值，如上升则保存当前模型，位于文件夹snapshot/；如果下降则从snapshot中加载当前最好的模型，并降低学习率。
69 | - 如果学习率低于config.py中设置的最低学习率，则终止训练。如果设置的最低学习率为1e-5，一般15个epoch左右后训练终止。
70 | 
71 | 
72 | ### 模型融合
73 | 尝试了多种模型融合方法后，只采用了最简单但有效的模型融合方法－概率等权重融合，代码见val_result/ensemble.py，修改代码中的
74 | ```
75 | model_list = ['GRU_word_rev_0.721350215541.pth.npy', 'GRU_word_0.771128871335.pth.npy',  'RCNN1_word_0.769018920031.pth.npy', 'LSTM_word_0.768979776301.pth.npy', 'TextCNN_word_0.760456622816.pth.npy', 'FastText_word_0.75425891649.pth.npy', 'GRU_article_0.747660923499.pth.npy', 'TextCNN_article_0.735192435177.pth.npy']
76 | ```
77 | 运行ensemble.py得到融合模型在验证集上的f1值，根据f1值选取参与融合的模型。
78 | 
79 | ### 生成提交csv
80 | 将上一步模型融合选取的模型在测试集上生成预测概率结果，进行等权重相加，将概率最大的类别作为预测类别生成提交csv，代码见result/ensemble.py。  
81 | 
82 | 更详细的比赛经验分享见知乎专栏文章https://zhuanlan.zhihu.com/p/45391378  
83 | 


--------------------------------------------------------------------------------
/config.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | @Author  : captain
 5 | @time    : 18-7-11 上午12:33
 6 | @ide     : PyCharm  
 7 | """
 8 | 
 9 | 
10 | class DefaultConfig(object):
11 |     '''
12 |     列出所有的参数，只根据模型需要获取参数
13 |     '''
14 |     env = 'default'  # visdom环境
15 |     seed = 777  # 设置随机数种子
16 |     best_score = 0
17 |     model = 'InCNN'  # 使用的模型，名字必须与models/__init__.py中的名字一致
18 |     model_path = None  # 如果有就加载
19 |     result_path = ''
20 |     save_dir = 'snapshot/'  # where to save the snapshot
21 |     id = 'default'
22 |     device = 0
23 |     boost = False  ## 是否使用adboost
24 |     bo_layers = 5  ## boost的层数
25 |     finetune = False  ## 是否对训练完成的模型进行finetune
26 |     aug = False  ## 是否进行数据增强
27 | 
28 |     # 数据集参数
29 |     data_path = '/data/yujun/captain/datasets/'
30 |     text_type = 'word'  # or 'article'  # 决定数据集位置
31 |     # train_data_path = '/data/yujun/datasets/daguanbei_data/new_split/new_train_set.csv'
32 |     # val_data_path = '/data/yujun/datasets/daguanbei_data/new_split/val_set.csv'
33 |     # test_data_path = '/data/yujun/datasets/daguanbei_data/test_set.csv'
34 |     # embedding_path = '/data/yujun/captain/emb'  # 使用的预训练词向量
35 |     # embedding_dim = 300  # number of embedding dimension
36 | 
37 |     # 在自己的PC上小数据集调试代码用
38 |     # train_data_path = 'D:/git/dataset/val_set.csv'
39 |     # test_data_path = 'D:/git/dataset/val_set.csv'
40 |     # val_data_path = 'D:/git/dataset/val_set.csv'
41 |     embedding_path = 'D:/git/emb'  # 预训练词向量的位置
42 |     embedding_dim = 100
43 | 
44 |     batch_size = 64
45 |     vocab_size = 10000  # 词库规模，配置中写的值没有意义，实际是预处理阶段获取
46 |     label_size = 19  # 分类类别数
47 |     max_text_len = 2000  # 之后会处理成变长的，这里的设置没有意义
48 | 
49 |     # 训练参数
50 |     lr1 = 1e-3  # learning rate
51 |     lr2 = 0  # embedding层的学习率
52 |     min_lr = 1e-5  # 当学习率低于这个nvi值时，就退出训练
53 |     lr_decay = 0.8  # 当一个epoch的损失开始上升时，lr ＝ lr*lr_decay
54 |     decay_every = 10000  # 每多少个batch  查看val acc，并修改学习率
55 |     weight_decay = 0  # 2e-5 # 权重衰减
56 |     max_epochs = 50
57 |     cuda = True
58 | 
59 |     # 模型通用
60 |     linear_hidden_size = 100  # 原来为2000(500)，之后还需要修改，感觉数值有点大
61 | 
62 |     # TextCNN
63 |     kernel_num = 200  # number of each kind of kernel
64 |     kernel_sizes = '3,4,5'  # kernel size to use for convolution
65 |     dropout_rate = 0.5  # the probability for dropout
66 | 
67 |     # LSTM
68 |     hidden_dim = 256
69 |     lstm_dropout = 0.5  # 只有当lstm_layers > 1时，设置lstm_dropout才有意义
70 |     lstm_layers = 1
71 |     kmax_pooling = 2
72 | 
73 |     # RCNN
74 |     rcnn_kernel = 512
75 | 
76 |     def parse(self, kwargs):
77 |         '''
78 |         根据字典kwargs 更新 config参数
79 |         '''
80 | 
81 |         # 更新配置参数
82 |         for k, v in kwargs.items():
83 |             if not hasattr(self, k):
84 |                 raise Exception("Warning: config has not attribute <%s>" % k)
85 |             setattr(self, k, v)
86 | 
87 |     def print_config(self):
88 |         # 打印配置信息
89 |         print('user config:')
90 |         for k, v in self.__class__.__dict__.items():
91 |             if not k.startswith('__') and k != 'parse' and k != 'print_config':
92 |                 print('    {} : {}'.format(k, getattr(self, k)))
93 | 


--------------------------------------------------------------------------------
/data.py:
--------------------------------------------------------------------------------
  1 | # -*- coding: utf-8 -*-
  2 | 
  3 | """
  4 | # @Author  : captain
  5 | # @Time    : 2018/8/28 16:15
  6 | # @Ide     : PyCharm
  7 | """
  8 | 
  9 | from torchtext import data
 10 | import pandas as pd
 11 | from torchtext.vocab import Vectors
 12 | from tqdm import tqdm
 13 | from torch.nn import init
 14 | import random
 15 | import os
 16 | import numpy as np
 17 | 
 18 | 
 19 | # 定义Dataset
 20 | class GrandDataset(data.Dataset):
 21 |     name = 'Grand Dataset'
 22 | 
 23 |     @staticmethod
 24 |     def sort_key(ex):
 25 |         return len(ex.text)
 26 | 
 27 |     def __init__(self, path, text_field, label_field, text_type='word', test=False, aug=False, **kwargs):
 28 |         fields = [('text', text_field), ('label', label_field)]
 29 |         examples = []
 30 |         csv_data = pd.read_csv(path)
 31 |         print('read data from {}'.format(path))
 32 | 
 33 |         if text_type == 'word':
 34 |             text_type = 'word_seg'
 35 | 
 36 |         if test:
 37 |             # 如果为测试集，则不加载label
 38 |             for text in tqdm(csv_data[text_type]):
 39 |                 examples.append(data.Example.fromlist([text, None], fields))
 40 |         else:
 41 |             for text, label in tqdm(zip(csv_data[text_type], csv_data['class'])):
 42 |                 if aug:
 43 |                     # do augmentation
 44 |                     rate = random.random()
 45 |                     if rate > 0.5:
 46 |                         text = self.dropout(text)
 47 |                     else:
 48 |                         text = self.shuffle(text)
 49 |                 examples.append(data.Example.fromlist([text, label - 1], fields))
 50 |         super(GrandDataset, self).__init__(examples, fields, **kwargs)
 51 | 
 52 |     def shuffle(self, text):
 53 |         text = np.random.permutation(text.strip().split())
 54 |         return ' '.join(text)
 55 | 
 56 |     def dropout(self, text, p=0.5):
 57 |         # random delete some text
 58 |         text = text.strip().split()
 59 |         len_ = len(text)
 60 |         indexs = np.random.choice(len_, int(len_ * p))
 61 |         for i in indexs:
 62 |             text[i] = ''
 63 |         return ' '.join(text)
 64 | 
 65 | 
 66 | def load_data(opt):
 67 |     # 不设置fix_length
 68 |     TEXT = data.Field(sequential=True, fix_length=opt.max_text_len)  # 词或者字符
 69 |     LABEL = data.Field(sequential=False, use_vocab=False)
 70 | 
 71 |     # load
 72 |     # word/ or article/
 73 |     train_path = opt.data_path + opt.text_type + '/train_set.csv'
 74 |     val_path = opt.data_path + opt.text_type + '/val_set.csv'
 75 |     test_path = opt.data_path + opt.text_type + '/test_set.csv'
 76 |     train_path = 'D:/git/dataset/val_set.csv'
 77 |     test_path = 'D:/git/dataset/val_set.csv'
 78 |     val_path = 'D:/git/dataset/val_set.csv'
 79 | 
 80 |     # aug for data augmentation
 81 |     if opt.aug:
 82 |         print('make augmentation datasets!')
 83 |     train = GrandDataset(train_path, text_field=TEXT, label_field=LABEL, text_type=opt.text_type, test=False,
 84 |                          aug=opt.aug)
 85 |     val = GrandDataset(val_path, text_field=TEXT, label_field=LABEL, text_type=opt.text_type, test=False)
 86 |     test = GrandDataset(test_path, text_field=TEXT, label_field=None, text_type=opt.text_type, test=True)
 87 | 
 88 |     cache = '.vector_cache'
 89 |     if not os.path.exists(cache):
 90 |         os.mkdir(cache)
 91 |     embedding_path = '{}/{}_{}.txt'.format(opt.embedding_path, opt.text_type, opt.embedding_dim)
 92 |     vectors = Vectors(name=embedding_path, cache=cache)
 93 |     print('load word2vec vectors from {}'.format(embedding_path))
 94 |     vectors.unk_init = init.xavier_uniform_  # 没有命中的token的初始化方式
 95 | 
 96 |     # 构建Vocab
 97 |     print('building {} vocabulary......'.format(opt.text_type))
 98 |     TEXT.build_vocab(train, val, test, min_freq=5, vectors=vectors)
 99 |     # LABEL.build_vocab(train)
100 | 
101 |     # 构建Iterator
102 |     # 在 test_iter, shuffle, sort, repeat一定要设置成 False, 要不然会被 torchtext 搞乱样本顺序
103 |     # 如果输入变长序列，sort_within_batch需要设置成true，使每个batch内数据按照sort_key降序进行排序
104 |     train_iter = data.BucketIterator(dataset=train, batch_size=opt.batch_size, shuffle=True, sort_within_batch=False,
105 |                                      repeat=False, device=opt.device)
106 |     # val_iter = data.BucketIterator(dataset=val, batch_size=opt.batch_size, sort_within_batch=False, repeat=False,
107 |     #                                device=opt.device)
108 |     # train_iter = data.Iterator(dataset=train, batch_size=opt.batch_size, train=True, repeat=False, device=opt.device)
109 |     val_iter = data.Iterator(dataset=val, batch_size=opt.batch_size, shuffle=False, sort=False, repeat=False,
110 |                              device=opt.device)
111 |     test_iter = data.Iterator(dataset=test, batch_size=opt.batch_size, shuffle=False, sort=False, repeat=False,
112 |                               device=opt.device)
113 | 
114 |     return train_iter, val_iter, test_iter, len(TEXT.vocab), TEXT.vocab.vectors
115 | 


--------------------------------------------------------------------------------
/emb_build/read_csv.py:
--------------------------------------------------------------------------------
 1 | import pandas as pd
 2 | 
 3 | print('loading datasets......')
 4 | train_data = pd.read_csv('/data/yujun/datasets/daguanbei_data/train_set.csv')
 5 | test_data = pd.read_csv('/data/yujun/datasets/daguanbei_data/test_set.csv')
 6 | 
 7 | print('{} lines in train datasets'.format(len(train_data)))
 8 | print('{} lines in test datasets'.format(len(test_data)))
 9 | 
10 | print('making raw_word.txt......')
11 | with open('raw_word.txt', 'w') as f:
12 |     f.writelines([text + '\n' for text in train_data['word_seg']])
13 |     f.writelines([text + '\n' for text in test_data['word_seg']])
14 | 
15 | print('making raw_article.txt......')
16 | with open('raw_article.txt', 'w') as f:
17 |     f.writelines([text + '\n' for text in train_data['article']])
18 |     f.writelines([text + '\n' for text in test_data['article']])
19 | 
20 | 


--------------------------------------------------------------------------------
/emb_build/tran_emb.py:
--------------------------------------------------------------------------------
 1 | import word2vec
 2 | import fire
 3 | 
 4 | paths = ['raw_word.txt', 'raw_article.txt']
 5 | sizes = [300]
 6 | 
 7 | 
 8 | def tran(path):
 9 |     model = word2vec.load(path)
10 |     vocab, vectors = model.vocab, model.vectors
11 |     print(path)
12 |     print('shape of word embeddings : ')
13 |     print(vectors.shape)
14 | 
15 |     new_path = path.split('.')[0] + '_.txt'
16 |     print('Transform start....')
17 |     f = open(new_path, 'w')
18 |     for word, vector in zip(vocab, vectors):
19 |         f.write(str(word) + ' ' + ' '.join(map(str, vector)) + '\n')
20 |     print('Transform Complete!\n')
21 | 
22 | 
23 | for path in paths:
24 |     for size in sizes:
25 |         emb_path = path.split('.')[0].split('_')[1] + '_' + str(size) + '.bin'
26 |         word2vec.word2vec(path, emb_path, min_count=5, size=size, verbose=True)
27 |         tran(emb_path)
28 | 


--------------------------------------------------------------------------------
/fine_tune.py:
--------------------------------------------------------------------------------
  1 | # -*- coding: utf-8 -*-
  2 | 
  3 | """
  4 | # @Author  : captain
  5 | # @Time    : 2018/8/29 23:06
  6 | # @Ide     : PyCharm
  7 | """
  8 | 
  9 | import torch
 10 | import time
 11 | import torch.nn.functional as F
 12 | import models
 13 | import data
 14 | import pandas as pd
 15 | import os
 16 | import fire
 17 | from sklearn import metrics
 18 | import numpy as np
 19 | 
 20 | t1 = time.time()
 21 | 
 22 | 
 23 | def tune(model_path=None, device=0):
 24 |     # 仅需设置model_path和device
 25 |     if model_path is None:
 26 |         print('model_path is needed!')
 27 |         return
 28 | 
 29 |     # load model and set small learning rate
 30 |     saved_model = torch.load(model_path)
 31 |     config = saved_model['config']
 32 |     config.model_path = model_path
 33 |     config.device = device
 34 |     config.lr1 = 5e-5
 35 |     config.lr2 = 5e-5
 36 |     config.max_epochs = 10
 37 | 
 38 |     config.print_config()
 39 | 
 40 |     train_iter, val_iter, test_iter, config.vocab_size, vectors = util.load_data(config)
 41 |     config.print_config()
 42 | 
 43 |     model = getattr(models, config.model)(config, None)
 44 |     model.load_state_dict(saved_model['state_dict'])
 45 |     print(model)
 46 |     best_score = config.best_score
 47 |     print('Load model from {}!'.format(config.model_path))
 48 |     print('Tmp best f1 score: {}'.format(best_score))
 49 | 
 50 |     # 模型保存位置
 51 |     save_path = os.path.join(config.save_dir, 'tune_{}_{}.pth'.format(config.model, config.id))
 52 |     torch.save(model, save_path)
 53 | 
 54 |     if config.cuda:
 55 |         torch.cuda.set_device(config.device)
 56 |         model.cuda()
 57 | 
 58 |     # 目标函数和优化器
 59 |     criterion = F.cross_entropy
 60 |     lr1, lr2 = config.lr1, config.lr2
 61 |     optimizer = model.get_optimizer(lr1, lr2, config.weight_decay)
 62 | 
 63 |     for i in range(config.max_epochs):
 64 |         total_loss = 0.0
 65 |         correct = 0
 66 |         total = 0
 67 | 
 68 |         model.train()
 69 | 
 70 |         for idx, batch in enumerate(train_iter):
 71 |             # 训练模型参数
 72 |             # 使用BatchNorm层时，batch size不能为1
 73 |             if len(batch) == 1:
 74 |                 continue
 75 |             text, label = batch.text, batch.label
 76 |             if config.cuda:
 77 |                 text, label = text.cuda(), label.cuda()
 78 | 
 79 |             optimizer.zero_grad()
 80 |             pred = model(text)
 81 |             loss = criterion(pred, label)
 82 |             loss.backward()
 83 |             optimizer.step()
 84 | 
 85 |             # 更新统计指标
 86 |             total_loss += loss.item()
 87 |             predicted = pred.max(1)[1]
 88 |             total += label.size(0)
 89 |             correct += predicted.eq(label).sum().item()
 90 | 
 91 |             if idx % 80 == 79:
 92 |                 print('[{}, {}] loss: {:.3f} | Acc: {:.3f}%({}/{})'.format(i + 1, idx + 1, total_loss / 20,
 93 |                                                                            100. * correct / total, correct, total))
 94 |                 total_loss = 0.0
 95 | 
 96 |         # 计算再验证集上的分数，并相应调整学习率
 97 |         f1score = val(model, val_iter, config)
 98 |         if f1score > best_score:
 99 |             best_score = f1score
100 |             torch.save(model, save_path)
101 |             print('Best tmp model f1score: {}'.format(best_score))
102 |         if f1score < best_score:
103 |             model = torch.load(save_path)
104 |             model = model.cuda()
105 |             lr1 *= config.lr_decay
106 |             lr2 *= 0.8
107 |             optimizer = model.get_optimizer(lr1, lr2, 0)
108 |             print('* load previous best model: {}'.format(best_score))
109 |             print('* model lr:{}  emb lr:{}'.format(lr1, lr2))
110 |             if lr1 < config.min_lr:
111 |                 print('* training over, best f1 score: {}'.format(best_score))
112 |                 break
113 | 
114 |     # 保存训练最终的模型
115 |     config.best_score = best_score
116 |     final_model = {
117 |         'state_dict': model.state_dict(),
118 |         'config': config
119 |     }
120 |     best_model_path = os.path.join(config.save_dir,
121 |                                    'tune_{}_{}_{}.pth'.format(config.model, config.text_type, best_score))
122 |     torch.save(final_model, best_model_path)
123 |     print('Best Final Model saved in {}'.format(best_model_path))
124 | 
125 |     # 在测试集上运行模型并生成概率结果和提交结果
126 |     if not os.path.exists('result/'):
127 |         os.mkdir('result/')
128 |     probs, test_pred = test(model, test_iter, config)
129 |     result_path = 'result/' + 'tune_{}_{}_{}'.format(config.model, config.id, config.best_score)
130 |     np.save('{}.npy'.format(result_path), probs)
131 |     print('Prob result {}.npy saved!'.format(result_path))
132 | 
133 |     test_pred[['id', 'class']].to_csv('{}.csv'.format(result_path), index=None)
134 |     print('Result {}.csv saved!'.format(result_path))
135 | 
136 |     t2 = time.time()
137 |     print('time use: {}'.format(t2 - t1))
138 | 
139 | 
140 | def test(model, test_data, args):
141 |     # 生成测试提交数据csv
142 |     # 将模型设为验证模式
143 |     model.eval()
144 | 
145 |     result = np.zeros((0,))
146 |     probs_list = []
147 |     with torch.no_grad():
148 |         for batch in test_data:
149 |             text = batch.text
150 |             if args.cuda:
151 |                 text = text.cuda()
152 |             outputs = model(text)
153 |             probs = F.softmax(outputs, dim=1)
154 |             probs_list.append(probs.cpu().numpy())
155 |             pred = outputs.max(1)[1]
156 |             result = np.hstack((result, pred.cpu().numpy()))
157 | 
158 |     # 生成概率文件npy
159 |     prob_cat = np.concatenate(probs_list, axis=0)
160 | 
161 |     test = pd.read_csv('/data/yujun/datasets/daguanbei_data/test_set.csv')
162 |     test_id = test['id'].copy()
163 |     test_pred = pd.DataFrame({'id': test_id, 'class': result})
164 |     test_pred['class'] = (test_pred['class'] + 1).astype(int)
165 | 
166 |     return prob_cat, test_pred
167 | 
168 | 
169 | def val(model, dataset, args):
170 |     # 计算模型在验证集上的分数
171 | 
172 |     # 将模型设为验证模式
173 |     model.eval()
174 | 
175 |     acc_n = 0
176 |     val_n = 0
177 |     predict = np.zeros((0,), dtype=np.int32)
178 |     gt = np.zeros((0,), dtype=np.int32)
179 |     with torch.no_grad():
180 |         for batch in dataset:
181 |             text, label = batch.text, batch.label
182 |             if args.cuda:
183 |                 text, label = text.cuda(), label.cuda()
184 |             outputs = model(text)
185 |             pred = outputs.max(1)[1]
186 |             acc_n += (pred == label).sum().item()
187 |             val_n += label.size(0)
188 |             predict = np.hstack((predict, pred.cpu().numpy()))
189 |             gt = np.hstack((gt, label.cpu().numpy()))
190 | 
191 |     acc = 100. * acc_n / val_n
192 |     f1score = np.mean(metrics.f1_score(predict, gt, average=None))
193 |     print('* Test Acc: {:.3f}%({}/{}), F1 Score: {}'.format(acc, acc_n, val_n, f1score))
194 |     return f1score
195 | 
196 | 
197 | if __name__ == '__main__':
198 |     fire.Fire()
199 | 


--------------------------------------------------------------------------------
/gen_result.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | # @Author  : captain
 5 | # @Time    : 2018/8/27 22:31
 6 | # @Ide     : PyCharm
 7 | """
 8 | 
 9 | import torch
10 | import models
11 | from config import DefaultConfig
12 | import data
13 | import fire
14 | import numpy as np
15 | import torch.nn.functional as F
16 | import pandas as pd
17 | 
18 | 
19 | def main(**kwargs):
20 |     args = DefaultConfig()
21 |     args.parse(kwargs)
22 | 
23 |     train_iter, val_iter, test_iter, args.vocab_size, vectors = util.load_data(args, args.text_type)
24 | 
25 |     args.print_config()
26 | 
27 |     # model
28 |     if args.model_path:
29 |         # 加载模型
30 |         saved_model = torch.load(args.model_path)
31 |         config = saved_model['config']
32 |         config.device = args.device
33 |         model = getattr(models, args.model)(args, vectors)
34 |         model.load_state_dict(saved_model['state_dict'])
35 |         best_score = saved_model['best_score']
36 |         print('Load model from {}!'.format(args.model_path))
37 |     else:
38 |         print("No trained model!")
39 | 
40 |     if not torch.cuda.is_available():
41 |         config.cuda = False
42 |         config.device = None
43 | 
44 |     if args.cuda:
45 |         torch.cuda.set_device(args.device)
46 |         model.cuda()
47 | 
48 |     probs = infer(model, test_iter, config)
49 |     result_path = 'result/' + '{}_{}_{}'.format(args.model, args.id, args.best_score)
50 |     np.save('{}.npy'.format(result_path), probs)
51 |     print('Prob result {}.npy saved!'.format(result_path))
52 |     # np.save('{}.npy'.format(args.model), result)
53 | 
54 |     # test = pd.read_csv('/data/yujun/datasets/daguanbei_data/test_set.csv')
55 |     # test_id = test['id'].copy()
56 |     #
57 |     # test_pred = pd.DataFrame({'id': test_id, 'class': result})
58 |     # test_pred['class'] = (test_pred['class'] + 1).astype(int)
59 |     # test_pred[['id', 'class']].to_csv('{}.csv'.format(args.model), index=None)
60 | 
61 | 
62 | def infer(model, test_iter, opt):
63 |     # 将模型设为验证模式
64 |     model.eval()
65 | 
66 |     result = np.zeros((0,))
67 |     probs_list = []
68 |     with torch.no_grad():
69 |         for batch in test_iter:
70 |             text = batch.text
71 |             if opt.cuda:
72 |                 text = text.cuda()
73 |             outputs = model(text)
74 |             probs = F.softmax(outputs, dim=1)
75 |             probs_list.append(probs.cpu().numpy())
76 | 
77 |     prob_cat = np.concatenate(probs_list, axis=0)
78 |     return prob_cat
79 | 
80 | 
81 | if __name__ == '__main__':
82 |     fire.Fire()
83 | 


--------------------------------------------------------------------------------
/main.py:
--------------------------------------------------------------------------------
  1 | # -*- coding: utf-8 -*-
  2 | 
  3 | """
  4 | @Author  : captain
  5 | @time    : 18-7-10 下午8:42
  6 | @ide     : PyCharm
  7 | """
  8 | 
  9 | import torch
 10 | import time
 11 | import torch.nn.functional as F
 12 | import models
 13 | import data
 14 | from config import DefaultConfig
 15 | import pandas as pd
 16 | import os
 17 | import fire
 18 | from sklearn import metrics
 19 | import numpy as np
 20 | 
 21 | best_score = 0.0
 22 | t1 = time.time()
 23 | 
 24 | 
 25 | def main(**kwargs):
 26 |     args = DefaultConfig()
 27 |     args.parse(kwargs)
 28 |     if not torch.cuda.is_available():
 29 |         args.cuda = False
 30 |         args.device = None
 31 |         torch.manual_seed(args.seed)  # set random seed for cpu
 32 | 
 33 |     train_iter, val_iter, test_iter, args.vocab_size, vectors = data.load_data(args)
 34 | 
 35 |     args.print_config()
 36 | 
 37 |     global best_score
 38 | 
 39 |     # init model
 40 |     model = getattr(models, args.model)(args, vectors)
 41 |     print(model)
 42 | 
 43 |     # 模型保存位置
 44 |     if not os.path.exists(args.save_dir):
 45 |         os.mkdir(args.save_dir)
 46 |     save_path = os.path.join(args.save_dir, '{}_{}.pth'.format(args.model, args.id))
 47 | 
 48 |     if args.cuda:
 49 |         torch.cuda.set_device(args.device)
 50 |         torch.cuda.manual_seed(args.seed)  # set random seed for gpu
 51 |         model.cuda()
 52 | 
 53 |     # 目标函数和优化器
 54 |     criterion = F.cross_entropy
 55 |     lr1, lr2 = args.lr1, args.lr2
 56 |     optimizer = model.get_optimizer(lr1, lr2, args.weight_decay)
 57 | 
 58 |     for i in range(args.max_epochs):
 59 |         total_loss = 0.0
 60 |         correct = 0
 61 |         total = 0
 62 | 
 63 |         model.train()
 64 | 
 65 |         for idx, batch in enumerate(train_iter):
 66 |             # 训练模型参数
 67 |             # 使用BatchNorm层时，batch size不能为1
 68 |             if len(batch) == 1:
 69 |                 continue
 70 |             text, label = batch.text, batch.label
 71 |             if args.cuda:
 72 |                 text, label = text.cuda(), label.cuda()
 73 | 
 74 |             optimizer.zero_grad()
 75 |             pred = model(text)
 76 |             loss = criterion(pred, label)
 77 |             loss.backward()
 78 |             optimizer.step()
 79 | 
 80 |             # 更新统计指标
 81 |             total_loss += loss.item()
 82 |             predicted = pred.max(1)[1]
 83 |             total += label.size(0)
 84 |             correct += predicted.eq(label).sum().item()
 85 | 
 86 |             if idx % 80 == 79:
 87 |                 print('[{}, {}] loss: {:.3f} | Acc: {:.3f}%({}/{})'.format(i + 1, idx + 1, total_loss / 20,
 88 |                                                                            100. * correct / total, correct, total))
 89 |                 total_loss = 0.0
 90 | 
 91 |         # 计算再验证集上的分数，并相应调整学习率
 92 |         f1score = val(model, val_iter, args)
 93 |         if f1score > best_score:
 94 |             best_score = f1score
 95 |             checkpoint = {
 96 |                 'state_dict': model.state_dict(),
 97 |                 'config': args
 98 |             }
 99 |             torch.save(checkpoint, save_path)
100 |             print('Best tmp model f1score: {}'.format(best_score))
101 |         if f1score < best_score:
102 |             model.load_state_dict(torch.load(save_path)['state_dict'])
103 |             lr1 *= args.lr_decay
104 |             lr2 = 2e-4 if lr2 == 0 else lr2 * 0.8
105 |             optimizer = model.get_optimizer(lr1, lr2, 0)
106 |             print('* load previous best model: {}'.format(best_score))
107 |             print('* model lr:{}  emb lr:{}'.format(lr1, lr2))
108 |             if lr1 < args.min_lr:
109 |                 print('* training over, best f1 score: {}'.format(best_score))
110 |                 break
111 | 
112 |     # 保存训练最终的模型
113 |     args.best_score = best_score
114 |     final_model = {
115 |         'state_dict': model.state_dict(),
116 |         'config': args
117 |     }
118 |     best_model_path = os.path.join(args.save_dir, '{}_{}_{}.pth'.format(args.model, args.text_type, best_score))
119 |     torch.save(final_model, best_model_path)
120 |     print('Best Final Model saved in {}'.format(best_model_path))
121 | 
122 |     # 在测试集上运行模型并生成概率结果和提交结果
123 |     if not os.path.exists('result/'):
124 |         os.mkdir('result/')
125 |     probs, test_pred = test(model, test_iter, args)
126 |     result_path = 'result/' + '{}_{}_{}'.format(args.model, args.id, args.best_score)
127 |     np.save('{}.npy'.format(result_path), probs)
128 |     print('Prob result {}.npy saved!'.format(result_path))
129 | 
130 |     test_pred[['id', 'class']].to_csv('{}.csv'.format(result_path), index=None)
131 |     print('Result {}.csv saved!'.format(result_path))
132 | 
133 |     t2 = time.time()
134 |     print('time use: {}'.format(t2 - t1))
135 | 
136 | 
137 | def test(model, test_data, args):
138 |     # 生成测试提交数据csv
139 |     # 将模型设为验证模式
140 |     model.eval()
141 | 
142 |     result = np.zeros((0,))
143 |     probs_list = []
144 |     with torch.no_grad():
145 |         for batch in test_data:
146 |             text = batch.text
147 |             if args.cuda:
148 |                 text = text.cuda()
149 |             outputs = model(text)
150 |             probs = F.softmax(outputs, dim=1)
151 |             probs_list.append(probs.cpu().numpy())
152 |             pred = outputs.max(1)[1]
153 |             result = np.hstack((result, pred.cpu().numpy()))
154 | 
155 |     # 生成概率文件npy
156 |     prob_cat = np.concatenate(probs_list, axis=0)
157 | 
158 |     test = pd.read_csv('/data/yujun/datasets/daguanbei_data/test_set.csv')
159 |     test_id = test['id'].copy()
160 |     test_pred = pd.DataFrame({'id': test_id, 'class': result})
161 |     test_pred['class'] = (test_pred['class'] + 1).astype(int)
162 | 
163 |     return prob_cat, test_pred
164 | 
165 | 
166 | def val(model, dataset, args):
167 |     # 计算模型在验证集上的分数
168 | 
169 |     # 将模型设为验证模式
170 |     model.eval()
171 | 
172 |     acc_n = 0
173 |     val_n = 0
174 |     predict = np.zeros((0,), dtype=np.int32)
175 |     gt = np.zeros((0,), dtype=np.int32)
176 |     with torch.no_grad():
177 |         for batch in dataset:
178 |             text, label = batch.text, batch.label
179 |             if args.cuda:
180 |                 text, label = text.cuda(), label.cuda()
181 |             outputs = model(text)
182 |             pred = outputs.max(1)[1]
183 |             acc_n += (pred == label).sum().item()
184 |             val_n += label.size(0)
185 |             predict = np.hstack((predict, pred.cpu().numpy()))
186 |             gt = np.hstack((gt, label.cpu().numpy()))
187 | 
188 |     acc = 100. * acc_n / val_n
189 |     f1score = np.mean(metrics.f1_score(predict, gt, average=None))
190 |     print('* Test Acc: {:.3f}%({}/{}), F1 Score: {}'.format(acc, acc_n, val_n, f1score))
191 |     return f1score
192 | 
193 | 
194 | if __name__ == '__main__':
195 |     fire.Fire()
196 | 


--------------------------------------------------------------------------------
/main_boost.py:
--------------------------------------------------------------------------------
  1 | # -*- coding: utf-8 -*-
  2 | 
  3 | """
  4 | # @Author  : captain
  5 | # @Time    : 2018/8/27 19:08
  6 | # @Ide     : PyCharm
  7 | """
  8 | 
  9 | import torch
 10 | import time
 11 | import torch.nn.functional as F
 12 | import models
 13 | import data
 14 | from config import DefaultConfig
 15 | import pandas as pd
 16 | import os
 17 | import fire
 18 | from sklearn import metrics
 19 | import numpy as np
 20 | 
 21 | t1 = time.time()
 22 | 
 23 | 
 24 | def main(**kwargs):
 25 |     args = DefaultConfig()
 26 |     args.parse(kwargs)
 27 |     # boost模型
 28 |     args.max_epochs = 5
 29 | 
 30 |     if not torch.cuda.is_available():
 31 |         args.cuda = False
 32 |         args.device = None
 33 |         torch.manual_seed(args.seed)  # set random seed for cpu
 34 | 
 35 |     train_iter, val_iter, test_iter, args.vocab_size, vectors = util.load_data(args)
 36 | 
 37 |     args.print_config()
 38 | 
 39 |     # 模型保存位置
 40 |     if not os.path.exists(args.save_dir):
 41 |         os.mkdir(args.save_dir)
 42 |     save_path = os.path.join(args.save_dir, '{}_{}_{}.pth'.format(args.model, args.text_type, args.id))
 43 | 
 44 |     if args.cuda:
 45 |         torch.cuda.set_device(args.device)
 46 |         torch.cuda.manual_seed(args.seed)  # set random seed for gpu
 47 | 
 48 |     for lay_i in range(args.bo_layers):
 49 |         print('-------------- lay {} ---------------'.format(lay_i))
 50 |         model = getattr(models, args.model)(args, vectors)
 51 |         model = model.cuda()
 52 |         print(model)
 53 |         best_score = 0.0
 54 | 
 55 |         # 目标函数和优化器
 56 |         criterion = F.cross_entropy
 57 |         lr1 = args.lr1
 58 |         lr2 = args.lr2
 59 |         optimizer = model.get_optimizer(lr1, lr2, args.weight_decay)
 60 | 
 61 |         if lay_i != 0:
 62 |             # 加载上一层模型的loss weight
 63 |             saved_model = torch.load(args.model_path)
 64 |             loss_weight = saved_model['loss_weight']
 65 |             print(list(enumerate(loss_weight)))
 66 |             loss_weight = loss_weight.cuda()
 67 | 
 68 |         for i in range(args.max_epochs):
 69 |             total_loss = 0.0
 70 |             correct = 0
 71 |             total = 0
 72 | 
 73 |             model.train()
 74 | 
 75 |             for idx, batch in enumerate(train_iter):
 76 |                 # 训练模型参数
 77 |                 # 使用BatchNorm层时，batch size不能为1
 78 |                 if len(batch) == 1:
 79 |                     continue
 80 |                 text, label = batch.text, batch.label
 81 |                 if args.cuda:
 82 |                     text, label = text.cuda(), label.cuda()
 83 | 
 84 |                 optimizer.zero_grad()
 85 |                 pred = model(text)
 86 |                 if lay_i != 0:
 87 |                     loss = criterion(pred, label, weight=loss_weight + 1 - loss_weight.mean())
 88 |                 else:
 89 |                     loss = criterion(pred, label)
 90 |                 loss.backward()
 91 |                 optimizer.step()
 92 | 
 93 |                 # 更新统计指标
 94 |                 total_loss += loss.item()
 95 |                 predicted = pred.max(1)[1]
 96 |                 total += label.size(0)
 97 |                 correct += predicted.eq(label).sum().item()
 98 | 
 99 |                 if idx % 80 == 79:
100 |                     print('[{}, {}] loss: {:.3f} | Acc: {:.3f}%({}/{})'.format(i + 1, idx + 1, total_loss / 20,
101 |                                                                                100. * correct / total, correct, total))
102 |                     total_loss = 0.0
103 | 
104 |             # 计算再验证集上的分数，并相应调整学习率
105 |             f1score, tmp_loss_weight = val(model, val_iter, args)
106 |             if f1score > best_score:
107 |                 best_score = f1score
108 |                 checkpoint = {
109 |                     'state_dict': model.state_dict(),
110 |                     'config': args
111 |                 }
112 |                 torch.save(checkpoint, save_path)
113 |                 print('Best tmp model f1score: {}'.format(best_score))
114 |             if f1score < best_score:
115 |                 model.load_state_dict(torch.load(save_path)['state_dict'])
116 |                 lr1 *= args.lr_decay
117 |                 lr2 = 2e-4 if lr2 == 0 else lr2 * 0.8
118 |                 optimizer = model.get_optimizer(lr1, lr2, 0)
119 |                 print('* load previous best model: {}'.format(best_score))
120 |                 print('* model lr:{}  emb lr:{}'.format(lr1, lr2))
121 |                 if lr1 < args.min_lr:
122 |                     print('* training over, best f1 score: {}'.format(best_score))
123 |                     break
124 | 
125 |         # 保存训练最终的模型
126 |         # 保存当前层的loss weight
127 |         loss_weight = tmp_loss_weight
128 |         args.best_score = best_score
129 |         final_model = {
130 |             'state_dict': model.state_dict(),
131 |             'config': args,
132 |             'loss_weight': loss_weight
133 |         }
134 |         args.model_path = os.path.join(args.save_dir,
135 |                                        '{}_{}_lay{}_{}.pth'.format(args.model, args.text_type, lay_i, best_score))
136 |         torch.save(final_model, args.model_path)
137 |         print('Best Final Model saved in {}'.format(args.model_path))
138 | 
139 |     t2 = time.time()
140 |     print('time use: {}'.format(t2 - t1))
141 | 
142 | 
143 | def val(model, dataset, args):
144 |     # 计算模型在验证集上的分数
145 | 
146 |     # 将模型设为验证模式
147 |     model.eval()
148 |     sample_per_class = torch.zeros(args.label_size)
149 |     error_per_class = torch.zeros(args.label_size)
150 | 
151 |     acc_n = 0
152 |     val_n = 0
153 |     predict = np.zeros((0,), dtype=np.int32)
154 |     gt = np.zeros((0,), dtype=np.int32)
155 |     with torch.no_grad():
156 |         for batch in dataset:
157 |             text, label = batch.text, batch.label
158 |             if args.cuda:
159 |                 text, label = text.cuda(), label.cuda()
160 |             outputs = model(text)
161 |             pred = outputs.max(1)[1]
162 |             acc_n += (pred == label).sum().item()
163 |             val_n += label.size(0)
164 |             predict = np.hstack((predict, pred.cpu().numpy()))
165 |             gt = np.hstack((gt, label.cpu().numpy()))
166 | 
167 |             for pre, la in zip(pred, label):
168 |                 sample_per_class[la] += 1
169 |                 if pre != la:
170 |                     error_per_class[la] += 1
171 | 
172 |     acc = 100. * acc_n / val_n
173 |     f1score = np.mean(metrics.f1_score(predict, gt, average=None))
174 |     print('* Test Acc: {:.3f}%({}/{}), F1 Score: {}'.format(acc, acc_n, val_n, f1score))
175 |     return f1score, error_per_class / sample_per_class
176 | 
177 | 
178 | if __name__ == '__main__':
179 |     fire.Fire()
180 | 


--------------------------------------------------------------------------------
/models/AttLSTM.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | # @Author  : captain
 5 | # @Time    : 2018/8/12 15:35
 6 | # @Ide     : PyCharm
 7 | """
 8 | 
 9 | import torch
10 | import numpy as np
11 | import torch.nn as nn
12 | from .BasicModule import BasicModule
13 | from config import DefaultConfig
14 | 
15 | 
16 | class AttLSTM(BasicModule):
17 |     def __init__(self, args, vectors):
18 |         super(AttLSTM, self).__init__()
19 |         self.args = args
20 | 
21 |         # LSTM
22 |         self.embedding = nn.Embedding(args.vocab_size, args.embedding_dim)
23 |         self.embedding.weight.data.copy_(vectors)
24 |         self.bilstm = nn.LSTM(
25 |             input_size=args.embedding_dim,
26 |             hidden_size=args.hidden_dim,
27 |             num_layers=args.lstm_layers,
28 |             batch_first=False,
29 |             dropout=args.lstm_dropout,
30 |             bidirectional=True
31 |         )
32 | 
33 |         self.fc = nn.Sequential(
34 |             nn.Linear(args.hidden_dim * 2, args.linear_hidden_size),
35 |             nn.BatchNorm1d(args.linear_hidden_size),
36 |             nn.ReLU(inplace=True),
37 |             nn.Linear(args.linear_hidden_size, args.label_size)
38 |         )
39 | 
40 |     def attention(self, rnn_out, state):
41 |         merged_state = torch.cat([s for s in state], 1)
42 |         merged_state = merged_state.unsqueeze(2)
43 |         # (batch, seq, hidden) * (batch, hidden, 1) = (batch, seq, 1)
44 |         weights = torch.bmm(rnn_out.permute(0, 2, 1), merged_state)
45 |         weights = torch.nn.functional.softmax(weights.squeeze(2), dim=1).unsqueeze(2)
46 |         # (batch, hidden, seq) * (batch, seq, 1) = (batch, hidden, 1)
47 |         return torch.bmm(rnn_out, weights).squeeze(2)
48 | 
49 |     def forward(self, text):
50 |         embed = self.embedding(text)  # seq * batch * emb
51 |         out, hidden = self.bilstm(embed)
52 |         out = out.permute(1, 2, 0)  # batch * hidden * seq
53 |         h_n, c_n = hidden
54 |         att_out = self.attention(out, h_n)
55 | 
56 |         logits = self.fc(att_out)
57 |         return logits
58 | 


--------------------------------------------------------------------------------
/models/BasicModule.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | @Author  : captain
 5 | @time    : 18-7-11 上午10:09
 6 | @ide     : PyCharm  
 7 | """
 8 | import torch
 9 | import time
10 | 
11 | 
12 | class BasicModule(torch.nn.Module):
13 |     # 封装了nn.Module， 主要是提供了save和load两个方法
14 | 
15 |     def __init__(self):
16 |         super(BasicModule, self).__init__()
17 |         self.model_name = str(type(self))  # 模型的默认名字
18 | 
19 |     # def load(self, path):
20 |     #     # 可加载指定路径的模型
21 |     #     data = torch.load(path)['state_dict']
22 |     #     self.load_state_dict(data)
23 |     #     return self.cuda()
24 |     #
25 |     # def save(self, name=None, new=False):
26 |     #     prefix = 'checkpoints/' + self.model_name + '_' + self.config.text_type + '_'
27 |     #     if name is None:
28 |     #         name = time.strftime('%m%d_%H:%M:%S.pth')
29 |     #     path = prefix + name
30 |     #
31 |     #     if new:
32 |     #         data = {'opt': self.opt.state_dict(), 'd': self.state_dict()}
33 |     #     else:
34 |     #         data = self.state_dict()
35 |     #
36 |     #     torch.save(data, path)
37 |     #     return path
38 | 
39 |     def get_optimizer(self, lr1, lr2=0, weight_decay=0):
40 |         embed_params = list(map(id, self.embedding.parameters()))
41 |         base_params = filter(lambda p: id(p) not in embed_params, self.parameters())
42 |         optimizer = torch.optim.Adam([
43 |             {'params': self.embedding.parameters(), 'lr': lr2},
44 |             {'params': base_params, 'lr': lr1, 'weight_decay': weight_decay}
45 |         ])
46 |         return optimizer
47 | 
48 |     # def get_optimizer(self, lr, lr_emb=0, weight_decay=0):
49 |     #     ignored_params = list(map(id, self.embedding.parameters()))
50 |     #     base_params = filter(lambda p: id(p) not in ignored_params,
51 |     #                          self.parameters())
52 |     #     if lr_emb is None:
53 |     #         lr_emb = lr * 0.5
54 |     #     optimizer = torch.optim.Adam([
55 |     #         dict(params=base_params, weight_decay=weight_decay, lr=lr),
56 |     #         {'params': self.embedding.parameters(), 'lr': lr_emb}
57 |     #     ])
58 |     #     return optimizer
59 | 


--------------------------------------------------------------------------------
/models/FastText.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | # @Time    : 18/7/16 下午9:17
 4 | # @Author  : Captain
 5 | # @Site    : 
 6 | # @File    : FastText.py
 7 | # @Software: PyCharm Community Edition
 8 | 
 9 | from .BasicModule import BasicModule
10 | from config import DefaultConfig
11 | import torch
12 | from torch import nn
13 | import torch.nn.functional as F
14 | import numpy as np
15 | 
16 | 
17 | class FastText(BasicModule):
18 |     def __init__(self, config, vectors=None):
19 |         super(FastText, self).__init__()
20 |         self.config = config
21 | 
22 |         self.embedding = nn.Embedding(config.vocab_size, config.embedding_dim)
23 |         if vectors is not None:
24 |             self.embedding.weight.data.copy_(vectors)
25 | 
26 |         self.pre = nn.Sequential(
27 |             nn.Linear(config.embedding_dim, config.embedding_dim * 2),
28 |             nn.BatchNorm1d(config.embedding_dim * 2),
29 |             nn.ReLU(True)
30 |         )
31 | 
32 |         self.fc = nn.Sequential(
33 |             nn.Linear(config.embedding_dim * 2, config.linear_hidden_size),
34 |             nn.BatchNorm1d(config.linear_hidden_size),
35 |             nn.ReLU(inplace=True),
36 |             nn.Linear(config.linear_hidden_size, config.label_size)
37 |         )
38 | 
39 |     def forward(self, sentence):
40 |         embed = self.embedding(sentence)  # seq * batch * emb
41 |         embed_size = embed.size()
42 |         out = self.pre(embed.contiguous().view(-1, self.config.embedding_dim)).view(embed_size[0], embed_size[1], -1)
43 |         mean_out = torch.mean(out, dim=0).squeeze()  # batch * 2emb
44 | 
45 |         logit = self.fc(mean_out)
46 |         return logit
47 | 


--------------------------------------------------------------------------------
/models/GRU.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | # @Author  : captain
 5 | # @Time    : 2018/8/25 22:08
 6 | # @Ide     : PyCharm
 7 | """
 8 | import torch
 9 | import torch.nn as nn
10 | import torch.nn.functional as F
11 | from .BasicModule import BasicModule
12 | from config import DefaultConfig
13 | import word2vec
14 | 
15 | 
16 | def kmax_pooling(x, dim, k):
17 |     index = x.topk(k, dim=dim)[1].sort(dim=dim)[0]  # torch.Tensor.topk()的输出有两项，后一项为索引
18 |     return x.gather(dim, index)
19 | 
20 | 
21 | class GRU(BasicModule):
22 |     def __init__(self, config, vectors=None):
23 |         super(GRU, self).__init__()
24 |         self.opt = config
25 |         self.kmax_pooling = config.kmax_pooling
26 | 
27 |         # GRU
28 |         self.embedding = nn.Embedding(config.vocab_size, config.embedding_dim)
29 |         if vectors is not None:
30 |             self.embedding.weight.data.copy_(vectors)
31 |         self.bigru = nn.GRU(
32 |             input_size=config.embedding_dim,
33 |             hidden_size=config.hidden_dim,
34 |             num_layers=config.lstm_layers,
35 |             batch_first=False,
36 |             dropout=config.lstm_dropout,
37 |             bidirectional=True)
38 | 
39 |         # self.fc = nn.Linear(args.hidden_dim * 2 * 2, args.label_size)
40 |         # 两层全连接层，中间添加批标准化层
41 |         # 全连接层隐藏元个数需要再做修改
42 |         self.fc = nn.Sequential(
43 |             nn.Linear(self.kmax_pooling * (config.hidden_dim * 2), config.linear_hidden_size),
44 |             nn.BatchNorm1d(config.linear_hidden_size),
45 |             nn.ReLU(inplace=True),
46 |             nn.Linear(config.linear_hidden_size, config.label_size)
47 |         )
48 | 
49 |     # 对LSTM所有隐含层的输出做kmax pooling
50 |     def forward(self, text):
51 |         embed = self.embedding(text)  # seq*batch*emb
52 |         out = self.bigru(embed)[0].permute(1, 2, 0)  # batch * hidden * seq
53 |         pooling = kmax_pooling(out, 2, self.kmax_pooling)  # batch * hidden * kmax
54 | 
55 |         # word+article
56 |         flatten = pooling.view(pooling.size(0), -1)
57 |         logits = self.fc(flatten)
58 | 
59 |         return logits
60 | 


--------------------------------------------------------------------------------
/models/Inception.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | # @Author  : captain
 5 | # @Time    : 2018/9/5 22:43
 6 | # @Ide     : PyCharm
 7 | """
 8 | 
 9 | from .BasicModule import BasicModule
10 | import torch as t
11 | import torch
12 | import numpy as np
13 | from torch import nn
14 | from collections import OrderedDict
15 | 
16 | 
17 | class Ince(nn.Module):
18 |     def __init__(self, cin, co, relu=True, norm=True):
19 |         super(Ince, self).__init__()
20 |         assert (co % 4 == 0)
21 |         cos = [co // 4] * 4
22 |         self.activa = nn.Sequential()
23 |         if norm:
24 |             self.activa.add_module('norm', nn.BatchNorm1d(co))
25 |         if relu:
26 |             self.activa.add_module('relu', nn.ReLU(True))
27 |         self.branch1 = nn.Sequential(OrderedDict([
28 |             ('conv1', nn.Conv1d(cin, cos[0], 1, stride=1)),
29 |         ]))
30 |         self.branch2 = nn.Sequential(OrderedDict([
31 |             ('conv1', nn.Conv1d(cin, cos[1], 1)),
32 |             ('norm1', nn.BatchNorm1d(cos[1])),
33 |             ('relu1', nn.ReLU(inplace=True)),
34 |             ('conv3', nn.Conv1d(cos[1], cos[1], 3, stride=1, padding=1)),
35 |         ]))
36 |         self.branch3 = nn.Sequential(OrderedDict([
37 |             ('conv1', nn.Conv1d(cin, cos[2], 3, padding=1)),
38 |             ('norm1', nn.BatchNorm1d(cos[2])),
39 |             ('relu1', nn.ReLU(inplace=True)),
40 |             ('conv3', nn.Conv1d(cos[2], cos[2], 5, stride=1, padding=2)),
41 |         ]))
42 |         self.branch4 = nn.Sequential(OrderedDict([
43 |             # ('pool',nn.MaxPool1d(2)),
44 |             ('conv3', nn.Conv1d(cin, cos[3], 3, stride=1, padding=1)),
45 |         ]))
46 | 
47 |     def forward(self, x):
48 |         branch1 = self.branch1(x)
49 |         branch2 = self.branch2(x)
50 |         branch3 = self.branch3(x)
51 |         branch4 = self.branch4(x)
52 |         result = self.activa(torch.cat((branch1, branch2, branch3, branch4), 1))
53 |         return result
54 | 
55 | 
56 | class InCNN(BasicModule):
57 |     def __init__(self, config, vectors=None):
58 |         super(InCNN, self).__init__()
59 |         self.opt = config
60 |         self.embedding = nn.Embedding(config.vocab_size, config.embedding_dim)
61 |         if vectors is not None:
62 |             self.embedding.weight.data.copy_(vectors)
63 | 
64 |         self.conv = nn.Sequential(
65 |             Ince(config.embedding_dim, 200),
66 |             Ince(200, 200),
67 |             nn.MaxPool1d(config.max_text_len)
68 |         )
69 | 
70 |         self.fc = nn.Sequential(
71 |             nn.Linear(200, config.linear_hidden_size),
72 |             nn.BatchNorm1d(config.linear_hidden_size),
73 |             nn.ReLU(inplace=True),
74 |             nn.Linear(100, config.label_size)
75 |         )
76 | 
77 |     def forward(self, text):
78 |         embed = self.embedding(text)  # seq*batch*emb
79 |         out = self.conv(embed.permute(1, 2, 0))  # batch*emb*seq
80 | 
81 |         flatten = out.view(out.size(0), -1)
82 |         logits = self.fc(flatten)
83 | 
84 |         return logits
85 | 


--------------------------------------------------------------------------------
/models/LSTM.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | @Author  : captain
 5 | @time    : 18-7-10 下午8:44
 6 | @ide     : PyCharm  
 7 | """
 8 | import torch
 9 | import torch.nn as nn
10 | import torch.nn.functional as F
11 | from .BasicModule import BasicModule
12 | from config import DefaultConfig
13 | 
14 | 
15 | # 有两个做法有待实验验证
16 | # 1、kmax_pooling的使用，对所有RNN的输出做最大池化
17 | # 2、分类器选用两层全连接层+BN层，还是直接使用一层全连接层
18 | # 3、是否需要init_hidden
19 | 
20 | def kmax_pooling(x, dim, k):
21 |     index = x.topk(k, dim=dim)[1].sort(dim=dim)[0]  # torch.Tensor.topk()的输出有两项，后一项为索引
22 |     return x.gather(dim, index)
23 | 
24 | 
25 | class LSTM(BasicModule):
26 |     def __init__(self, config, vectors):
27 |         super(LSTM, self).__init__()
28 |         self.config = config
29 |         self.kmax_pooling = config.kmax_pooling
30 | 
31 | 
32 |         # LSTM
33 |         self.embedding = nn.Embedding(config.vocab_size, config.embedding_dim)
34 |         self.embedding.weight.data.copy_(vectors)
35 |         self.bilstm = nn.LSTM(
36 |             input_size=config.embedding_dim,
37 |             hidden_size=config.hidden_dim,
38 |             num_layers=config.lstm_layers,
39 |             batch_first=False,
40 |             dropout=config.lstm_dropout,
41 |             bidirectional=True)
42 | 
43 |         # self.fc = nn.Linear(args.hidden_dim * 2 * 2, args.label_size)
44 |         # 两层全连接层，中间添加批标准化层
45 |         # 全连接层隐藏元个数需要再做修改
46 |         self.fc = nn.Sequential(
47 |             nn.Linear(self.kmax_pooling * (config.hidden_dim * 2), config.linear_hidden_size),
48 |             nn.BatchNorm1d(config.linear_hidden_size),
49 |             nn.ReLU(inplace=True),
50 |             nn.Linear(config.linear_hidden_size, config.label_size)
51 |         )
52 | 
53 |     # 对LSTM所有隐含层的输出做kmax pooling
54 |     def forward(self, text):
55 |         embed = self.embedding(text)  # seq*batch*emb
56 |         out = self.bilstm(embed)[0].permute(1, 2, 0)  # batch * hidden * seq
57 |         pooling = kmax_pooling(out, 2, self.kmax_pooling)  # batch * hidden * kmax
58 | 
59 |         # word+article
60 |         flatten = pooling.view(pooling.size(0), -1)
61 |         logits = self.fc(flatten)
62 | 
63 |         return logits
64 | 


--------------------------------------------------------------------------------
/models/RCNN.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | @Author  : captain
 5 | @time    : 18-8-10 下午2:58
 6 | @ide     : PyCharm  
 7 | """
 8 | 
 9 | import torch
10 | import torch.nn as nn
11 | import torch.nn.functional as F
12 | from .BasicModule import BasicModule
13 | from config import DefaultConfig
14 | 
15 | 
16 | def kmax_pooling(x, dim, k):
17 |     index = x.topk(k, dim=dim)[1].sort(dim=dim)[0]
18 |     return x.gather(dim, index)
19 | 
20 | 
21 | class RCNN(BasicModule):
22 |     def __init__(self, args, vectors=None):
23 |         super(RCNN, self).__init__()
24 |         self.kmax_k = args.kmax_pooling
25 |         self.config = args
26 | 
27 |         #
28 |         self.embedding = nn.Embedding(args.vocab_size, args.embedding_dim)
29 |         if vectors is not None:
30 |             self.embedding.weight.data.copy_(vectors)
31 |         self.lstm = nn.LSTM(
32 |             input_size=args.embedding_dim,
33 |             hidden_size=args.hidden_dim,
34 |             num_layers=args.lstm_layers,
35 |             batch_first=False,
36 |             dropout=args.lstm_dropout,
37 |             bidirectional=True
38 |         )
39 |         self.conv = nn.Sequential(
40 |             nn.Conv1d(in_channels=args.hidden_dim * 2 + args.embedding_dim, out_channels=200, kernel_size=3),
41 |             nn.BatchNorm1d(200),
42 |             nn.ReLU(inplace=True),
43 |             nn.Conv1d(in_channels=200, out_channels=200, kernel_size=3),
44 |             nn.BatchNorm1d(200),
45 |             nn.ReLU(inplace=True)
46 |         )
47 | 
48 |         # classifer
49 |         # self.fc = nn.Linear(2 * (100 + 100), args.label_size)
50 |         self.fc = nn.Sequential(
51 |             nn.Linear(2 * 200, args.linear_hidden_size),
52 |             nn.BatchNorm1d(args.linear_hidden_size),
53 |             nn.ReLU(inplace=True),
54 |             nn.Linear(args.linear_hidden_size, args.label_size)
55 |         )
56 | 
57 |     def forward(self, text):
58 |         embed = self.embedding(text)
59 |         out = self.lstm(embed)[0].permute(1, 2, 0)
60 |         out = torch.cat((out, embed.permute(1, 2, 0)), dim=1)
61 |         conv_out = kmax_pooling(self.conv(out), 2, self.kmax_k)
62 | 
63 |         flatten = conv_out.view(conv_out.size(0), -1)
64 |         logits = self.fc(flatten)
65 |         return logits
66 | 


--------------------------------------------------------------------------------
/models/RCNN1.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | # @Author  : captain
 5 | # @Time    : 2018/9/3 22:23
 6 | # @Ide     : PyCharm
 7 | """
 8 | 
 9 | import torch
10 | import torch.nn as nn
11 | import torch.nn.functional as F
12 | from .BasicModule import BasicModule
13 | from config import DefaultConfig
14 | 
15 | 
16 | def kmax_pooling(x, dim, k):
17 |     index = x.topk(k, dim=dim)[1].sort(dim=dim)[0]
18 |     return x.gather(dim, index)
19 | 
20 | 
21 | class RCNN1(BasicModule):
22 |     def __init__(self, args, vectors=None):
23 |         super(RCNN1, self).__init__()
24 |         self.kmax_k = args.kmax_pooling
25 |         self.config = args
26 | 
27 |         #
28 |         self.embedding = nn.Embedding(args.vocab_size, args.embedding_dim)
29 |         if vectors is not None:
30 |             self.embedding.weight.data.copy_(vectors)
31 |         self.gru = nn.GRU(
32 |             input_size=args.embedding_dim,
33 |             hidden_size=args.hidden_dim,
34 |             num_layers=args.lstm_layers,
35 |             batch_first=False,
36 |             dropout=args.lstm_dropout,
37 |             bidirectional=True
38 |         )
39 |         self.conv = nn.Sequential(
40 |             nn.Conv1d(in_channels=args.hidden_dim * 2 + args.embedding_dim, out_channels=args.rcnn_kernel, kernel_size=3),
41 |             nn.BatchNorm1d(args.rcnn_kernel),
42 |             nn.ReLU(inplace=True),
43 |             nn.Conv1d(in_channels=args.rcnn_kernel, out_channels=args.rcnn_kernel, kernel_size=3),
44 |             nn.BatchNorm1d(args.rcnn_kernel),
45 |             nn.ReLU(inplace=True)
46 |         )
47 | 
48 |         # classifer
49 |         # self.fc = nn.Linear(2 * (100 + 100), args.label_size)
50 |         self.fc = nn.Sequential(
51 |             nn.Linear(args.kmax_pooling * args.rcnn_kernel, args.linear_hidden_size),
52 |             nn.BatchNorm1d(args.linear_hidden_size),
53 |             nn.ReLU(inplace=True),
54 |             nn.Linear(args.linear_hidden_size, args.label_size)
55 |         )
56 | 
57 |     def forward(self, text):
58 |         embed = self.embedding(text)
59 |         out = self.gru(embed)[0].permute(1, 2, 0)
60 |         out = torch.cat((out, embed.permute(1, 2, 0)), dim=1)
61 |         conv_out = kmax_pooling(self.conv(out), 2, self.kmax_k)
62 | 
63 |         flatten = conv_out.view(conv_out.size(0), -1)
64 |         logits = self.fc(flatten)
65 |         return logits
66 | 


--------------------------------------------------------------------------------
/models/TextCNN.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | @Author  : captain
 5 | @time    : 18-7-10 下午8:55
 6 | @ide     : PyCharm  
 7 | """
 8 | 
 9 | import torch
10 | import torch.nn as nn
11 | import numpy as np
12 | from .BasicModule import BasicModule
13 | from config import DefaultConfig
14 | 
15 | kernal_sizes = [1, 2, 3, 4, 5]
16 | 
17 | 
18 | class TextCNN(BasicModule):
19 |     def __init__(self, opt, vectors=None):
20 |         super(TextCNN, self).__init__()
21 | 
22 |         '''Embedding Layer'''
23 |         # 使用预训练的词向量
24 |         self.embedding = nn.Embedding(opt.vocab_size, opt.embedding_dim)
25 |         if vectors is not None:
26 |             self.embedding.weight.data.copy_(vectors)
27 | 
28 |         convs = [
29 |             nn.Sequential(
30 |                 nn.Conv1d(in_channels=opt.embedding_dim,
31 |                           out_channels=opt.kernel_num,
32 |                           kernel_size=kernel_size),
33 |                 nn.BatchNorm1d(opt.kernel_num),
34 |                 nn.ReLU(inplace=True),
35 | 
36 |                 nn.Conv1d(in_channels=opt.kernel_num,
37 |                           out_channels=opt.kernel_num,
38 |                           kernel_size=kernel_size),
39 |                 nn.BatchNorm1d(opt.kernel_num),
40 |                 nn.ReLU(inplace=True),
41 |                 nn.MaxPool1d(kernel_size=(opt.max_text_len - kernel_size*2 + 2))
42 |             )
43 |             for kernel_size in kernal_sizes
44 |         ]
45 | 
46 |         self.convs = nn.ModuleList(convs)
47 | 
48 |         self.fc = nn.Sequential(
49 |             nn.Linear(5 * opt.kernel_num, opt.linear_hidden_size),
50 |             nn.BatchNorm1d(opt.linear_hidden_size),
51 |             nn.ReLU(inplace=True),
52 |             nn.Linear(opt.linear_hidden_size, opt.label_size)
53 |         )
54 | 
55 |     def forward(self, inputs):
56 |         embeds = self.embedding(inputs)  # seq * batch * embed
57 |         # 进入卷积层前需要将Tensor第二个维度变成emb_dim，作为卷积的通道数
58 |         conv_out = [conv(embeds.permute(1, 2, 0)) for conv in self.convs]
59 |         conv_out = torch.cat(conv_out, dim=1)
60 | 
61 |         flatten = conv_out.view(conv_out.size(0), -1)
62 |         logits = self.fc(flatten)
63 |         return logits
64 | 


--------------------------------------------------------------------------------
/models/__init__.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | @Author  : captain
 5 | @time    : 18-7-10 下午8:42
 6 | @ide     : PyCharm  
 7 | """
 8 | 
 9 | from .TextCNN import TextCNN
10 | from .LSTM import LSTM
11 | from .FastText import FastText
12 | from .RCNN import RCNN
13 | from .RCNN1 import RCNN1
14 | from .AttLSTM import AttLSTM
15 | from .GRU import GRU
16 | from .Inception import InCNN
17 | from .bigru_att import bigru_attention
18 | 


--------------------------------------------------------------------------------
/models/bigru_att.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | @Author  : captain
 5 | @time    : 18-7-10 下午8:44
 6 | @ide     : PyCharm  
 7 | """
 8 | import torch
 9 | import torch.nn as nn
10 | import torch.nn.functional as F
11 | from .BasicModule import BasicModule
12 | 
13 | class bigru_attention(BasicModule):
14 |     def __init__(self, args, vectors=None):
15 |         self.args = args
16 |         super(bigru_attention, self).__init__()
17 |         self.hidden_dim = args.hidden_dim
18 |         self.gru_layers = args.lstm_layers
19 | 
20 |         self.embedding = nn.Embedding(args.vocab_size, args.embedding_dim)
21 |         if vectors is not None:
22 |             self.embedding.weight.data.copy_(vectors)
23 |         
24 |         self.bigru = nn.GRU(args.embedding_dim, self.hidden_dim // 2, num_layers=self.gru_layers, bidirectional=True)
25 |         self.weight_W = nn.Parameter(torch.Tensor(self.hidden_dim, self.hidden_dim))
26 |         self.weight_proj = nn.Parameter(torch.Tensor(self.hidden_dim, 1))
27 |         self.fc = nn.Linear(self.hidden_dim, args.label_size)
28 | 
29 |         nn.init.uniform_(self.weight_W, -0.1, 0.1)
30 |         nn.init.uniform_(self.weight_proj, -0.1, 0.1)
31 | 
32 |     def forward(self, sentence):
33 |         embeds = self.embedding(sentence) # [seq_len, bs, emb_dim]
34 |         gru_out, _ = self.bigru(embeds) # [seq_len, bs, hid_dim]
35 |         x = gru_out.permute(1, 0, 2)
36 |         u = torch.tanh(torch.matmul(x, self.weight_W))
37 |         att = torch.matmul(u, self.weight_proj)
38 |         att_score = F.softmax(att, dim=1)
39 |         scored_x = x * att_score
40 |         feat = torch.sum(scored_x, dim=1)
41 |         y = self.fc(feat)
42 |         return y
43 | 


--------------------------------------------------------------------------------
/models/bilstm_conv.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | @Author  : captain
 5 | @time    : 18-7-10 下午8:44
 6 | @ide     : PyCharm  
 7 | """
 8 | import torch
 9 | import torch.nn as nn
10 | import torch.nn.functional as F
11 | from .BasicModule import BasicModule
12 | 
13 | 
14 | class bilstm_conv(BasicModule):
15 |     def __init__(self, args, vectors):
16 |         self.args = args
17 |         super(bilstm_conv, self).__init__()
18 |         self.hidden_dim = args.hidden_dim
19 |         self.batch_size = args.batch_size
20 |         self.use_gpu = args.use_gpu
21 |         self.lstm_layers = args.lstm_layers
22 | 
23 |         self.embedding = nn.Embedding(args.vocab_size, args.embedding_dim)
24 |         self.embedding.weight.data.copy_(vectors)
25 | 
26 |         self.bidirectional = True
27 |         if self.lstm_layers > 1:
28 |             self.dropout = args.lstm_dropout
29 |             self.bilstm = nn.LSTM(args.embedding_dim, self.hidden_dim // 2, num_layers=self.lstm_layers,
30 |                                   dropout=self.dropout, bidirectional=True)
31 |         else:
32 |             self.bilstm = nn.LSTM(args.embedding_dim, self.hidden_dim // 2, num_layers=self.lstm_layers,
33 |                                   bidirectional=True)
34 |         self.conv1 = nn.Conv1d(self.hidden_dim, 64, kernel_size=3)
35 |         self.fc = nn.Linear(64 * 2, args.label_size)
36 | 
37 |     def forward(self, sentence):
38 |         embed = self.embedding(sentence)  # [seq_len, bs, emb_dim]
39 | 
40 |         lstm_out, _ = self.bilstm(embed)  # [seq_len, bs, hid_dim]
41 |         x = F.relu(self.conv1(lstm_out.permute(1, 2, 0)))
42 |         avg_pool = torch.mean(x, dim=2)
43 |         max_pool, _ = torch.max(x, dim=2)
44 |         feat = torch.cat((avg_pool, max_pool), dim=1)
45 |         y = self.fc(feat)
46 |         return y
47 | 


--------------------------------------------------------------------------------
/result/ensemble.py:
--------------------------------------------------------------------------------
 1 | import numpy as np
 2 | import pandas as pd
 3 | 
 4 | mlist = ['GRU_word_rev_0.721350215541.npy', 'GRU_k3_0.771128871335.npy', 'GRU_lay2_0.770047017921.npy', 'GRU_k4_0.769904744261.npy', 'RCNN1_512_0.769018920031.npy', 'LSTM_word_0.768979776301.npy', 'RCNN1_200_0.767783682451.npy', 'TextCNN_base_0.760456622816.npy', 'FastText_word_0.75425891649.npy', 'GRU_art2_0.747660923499.npy', 'GRU_art_0.736854236207.npy', 'TextCNN_art_0.735192435177.npy', 'GRU_e4_1_0.771337563812.npy']
 5 | 
 6 | prob_list = [np.load(model) for model in mlist]
 7 | pre_res = np.zeros(prob_list[0].shape)
 8 | for prob in prob_list:
 9 |     pre_res += prob
10 | dl_res = pre_res / len(mlist)
11 | 
12 | n_dl = np.load('n_dl3.npy')
13 | 
14 | # svm_res = np.load('ensemble_lgb_svc_preds2.npy')
15 | svm_res = np.load('ensemble_lgb_svc_preds10.npy')
16 | all_res = 0.9*dl_res + svm_res + 0.9*n_dl
17 | 
18 | # new_res = np.load('n_RCNN1_e4_0.767059778648.npy')
19 | # all_res += new_res
20 | 
21 | res = np.argmax(all_res, axis=1)
22 | 
23 | test = pd.read_csv('/data/yujun/datasets/daguanbei_data/test_set.csv')
24 | test_id = test['id'].copy()
25 | 
26 | test_pred = pd.DataFrame({'id': test_id, 'class': res})
27 | test_pred['class'] = (test_pred['class']+1).astype(int)
28 | test_pred[['id', 'class']].to_csv('esb_3.csv', index=None)
29 | print('ensemble csv saved in {}'.format('esb_3.csv'))
30 | 


--------------------------------------------------------------------------------
/script/boost.sh:
--------------------------------------------------------------------------------
1 | nohup python main_boost.py main --model='GRU' --id='bo' --device=1 > gru_word_bo.log &
2 | 


--------------------------------------------------------------------------------
/script/run.sh:
--------------------------------------------------------------------------------
 1 | nohup python main.py main --model='bigru_attention' --device=0 --id='word' > att_word.log &
 2 | nohup python main.py main --model='bigru_attention' --device=1 --id='art' --max_text_len=3200 --text_type='article' > att_art.log &
 3 | nohup python main.py main --model='bigru_attention' --device=2 --id='word_n' --data_path='/data/yujun/captain/datasets1/' > n_att_word.log &
 4 | nohup python main.py main --model='bigru_attention' --device=3 --id='art_n' --max_text_len=3200 --text_type='article' --data_path='/data/yujun/captain/datasets1/' > n_att_art.log &
 5 | 
 6 | # nohup python main.py main --model='LSTM' --device=5 --id='word4' --emb=4 --data_path='/data/yujun/captain/datasets1/' > n_lstm_word4.log &
 7 | # nohup python main.py main --model='GRU' --device=6 --id='word4' --emb=4 --data_path='/data/yujun/captain/datasets1/' > n_gru_word4.log &
 8 | # nohup python main.py main --model='RCNN1' --device=4 --id='word4' --emb=4 --data_path='/data/yujun/captain/datasets1/' > n_rcnn_word4.log &
 9 | # nohup python main.py main --model='GRU' --device=8 --id='word41' --emb=4 --data_path='/data/yujun/captain/datasets1/' > n_gru_word41.log &
10 | # nohup python main.py main --model='TextCNN' --device=10 --id='rev4' --emb=4 --data_path='/data/yujun/captain/datasets1/' > n_cnn_4.log &
11 | 
12 | # new datasets:
13 | # nohup python main.py main --model='GRU' --device=1 --id='word_new' > n_gru_word.log &
14 | # nohup python main.py main --model='GRU' --device=2 --id='art_new' --text_type='article' --max_text_len=3200 > n_gru_art.log &
15 | # nohup python main.py main --model='TextCNN' --device=3 --id='word_new' > n_cnn_word.log &
16 | # nohup python main.py main --model='TextCNN' --device=4 --id='art_new' --text_type='article' --max_text_len=3200 > n_cnn_art.log &
17 | # nohup python main.py main --model='RCNN1' --device=6 --id='art_new' --text_type='article' --max_text_len=3200 > n_rcnn_art.log &
18 | # nohup python main.py main --model='FastText' --device=9 --id='word_new' > n_fast_word.log &
19 | 
20 | 
21 | # nohup python main.py main --model='TextCNN' --device=2 --id='word' > cnn_word &
22 | # nohup python main.py main --model='GRU' --device=9 --id='e3' > gru_e3.log &
23 | # nohup python main.py main --model='TextCNN' --device=2 --kernel_sizes='1 3 5 7 9' --id='e41' > cnn_e41.log &
24 | # nohup python main.py main --model='TextCNN' --device=3 --kernel_sizes='2 4 6 8 10' --id='e42' > cnn_e42.log &
25 | # nohup python main.py main --model='RCNN1' --device=4 --id='e4' > rcnn_e4.log &
26 | # nohup python main.py main --model='RCNN1' --device=10 --id='e4_rev' --rev=True > rcnn_e4_rev.log &
27 | # nohup python main.py main --model='RCNN1' --device=5 --data_path='/data/yujun/captain/datasets1/' --id='e4_new' > rcnn_e4_new.log &
28 | # nohup python main.py main --model='GRU' --device=6 --max_text_len=3200 --text_type='article' --id='ar_e4' > gru_ar_e4.log &
29 | # nohup python main.py main --model='TextCNN' --device=6 --id='art1' --text_type='article' --max_text_len=3200 > cnn_art1.log &
30 | # nohup python main.py main --model='TextCNN' --device=9 --id='art2' --text_type='article' --max_text_len=3200 --kernel_sizes='4 5 6 7 8' > cnn_art2.log &
31 | # nohup python main.py main --model='GRU' --device=10 --id='word_rev' --rev=True > gru_rev.log &
32 | # nohup python main.py main --model='GRU' --device=2 --embedding_dim=100 --id='100' > gru_100.log &
33 | # nohup python main.py main --model='GRU' --device=3 --embedding_dim=200 --id='200' > gru_200.log &
34 | # nohup python main.py main --model='InCNN' --device=10 --id='' > in.log &
35 | # nohup python main.py main --model='GRU' --max_epochs=0 --device=0 --id='del' --max_text_len=2000 --text_type='word' > gru_del.log &
36 | # nohup python main.py main --model='LSTM' --model_path=None --device=1 --id='new1' --max_text_len=2000 --text_type='word' > lstm_new1.log &
37 | # nohup python main.py main --model='GRU' --model_path='snapshot/GRU_base.pth' --device=0 --id='base' --max_text_len=2000 --text_type='word' --lr1=5e-5 --lr2=5e-5 > gru_tune.log &
38 | # nohup python main.py main --model='GRU' --model_path=None --device=2 --id='100' --max_text_len=2000 --text_type='word' --linear_hidden_size=100 > gru_100.log &
39 | # nohup python main.py main --model='GRU' --device=4 --id='emb4' > gru_emb4.log &
40 | # nohup python main.py main --model='GRU' --device=1 --id='art2_l' --text_type='article' --max_text_len=4000 --kmax_pooling=2 > gru_art2_l.log &
41 | # nohup python main.py main --model='GRU' --device=2 --id='art3_l' --text_type='article' --max_text_len=4000 --kmax_pooling=3 > gru_art3_l.log &
42 | # nohup python main.py main --model='GRU' --device=5 --id='art3' --text_type='article' --max_text_len=3200 --kmax_pooling=3 > gru_art3.log &
43 | # nohup python main.py main --model='GRU' --device=6 --id='art4' --text_type='article' --max_text_len=3200 --kmax_pooling=4 > gru_art4.log &
44 | # nohup python main.py main --model='FastText' --device=1 --id='word1' --batch_size=128  > fast_word1.log &
45 | # nohup python main.py main --model='FastText' --device=3 --id='art1' --batch_size=128 --max_text_len=4000 --text_type='article'> fast_art1.log &
46 | # nohup python main.py main --model='GRU' --device=9 --id='art_lay2' --text_type='article' --max_text_len=3200 -lstm_layers=2 --lstm_dropout=0.5 > gru_art_lay2.log &
47 | # nohup python main.py main --model='GRU' --device=5 --id='art' --text_type='article' --max_text_len=5000 > gru_art.log &
48 | # nohup python main.py main --model='LSTM' --device=6 --id='art' --text_type='article' --max_text_len=5000 > lstm_art.log &
49 | # nohup python main.py main --model='LSTM' --device=7 --id='base' --text_type='word' --max_text_len=2000 > lstm_base.log &
50 | # nohup python main.py main --model='TextCNN' --device=8 --id='new' > cnn_new.log &
51 | # nohup python main.py main --model='TextCNN' --device=9 --id='art' --text_type='article' --max_text_len=5000 > cnn_art.log &
52 | # nohup python main.py main --model='LSTM' --device=3 --id='100' --linear_hidden_size=100 > lstm_100.log &
53 | # nohup python main.py main --model='GRU' --device=1 --id='k1' --kmax_pooling=1 > gru_k1.log &
54 | # nohup python main.py main --model='GRU' --device=2 --id='k3' --kmax_pooling=3 > gru_k3.log &
55 | # nohup python main.py main --model='GRU' --device=6 --id='k4' --kmax_pooling=4 > gru_k4.log &
56 | # nohup python main.py main --model='GRU' --device=4 --hidden_dim=400 --id='h400' > gru_h400.log &
57 | # nohup python main.py main --model='GRU' --device=5 --kmax_pooling=4 --max_text_len=4000 --text_type='article' --id='art' > gru_art.log &
58 | # nohup python main.py main --model='GRU' --device=3 --batch_size=32 --lstm_dropout=0.5 --lstm_layers=2 --id='lay2' > gru_lay2.log &
59 | # nohup python main.py main --model='RCNN1' --device=0 --kmax_pooling=1 --rcnn_kernel=200 --id=200 > rcnn_200.log &
60 | # nohup python main.py main --model='RCNN1' --device=5 --id='512' > rcnn_512.log &
61 | # nohup python main.py main --model='RCNN1' --device=4 --id='art' --text_type='article' --max_text_len=3200 > rcnn_art.log &
62 | # nohup python main.py main --model='RCNN' --device=7 --id='art' --text_type='article' --max_text_len=4000 > rcnn_art.log &
63 | # nohup python main.py main --model='TextCNN' --device=8 --id='base' > cnn_base.log &
64 | # nohup python main.py main --model='TextCNN' --device=9 --id='art' --text_type='article' --max_text_len=4000 > cnn_art.log &
65 | # nohup python main.py main --model='GRU' --device=4 --aug=True --id='aug' > gru_aug.log &
66 | # nohup python main.py main --model='GRU' --device=2 --aug=True --max_text_len=4000 --text_type='article' --id='art_aug' > gru_art_aug.log &
67 | # nohup python main.py main --model='GRU' --device=2 --id='seed1' > gru_seed1.log &
68 | # nohup python main.py main --model='GRU' --device=3 --id='seed2' > gru_seed2.log &
69 | 
70 | 


--------------------------------------------------------------------------------
/script/tune.sh:
--------------------------------------------------------------------------------
1 | nohup python fine_tune.py tune --model_path='snapshot/RCNN_word_0.75951625072.pth' --device=0 > rcnn_tune.log &
2 | 
3 | 


--------------------------------------------------------------------------------
/script/val_gen.sh:
--------------------------------------------------------------------------------
1 | nohup python test_ensemble.py gen_test --model_path='result/bigru_attention_article_0.74793217006.pth' --device=0 &
2 | nohup python test_ensemble.py gen_test --model_path='result/RCNN1_word_0.763518737493.pth' --device=1 &
3 | # nohup python test_ensemble.py gen_test --model_path='result/LSTM_word_0.767802666207.pth' --device=1 &
4 | # nohup python test_ensemble.py gen_test --model_path='result/GRU_word_0.773231866725.pth' --device=0 &
5 | # nohup python test_ensemble.py gen_test --model_path='result/TextCNN_word_0.764570136444.pth' --device=2 &
6 | #nohup python test_ensemble.py gen_test --model_path='result/GRU_article_0.734679227475.pth' --device=1 &
7 | #nohup python test_ensemble.py gen_test --model-path='result/RCNN1_article_0.744593116562.pth' --device=10 &
8 | #nohup python test_ensemble.py gen_test --model_path='result/n_RCNN1_word_0.767059778648.pth' --device=10 & 
9 | 


--------------------------------------------------------------------------------
/test_ensemble.py:
--------------------------------------------------------------------------------
 1 | # -*- coding: utf-8 -*-
 2 | 
 3 | """
 4 | # @Author  : captain
 5 | # @Time    : 2018/8/31 20:21
 6 | # @Ide     : PyCharm
 7 | """
 8 | 
 9 | import torch
10 | import models
11 | import data
12 | import numpy as np
13 | import torch.nn.functional as F
14 | import pandas as pd
15 | from sklearn import metrics
16 | import fire
17 | from config import DefaultConfig
18 | 
19 | 
20 | 
21 | def val_ensemble():
22 |     files = []
23 |     ensmble = np.zeros([10000, 19])
24 | 
25 |     for file in files:
26 |         prob = np.load(file)
27 |         ensmble += prob
28 | 
29 |     result = np.argmax(ensmble, axis=1)
30 |     cal_f1(result)
31 | 
32 | 
33 | def cal_f1(result):
34 |     val = pd.read_csv('/data/yujun/captain/datasets/word/val_set.csv')
35 |     val_label = val['class'].as_matrix()
36 |     val_label -= 1
37 |     f1score = np.mean(metrics.f1_score(result, val_label, average=None))
38 |     print('F1 Score: {}'.format(f1score))
39 | 
40 | 
41 | def gen_test(model_path=None, device=0):
42 |     saved_model = torch.load(model_path)
43 |     config = saved_model['config']
44 |     config.device = int(device)
45 |     print('Load model from {}'.format(model_path))
46 | 
47 |     _, val_iter, _, _, _ = util.load_data(config)
48 | 
49 |     model = getattr(models, config.model)(config)
50 |     model.load_state_dict(saved_model['state_dict'])
51 |     torch.cuda.set_device(config.device)
52 |     model = model.cuda()
53 |     print(model)
54 | 
55 |     model.eval()
56 |     probs_list = []
57 |     with torch.no_grad():
58 |         for batch in val_iter:
59 |             text, label = batch.text, batch.label
60 |             text, label = text.cuda(), label.cuda()
61 |             outputs = model(text)
62 |             probs = F.softmax(outputs, dim=1)
63 |             probs_list.append(probs.cpu().numpy())
64 |     prob_cat = np.concatenate(probs_list, axis=0)
65 |     np.save('val_{}.npy'.format(model_path), prob_cat)
66 |     print('Prob result val_{}.npy saved!'.format(model_path))
67 | 
68 | 
69 | if __name__ == '__main__':
70 |     fire.Fire()
71 | 


--------------------------------------------------------------------------------
/util/split_val.py:
--------------------------------------------------------------------------------
 1 | import pandas as pd
 2 | import numpy as np
 3 | import random
 4 | 
 5 | train_data = pd.read_csv('/data/yujun/captain/datasets1/train_set.csv')
 6 | val_data = pd.read_csv('/data/yujun/captain/datasets1/val_set.csv')
 7 | test_data = pd.read_csv('/data/yujun/datasets/daguanbei_data/test_set.csv')
 8 | print('raw data loaded!')
 9 | 
10 | train_data[['word_seg', 'class']].to_csv('word/train_set.csv')
11 | val_data[['word_seg', 'class']].to_csv('word/val_set.csv')
12 | test_data[['id', 'word_seg']].to_csv('word/test_set.csv')
13 | print('word data made!')
14 | 
15 | train_data[['article', 'class']].to_csv('article/train_set.csv')
16 | val_data[['article', 'class']].to_csv('article/val_set.csv')
17 | test_data[['id', 'article']].to_csv('article/test_set.csv')
18 | print('article data made!')
19 | 


--------------------------------------------------------------------------------
/val_result/avg_ens.py:
--------------------------------------------------------------------------------
 1 | import numpy as np
 2 | import pandas as pd
 3 | from sklearn import metrics
 4 | 
 5 | val = pd.read_csv('/data/yujun/captain/datasets/val_set.csv')
 6 | val_target = (val['class']-1).astype(int)
 7 | 
 8 | # model_list = ['GRU_word_rev_0.721350215541.pth.npy', 'GRU_word_0.771128871335.pth.npy', 'GRU_word_0.770047017921.pth.npy', 'GRU_word_0.769904744261.pth.npy', 'RCNN1_word_0.769018920031.pth.npy', 'LSTM_word_0.768979776301.pth.npy', 'RCNN1_word_0.767783682451.pth.npy', 'TextCNN_word_0.760456622816.pth.npy', 'FastText_word_0.75425891649.pth.npy', 'GRU_article_0.747660923499.pth.npy', 'GRU_article_0.736854236207.pth.npy', 'TextCNN_article_0.735192435177.pth.npy', 'GRU_word_e4_0.77133756.pth.npy']
 9 | 
10 | model_list = ['GRU_word_0.771128871335.pth.npy', 'bigru_attention_word_0.775937678128.pth.npy']
11 | 
12 | # model_list = model_list[::-1]
13 | 
14 | prob_list = [np.load(model) for model in model_list]
15 | weight = [1.0] * len(model_list)
16 | 
17 | avg_res = np.zeros(prob_list[0].shape)
18 | for prob in prob_list:
19 |     avg_res += prob
20 | avg_res = np.argmax(avg_res, axis=1)
21 | avg_f1 = np.mean(metrics.f1_score(val_target, avg_res, average=None))
22 | print('Avg F1score: {}'.format(avg_f1))
23 | 
24 | 


--------------------------------------------------------------------------------
/val_result/ensemble.py:
--------------------------------------------------------------------------------
 1 | import numpy as np
 2 | import pandas as pd
 3 | from sklearn import metrics
 4 | 
 5 | val = pd.read_csv('/data/yujun/captain/datasets/val_set.csv')
 6 | val_target = (val['class']-1).astype(int)
 7 | 
 8 | # model_list = ['bigru_attention_word_0.775937678128.pth.npy', 'GRU_word_0.771128871335.pth.npy', 'GRU_word_e4_0.77133756.pth.npy', 'GRU_word_0.770047017921.pth.npy', 'RCNN1_word_0.769018920031.pth.npy', 'LSTM_word_0.768979776301.pth.npy', 'RCNN1_word_0.767783682451.pth.npy', 'TextCNN_word_0.760456622816.pth.npy', 'FastText_word_0.75425891649.pth.npy', 'GRU_article_0.747660923499.pth.npy', 'GRU_article_0.736854236207.pth.npy', 'TextCNN_article_0.735192435177.pth.npy',  'GRU_word_rev_0.721350215541.pth.npy']
 9 | 
10 | # model_list = ['GRU_article_0.740587991389.pth.npy', 'GRU_word_aug_0.7531487.pth.npy', 'RCNN_article_0.738300624533.pth.npy', 'GRU_article_0.742706978708.pth.npy', 'GRU_word_aug_0.75959.pth.npy', 'RCNN_word_0.75951625072.pth.npy', 'GRU_article_0.746959892485.pth.npy',  'GRU_word_e2_0.766984819683.pth.npy', 'TextCNN_article_0.732215582073.pth.npy', 'GRU_article_0.747660923499.pth.npy', 'GRU_word_e3_0.770530528299.pth.npy', 'TextCNN_article_0.735192435177.pth.npy', 'GRU_article_e4_0.731274858243.pth.npy', 'GRU_word_e4_0.77133756.pth.npy', 'TextCNN_word_0.759799295233.pth.npy', 'GRU_word_0.766268382394.pth.npy', 'GRU_word_rev_0.721350215541.pth.npy', 'TextCNN_word_0.760456622816.pth.npy', 'FastText_article_0.700147562111.pth.npy', 'GRU_word_0.76633007379.pth.npy', 'TextCNN_word_0.760762381324.pth.npy', 'FastText_article_0.716388158416.pth.npy', 'GRU_word_0.769904744261.pth.npy', 'LSTM_word_0.768979776301.pth.npy', 'TextCNN_word_0.761957574957.pth.npy', 'FastText_word_0.739320289459.pth.npy', 'GRU_word_0.771128871335.pth.npy', 'RCNN1_word_0.768352769502.pth.npy',  'GRU_word_100_0.765372.pth.npy', 'RCNN1_word_0.769018920031.pth.npy', 'GRU_article_0.740203015091.pth.npy', 'GRU_word_200_0.76332.pth.npy', 'RCNN1_word_e4_0.773231775419.pth.npy']
11 | 
12 | model_list = ['bigru_attention_word_0.775937678128.pth.npy', 'GRU_word_rev_0.721350215541.pth.npy', 'GRU_word_0.771128871335.pth.npy', 'GRU_word_0.770047017921.pth.npy', 'GRU_word_0.769904744261.pth.npy', 'RCNN1_word_0.769018920031.pth.npy', 'LSTM_word_0.768979776301.pth.npy', 'RCNN1_word_0.767783682451.pth.npy', 'TextCNN_word_0.760456622816.pth.npy', 'FastText_word_0.75425891649.pth.npy', 'GRU_article_0.747660923499.pth.npy', 'GRU_article_0.736854236207.pth.npy', 'TextCNN_article_0.735192435177.pth.npy', 'GRU_word_e4_0.77133756.pth.npy']
13 | 
14 | # model_list = model_list[::-1]
15 | 
16 | prob_list = [np.load(model) for model in model_list]
17 | weight = [1.0] * len(model_list)
18 | 
19 | avg_res = np.zeros(prob_list[0].shape)
20 | for prob in prob_list:
21 |     avg_res += prob
22 | avg_res = np.argmax(avg_res, axis=1)
23 | avg_f1 = np.mean(metrics.f1_score(val_target, avg_res, average=None))
24 | print('Avg F1score: {}'.format(avg_f1))
25 | 
26 | 
27 | score = [0.0] * len(model_list)
28 | score[0] = float(model_list[0].split('_')[-1][:-8])
29 | print('-' * 60)
30 | print('first model {}...'.format(model_list[0]))
31 | 
32 | for idx, model in enumerate(model_list):
33 |     if idx == 0:
34 |         continue
35 |     print('-' * 60)
36 |     print('ensemble model {}...'.format(model_list[idx]))
37 |     pre_res = np.zeros(prob_list[0].shape)
38 |     sum_weight = 0.0
39 |     for i in range(idx):
40 |         pre_res += prob_list[i] * weight[i]
41 |         sum_weight += weight[i]
42 | 
43 |     m_prob = np.load(model)
44 |     best_score = score[idx-1]
45 |     best_weight = 0.0
46 |     for w in np.linspace(0, 3, 300):
47 |         val_res = (pre_res + m_prob * w) / (sum_weight + w)
48 |         val_res = np.argmax(val_res, axis=1)
49 |         f1score = np.mean(metrics.f1_score(val_target, val_res, average=None))
50 |         # print('for weight {:.2f}, f1score is {}'.format(w, f1score))
51 |         if f1score > best_score:
52 |             best_score = f1score
53 |             best_weight = w
54 |     score[idx] = best_score
55 |     weight[idx] = best_weight
56 |     print('best weight is {:.2f}, best f1score is {}\n'.format(best_weight, best_score))
57 | 
58 | print(weight)
59 | print(score)
60 | 


--------------------------------------------------------------------------------