30 |
31 |
32 |
33 |
--------------------------------------------------------------------------------
/core/LRScheduler.py:
--------------------------------------------------------------------------------
1 | # Copyright (c) Microsoft Corporation. All rights reserved.
2 | # Licensed under the MIT license.
3 |
4 | import numpy as np
5 |
6 | class LRScheduler():
7 | def __init__(self, optimizer, decay_rate=1, minimum_lr=0, epoch_start_decay=1):
8 | """
9 |
10 | Args:
11 | optimizer:
12 | decay_rate:
13 | minimum_lr: if lr < minimum_lr, stop lr decay
14 | """
15 | self.optimizer = optimizer
16 | self.decay_rate = decay_rate
17 | self.minimum_lr = minimum_lr
18 | self.epoch_cnt = 0
19 | self.epoch_start_decay = epoch_start_decay
20 |
21 | def step(self):
22 | """ adjust learning rate
23 |
24 | Args:
25 | optimizer:
26 | decay_rate:
27 | minimum_lr:
28 |
29 | Returns:
30 | None
31 |
32 | """
33 | self.epoch_cnt += 1
34 |
35 | if self.epoch_cnt >= self.epoch_start_decay:
36 | for param_group in self.optimizer.param_groups:
37 | if param_group['lr'] * self.decay_rate >= self.minimum_lr:
38 | param_group['lr'] = param_group['lr'] * self.decay_rate
39 | else:
40 | param_group['lr'] = self.minimum_lr
41 |
42 |
43 | def get_lr(self):
44 | """ get average learning rate of optimizer.param_groups
45 |
46 | Args:
47 | optimizer:
48 |
49 | Returns:
50 |
51 | """
52 | lr_total = []
53 | for param_group in self.optimizer.param_groups:
54 | lr_total.append(param_group['lr'])
55 | return np.mean(lr_total)
56 |
--------------------------------------------------------------------------------
/core/ChineseTokenizer.py:
--------------------------------------------------------------------------------
1 | # Copyright (c) Microsoft Corporation. All rights reserved.
2 | # Licensed under the MIT license.
3 |
4 | import jieba
5 | import logging
6 | jieba.setLogLevel(logging.INFO)
7 | from nltk.tokenize.util import align_tokens
8 | from .Stopwords import Stopwords
9 |
10 | class ChineseTokenizer(object):
11 | def __init__(self, tokenizer='jieba', remove_stopwords=False):
12 | self.__tokenizer = tokenizer
13 | self.__remove_stopwords = remove_stopwords
14 | if self.__remove_stopwords:
15 | self.__stop_words = Stopwords.chinese_stopwords
16 | else:
17 | self.__stop_words = None
18 |
19 | def tokenize(self, string):
20 | if self.__tokenizer == 'jieba':
21 | tokens = list(jieba.cut(string))
22 |
23 | if self.__remove_stopwords:
24 | tokens = [word for word in tokens if word not in self.__stop_words]
25 | return tokens
26 |
27 | def span_tokenize(self, string):
28 | if self.__tokenizer == 'jieba':
29 | tokens = self.tokenize(string)
30 | spans = align_tokens(tokens, string)
31 | return spans
32 |
33 |
34 | if __name__ == '__main__':
35 | import os
36 | # nltk.data.path.append(r"C:\Users\wutlin\OneDrive - Microsoft\workspace\DNNMatchingToolkit\dataset\nltk_data")
37 | tokenizer = ChineseTokenizer(tokenizer='jieba', remove_stopwords=True)
38 | print(tokenizer.tokenize("我爱北京天安门，天安门上太阳升。"))
39 | print(tokenizer.span_tokenize("我爱北京天安门，天安门上太阳升。"))
40 | print(tokenizer.tokenize("给每一条河每一座山取一个温暖的名字；陌生人，我也为你祝福；愿你有一个灿烂的前程；愿你有情人终成眷属；愿你在尘世获得幸福；我只愿面朝大海，春暖花开。"))
41 |
--------------------------------------------------------------------------------
/docs/build/html/_sources/block_zoo.math.rst.txt:
--------------------------------------------------------------------------------
1 | block\_zoo.math package
2 | =======================
3 |
4 | Submodules
5 | ----------
6 |
7 | block\_zoo.math.Add2D module
8 | ----------------------------
9 |
10 | .. automodule:: block_zoo.math.Add2D
11 | :members:
12 | :undoc-members:
13 | :show-inheritance:
14 |
15 | block\_zoo.math.Add3D module
16 | ----------------------------
17 |
18 | .. automodule:: block_zoo.math.Add3D
19 | :members:
20 | :undoc-members:
21 | :show-inheritance:
22 |
23 | block\_zoo.math.ElementWisedMultiply2D module
24 | ---------------------------------------------
25 |
26 | .. automodule:: block_zoo.math.ElementWisedMultiply2D
27 | :members:
28 | :undoc-members:
29 | :show-inheritance:
30 |
31 | block\_zoo.math.ElementWisedMultiply3D module
32 | ---------------------------------------------
33 |
34 | .. automodule:: block_zoo.math.ElementWisedMultiply3D
35 | :members:
36 | :undoc-members:
37 | :show-inheritance:
38 |
39 | block\_zoo.math.MatrixMultiply module
40 | -------------------------------------
41 |
42 | .. automodule:: block_zoo.math.MatrixMultiply
43 | :members:
44 | :undoc-members:
45 | :show-inheritance:
46 |
47 | block\_zoo.math.Minus2D module
48 | ------------------------------
49 |
50 | .. automodule:: block_zoo.math.Minus2D
51 | :members:
52 | :undoc-members:
53 | :show-inheritance:
54 |
55 | block\_zoo.math.Minus3D module
56 | ------------------------------
57 |
58 | .. automodule:: block_zoo.math.Minus3D
59 | :members:
60 | :undoc-members:
61 | :show-inheritance:
62 |
63 |
64 | Module contents
65 | ---------------
66 |
67 | .. automodule:: block_zoo.math
68 | :members:
69 | :undoc-members:
70 | :show-inheritance:
71 |
--------------------------------------------------------------------------------
/core/EnglishTextPreprocessor.py:
--------------------------------------------------------------------------------
1 | # Copyright (c) Microsoft Corporation. All rights reserved.
2 | # Licensed under the MIT license.
3 |
4 | import ftfy
5 |
6 |
7 | class EnglishTextPreprocessor(object):
8 | def __init__(self, DBC2SBC=True, unicode_fix=True):
9 | self.__DBC2SBC = DBC2SBC
10 | self.__unicode_fix = unicode_fix
11 |
12 | def preprocess(self, string):
13 | if self.__unicode_fix:
14 | string = ftfy.fix_text(string)
15 | if self.__DBC2SBC:
16 | string = self.DBC2SBC(string)
17 | return string
18 |
19 | def DBC2SBC(self, ustring):
20 | """ DBC characters to SBC
21 |
22 | Args:
23 | ustring:
24 |
25 | Returns:
26 |
27 | """
28 | rstring = ""
29 | for uchar in ustring:
30 | inside_code = ord(uchar)
31 | if inside_code == 0x3000:
32 | inside_code = 0x0020
33 | else:
34 | inside_code -= 0xfee0
35 | if not (0x0021 <= inside_code and inside_code <= 0x7e):
36 | rstring += uchar
37 | continue
38 | rstring += chr(inside_code)
39 | return rstring
40 |
41 | def SBC2DBC(ustring):
42 | """ SBC to DBC
43 |
44 | Returns:
45 |
46 | """
47 | rstring = ""
48 | for uchar in ustring:
49 | inside_code = ord(uchar)
50 | if inside_code == 0x0020:
51 | inside_code = 0x3000
52 | else:
53 | if not (0x0021 <= inside_code and inside_code <= 0x7e):
54 | rstring += uchar
55 | continue
56 | inside_code += 0xfee0
57 | rstring += chr(inside_code)
58 | return rstring
--------------------------------------------------------------------------------
/core/EnglishTokenizer.py:
--------------------------------------------------------------------------------
1 | # Copyright (c) Microsoft Corporation. All rights reserved.
2 | # Licensed under the MIT license.
3 |
4 | import nltk
5 | import re
6 | from nltk.tokenize.util import align_tokens
7 | from .Stopwords import Stopwords
8 |
9 | class EnglishTokenizer(object):
10 | def __init__(self, tokenizer='nltk', remove_stopwords=False):
11 | self.__tokenizer = tokenizer
12 | self.__remove_stopwords = remove_stopwords
13 | if self.__remove_stopwords:
14 | self.__stop_words = Stopwords.english_stopwords
15 | else:
16 | self.__stop_words = None
17 |
18 | def tokenize(self, string):
19 | if self.__tokenizer == 'nltk':
20 | tokens = nltk.word_tokenize(string)
21 |
22 | if self.__remove_stopwords:
23 | tokens = [word for word in tokens if word not in self.__stop_words]
24 | return tokens
25 |
26 | def span_tokenize(self, string):
27 | if self.__tokenizer == 'nltk':
28 | raw_tokens = nltk.word_tokenize(string)
29 | if ('"' in string) or ("''" in string):
30 | matched = [m.group() for m in re.finditer(r"``|'{2}|\"", string)]
31 | tokens = [matched.pop(0) if tok in ['"', "``", "''"] else tok for tok in raw_tokens]
32 | else:
33 | tokens = raw_tokens
34 | spans = align_tokens(tokens, string)
35 | return spans
36 |
37 |
38 | if __name__ == '__main__':
39 | import os
40 | # nltk.data.path.append(r"C:\Users\wutlin\OneDrive - Microsoft\workspace\DNNMatchingToolkit\dataset\nltk_data")
41 | tokenizer = EnglishTokenizer(tokenizer='nltk', remove_stopwords=True)
42 | print(tokenizer.span_tokenize("""What singer did Beyonce record a song with for the movie, ''The Best Man"?"""))
--------------------------------------------------------------------------------
/losses/BaseLossConf.py:
--------------------------------------------------------------------------------
1 | # Copyright (c) Microsoft Corporation. All rights reserved.
2 | # Licensed under the MIT license.
3 |
4 | import torch
5 | from torch.nn import CrossEntropyLoss
6 | import copy
7 | import logging
8 |
9 | class BaseLossConf(object):
10 | @staticmethod
11 | def get_conf(**kwargs):
12 | # judge loss num & setting
13 | kwargs['multiLoss'] = True if len(kwargs['losses']) > 1 else False
14 | # loss = copy.deepcopy(kwargs['losses'])
15 | if kwargs['multiLoss']:
16 | if kwargs.get('multi_loss_op', '') is None:
17 | kwargs['multi_loss_op'] = 'weighted_sum'
18 | logging.info('model has multi-loss but no multi_loss_op, we set default option {0}.'.format('weighted_sum'))
19 | if kwargs.get('weights', None) is None:
20 | kwargs['weights'] = [1] * len(kwargs['losses'])
21 | logging.warning("MultiLoss have no weights, set the weights to 1.")
22 | assert len(kwargs['weights']) == len(kwargs['losses']), "The number of loss is inconsistent with loss weights!"
23 |
24 |
25 | # IF NEEDED, TRANSFORM SOME INT OR FLOAT, OR NUMPY ARRAY TO TENSORS.
26 | for single_loss in kwargs['losses']:
27 | if 'inputs' not in single_loss:
28 | raise Exception("Each loss must have inputs")
29 | if not isinstance(single_loss['inputs'], list):
30 | raise Exception('The inputs of loss must be list')
31 | if len(single_loss['inputs']) != 2:
32 | raise Exception('The length of loss inputs must be 2')
33 | if 'weight' in single_loss['conf']:
34 | single_loss['conf']['weight'] = torch.FloatTensor(single_loss['conf']['weight'])
35 |
36 | return kwargs
37 |
38 |
39 |
--------------------------------------------------------------------------------
/utils/DocInherit.py:
--------------------------------------------------------------------------------
1 | # Copyright (c) Microsoft Corporation. All rights reserved.
2 | # Licensed under the MIT license.
3 |
4 | from functools import wraps
5 |
6 | class DocInherit(object):
7 | """
8 | Docstring inheriting method descriptor
9 | The class itself is also used as a decorator doc_inherit decorator
10 |
11 | """
12 |
13 | def __init__(self, mthd):
14 | self.mthd = mthd
15 | self.name = mthd.__name__
16 |
17 | def __get__(self, obj, cls):
18 | if obj:
19 | return self.get_with_inst(obj, cls)
20 | else:
21 | return self.get_no_inst(cls)
22 |
23 | def get_with_inst(self, obj, cls):
24 | #overridden = getattr(super(cls, obj), self.name, None)
25 | for parent in cls.__mro__[1:]:
26 | overridden = getattr(parent, self.name, None)
27 | if overridden:
28 | break
29 |
30 | @wraps(self.mthd, assigned=('__name__', '__module__', '__doc__'))
31 | def f(*args, **kwargs):
32 | return self.mthd(obj, *args, **kwargs)
33 |
34 | return self.use_parent_doc(f, overridden)
35 |
36 | def get_no_inst(self, cls):
37 | for parent in cls.__mro__[1:]:
38 | overridden = getattr(parent, self.name, None)
39 | if overridden:
40 | break
41 |
42 | @wraps(self.mthd, assigned=('__name__', '__module__', '__doc__'))
43 | def f(*args, **kwargs):
44 | return self.mthd(*args, **kwargs)
45 |
46 | return self.use_parent_doc(f, overridden)
47 |
48 | def use_parent_doc(self, func, source):
49 | if source is None:
50 | raise NameError("Can't find '%s' in parents" % self.name)
51 | if func.__doc__ is None:
52 | func.__doc__ = source.__doc__
53 | return func
54 |
55 |
--------------------------------------------------------------------------------
/model_visualizer/README.md:
--------------------------------------------------------------------------------
1 | # ***NeuronBlocks*** Model Visualizer
2 |
3 | In ***NeuronBlocks***, a model visualizer is provided for visualization and configuration correctness checking,
4 | which helps users to visualize the model architecture easily during debugging.
5 |
6 | ## Installation
7 | Two libraries are needed for ***NeuronBlocks*** Model Visualizer: graphviz, web.py.
8 |
9 | You can install them via pip:
10 |
11 | ```bash
12 | pip install graphviz
13 | pip install web.py==0.40.dev0
14 | ```
15 |
16 | *Note: For Windows, please visit [Graphviz Windows Packages](https://graphviz.gitlab.io/_pages/Download/Download_windows.html) to
17 | download graphviz, then install it and set the PATH variable.*
18 |
19 | ## Usage
20 |
21 | Model Visualizer has 2 modes:
22 | - **Command Line Mode**: View model architecture via command line.
23 | - **Browser Mode**: Firstly build a model visualizer server, then view model architecture via a browser.
24 |
25 | ### Command Line Mode
26 |
27 | View model architecture via command line mode, by running:
28 | ```bash
29 | python get_model_graph.py --conf_path ../model_zoo/demo/conf.json --graph_path ./graph
30 | ```
31 | Arguments:
32 | ```bash
33 | --conf_path: [necessary] Path of the input JSON config file.
34 | --graph_path: [optional, default: './graph'] Path of the ouput model graph file.
35 | ```
36 | You will get two file: *graph.gv* and *graph.gv.svg*.
37 | Open *graph.gv.svg*, then view the model architecture.
38 |
39 | ### Browser Mode
40 |
41 | Firstly, start Model Visualizer server:
42 |
43 | ```bash
44 | cd server/
45 | python main.py 8080
46 | ```
47 | Then, you can access a model visualizer in your browser.
48 | For Linux, visit http://