-
Notifications
You must be signed in to change notification settings - Fork 9
/
main.py
442 lines (378 loc) · 24 KB
/
main.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
import argparse
from collections import Counter
from copy import deepcopy
import numpy as np
import yaml
from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras.preprocessing.sequence import pad_sequences
from keras.utils import np_utils
from sklearn.preprocessing import LabelEncoder
from src import extract_word_root_and_feature, cnn_rnn_with_context, evaluate_and_plot
from src import handle_pickles, process_words, extract_phonetic_features
def str2bool(v):
if v.lower() in ('yes', 'true', 't', 'y', '1'):
return True
elif v.lower() in ('no', 'false', 'f', 'n', '0'):
return False
else:
raise argparse.ArgumentTypeError('Boolean value expected.')
parser = argparse.ArgumentParser(description="Enter --lang = 'hindi' for Hindi and 'urdu' for Urdu; "
"--mode = 'train, test, or predict'")
parser.add_argument("--lang", required=True)
parser.add_argument("--mode", required=True, default='test')
parser.add_argument("--phonetic", type=str2bool, nargs='?')
parser.add_argument('--freezing', type=str2bool, nargs='?')
args = vars(parser.parse_args())
pickle_handler= handle_pickles.PickleHandler()
LANG, MODE = args['lang'], args['mode']
PHONETIC_FLAG = args['phonetic'] if args['phonetic'] is not None else False
FREEZER_FLAG = args['freezing'] if args['freezing'] is not None else False
CONFIG_PATH = 'config/'
VOCAB_SIZE = 89
CONTEXT_WINDOW = 4
FEATURE_NUMS = 6
def read_path_configs(filename):
with open(CONFIG_PATH + filename, 'r') as stream:
try:
res = yaml.load(stream)
except yaml.YAMLError as e:
res = None
print("Error while reading yaml: ", e)
return res
class ProcessAndTokenizeData():
def __init__(self, n_features, words, roots, features):
self.n_features = n_features
self.all_words, self.all_roots, self.all_segregated_features = words, roots, features
@staticmethod
def get_counters_for_features(all_features, flag='original'):
list_of_counters = [Counter(each) for each in all_features]
class_labels = [list(each.keys()) for each in list_of_counters]
if flag == 'original':
return class_labels
elif flag == 'transformed':
num_of_indiv_feature_tags = [max(each_cnt, key=int) + 1 for each_cnt in list_of_counters]
return class_labels, num_of_indiv_feature_tags
def process_features(self):
# list_of_counters = [Counter(each) for each in self.all_segregated_features]
# labels = [list(each.keys()) for each in list_of_counters]
if MODE == 'train':
dict_of_encoders = {i: LabelEncoder() for i in range(self.n_features)}
encoded_features = [value.fit_transform(self.all_segregated_features[idx]) for idx, value in
dict_of_encoders.items()]
class_labels_orig = self.get_counters_for_features(self.all_segregated_features, flag='original')
class_labels_transformed, num_of_indiv_feature_tags = self.get_counters_for_features(encoded_features,
flag='transformed')
categorical_features = [np_utils.to_categorical(feature, num_classes=n) for feature, n in
zip(encoded_features, num_of_indiv_feature_tags)]
_ = [pickle_handler.pickle_dumper(obj, name+'_'+LANG) for obj, name in zip([dict_of_encoders,
num_of_indiv_feature_tags,
categorical_features,
class_labels_orig,
class_labels_transformed],
["dict_of_encoders", "num_of_indiv_features",
"categorized_features", 'class_labels_orig',
'class_labels_transformed'])]
return categorical_features, num_of_indiv_feature_tags
elif MODE == 'test':
dict_of_encoders, num_of_indiv_feature_tags = [pickle_handler.pickle_loader(name+'_'+LANG) for name in
["dict_of_encoders", "num_of_indiv_features"]]
encoded_features_test = [dict_of_encoders[i].transform(self.all_segregated_features[i]) \
for i in range(self.n_features)]
categorical_features_test = [np_utils.to_categorical(feature, num_classes=n) for feature, n in
zip(encoded_features_test, num_of_indiv_feature_tags)]
return categorical_features_test, num_of_indiv_feature_tags
def process_words_and_roots(self, context_window=4):
X = [item[::-1] for item in self.all_words]
y = deepcopy(self.all_roots)
if MODE == 'train':
X_indexed = process_words.get_indexed_words(X, mode='build_vocab', vocab_size=VOCAB_SIZE, lang=LANG)
else:
X_indexed = process_words.get_indexed_words(X, mode='use_vocab', vocab_size=VOCAB_SIZE, lang=LANG)
y_indexed = process_words.get_indexed_words(y, mode='use_vocab', vocab_size=VOCAB_SIZE, lang=LANG)
input_shifter = process_words.ShiftWordsPerCW(X=X, cw=context_window, vocab_size=VOCAB_SIZE, lang=LANG)
X_indexed_left, X_indexed_right = input_shifter.shift_input()
all_inputs = list()
all_inputs.append(X_indexed)
all_inputs += X_indexed_left
all_inputs += X_indexed_right
all_inputs.append(y_indexed)
return all_inputs
def sequence_padder(in_list, maxlen):
out_list = pad_sequences(in_list, maxlen=maxlen, dtype='int32', padding='post')
return out_list
def pad_all_sequences(indexed_outputs):
max_word_len = max(max([len(word) for word in indexed_outputs[0]]), max([len(word) for word in indexed_outputs[-1]]))
all_padded_inputs = [sequence_padder(each, max_word_len) for each in indexed_outputs]
return all_padded_inputs, max_word_len
def _create_model(max_word_len, embed_dim, n, phonetic_feature_nums, freezing_call=False):
model_instance = cnn_rnn_with_context.MorphAnalyzerModels(max_word_len=max_word_len, vocab_len=VOCAB_SIZE+2,
embedding_dim=embed_dim, list_of_feature_nums=n,
cw=CONTEXT_WINDOW, use_phonetic_features=PHONETIC_FLAG,
phonetic_dims=phonetic_feature_nums)
compiled_model = model_instance.create_and_compile_model(freezer=freezing_call)
return compiled_model
def split_train_val(all_data, train_size):
train_data = [x[:train_size] for x in all_data]
val_data = [x[train_size:] for x in all_data]
return train_data, val_data
def get_decoder_input(x_train):
x_decoder_input = np.zeros_like(x_train)
x_decoder_input[:, 1:] = x_train[:, :-1]
x_decoder_input[:, 0] = 1
return x_decoder_input
def segregate_inputs_and_outputs(words_and_roots, features, decoder_inputs, phonetic_features=None):
roots = words_and_roots[-1]
inputs = words_and_roots[:-1]
inputs.append(decoder_inputs)
num_of_optimized_features = list()
if PHONETIC_FLAG is True:
tag_grouped_phonetic_features = [list(zip(*phonetic_features))[idx] for idx in range(len(phonetic_features[0]))]
_ = [inputs.append(np.array(each)) for each in tag_grouped_phonetic_features]
num_of_optimized_features = [len(each) for each in phonetic_features[0]]
outputs = [roots]
outputs += features
return inputs, outputs, num_of_optimized_features
def write_features_to_file(words, orig_features, pred_features, output_path):
encoders = pickle_handler.pickle_loader('dict_of_encoders'+'_'+LANG)
orig_features = [[np.where(idx==1)[0][0] for idx in each] for each in orig_features]
pred_features = [each.tolist() for each in pred_features]
orig_transformed_features = [encoders[i].inverse_transform(orig_features[i]) for i in range(FEATURE_NUMS)]
pred_transformed_features = [encoders[i].inverse_transform(pred_features[i]) for i in range(FEATURE_NUMS)]
for idx in range(FEATURE_NUMS):
filename = output_path+'feature_'+str(idx)+'.txt'
with open(filename, 'w', encoding='utf-8') as f:
f.write("Word\t\tOriginal_feature\t\tPredicted_feature\n")
for i,j,k in zip(words, orig_transformed_features[idx], pred_transformed_features[idx]):
f.write(i + '\t\t' + str(j) + '\t\t' + str(k) + '\n')
f.close()
def write_roots_to_file(words, orig_roots, pred_roots, output_path):
idx_to_char_mapping = pickle_handler.pickle_loader('index_to_char_mapping'+'_'+LANG)
pred_sequences = list()
for each in pred_roots:
list_of_chars = list()
list_of_chars += [idx_to_char_mapping[idx] for idx in each if idx > 0]
sequence = ''.join(list_of_chars)
pred_sequences.append(sequence)
out_file = output_path+'_words.txt'
with open(out_file, 'w', encoding='utf-8') as f:
f.write("Word\t\tOriginal_root\t\tPredicted_root\n")
for i,j,k in zip(words, orig_roots, pred_sequences):
f.write(i + '\t\t' + j + '\t\t' + str(k) + '\n')
f.close()
return orig_roots, pred_sequences
def write_predicted_roots_and_features(sentences, predictions, output_path):
encoders = pickle_handler.pickle_loader('dict_of_encoders'+'_'+LANG)
idx_to_char_mapping = pickle_handler.pickle_loader('index_to_char_mapping'+'_'+LANG)
with open(output_path + 'predictions.txt', 'w', encoding='utf-8') as f:
f.write("Word\t\tRoot\t\tPOS\t\tGender\t\tNumber\t\tPerson\t\tCase\t\tTAM\n")
for sentence, prediction in zip(sentences, predictions):
pred_features = [each.tolist() for each in predictions[1:]]
pred_transformed_features = [encoders[i].inverse_transform(pred_features[i]) for i in range(FEATURE_NUMS)]
pred_sequences = list()
for word in prediction:
list_of_chars = list()
list_of_chars += [idx_to_char_mapping[idx] for idx in word if idx > 0]
sequence = ''.join(list_of_chars)
pred_sequences.append(sequence)
all_outputs = list()
all_outputs.append(sentence)
all_outputs.append(pred_sequences)
all_outputs += pred_transformed_features
for each in zip(*all_outputs):
f.write('\t\t'.join(each) + '\n')
f.write('\n')
f.close()
def get_model_path(paths):
if PHONETIC_FLAG is True and FREEZER_FLAG is True:
key = 4
elif PHONETIC_FLAG is False and FREEZER_FLAG is True:
key = 3
elif PHONETIC_FLAG is True and FREEZER_FLAG is False:
key = 2
else:
key = 1
return paths['model_weights'][key]+'_'+LANG+'.hdf5'
def get_frozen_layer_names():
layers = ['drop0', 'drop1', 'drop2', 'drop3', 'drop4', 'drop5', 'drop6', 'drop7', 'drop8', 'drop9', 'noise0',
'noise1', 'noise2', 'noise3', 'noise4', 'noise5', 'noise6', 'noise7', 'noise8', 'noise9', 'Conv4_0',
'Conv4_1', 'Conv4_2', 'Conv4_3', 'Conv4_4', 'Conv4_5', 'Conv4_6', 'Conv4_7', 'Conv4_8', 'Conv4_9',
'Conv5_0', 'Conv5_1', 'Conv5_2', 'Conv5_3', 'Conv5_4', 'Conv5_5', 'Conv5_6', 'Conv5_7', 'Conv5_8',
'Conv5_9', 'MaxPool4_0', 'AvgPool4_0', 'MaxPool4_1', 'AvgPool4_1', 'MaxPool4_2', 'AvgPool4_2', 'MaxPool4_3',
'AvgPool4_3', 'MaxPool4_4', 'AvgPool4_4', 'MaxPool4_5', 'AvgPool4_5', 'MaxPool4_6', 'AvgPool4_6', 'MaxPool4_7',
'AvgPool4_7', 'MaxPool4_8', 'AvgPool4_8', 'MaxPool4_9', 'AvgPool4_9', 'MaxPool5_0', 'AvgPool5_0', 'MaxPool5_1',
'AvgPool5_1', 'MaxPool5_2', 'AvgPool5_2', 'MaxPool5_3', 'AvgPool5_3', 'MaxPool5_4', 'AvgPool5_4', 'MaxPool5_5',
'AvgPool5_5', 'MaxPool5_6', 'AvgPool5_6', 'MaxPool5_7', 'AvgPool5_7', 'MaxPool5_8', 'AvgPool5_8', 'MaxPool5_9',
'AvgPool5_9', 'Merge_4_0', 'Merge_4_1', 'Merge_4_2', 'Merge_4_3', 'Merge_4_4', 'Merge_4_5', 'Merge_4_6',
'Merge_4_7', 'Merge_4_8', 'Merge_4_9', 'Merge_5_0', 'Merge_5_1', 'Merge_5_2', 'Merge_5_3', 'Merge_5_4',
'Merge_5_5', 'Merge_5_6', 'Merge_5_7', 'Merge_5_8', 'Merge_5_9', 'main_merge', 'gru_1', 'phonetic_merge_0',
'phonetic_merge_1', 'phonetic_merge_2', 'phonetic_merge_3', 'phonetic_merge_4', 'phonetic_merge_5',
'dense_phonetic_0', 'dense_phonetic_1', 'dense_phonetic_2', 'dense_phonetic_3', 'dense_phonetic_4',
'dense_phonetic_5', 'dot2', 'dropout_phonetic_0', 'dropout_phonetic_1', 'dropout_phonetic_2',
'dropout_phonetic_3', 'dropout_phonetic_4', 'dropout_phonetic_5', 'dense1_0', 'dense1_1', 'dense1_2',
'dense1_3', 'dense1_4', 'dense1_5', 'drop_2_0', 'drop_2_1', 'drop_2_2', 'drop_2_3', 'drop_2_4', 'drop_2_5',
'output0', 'output1', 'output2', 'output3', 'output4', 'output5']
return layers
class RemoveErroneousIndices():
def __init__(self, test_file_contents):
self.contents = test_file_contents
self.class_labels = pickle_handler.pickle_loader('class_labels_orig_'+LANG)
self.erroneous_indices = self.get_erroneous_indices()
def filter_erroneous_indices(self, _list):
_list = [each for i, each in enumerate(_list) if i not in self.erroneous_indices]
return _list
def get_erroneous_indices(self):
erroneous_indices = list()
features = self.contents[-1]
for i, (feature, label) in enumerate(zip(features, self.class_labels)):
for idx in range(len(feature)):
# print("{} {}".format(feature[idx], label))
if feature[idx] not in label or (i == 6 and feature[idx] == 'kI'):
erroneous_indices.append(idx)
return erroneous_indices
def remove_unknown_feature_labels(self): # file_contents = [words, roots, features]
words, roots = [self.filter_erroneous_indices(each) for each in self.contents[:2]]
features = [self.filter_erroneous_indices(each) for each in self.contents[-1]]
return words, roots, features
class ProcessDataForModel():
def __init__(self, words, roots, features):
self.words = words
self.roots = roots
self.features = features
def phonetic_features_extractor(self):
extractor = extract_phonetic_features.PhoneticFeatures(self.words)
features = extractor.get_features()
features = [word_feature[:FEATURE_NUMS] for word_feature in features]
return features
def process_end_to_end(self):
data_processor = ProcessAndTokenizeData(n_features=FEATURE_NUMS, words=self.words,
roots=self.roots,
features = self.features)
categorized_features, n = data_processor.process_features()
indexed_inputs = data_processor.process_words_and_roots(CONTEXT_WINDOW)
padded_indexed_inputs, max_word_len = pad_all_sequences(indexed_inputs)
padded_indexed_inputs[-1] = process_words.one_hot_encode_output_data(
padded_indexed_inputs[-1], max_word_len, VOCAB_SIZE+2
)
decoder_input = get_decoder_input(padded_indexed_inputs[0])
phonetic_features = list()
if PHONETIC_FLAG is True:
phonetic_features = self.phonetic_features_extractor()
all_inputs, all_outputs, num_of_optimized_features = segregate_inputs_and_outputs(padded_indexed_inputs,
categorized_features,
decoder_inputs=decoder_input,
phonetic_features=phonetic_features)
return [all_inputs, all_outputs, max_word_len, n, num_of_optimized_features]
def main():
paths = read_path_configs('data_paths.yaml')
if MODE == 'train':
train_data_dir = paths[LANG][MODE]
train_words, train_roots, train_features = \
extract_word_root_and_feature.get_words_roots_and_features(train_data_dir, n_features=FEATURE_NUMS,
lang=LANG, get_stats=False)
val_data_dir = paths[LANG]['validation']
val_words, val_roots, val_features = \
extract_word_root_and_feature.get_words_roots_and_features(val_data_dir, n_features=FEATURE_NUMS,
lang=LANG, get_stats=False)
assert len(train_words) == len(train_roots) == len(train_features[1]), \
"Length mismatch while flattening train features"
assert len(val_words) == len(val_roots) == len(val_features[1]),\
"Length mismatch while flattening val features"
# print(len(train_words), len(train_roots), len(train_features[0]))
# print("words: {}, roots: {}, features: {}".format(train_words[:5], train_roots[:5], train_features[:5]))
train_size, val_size = [len(each) for each in [train_words, val_words]]
train_val_words, train_val_roots = [train_words + val_words, train_roots + val_roots]
train_val_features = [i+j for i,j in zip(train_features, val_features)]
train_data_generator = ProcessDataForModel(words=train_val_words, roots=train_val_roots,
features=train_val_features)
all_inputs, all_outputs, max_word_len, n, phonetic_feature_num = train_data_generator.process_end_to_end()
params = read_path_configs('model_params.yaml')
model = _create_model(max_word_len, params['EMBED_DIM'], n, phonetic_feature_num)
train_inputs, val_inputs = split_train_val(all_inputs, train_size)
train_outputs, val_outputs = split_train_val(all_outputs, train_size)
hist = model.fit(train_inputs, train_outputs, validation_data=(val_inputs, val_outputs),
batch_size = params['BATCH_SIZE'], epochs=params['EPOCHS'],
callbacks=[EarlyStopping(patience=10),
ModelCheckpoint(filepath= get_model_path(paths=paths),
save_best_only=True,
verbose=1, save_weights_only=True)
])
if FREEZER_FLAG is True:
frozen_model = _create_model(max_word_len, params['EMBED_DIM'], n, phonetic_feature_num, freezing_call=True)
model.load_weights(get_model_path(paths=paths))
layers_to_be_frozen = get_frozen_layer_names()
for layer in model.layers:
frozen_model.get_layer(layer.name).set_weights(model.get_layer(layer.name).get_weights())
for layer_to_be_frozen in layers_to_be_frozen:
try:
frozen_model.get_layer(layer_to_be_frozen).trainable = False
except KeyError:
pass
frozen_model.compile(optimizer='adadelta', loss='categorical_crossentropy', metrics=['accuracy'])
hist = frozen_model.fit(train_inputs, train_outputs, validation_data=(val_inputs, val_outputs),
batch_size=params['BATCH_SIZE'], epochs=params['EPOCHS'],
callbacks=[EarlyStopping(patience=10),
ModelCheckpoint(filepath=get_model_path(paths=paths),
save_best_only=True,
verbose=1, save_weights_only=True)
])
elif MODE == 'test':
test_data_dir = paths[LANG][MODE]
contents = extract_word_root_and_feature.get_words_roots_and_features(test_data_dir, n_features=FEATURE_NUMS,
lang=LANG, get_stats=False)
index_identifier = RemoveErroneousIndices(contents)
test_words, test_roots, test_features = index_identifier.remove_unknown_feature_labels()
test_data_generator = ProcessDataForModel(words=test_words, roots=test_roots,
features=test_features)
all_inputs, all_outputs, max_word_len, n, phonetic_feature_num = test_data_generator.process_end_to_end()
params = read_path_configs('model_params.yaml')
model = _create_model(max_word_len, params['EMBED_DIM'], n, phonetic_feature_num)
model.load_weights(get_model_path(paths=paths))
pred_outputs = model.predict(all_inputs)
predicted_char_indices = np.argmax(pred_outputs[0], axis=2)
predicted_features = [np.argmax(each, axis=1) for each in pred_outputs[1:]]
_ = write_features_to_file(test_words, all_outputs[1:], predicted_features, paths['output_'+LANG])
root_outputs = write_roots_to_file(test_words, test_roots, predicted_char_indices, paths['output_'+LANG])
evaluator = evaluate_and_plot.EvaluatePerformance(test_words, root_outputs, all_outputs[1:], pred_outputs[1:],
pickle_handler.pickle_loader('class_labels_transformed'+'_'+
LANG))
_ = evaluator.p_r_curve_plotter(lang=LANG)
elif MODE == 'predict':
test_data_dir = paths[LANG+'_'+MODE+'_input']
sentences = extract_word_root_and_feature.get_words_for_predictions(test_data_dir)
predictions = list()
for sentence in sentences:
words_reversed = [item[::-1] for item in sentence]
X_indexed = process_words.get_indexed_words(words_reversed, mode='use_vocab', vocab_size=VOCAB_SIZE,
lang=LANG)
input_shifter = process_words.ShiftWordsPerCW(X=words_reversed, cw=CONTEXT_WINDOW, vocab_size=VOCAB_SIZE,
lang=LANG)
X_indexed_left, X_indexed_right = input_shifter.shift_input()
all_inputs = list()
all_inputs.append(X_indexed)
all_inputs += X_indexed_left
all_inputs += X_indexed_right
padded_indexed_inputs, max_word_len = pad_all_sequences(all_inputs)
n = pickle_handler.pickle_loader('num_of_indiv_features'+'_'+LANG)
decoder_input = get_decoder_input(padded_indexed_inputs[0])
padded_indexed_inputs.append(decoder_input)
num_of_optimized_features = list()
if PHONETIC_FLAG is True:
extractor = extract_phonetic_features.PhoneticFeatures(sentence)
features = extractor.get_features()
features = [word_feature[:FEATURE_NUMS] for word_feature in features]
tag_grouped_phonetic_features = [list(zip(*features))[idx] for idx in
range(len(features[0]))]
_ = [padded_indexed_inputs.append(np.array(each)) for each in tag_grouped_phonetic_features]
num_of_optimized_features = [len(each) for each in features[0]]
params = read_path_configs('model_params.yaml')
model = _create_model(max_word_len, params['EMBED_DIM'], n, num_of_optimized_features)
model.load_weights(get_model_path(paths=paths))
pred_outputs = model.predict(padded_indexed_inputs)
predicted_char_indices = np.argmax(pred_outputs[0], axis=2)
predicted_features = [np.argmax(each, axis=1) for each in pred_outputs[1:]]
predictions = list()
predictions.append(predicted_char_indices)
predictions += predicted_features
_ = write_predicted_roots_and_features(sentences, predictions, paths['output_'+LANG])
if __name__ == "__main__":
main()