-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathamazon_kdd_bert_mlp_adamw_optimizer_8_epochs_200k.py
278 lines (190 loc) · 8.51 KB
/
amazon_kdd_bert_mlp_adamw_optimizer_8_epochs_200k.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
# -*- coding: utf-8 -*-
"""Amazon_KDD_bert_mlp_adamw_optimizer_8_epochs_200k.ipynb
Automatically generated by Colaboratory.
Original file is located at
https://colab.research.google.com/drive/1w3hUOyD317uEJpzdNsYevgpkygajx1xX
## Install dependencies
"""
"""# Import Dependencies
## Import dependencies
"""
import os
import math
import numpy as np
import pandas as pd
from tqdm.auto import tqdm
import tensorflow as tf
from transformers import BertTokenizer
from transformers import TFBertForSequenceClassification
from official.nlp import optimization
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.666)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
"""# Cleaned Dataset Import"""
train_df = pd.read_csv('import/train/train_final_200k.csv')
train_df.fillna('', inplace=True)
train_df.info()
"""# BERT Input Like Sentences generation"""
train_df['text_a'] = train_df['query']
train_df['text_b'] = train_df['product_title'] + '. ' + train_df['product_brand'] + '. ' + train_df['product_color_name']
train_df = train_df.drop(train_df.columns[[1,2,4,5,6]], axis=1)
train_df.head()
"""# Factorizing Classes
###The category labels are:
* *0* - exact
* 1 - substitute
* 2 - complement
* 3 - irrelevant
"""
train_df['esci_label_code'] = train_df['esci_label'].map( {'exact':0, 'substitute':1, 'complement':2, 'irrelevant':3})
train_df.head()
seq_len = [len(i.split()) for i in train_df['text_a']]
pd.Series(seq_len).hist(bins = 30, log=True)
count = 0
for x in seq_len:
count = count + x
print("Avg: "+ str(count/len(seq_len)))
print("Max: "+ str(max(seq_len)))
seq_len = [len(i.split()) for i in train_df['text_b']]
pd.Series(seq_len).hist(bins = 30, log=True)
count = 0
for x in seq_len:
count = count + x
print("Avg: "+ str(count/len(seq_len)))
print("Max: "+ str(max(seq_len)))
"""# Check factorized classes correspondences"""
train_df['esci_label'].value_counts(normalize=True).mul(100)
train_df['esci_label_code'].value_counts(normalize=True).mul(100)
"""# Class weights
Since the train dataset is not homogeneous and the goal of the model is to identify complement products matches, but we don't have very many of those samples to work with, so I would want to have the classifier heavily weight the few examples that are available. I am going to do this by passing Keras weights for each class through a parameter. These will cause the model to "pay more attention" to examples from an under-represented class.
Each class weight is calculated as the frequency of that class multiplied by the total number of element divided by the class number.
"""
NUM_ELEMENTS = len(train_df)
weight_for_0 = (1 / train_df['esci_label'].value_counts()['exact'])*(NUM_ELEMENTS/4.0)
weight_for_1 = (1 / train_df['esci_label'].value_counts()['substitute'])*(NUM_ELEMENTS/4.0)
weight_for_2 = (1 / train_df['esci_label'].value_counts()['complement'])*(NUM_ELEMENTS/4.0)/2
weight_for_3 = (1 / train_df['esci_label'].value_counts()['irrelevant'])*(NUM_ELEMENTS/4.0)
class_weight = {0: weight_for_0, 1: weight_for_1, 2: weight_for_2, 3: weight_for_3}
class_weight
"""### Using 10000 rows the model accuracy on test set was 0.626:
Change train size to 10000 rows
"""
#train_df = train_df[:10000]
"""# Initialize bert base multilingual uncased model
## Load Bert Base Multilingual Uncased
"""
tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-uncased')
"""## Initialize Input Token Identification variable
This argument indicates to the model token indices, numerical representations of tokens building the sequences.
"""
X_input_ids = np.zeros((len(train_df), 100))
"""## Initialize Attention Mask variable
This argument indicates to the model which tokens should be attended to, and which should not.
"""
X_attn_masks = np.zeros((len(train_df), 100))
X_token_ids = np.zeros((len(train_df), 100))
train_df.info()
"""# Fine Tuning
## Function to generate the Bert understandable dataset
"""
def generate_training_data(df, ids, masks, token, tokenizer):
for i in tqdm(range(0, len(train_df))):
tokenized_text = tokenizer(
text=train_df.iloc[i,2],
text_pair=train_df.iloc[i,3],
max_length=100,
truncation= 'only_second',
padding='max_length',
add_special_tokens=True,
return_tensors='tf'
)
ids[i, :] = tokenized_text.input_ids
masks[i, :] = tokenized_text.attention_mask
token[i, :] = tokenized_text.token_type_ids
return ids, masks, token
X_input_ids, X_attn_masks,X_token_ids = generate_training_data(train_df, X_input_ids, X_attn_masks,X_token_ids, tokenizer)
X_input_ids[2012]
X_attn_masks[2012]
X_token_ids[2012]
"""## Initialize train labels variable"""
labels = np.zeros((len(train_df), 4))
labels.shape
labels
"""## One Hot encoding"""
labels[np.arange(len(train_df)), train_df['esci_label_code'].values] = 1 # one-hot encoded target tensor
"""## Build Tensorflow input"""
dataset = tf.data.Dataset.from_tensor_slices((X_input_ids, X_attn_masks, X_token_ids, labels))
def ESCIDatasetMapFunction(input_ids, attn_masks,token_ids, labels):
return {
'input_ids': input_ids,
'token_type_ids':token_ids,
'attention_mask': attn_masks
}, labels
"""We are using drop_remainder argument to ignore that last batch, and get full shape propagation."""
dataset = dataset.map(ESCIDatasetMapFunction) # converting to required format for tensorflow dataset
dataset = dataset.shuffle(len(train_df)).batch(16, drop_remainder=True)
p = 0.80
train_size = int((len(train_df)//16)*p)
train_size
"""## Dividing train_df into Train and Validation Dataset"""
train_dataset = dataset.take(train_size)
val_dataset = dataset.skip(train_size)
"""## Load pretrained Bert model"""
model = TFBertForSequenceClassification.from_pretrained('bert-base-multilingual-uncased') # bert base model with pretrained weights
model.layers[0].trainable = True
"""## Define model sequential layers
Output layer has softmax activation in order to set initial bias to a value that correctly represents the imbalanced training dataset, some classes has much more instances than others ('exact' = 65% of train dataset)
"""
input_ids = tf.keras.layers.Input(shape=(100,), name='input_ids', dtype='int32')
attn_masks = tf.keras.layers.Input(shape=(100,), name='attention_mask', dtype='int32')
token_ids = tf.keras.layers.Input(shape=(100,), name='token_type_ids', dtype='int32')
#da tf documentazione
bert_embds = model.bert(input_ids=input_ids, attention_mask=attn_masks, token_type_ids=token_ids)[1] # 0 -> activation layer (3D), 1 -> pooled output layer (2D)
#intermediate layer
hidden_layer_1 = tf.keras.layers.Dense(100,activation='relu', name='hidden_layer_1')(bert_embds)
output_layer = tf.keras.layers.Dense(4, activation='softmax', name='output_layer')(hidden_layer_1) # softmax -> calcs probs of classes
esci_model = tf.keras.Model(inputs=[input_ids, attn_masks, token_ids], outputs=output_layer)
esci_model.summary()
"""## Define Hyperparameters"""
# hyperparameters
#optim = AdamW(model.parameters(), lr=1e-5, eps=1e-8)
epochs = 4
steps_per_epoch = 3000
num_train_steps = steps_per_epoch * epochs
num_warmup_steps = int(0.1*num_train_steps)
init_lr = 3e-5
optimizer = optimization.create_optimizer(init_lr=init_lr,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
optimizer_type='adamw')
loss_func = tf.keras.losses.CategoricalCrossentropy()
acc = tf.keras.metrics.CategoricalAccuracy('accuracy')
"""## Compile the model"""
esci_model.compile(optimizer=optimizer, loss=loss_func, metrics=[acc])
"""## Train the model"""
hist = esci_model.fit(
train_dataset,
validation_data=val_dataset,
epochs=epochs,
class_weight=class_weight
)
"""## Save the trained model for inference"""
esci_model.save('models/esci_model_mlp_adamw_no_optimizer_8_epochs_200k', include_optimizer=False)
esci_model.save('models/esci_model_mlp_adamw_with_optimizer_8_epochs_200k', include_optimizer=True)
"""# Plot Model Training"""
import matplotlib.pyplot as plt
# summarize history for accuracy
plt.plot(hist.history['accuracy'])
plt.plot(hist.history['val_accuracy'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()
# summarize history for loss
plt.plot(hist.history['loss'])
plt.plot(hist.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()