-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathfine-tune_mT5.py
171 lines (125 loc) · 5.03 KB
/
fine-tune_mT5.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
# from transformers.keras_callbacks import KerasMetricCallback
from transformers import TFAutoModelForSeq2SeqLM
from transformers import AdamWeightDecay
# from transformers import DataCollatorForSeq2Seq
from transformers import AutoTokenizer
# import datasets
import tensorflow as tf
# import keras_nlp
# from nltk.translate.bleu_score import sentence_bleu
# import evaluate
# import numpy as np
import time
# ---
split_char = '⫯'
data_name = 'CWIKI_2023_09_27'
# model_name = f'mT5_small_PTT_2023_08_06_Thu_Aug_17_081645_2023.ckpt'
model_name = f'mT5_small_{data_name}_{time.ctime().replace(" ", "_").replace(":", "")}.ckpt'
checkpoint = "google/mt5-small"
# ---
tokenizer = AutoTokenizer.from_pretrained('models/mT5_pretrained_tokenizer')
# ---
prefix = "translate engTyping to Traditional Chinese:".split()
# ---
# ---
strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()
with strategy.scope():
model = TFAutoModelForSeq2SeqLM.from_pretrained('models/mT5_pretrained_model')
optimizer = AdamWeightDecay(learning_rate=2e-5, weight_decay_rate=0.01)
model.compile(optimizer=optimizer, metrics=["accuracy"]) # No loss argument!
# ---
# model.load_weights('models/mT5_small_PTT_2023_08_06_Thu_Aug_17_081645_2023.ckpt')
# ---
# ---
# ---
# ---
# tf.debugging.disable_traceback_filtering()
# ---
# class BLEU(tf.keras.metrics.Metric):
# def __init__(self, tokenizer, name="bleu", **kwargs):
# super().__init__(name=name, **kwargs)
# self.tokenizer = tokenizer
# self.bleu = sentence_bleu
# self.bleu_score = self.add_weight(
# shape=(),
# initializer='zeros',
# dtype=self.dtype,
# name = 'bleu_score'
# )
# def update_state(self, y_true, y_pred, sample_weight=None):
# # print(dir(y_true))
# # print(dir(y_pred))
# # print(y_true.value_index)
# # print(y_pred.value_index)
# # print(y_true[2])
# # print(y_pred[0])
# # print(dir(y_true[2]))
# # print(dir(y_pred[0]))
# # print(y_true[2].value_index)
# # print(y_pred[0].value_index)
# # print(y_true[2][0])
# # print(y_pred[0][0])
# # print(y_true[2][0][0])
# # print(y_pred[0][0][0])
# # decoded_pred = self.tokenizer.batch_decode(y_pred, skip_special_tokens=True)
# # y_true = np.where(y_true != -100, y_true, tokenizer.pad_token_id)
# # decoded_true = self.tokenizer.batch_decode(y_true, skip_special_tokens=True)
# # decoded_pred, decoded_true = postprocess_text(decoded_pred, decoded_true)
# result = self.bleu(y_true, y_pred)
# self.bleu_score.assign(result)
# def result(self):
# return self.bleu_score
# def reset_state(self):
# self.bleu_score.assign(0.0)
# def get_config(self):
# config = super().get_config()
# config.update(
# {
# "tokenizer": self.tokenizer,
# }
# )
# return config
# ---
# ---
tf_train_set = tf.data.Dataset.load(f'datasets/mT5_{data_name}_tf_train_dataset.tfrecord')
tf_val_set = tf.data.Dataset.load(f'datasets/mT5_{data_name}_tf_val_dataset.tfrecord')
tf_test_set = tf.data.Dataset.load(f'datasets/mT5_{data_name}_tf_test_dataset.tfrecord')
options = tf.data.Options()
options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.DATA
tf_train_set = tf_train_set.with_options(options)
tf_val_set = tf_val_set.with_options(options)
tf_test_set = tf_test_set.with_options(options)
# ---
# tf.config.run_functions_eagerly(True)
# ---
# metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_val_set)
callbacks = [
tf.keras.callbacks.ModelCheckpoint(f'models/{model_name}',save_best_only=True, save_weights_only=True),
tf.keras.callbacks.TensorBoard(log_dir=f'logs/{model_name}_logs')
]
model.fit(x=tf_train_set, validation_data=tf_val_set, epochs=5, callbacks=callbacks)
model.evaluate(tf_test_set)
# from google.colab import drive
# import os
# drive.mount('/content/drive', force_remount=True)
# os.chdir('/content/drive/My Drive/translate')
# model.save_pretrained(f'models/{model_name}')
# ---
# tf.keras.utils.plot_model(model, 'model.png')
# ---
# model.save(f'models/{model_name}', save_format='tf')
# ---
# split_char = '⫯'
# punctuations = '、，。？！：；'
# def engTyping_insert_split_char(sentence: str, split_char: str) -> str:
# insert_times = 0
# sentence_list = list(sentence)
# for i, char in enumerate(sentence):
# if char in ' 6347' + punctuations:
# sentence_list.insert(i + insert_times + 1, split_char)
# insert_times += 1
# return ''.join(sentence_list[:-1])
# # ---
# inputs = tokenizer(prefix + engTyping_insert_split_char(input('?:'), split_char).split(split_char), is_split_into_words=True, return_tensors="tf", truncation=True).input_ids
# outputs = model.generate(inputs, max_new_tokens=40, do_sample=True, top_k=30, top_p=0.95)
# print(tokenizer.decode(outputs[0], skip_special_tokens=True))