-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
204 lines (154 loc) · 8.02 KB
/
train.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
import os, json
import numpy as np
import transformers
from transformers import AutoTokenizer, AutoTokenizer
from transformers import Seq2SeqTrainingArguments, DataCollatorForSeq2Seq,Seq2SeqTrainer
from nltk.tokenize import sent_tokenize
import torch
import evaluate
rouge_score = evaluate.load("rouge")
os.environ['TRANSFORMERS_NO_ADVISORY_WARNINGS'] = 'true'
from data.data import LECR_prepare_data
from model.model import give_model
params = {
"load_tokenized_datasets":False,
"save_dataset": True,
"DEBUG" : False,
"model_checkpoint":"google/flan-t5-large",
"SEMANTIC_EMBEDING": "tools/semantic_embeding_v3_FlanEncoder_large_128.pkl", # see tools/readme.txt
"content_multiplier" : 10,
"topic_multiplier" : 2,
"topic_tree_multiplier":20,
"topic_train_explode":False,
"reduce_topics": False,
"max_input_length": 128,
"max_target_length":128,
"generation_max_length":128,
"output_dir": f"output/large_flan_v10_1_v2",
"dropout_rate":0,
"freeze_encoder":False,
"num_decoder_layers":0,
"freze_embed":False,
"freze_layers": 0, #["encoder.block.2","encoder.block.3","encoder.block.4","encoder.block.5","encoder.block.6","decoder.block.1","decoder.block.2","decoder.block.3","decoder.block.4"], #0, #["shared"]+["encoder.block."+x for x in map(str,list(range(8)))], #["shared","encoder.block.0","encoder.block.1","block.3","block.4","block.5"]
"eval_steps":0.05,
"per_device_train_batch_size" : 16,
"per_device_eval_batch_size" : 64,
"num_train_epochs" : 1,
"gradient_accumulation_steps" : 8,
"gradient_checkpointing":False,
"learning_rate":2e-4,
"sch" :"cosine",
"cosine_warmup": 0.1,
}
class paramss:
def __init__(self, **kw):
self.__dict__.update(kw)
self.params_dict = params
params = paramss(**params)
def compute_metrics(eval_pred):
predictions, labels = eval_pred
# Decode generated summaries into text
decoded_preds = tokenizer.batch_decode(predictions, skip_special_tokens=True)
# Replace -100 in the labels as we can't decode them
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
# Decode reference summaries into text
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
# ROUGE expects a newline after each sentence
decoded_preds = ["\n".join(sent_tokenize(pred.strip())) for pred in decoded_preds]
decoded_labels = ["\n".join(sent_tokenize(label.strip())) for label in decoded_labels]
# Compute ROUGE scores
result = rouge_score.compute(
predictions=decoded_preds, references=decoded_labels, use_stemmer=True
)
# Extract the median scores
result = {key: value * 100 for key, value in result.items()}
return {k: round(v, 4) for k, v in result.items()}
def train(params):
torch.backends.cuda.matmul.allow_tf32 = True
global tokenizer
tokenizer = AutoTokenizer.from_pretrained(params.model_checkpoint)
model = give_model(model_checkpoint=params.model_checkpoint,
num_decoder_layers=params.num_decoder_layers,
dropout_rate=params.dropout_rate,
freze_embed=params.freze_embed,
freeze_encoder=params.freeze_encoder,
freze_layers=params.freze_layers)
dataset = LECR_prepare_data(tokenizer = tokenizer,
DEBUG = params.DEBUG ,
content_multiplier = params.content_multiplier,
max_input_length = params.max_input_length,
max_target_length = params.max_target_length,
topic_multiplier = params.topic_multiplier,
topic_tree_multiplier = params.topic_tree_multiplier,
topic_train_explode = params.topic_train_explode,
load_tokenized_datasets=params.load_tokenized_datasets,
SEMANTIC_EMBEDING = params.SEMANTIC_EMBEDING,
save_dataset = params.save_dataset,
reduce_topics = params.reduce_topics
)
tokenized_datasets = dataset.tokenized_Dataset
data_collator = DataCollatorForSeq2Seq(tokenizer, model=model)
num_train_samples = len(tokenized_datasets["train"])
number_of_batches_per_epoch = num_train_samples//params.per_device_train_batch_size
number_of_steps_per_epoch = number_of_batches_per_epoch//params.gradient_accumulation_steps
number_of_total_steps = number_of_steps_per_epoch*params.num_train_epochs
optimizerr = transformers.Adafactor(filter(lambda p: p.requires_grad, model.parameters()),
lr=params.learning_rate,
relative_step=False,
# weight_decay=0.01,
# beta1=0.9,
# clip_threshold=0.5
)
if params.sch == "one_cycle":
lr_sch = torch.optim.lr_scheduler.OneCycleLR(optimizerr, params.learning_rate, total_steps=None, epochs=params.num_train_epochs, steps_per_epoch=number_of_steps_per_epoch, pct_start=0.3,
anneal_strategy='cos', cycle_momentum=False, base_momentum=0.85, max_momentum=0.95, div_factor=200.0,
final_div_factor=10000.0, three_phase=False, last_epoch=- 1, verbose=False)
elif params.sch == "cosine":
lr_sch = transformers.get_cosine_schedule_with_warmup(optimizer=optimizerr,
num_warmup_steps=int(params.cosine_warmup*number_of_total_steps),
num_training_steps=number_of_total_steps)
optimizers = (optimizerr,lr_sch)
model_name = params.model_checkpoint.split("/")[-1]
args = Seq2SeqTrainingArguments(
output_dir=params.output_dir,
evaluation_strategy="steps",
eval_steps= int(params.eval_steps* number_of_steps_per_epoch) ,
logging_steps= int(params.eval_steps* number_of_steps_per_epoch) ,
save_steps= int(params.eval_steps* number_of_steps_per_epoch) ,
save_total_limit=1,
per_device_train_batch_size=params.per_device_train_batch_size,
per_device_eval_batch_size=params.per_device_eval_batch_size,
# weight_decay=0.01,
gradient_accumulation_steps=params.gradient_accumulation_steps,
num_train_epochs=params.num_train_epochs,
predict_with_generate=True,
gradient_checkpointing=params.gradient_checkpointing,
# fp16=True,
bf16=True,
tf32=True,
generation_max_length=params.generation_max_length,
report_to="tensorboard",
dataloader_drop_last = True,
dataloader_num_workers=32
)
os.environ["TOKENIZERS_PARALLELISM"] = "false"
trainer = Seq2SeqTrainer(
model,
args,
train_dataset=tokenized_datasets["train"],
eval_dataset={"topic_valid":tokenized_datasets["topic_valid"],
"content_train_sample":tokenized_datasets["content_train_sample"],
"topic_train_sample":tokenized_datasets["topic_train_sample"],
"topic_tree_train_sample":tokenized_datasets["topic_tree_train_sample"]}, #[tokenized_datasets["validation"],tokenized_datasets["mem_validation"]]
data_collator=data_collator,
tokenizer=tokenizer,
compute_metrics=compute_metrics,
optimizers = optimizers
)
params_dict = params.params_dict
with open(os.path.join(params_dict["output_dir"],"params.json"), "w") as write_file:
json.dump(params_dict, write_file, indent=4)
trainer.train()
return model, dataset
if __name__ == "__main__":
model, dataset = train(params)