-
Notifications
You must be signed in to change notification settings - Fork 0
/
distilgpt2-train-RespGen.py
324 lines (266 loc) · 12.4 KB
/
distilgpt2-train-RespGen.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
# Importing stock libraries
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader, RandomSampler, SequentialSampler
from sklearn.model_selection import train_test_split
from itertools import chain
import random
import json
from datetime import datetime
# Importing the T5 modules from huggingface/transformers
from transformers import GPT2Tokenizer, GPT2LMHeadModel, get_polynomial_decay_schedule_with_warmup
from torch.utils.tensorboard import SummaryWriter
# from transformers import T5Tokenizer, T5ForConditionalGeneration
# from transformers import T5Tokenizer, AutoConfig, AutoModelForSeq2SeqLM
import csv
from tqdm import tqdm
import pickle
import gzip
import os
import sys
# WandB – Import the wandb library
import wandb
import math
# # Setting up the device for GPU usage
from torch import cuda
device = 'cuda' if cuda.is_available() else 'cpu'
# Creating a custom dataset for reading the dataframe and loading it into the dataloader to pass it to the neural network at a later stage for finetuning the model and to prepare it for predictions
import warnings
warnings.filterwarnings("ignore")
space = 'Ġ'
pre_quote = '’'
end_marks = ['.', ',', '?', '!', '...']
quotes = ['"', '\'']
abbreviations = ['s', 'd', 't', 'm', 're', 'll', 've', 'S', 'D', 'T', 'M', 'Re', 'Ll', 'Ve']
class CustomDataset(Dataset):
def __init__(self, dialogues, tokenizer, config):
self.input_ids = [] # (N, L)
self.token_type_ids = [] # (N, L)
self.labels = [] # (N, L)
self.tokenizer = tokenizer
print(f"Processing data...")
for dial in tqdm(dialogues):
hists = []
# ignore the task title and intro that give the context
dial = dial.split('[SEP]')
for u, utter in enumerate(dial):
tokens = tokenizer.tokenize(utter.strip().replace(pre_quote, quotes[1]))
# token_list = process_token_list(token_list)
# text = tokenizer.convert_tokens_to_string(token_list)
# tokens = tokenizer.tokenize(utter)
token_ids = tokenizer.convert_tokens_to_ids(tokens)
if u % 2 == 0:
hists.append([0] + token_ids)
else:
hists.append([1] + token_ids)
for h in range(0, len(hists)):
if h % 2 == 1:
for s in range(h):
contexts = hists[s:h+1] # qa
input_ids = [config.bos_id] + list(chain.from_iterable(contexts)) + [config.eos_id]
if len(input_ids) <= config.MAX_LEN:
start_sp_id, next_sp_id = contexts[0][0], contexts[1][0]
token_type_ids = [[start_sp_id] * len(ctx) if c % 2 == 0 else [next_sp_id] * len(ctx) for c, ctx in enumerate(contexts)]
assert token_type_ids[-1][0] == 1
token_type_ids = [start_sp_id] + list(chain.from_iterable(token_type_ids)) + [1]
assert len(input_ids) == len(token_type_ids)
labels = [[-100] * len(ctx) if c < len(contexts)-1 else [-100] + ctx[1:] for c, ctx in enumerate(contexts)]
assert labels[-1][1:] == contexts[-1][1:]
labels = [-100] + list(chain.from_iterable(labels)) + [config.eos_id]
assert len(input_ids) == len(labels)
self.input_ids.append(input_ids)
self.token_type_ids.append(token_type_ids)
self.labels.append(labels)
break
def __len__(self):
return len(self.input_ids)
def __getitem__(self, idx):
return self.input_ids[idx], self.token_type_ids[idx], self.labels[idx]
class PadCollate():
def __init__(self, eos_id):
self.eos_id = eos_id
def pad_collate(self, batch):
input_ids, token_type_ids, labels =[], [], []
for idx, seqs in enumerate(batch):
input_ids.append(torch.LongTensor(seqs[0]))
token_type_ids.append(torch.LongTensor(seqs[1]))
labels.append(torch.LongTensor(seqs[2]))
input_ids = torch.nn.utils.rnn.pad_sequence(
input_ids, batch_first=True, padding_value=self.eos_id
)
token_type_ids = torch.nn.utils.rnn.pad_sequence(
token_type_ids, batch_first=True, padding_value=self.eos_id
)
labels = torch.nn.utils.rnn.pad_sequence(
labels, batch_first=True, padding_value=-100
)
return input_ids, token_type_ids, labels
def validation(model, valid_loader):
print("Validation processing...")
model.eval()
valid_losses = []
valid_ppls = []
with torch.no_grad():
for i, batch in enumerate(tqdm(valid_loader)):
input_ids, token_type_ids, labels = batch
input_ids, token_type_ids, labels = \
input_ids.to(device), token_type_ids.to(device), labels.to(device)
outputs = model(
input_ids=input_ids,
token_type_ids = token_type_ids,
labels = labels
)
loss, logits = outputs[0], outputs[1]
valid_losses.append(loss.detach())
ppl = torch.exp(loss.detach())
valid_ppls.append(ppl)
valid_losses = [loss.item() for loss in valid_losses]
valid_ppls = [ppl.item() if not math.isinf(ppl.item()) else 1e+8 for ppl in valid_ppls]
valid_loss = np.mean(valid_losses)
valid_ppl = np.mean(valid_ppls)
if math.isnan(valid_ppl):
valid_ppl = 1e+8
return valid_loss, valid_ppl
def main():
# WandB – Initialize a new run
wandb.init(project="TOC-distilgpt")
# WandB – Config is a variable that holds and saves hyperparameters and inputs
# Defining some key variables that will be used later on in the training
config = wandb.config # Initialize config
config.TRAIN_BATCH_SIZE = 8 # input batch size for training (default: 64)
config.VALID_BATCH_SIZE = 8 # input batch size for testing (default: 1000)
config.TRAIN_EPOCHS = 10 # number of epochs to train (default: 10)
config.VAL_EPOCHS = 1
config.LEARNING_RATE = 2e-5 # learning rate (default: 2e-5)
config.SEED = 42 # random seed (default: 42)
config.MAX_LEN = 1024
config.OUTPUT_LEN = 200
config.warmup_ratio = 0.1
config.ckpt_dir = './distilgpt2-ckpt/'
# Set random seeds and deterministic pytorch for reproducibility
torch.manual_seed(config.SEED) # pytorch random seed
np.random.seed(config.SEED) # numpy random seed
torch.backends.cudnn.deterministic = True
random.seed(config.SEED)
# Tokenizer & Vocab
print("Loading the tokenizer...")
tokenizer = GPT2Tokenizer.from_pretrained('distilgpt2')
config.eos_token = tokenizer.eos_token
config.bos_token = tokenizer.bos_token
vocab = tokenizer.get_vocab()
config.vocab_size = len(vocab)
config.bos_id = vocab[config.bos_token]
config.eos_id = vocab[config.eos_token]
model = GPT2LMHeadModel.from_pretrained('distilgpt2')
model = model.to(device)
optim = torch.optim.AdamW(model.parameters(), lr=config.LEARNING_RATE)
model.resize_token_embeddings(config.vocab_size)
# load from checkpoint
load_ckpt = False
if load_ckpt:
checkpoint = torch.load(config.ckpt_dir + 'best_ckpt_epoch=5_valid_loss=2.4685.ckpt')
model.load_state_dict(checkpoint['model_state_dict'])
optim.load_state_dict(checkpoint['optim_state_dict'])
print("loading data... ")
data_categories = ['art', 'car', 'computer', 'education', 'family', 'finance', 'food', 'health', 'hobby', 'holiday', 'home', 'pet', 'philosophy', 'relationship', 'sport', 'style', 'travel', 'work', 'youth']
dialogues = []
for category in data_categories:
data = json.load(open(f'INST2DIAL-Auto/{category}_multiwoz_convqa_flan_t5_large_2qa_3sent.json'))
keys = list(data.keys())
for key in keys:
dialogues.append(data[key])
train, valid = train_test_split(dialogues, test_size=0.01, random_state=config.SEED)
train_dataset = CustomDataset(train, tokenizer, config)
valid_dataset = CustomDataset(valid, tokenizer, config)
print("train: ", len(train_dataset))
print("valid: ", len(valid_dataset))
ppd = PadCollate(eos_id=config.eos_id)
train_loader = DataLoader(train_dataset,
collate_fn=ppd.pad_collate,
shuffle=True,
batch_size=config.TRAIN_BATCH_SIZE,
pin_memory=True)
valid_loader = DataLoader(valid_dataset,
collate_fn=ppd.pad_collate,
batch_size=config.VALID_BATCH_SIZE,
pin_memory=True)
if not os.path.exists(config.ckpt_dir):
os.makedirs(config.ckpt_dir)
# Calculate total training steps
num_batches = len(train_loader)
config.total_train_steps = config.TRAIN_EPOCHS * num_batches
config.warmup_steps = int(config.warmup_ratio * config.total_train_steps)
sched = get_polynomial_decay_schedule_with_warmup(
optim,
num_warmup_steps=config.warmup_steps,
num_training_steps=config.total_train_steps,
power=2
)
if load_ckpt:
sched.load_state_dict(checkpoint['sched_state_dict'])
writer = SummaryWriter()
print("Setting finished.")
print("Training starts.")
best_loss = sys.float_info.max
last_epoch = 0
start_epoch = 1
for epoch in range(start_epoch, start_epoch + config.TRAIN_EPOCHS):
model.train()
print(f"#"*50 + f"Epoch: {epoch}" + "#"*50)
train_losses = []
train_ppls = []
for i, batch in enumerate(tqdm(train_loader)):
input_ids, token_type_ids, labels = batch
input_ids, token_type_ids, labels = \
input_ids.to(device), token_type_ids.to(device), labels.to(device)
outputs = model(
input_ids=input_ids,
token_type_ids=token_type_ids,
labels=labels
)
loss, logits = outputs[0], outputs[1]
optim.zero_grad()
loss.backward()
optim.step()
sched.step()
train_losses.append(loss.detach())
ppl = torch.exp(loss.detach())
train_ppls.append(ppl)
train_losses = [loss.item() for loss in train_losses]
train_ppls = [ppl.item() if not math.isinf(ppl.item()) else 1e+8 for ppl in train_ppls]
train_loss = np.mean(train_losses)
train_ppl = np.mean(train_ppls)
print(f"Train loss: {train_loss} || Train perplexity: {train_ppl}")
writer.add_scalar("Loss/train", train_loss, epoch)
writer.add_scalar("PPL/train", train_ppl, epoch)
last_epoch += 1
valid_loss, valid_ppl = validation(model, valid_loader)
if valid_loss < best_loss:
best_loss = valid_loss
state_dict = {
'model_state_dict': model.state_dict(),
'optim_state_dict': optim.state_dict(),
'sched_state_dict': sched.state_dict(),
'loss': best_loss,
'epoch': last_epoch
}
torch.save(state_dict, f"{config.ckpt_dir}/best_ckpt_epoch={epoch}_valid_loss={round(best_loss, 4)}.ckpt")
print("*"*10 + "Current best checkpoint is saved." + "*"*10)
print(f"{config.ckpt_dir}/best_ckpt_epoch={epoch}_valid_loss={round(best_loss, 4)}.ckpt")
print(f"Best valid loss: {best_loss}")
print(f"Valid loss: {valid_loss} || Valid perplexity: {valid_ppl}")
writer.add_scalar("Loss/valid", valid_loss, epoch)
writer.add_scalar("PPL/valid", valid_ppl, epoch)
writer.add_scalars("Losses", {
'train': train_loss,
'valid': valid_loss,
}, epoch)
writer.add_scalars("PPLs", {
'train': train_ppl,
'valid': valid_ppl,
}, epoch)
if __name__ == '__main__':
main()