forked from giobin/Applied_Intelligence_sparsity
-
Notifications
You must be signed in to change notification settings - Fork 0
/
extract_info_from_ckpt.py
271 lines (227 loc) · 11.7 KB
/
extract_info_from_ckpt.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
import argparse
import math
import os
import time
from datetime import datetime
import torch
from tensor2tensor.utils.bleu_hook import compute_bleu
from torch.nn.utils import prune
from torch.optim.lr_scheduler import CosineAnnealingLR
from tqdm import tqdm
from transformers import BartConfig, BartForConditionalGeneration
import wandb
from utils.feedback import Logger, get_stats
from utils.file_manager import get_dataset
from utils.pruning import select_pruning_params, gamma_decay, get_regularizer_for_extracting_info, sum_masks, crop
from utils.utils import build_vocab, generate_iterator_bucket, get_spec_tokens, \
init_seeds, count_parameters, bleu_tokenize, pruned, get_configuration, check_max_n_ckpt, get_hyper, \
get_param_optimizer
def extract(model, device, data_train, optimizer, args, logger,
parameters_to_prune, step_num_for_epoch, l2_vocab, workdir):
gamma_cntr = 0
model.train()
pad_id, _, _ = get_spec_tokens(l2_vocab)
for epoch in range(args.epochs):
for i, inputs in enumerate(data_train, 0):
gamma_cntr += 1
inputs = inputs.to(device)
optimizer.zero_grad()
loss = model(**inputs)["loss"]
loss.backward(retain_graph=True)
info_dict = get_regularizer_for_extracting_info(model, alpha=args.exp_alpha, step=i)
optimizer.zero_grad()
break
break
return info_dict
def evaluate(model, iterator, device):
epoch_loss = 0
model.eval()
with torch.no_grad():
for inputs in tqdm(iterator):
inputs = inputs.to(device)
loss = model(**inputs)['loss']
epoch_loss += loss.item()
model.train()
return (epoch_loss / len(iterator))
def calculate_bleu(model, iterator, tokenizer, device, args, verbose=None):
model.eval()
generated_bleu = []
trg_bleu = []
logger = None
PAD_ID, BOS_ID, EOS_ID = get_spec_tokens(tokenizer)
beam_size, length_penalty, length = get_hyper(args)
if verbose is not None: logger = Logger(verbose + "/translated.txt")
for inputs in tqdm(iterator):
input_token = inputs['input_ids'].to(device)
attention_mask = inputs['attention_mask'].to(device)
gen = model.generate(input_token, attention_mask=attention_mask, pad_token_id=PAD_ID, bos_token_id=BOS_ID,
eos_token_id=EOS_ID,
max_length=length, num_beams=beam_size, length_penalty=length_penalty)
translated = tokenizer.batch_decode(gen, skip_special_tokens=True)
target = tokenizer.batch_decode(inputs['labels'], skip_special_tokens=True)
for elem in translated:
generated_bleu.append(bleu_tokenize(elem))
for elem in target:
trg_bleu.append(bleu_tokenize(elem))
if logger is not None:
logger.logging(f'Generated:{translated}', print_=False)
logger.logging(f'Target:{target}', print_=False)
logger.logging("\n", print_=False)
model.train()
return compute_bleu(trg_bleu, generated_bleu)
def generate_model(tokenizer, args):
PAD_ID, BOS_ID, EOS_ID = get_spec_tokens(tokenizer)
vocab_size = len(tokenizer)
MPE, EL, DL, FFN_DIM, AH, DM, DR = get_configuration(args)
configuration = BartConfig(vocab_size=vocab_size,
activation_function="relu",
max_position_embeddings=MPE,
encoder_layers=EL,
encoder_ffn_dim=FFN_DIM,
encoder_attention_heads=AH,
decoder_layers=DL,
decoder_ffn_dim=FFN_DIM,
decoder_attention_heads=AH,
pad_token_id=PAD_ID,
bos_token_id=BOS_ID,
decoder_start_token_id=BOS_ID,
eos_token_id=EOS_ID,
forced_eos_token_id=EOS_ID,
forced_bos_token_id=BOS_ID,
d_model=DM,
scale_embedding=True,
add_bias_logits=False,
dropout=DR)
if args.configuration is not None:
print("Load Config...")
path_config = os.path.abspath(os.getcwd()) + "/" + args.configuration
configuration = configuration.from_json_file(path_config)
model = BartForConditionalGeneration(configuration)
for p in model.parameters():
if p.dim() > 1:
torch.nn.init.xavier_uniform_(p)
if args.ckpt is not None:
print("Load CKPT...")
path_model = os.path.abspath(os.getcwd()) + "/" + args.ckpt
checkpoint_pruned = pruned(path_model)
if checkpoint_pruned is not None:
print("Load Pruned CKPT...")
model_params = select_pruning_params(model)
for p in model_params:
prune.identity(p[0], p[1])
model.load_state_dict(checkpoint_pruned)
model.tie_weights()
model.eval()
else:
model = model.from_pretrained(path_model, config=configuration)
return model
def main(args):
init_seeds()
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
workdir = f'{args.work_dir}/{datetime.now().strftime("%d-%m-%Y_%H-%M-%S")}_extracting_info'
if not os.path.exists(workdir):
os.makedirs(workdir)
path_log = workdir + "/log.txt"
logger = Logger(path_log)
logger.logging('===================')
for ar in vars(args):
logger.logging(f'--{ar}, {getattr(args, ar)}')
logger.logging('===================')
# Train Dataset
path_train = os.path.abspath(os.getcwd()) + "/" + args.train_path
train_dataset = get_dataset(path_train)
# Validation Dataset
path_validation = os.path.abspath(os.getcwd()) + "/" + args.validation_path
validation_dataset = get_dataset(path_validation)
# Test Dataset
path_test = os.path.abspath(os.getcwd()) + "/" + args.test_path
test_dataset = get_dataset(path_test)
logger.logging(f'Train:{len(train_dataset)}')
logger.logging(f'Validation:{len(validation_dataset)}')
logger.logging(f'Test:{len(test_dataset)}')
# Generate Vocabs
path = path_train
if args.single_vocab: # One language
logger.logging('Generating single vocab ...')
l1_vocab = build_vocab(path + "/train.L1", workdir, "L", path_vocab=args.path_vocab1)
l2_vocab = l1_vocab
else: # Different Languages
logger.logging('Generating two vocabs ...')
l1_vocab = build_vocab(path + "/train.L1", workdir, "L1", path_vocab=args.path_vocab1)
l2_vocab = build_vocab(path + "/train.L2", workdir, "L2", path_vocab=args.path_vocab2)
logger.logging(f'L1 VOCAB Size:{len(l1_vocab)}')
logger.logging(f'L2 VOCAB Size:{len(l1_vocab)}')
step_num_for_epoch = len(train_dataset) / args.batch_size
# Generate Iterators
iterator_train = generate_iterator_bucket(train_dataset, l1_vocab, l2_vocab, args)
iterator_valid = generate_iterator_bucket(validation_dataset, l1_vocab, l2_vocab, args)
iterator_test = generate_iterator_bucket(test_dataset, l1_vocab, l2_vocab, args)
# Generate Model
model = generate_model(l1_vocab, args)
model.to(device)
#wandb.watch(model, log="all")
logger.logging(str(model))
# Get Pruning Params
parameters_to_prune = select_pruning_params(model)
#anyway make it as it was pruned
for p in parameters_to_prune:
prune.identity(p[0], p[1])
num_all_params = count_parameters(model, trainable=False)
num_trainable_params = count_parameters(model)
non_zero_params = sum_masks(parameters_to_prune)
logger.logging(f'there are {len(parameters_to_prune)} parameters_to_prune: {parameters_to_prune}')
logger.logging(f'num of model parameters: {num_all_params}')
logger.logging(f'num of trainable model parameters: {num_trainable_params}')
logger.logging(f'num of non zero parameters: {non_zero_params}')
# Optimizer and scheduler
beta1, beta2, eps = get_param_optimizer(args)
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate, betas=(beta1, beta2), eps=eps)
if args.opt is not None:
logger.logging('Loading OPT...')
opt = torch.load(args.opt)
optimizer.load_state_dict(opt)
optimizer.defaults["lr"] = args.learning_rate
for g in optimizer.param_groups:
g['lr'] = args.learning_rate
t_max = int((len(train_dataset) / args.batch_size) * args.epochs)
# Start Training
info_dict = extract(model, device, iterator_train, optimizer,
args, logger,
parameters_to_prune, step_num_for_epoch, l2_vocab, workdir)
torch.save(info_dict, os.path.join(workdir, "info_dict.pt"))
print(f"saved info_dict")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('-tr', '--train_path', default='data/wmt/train', type=str)
parser.add_argument('-va', '--validation_path', default='data/wmt/dev', type=str)
parser.add_argument('-te', '--test_path', default='data/wmt/test', type=str)
parser.add_argument('--type_dataset', default='wmt', choices=['tsk1', 'wmt'])
parser.add_argument('-single_vocab', action='store_true')
parser.add_argument('--path_vocab1', default=None, help="path of vocab1, for ex: workdir/L1.vocab")
parser.add_argument('--path_vocab2', default=None, help="path of vocab2, for ex: workdir/L2.vocab")
parser.add_argument('--configuration', default=None,
help="path of model configuration, for ex: workdir/config.json")
parser.add_argument('--ckpt', default=None,
help="dir of model ckpt, for ex: workdir/ckpt.bin")
parser.add_argument('--opt', default=None,
help="optimizer state dict path")
parser.add_argument('--work_dir', type=str, default='workdir', help="dir where to store model ckpt")
parser.add_argument('-e', '--epochs', type=int, default=10, help="number of epochs to train")
parser.add_argument('-b', '--batch_size', type=int, default=100, help="batch dimension")
parser.add_argument('-ga', '--gradient_accumulation', type=int, default=3, help="number of gradient accumulation")
parser.add_argument('-lr', '--learning_rate', type=float, default=0.001, help="learning rate")
parser.add_argument('--log_interval', type=int, default=50, help="log interval in number of batches")
parser.add_argument('--eval_interval', type=int, default=400, help="eval interval in number of batches")
parser.add_argument('-t', '--threshold', type=float, default=0.10, help="threshold to use to crop weights")
parser.add_argument('--bleu_lower_bound', type=float, default=0.2720, help="lower acceptable bleu for cropping")
parser.add_argument('-r', '--regularize', action='store_true', help="regularize")
parser.add_argument('-rt', '--rtype', type=str, help="the regularization type", default="relevance", choices=['relevance', 'l1', 'l2'])
parser.add_argument('--save_eval', action='store_true', help="save model after evaluation")
parser.add_argument('--reg_gamma', type=float, default=10 ** -5, help="regularizer term weight")
parser.add_argument('--gamma_decay', action='store_true', help="if you want gamma decay or not")
parser.add_argument('--exp_alpha', type=float, default=1., help="weight to multiply exp argument")
parser.add_argument('-p', '--patience', type=int, default=5, help="last [patience] checkpoint saved")
parser.add_argument('--numworkers', type=int, default=8, help="num of data workers")
args = parser.parse_args()
wandb.init(project="tsk1_l2", entity='giobin')
main(args)