-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtext_sample_sum.py
214 lines (181 loc) · 8.81 KB
/
text_sample_sum.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
"""
Generate a large batch of image samples from a model and save them as a large
numpy array. This can be used to produce samples for FID evaluation.
"""
import argparse
import os, json
from functools import partial
import numpy as np
import torch
import torch as th
import torch.distributed as dist
from torch.nn.parallel.distributed import DistributedDataParallel as DDP
import torch.nn.functional as F
from tqdm import tqdm
from imp_diff.improved_diffusion import dist_util, logger
from imp_diff.improved_diffusion.nn import mean_flat
from imp_diff.improved_diffusion.script_util import create_model_and_diffusion, args_to_dict, \
model_and_diffusion_defaults, add_dict_to_argparser
from imp_diff.improved_diffusion.summarization_datasets_exp import load_data_summarization
from imp_diff.improved_diffusion.test_util import get_weights, compute_logp
from transformers import set_seed
def main():
set_seed(101)
args = create_argparser().parse_args()
dist_util.setup_dist()
logger.configure()
# load configurations.
config_path = os.path.join(os.path.split(args.model_path)[0], "training_args.json")
print(config_path)
# sys.setdefaultencoding('utf-8')
with open(config_path, 'rb', ) as f:
training_args = json.load(f)
training_args['batch_size'] = args.batch_size
args.__dict__.update(training_args)
args.sigma_small = True
# args.diffusion_steps = 200 #500 # DEBUG
logger.log("creating model and diffusion...")
model, diffusion = create_model_and_diffusion(
**args_to_dict(args, model_and_diffusion_defaults().keys())
)
print("Loading Model from: {}".format(args.model_path))
model.load_state_dict(
torch.load(args.model_path)
)
model.to(dist_util.dev())
model.eval() # DEBUG
print("Model Type: {}".format(type(model))) # Transformer 3 Model
pytorch_total_params = sum(p.numel() for p in model.parameters())
logger.log(f'the parameter count is {pytorch_total_params}')
print(diffusion.rescale_timesteps, 'a marker for whether we are in the debug mode')
batch_size = 128
data_test = load_data_summarization(batch_size, 8, training_args['roc_train'], split=['test'],
sent_encoder_type='sbert', summary_type='oracle',
summary_level='sen', shuffle=False)
logger.log("sampling...")
all_images = []
all_article_sen_embs = []
all_article_sen_masks = []
all_article_sens = []
all_summary_sens = []
all_cand_sens = []
all_cand_combs = []
with torch.no_grad():
for i, (batch, model_kwargs) in tqdm(enumerate(data_test)):
summary_sens = model_kwargs['summary_sens']
article_sens = model_kwargs['article_sens']
all_article_sens += article_sens
all_summary_sens += summary_sens
article_sen_embs = model_kwargs['article_sen_embeds']
article_sen_masks = model_kwargs['article_sen_masks']
all_article_sen_masks.append(article_sen_masks)
all_article_sen_embs.append(article_sen_embs)
all_cand_sens += (model_kwargs['cand_sens'])
all_cand_combs += (model_kwargs['cand_combs'])
sample_fn = (
diffusion.p_sample_loop
)
sample_shape = (batch_size, args.image_size ** 2, args.in_channel)
sample = sample_fn(
model,
sample_shape,
clip_denoised=args.clip_denoised,
denoised_fn= None ,
model_kwargs=model_kwargs,
top_p =-1,
)
gathered_samples = [th.zeros_like(sample) for _ in range(dist.get_world_size())]
dist.all_gather(gathered_samples, sample) # gather not supported with NCCL
all_images.extend([sample.cpu().numpy() for sample in gathered_samples])
arr = np.concatenate(all_images, axis=0)
print(arr.shape, 'full shape')
model_base_name = os.path.basename(os.path.split(args.model_path)[0]) + f'.{os.path.split(args.model_path)[1]}'
out_path = os.path.join(args.out_dir, f"{model_base_name}.arr.npy")
print("Saved Arr to :{}".format(out_path))
if diffusion.training_mode.startswith('e2e'):
preds = []
summaries = []
x_t = th.tensor(arr).cuda()
all_article_cand_embs = torch.zeros_like(x_t[:, :article_sen_embs.shape[1], :]).cuda()
all_artice_cand_mask = torch.zeros((all_article_cand_embs.shape[0], 1, all_article_cand_embs.shape[1])).cuda()
for bid, cands in enumerate(all_cand_sens):
for j, cand_idx in enumerate(cands):
all_article_cand_embs[bid, cand_idx, :] = x_t[bid, cand_idx, :]
all_artice_cand_mask[bid, :, cand_idx] = 1
x_sum = x_t[:, article_sen_embs.shape[1]:, :]
if args.model_arch == 'conv-unet':
reshaped_x_t = x_t.view(x_t.size(0), -1, x_t.size(-1))
else:
reshaped_x_t = x_sum
logits = th.softmax(torch.bmm(reshaped_x_t, x_t[:, :article_sen_embs.shape[1], :].permute(0, 2, 1)),
dim=-1).squeeze() # bsz, seqlen, vocab
logits *= all_artice_cand_mask.cuda()
cands = th.topk(logits, k=5, dim=-1)
sample = cands.indices
for i, (indice, value) in enumerate(zip(cands.indices, cands.values)):
end_idx = article_sen_embs.shape[1]
pred_idxs =[]
if training_args['roc_train'] == 'cnn_ext':
max_summary_sen_num = 3
elif training_args['roc_train'] == 'pubmed':
max_summary_sen_num = 6
elif training_args['roc_train'] == 'wikihow':
max_summary_sen_num = 4
elif training_args['roc_train'] == 'multinews':
max_summary_sen_num = 9
elif training_args['roc_train'] == 'xsum':
max_summary_sen_num = 2
for candis, cands_val in zip(indice, value):
for idx, val in zip(candis, cands_val):
if idx not in pred_idxs:
pred_idxs.append(int(idx))
break
if len(pred_idxs) == max_summary_sen_num:
break
pred_idxs = sorted([idx for idx in pred_idxs if idx != end_idx])
tokens = " ".join([all_article_sens[i][x] for x in pred_idxs if x < len(all_article_sens[i])])
summary = " ".join(all_summary_sens[i])
preds.append(tokens)
summaries.append(summary)
dict = {'preds': preds, 'summaries': summaries, 'cands': sample}
model_base_name = os.path.basename(os.path.split(args.model_path)[0]) + f'.{os.path.split(args.model_path)[1]}'
out_path = os.path.join(args.out_dir, f"{model_base_name}.samples_{args.top_p}.npy")
logger.log(f"saving to {out_path}")
np.save(out_path, dict)
logger.log(f"Finished saving to {out_path}")
def create_argparser():
defaults = dict(
clip_denoised=False,
num_samples=50,#10000,
batch_size=64,
use_ddim=False,
mbr_sample=1,
model_path="",
model_arch='conv-unet',
verbose='yes',
out_dir="diffusion_lm/improved_diffusion/out_gen",
gradient_clipping=-1.0,
use_fp16=False,
fp16_scale_growth=1e-3,
)
text_defaults = dict(modality='text',
dataset_name='wikitext',
dataset_config_name='wikitext-2-raw-v1',
config='diffusion_lm/synthetic_data/configs/emnlp2020/experiments/difflm_seed0_m3_k128_trainc20000.yaml',
model_name_or_path='predictability/diff_models/compress_e=5_b=60_m=gpt2_wikitext-103-raw-v1_None',
experiment='gpt2_pre_compress', model_arch='conv-unet',
roc_train='diffusion_lm/ROCstory', # 'diffusion_lm/ROCstory/ROCstory17.csv',
wiki_train='diffusion_lm/simple_wiki/data.v1.split/simple.training.txt',
e2e_train='e2e_data',
yelp_train='diffusion_lm/yelpnlg-resources/yelpnlg-corpus',
commonGen_train='diffusion_lm/common-gen/commongen_data',
emb_scale_factor=1.0, noise_level=0.0, cache_mode='no', use_bert_tokenizer='no',
padding_mode='block',
preprocessing_num_workers=1, top_p=-1.0)
defaults.update(model_and_diffusion_defaults())
defaults.update(text_defaults)
parser = argparse.ArgumentParser()
add_dict_to_argparser(parser, defaults)
return parser
if __name__ == "__main__":
main()