-
Notifications
You must be signed in to change notification settings - Fork 3
/
inference_task.py
398 lines (331 loc) · 17.1 KB
/
inference_task.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
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
import os
import json
import argparse
import logging
import random
import time
import numpy as np
from tqdm import tqdm
import torch
import transformers
from transformers import AutoConfig, AutoTokenizer
from src.model.GPT2modeling import GPT2LMHeadModel
from src.data_utils.dialog_reader import get_wow_dataloader
from src.data_utils.cmu_dog_reader import get_cmu_dog_dataloader
from src.data_utils.moe_reader import get_wow_topic_dataloaders, get_cmu_dog_topic_dataloaders
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO,
)
logger = logging.getLogger(__name__)
def prepare_inputs_for_generation(config, batch):
if args.model_type=="decoder_only":
inputs = batch["context"]
masks = batch["response"] # context 1 response 0 padding 0
label_masks = batch["context_mask"] # context 1 response 1 padding 0
label_starts = torch.sum(masks, 1)
label_idxs = torch.sum(label_masks, 1)
inputs = inputs[:,:label_starts[0]]
masks = masks[:,:label_starts[0]]
if args.history_in_context and "token_type" in batch:
token_type_ids = batch["token_type"][:,:label_starts[0]]
else:
token_type_ids = None
assert inputs.shape[0] == 1
elif args.model_type=="seq2seq":
inputs = batch["context"]
masks = batch["context_mask"]
labels = batch["response"]
label_masks = batch["response_mask"]
token_type_ids = None
label_starts = labels
label_idxs = torch.sum(label_masks, 1)
else:
raise ValueError("Invalid model type when preparing inputs for generation!")
if args.moe:
experts = batch["topic_map"]
else:
experts = None
return inputs, masks, token_type_ids, experts, label_starts, label_idxs
def prediction_step(args, model, batch, tokenizer, stop_token, pad_token_id, lens=None):
inputs, masks, token_type_ids, experts, label_starts, label_idxs = prepare_inputs_for_generation(args, batch)
kn_sent = batch["chosen_sentence"]
knowledges = batch["knowledge"]
if experts is not None:
experts = experts.to(args.device)
if token_type_ids is not None:
token_type_ids = token_type_ids.to(args.device)
# max length
max_length = args.length + inputs.shape[-1]
start = time.time()
generated_sequence = model.generate(
input_ids=inputs.to(args.device),
max_length=max_length,
temperature=args.temperature,
top_k=args.k,
top_p=args.p,
repetition_penalty=args.repetition_penalty,
do_sample=args.sampling,
num_return_sequences=args.num_return_sequences,
attention_mask=masks.to(args.device),
experts=experts,
token_type_ids=token_type_ids,
pad_token_id=pad_token_id,
)
generated_sequence = generated_sequence[0] # only take the first generated sentence
if args.model_type=="decoder_only":
generated_sequence = generated_sequence[inputs.shape[-1]:]
gen_len = generated_sequence.size()[0]
if args.debug:
logger.info(f"The shape of the output sequences {len(generated_sequence)}.")
# Decode text
text = tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True) # DO NOT skip_special_tokens
if not args.debug:
# Remove all text after the stop token
text = text[: text.find(stop_token) if stop_token and text.find(stop_token)>0 else None].replace("</s>", "").replace("<s>", "")
interval = time.time() - start
response = batch["context"][0,label_starts[0]:label_idxs[0]-1] if args.model_type=="decoder_only" else label_starts[0,:label_idxs[0]-1]
response_text = tokenizer.decode(response, clean_up_tokenization_spaces=True, skip_special_tokens=True)
kn_text = tokenizer.decode(kn_sent[0], clean_up_tokenization_spaces=True)
kn_text = kn_text[: kn_text.find(stop_token) if stop_token else None]
if args.debug:
print(f"The generated sentence is: {text}")
print(f"The golden sentence is: {response_text}")
print(f"The knowledge sentence is: {kn_text}")
print(f"The whole input is: {tokenizer.decode(inputs[0,:label_starts[0]], clean_up_tokenization_spaces=True)}")
print(f"The knowledge candidates are:")
for kn in knowledges:
print(kn[0].split("__knowledge__"))
print("="*80)
input()
return text, response_text, kn_text, interval, gen_len
def Inference(args):
os.makedirs(f"{args.save_path}/{args.exp}", exist_ok=True)
# Initialize the model and tokenizer
model_name_or_path = os.path.join(args.model_folder, args.exp) + "/checkpoint-" + args.checkpoint if "best" not in args.checkpoint else os.path.join(args.model_folder, args.exp) + "/model-" + args.checkpoint
config = AutoConfig.from_pretrained(model_name_or_path)
config.lm = args.lm
config.task_adapter = args.task_adapter
config.n_neck = args.t_neck
config.kadapter = args.kadapter
config.kn_neck = args.kn_neck
config.num_kadapter = args.n_experts if args.kadapter else 0
config.kadapter_one_hot = args.kadapter_one_hot
config.num_beams = 1 if not args.beam_search else 4
config.dual_kadapter = args.dual_kadapter
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
if "gpt2" in args.pretrained_model:
model, loading_info = GPT2LMHeadModel.from_pretrained(model_name_or_path, config=config, output_loading_info=True)
print("loading_info[missing_keys]", sorted(loading_info["missing_keys"]))
assert len(loading_info["missing_keys"]) == len(loading_info["unexpected_keys"])
assert len(loading_info["missing_keys"]) == 0
# Load ckpt
if args.ckpt != '':
logger.info("Load the fine-tuned model...")
model.load_state_dict(torch.load(args.ckpt),strict=False)
model.to(args.device)
# get dataloaders
args.inference = True
if args.dataset == "wow":
if args.moe:
dataloaders = get_wow_topic_dataloaders(args, tokenizer, train=False if args.split!="train" else True, valid=True if "valid" in args.split else False, cal_time=args.cal_time)
else:
dataloaders = get_wow_dataloader(args, tokenizer, train=False if args.split!="train" else True)
elif args.dataset == "cmu_dog":
if args.moe:
dataloaders = get_cmu_dog_topic_dataloaders(args, tokenizer, train=False if args.split!="train" else True, valid=True if "valid" in args.split else False)
else:
dataloaders = get_cmu_dog_dataloader(args, tokenizer, train=False if args.split!="train" else True)
else:
raise ValueError("Invalid dataset!")
if tokenizer.sep_token is None:
stop_token = tokenizer.eos_token
else:
stop_token = tokenizer.sep_token
pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
print(f"The stop token is {stop_token}")
if args.dataset == "wow":
splits = [args.split, args.split+"_unseen"]
elif args.dataset == "cmu_dog":
splits = [args.split]
else:
raise NotImplementedError
for split in splits:
if os.path.exists(f'{args.save_path}/{args.exp}/{split}_{args.checkpoint}_generation.txt') and os.path.exists(f'{args.save_path}/{args.exp}/{split}_{args.checkpoint}_gold.txt') and not args.overwrite and not args.debug and not args.cal_time:
logger.info(f"The result already exists and saved in '{args.save_path}/{args.exp}/{split}_{args.checkpoint}_generation.txt'! Skip inference!")
continue
logger.info(f"Doing inference for split {split}. Will save the results in '{args.save_path}/{args.exp}/{split}_{args.checkpoint}_generation.txt'.")
loader = dataloaders[split]
lens = None
avg_lens = []
total_time = 0
generated_sequences = []
golden_sequences = []
kn_sequences = []
for i, batch in tqdm(enumerate(loader), desc=f'Inference', total=len(loader), ncols=100):
gen_text, response_text, kn_text, interval, gen_len = prediction_step(args, model, batch, tokenizer, stop_token, pad_token_id, lens=lens if lens is not None else None)
total_time += interval
avg_lens.append(gen_len)
generated_sequences.append(gen_text)
golden_sequences.append(response_text)
kn_sequences.append(kn_text)
if not args.debug and not args.cal_time:
if args.expert_case:
with open(f'{args.save_path}/{args.exp}/{split}_{args.checkpoint}_{args.expert_idx}_generation.txt', "w") as f:
for line in generated_sequences:
f.write(line.replace("\n", " ")+"\n")
else:
with open(f'{args.save_path}/{args.exp}/{split}_{args.checkpoint}_generation.txt', "w") as f:
for line in generated_sequences:
f.write(line.replace("\n", " ")+"\n")
with open(f'{args.save_path}/{args.exp}/{split}_{args.checkpoint}_gold.txt', "w") as f:
for line in golden_sequences:
f.write(line.replace("\n", " ")+"\n")
with open(f'{args.save_path}/{args.exp}/{split}_kn.txt', "w") as f:
for line in kn_sequences:
f.write(line.replace("\n", " ")+"\n")
logger.info(f"Inference process takes {total_time} seconds.")
logger.info(f"Generation average length is {np.mean(avg_lens)} tokens.")
def Comparison(args):
## Get a subset of experiment examples for comparison RANDOM INDEX
l = random.sample(range(900), 100)
l.sort()
print(f"The selected sample indexes are {l}")
results = [] # the item stored in is DICT: model name & index & generations
# Initialize the model and tokenizer -> tokenizer + a list of results from different models
tokenizer = AutoTokenizer.from_pretrained(args.pretrained_model)
if tokenizer.sep_token is None:
stop_token = tokenizer.eos_token
else:
stop_token = tokenizer.sep_token
print(f"The stop token is {stop_token}")
args.inference = True
dataloaders = get_wow_dataloader(args, tokenizer, train=False if args.split != "train" else True, valid=True if "valid" in args.split else False, shuffle_train=False)
split = args.split
dataloader = dataloaders[split]
for exp, ckpt in zip(args.exps, args.checkpoints):
file_path = os.path.join(args.save_path, exp, f"{split}_{ckpt}_generation.txt")
print(f"Loading file '{file_path}'......")
with open(file_path, "r") as f:
lines = f.readlines()
generated_sequences = {}
for i, line in enumerate(lines):
generated_sequences[i] = line.strip()
result = {
"name": exp+"-"+ckpt,
"text": generated_sequences,
}
results.append(result)
for idx, batch in enumerate(dataloader):
if idx not in l:
continue
inputs = batch["context"]
masks = batch["response"] # context 1 response 0 padding 0
label_masks = batch["context_mask"] # context 1 response 1 padding 0
label_starts = torch.sum(masks, 1)
label_idxs = torch.sum(label_masks, 1)
kn_sent = batch["chosen_sentence"]
knowledges = batch["knowledge"]
topic = batch["title"]
response = inputs[0,label_starts[0]:label_idxs[0]-1]
response_text = tokenizer.decode(response, clean_up_tokenization_spaces=True)
kn_text = tokenizer.decode(kn_sent[0], clean_up_tokenization_spaces=True)
kn_text = kn_text[: kn_text.find(stop_token) if stop_token else None]
topic_text = tokenizer.decode(topic[0], clean_up_tokenization_spaces=True)
print(f"The knowledge candidates are:")
for kn in knowledges:
print(kn[0].split("__knowledge__"))
print(f"The knowledge sentence is: {topic_text.strip()}\t{kn_text}")
print(f"The whole input is: ")
input_turns = tokenizer.decode(inputs[0,:label_starts[0]], clean_up_tokenization_spaces=True).split("<|endoftext|>")
for turn in input_turns:
print("\t" + turn.replace("\n","\t"))
print(f"The golden response is: {response_text}")
for result in results:
# print the results
print(f"Model {result['name']}: {result['text'][idx]}")
print("="*80)
input()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# model settings
parser.add_argument(
"--model_folder",
default=None,
type=str,
required=True,
)
parser.add_argument(
"--ckpt",
default="",
type=str,
)
parser.add_argument('--exp', type=str, default="wow-history")
parser.add_argument('--checkpoint', type=str, default="9")
parser.add_argument("--bsz", type=int, default=1)
parser.add_argument("--eval_bsz", type=int, default=1)
# generation settings
parser.add_argument("--length", type=int, default=100)
parser.add_argument(
"--temperature",
type=float,
default=1.0,
help="temperature of 1.0 has no effect, lower tend toward greedy sampling",
)
parser.add_argument(
"--repetition_penalty", type=float, default=1.0, help="primarily useful for CTRL model; in that case, use 1.2"
)
parser.add_argument("--beam_search", action="store_true")
parser.add_argument("--sampling", action="store_true")
parser.add_argument("--k", type=int, default=0)
parser.add_argument("--p", type=float, default=0.9)
# data settings
parser.add_argument('--dataset', type=str, default="wow")
parser.add_argument('--data_dir', type=str, default="./data")
parser.add_argument('--save_path', type=str, default="./save/results")
parser.add_argument('--pretrained_model', type=str, default="gpt2")
parser.add_argument('--kadapter', action="store_true")
parser.add_argument('--task_adapter', action="store_true")
parser.add_argument('--dual_kadapter', action="store_true")
parser.add_argument('--n_experts', type=int, default=1) # number of kadapter
parser.add_argument('--kadapter_one_hot', action='store_true') # if set True, top 1 kadapter else weighted sum
parser.add_argument('--kadapter_equal', action='store_true') # if set True, the kadapters will be equally distributed
parser.add_argument('--cluster_path', type=str, default='save/models/topic_models/ctm_4')
parser.add_argument('--kadapter_path', type=str, default='')
parser.add_argument('--kadapter_ckp', type=int, default=29)
parser.add_argument('--moe', action="store_true") # moe mode
parser.add_argument("--lm", action="store_true", help="Train with language modeling.")
parser.add_argument('--t_neck', help='the dimension of the bottleneck of the task adapter', type=int, default=256)
parser.add_argument('--kn_neck', help='the dimension of the bottleneck of the knowledge adapter', type=int, default=256)
parser.add_argument('--max_length', type=int, default=1024)
parser.add_argument('--max_context_length', type=int, default=1024) # 256 if --history_in_context
parser.add_argument('--max_kn_length', type=int, default=1024)
parser.add_argument('--max_episode_length', type=int, default=1) # history length
parser.add_argument('-hic','--history_in_context', action="store_true")
parser.add_argument('-kic','--kn_in_context', action="store_true")
parser.add_argument('--mode', type=str, default="wizard")
parser.add_argument('--split', type=str, choices=["train", "valid", "test"], default="valid")
parser.add_argument('--model_type', type=str, default="decoder_only")
# ablation study
parser.add_argument('--expert_case', action="store_true")
parser.add_argument('--expert_idx', type=int, default=0)
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
parser.add_argument("--num_return_sequences", type=int, default=1, help="The number of samples to generate.")
# for debug
parser.add_argument("--debug", action="store_true", help="Enter DEBUG mode")
parser.add_argument("--overwrite", help="Overwrite the inference results even though it exists already", type=bool, default=False)
# model comparison
parser.add_argument("--compare", action="store_true", help="Enter COMPARE mode")
parser.add_argument('--exps', help="The path to the experiment results", action="append")
parser.add_argument('--checkpoints', help="The checkpoint corresponding the experiment results", action="append")
parser.add_argument("--cal_time", action="store_true", help="Calculate the inference time.")
parser.add_argument('--gold_cluster', action="store_true")
args = parser.parse_args()
args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
np.random.seed(args.seed)
torch.manual_seed(args.seed)
os.environ["TOKENIZERS_PARALLELISM"] = "false"
if args.compare:
Comparison(args)
else:
Inference(args)