-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_many_eval.py
260 lines (231 loc) · 12.2 KB
/
run_many_eval.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
import argparse
import inspect
import os
import random
import re
from copy import deepcopy
from typing import Any, Dict, List
from omegaconf import OmegaConf
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from llmog import METHOD_MAPPING, METRIC_MAPPER, RUN_MAPPER, TORCH_DTYPE_MAPPING
from llmog.models.fid_model import FiDModel
from llmog.models.rag_model import RagSequenceModel, RagTokenModel
from llmog.prompting import get_template_mappings
from llmog.utils import fix_seed, get_train_valid_dataset, write_results
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument_group(title="plm")
parser.add_argument("--model_path", type=str, required=True)
parser.add_argument("--model_revision", type=str, default=None, help="The revision for model checkpoint")
parser.add_argument("--torch_dtype", type=str, default="torch.float32", help="The dtype for the model")
parser.add_argument("--first_sentinel_token", type=str, default="<extra_id_0>")
parser.add_argument("--denoiser_prefix", type=str, default=None, choices=["[NLU]", "[NLG]", "[S2S]"])
parser.add_argument("--model_cache_dir", type=str, default=None)
parser.add_argument_group(title="env")
parser.add_argument("--seed", type=int, nargs="*")
parser.add_argument("--num_gpus", type=int, default=1, help="Num gpus in your node for model parallel")
parser.add_argument("--logging_samples", action="store_true", help="Print I/O text or not")
parser.add_argument("--output_file", type=str, nargs="*", help="Path to save output results")
parser.add_argument_group(title="task")
parser.add_argument(
"--type", type=str, choices=["k-shot", "fid-k-shot", "rag-sequence-k-shot", "rag-token-k-shot"]
)
parser.add_argument("--num_k", type=int, nargs="*", required=True)
parser.add_argument("--dataset_name", type=str, nargs="*", required=True)
parser.add_argument(
"--subtask_name",
type=lambda x: None if x == "None" else str(x),
nargs="*",
required=True,
help="For the bundle-type benchmark like superGLUE or KLUE",
)
parser.add_argument(
"--train_path",
type=lambda x: None if x == "None" else str(x),
nargs="*",
help="Train data path for tasks not uploaded to huggingface datasets",
)
parser.add_argument(
"--valid_path",
type=lambda x: None if x == "None" else str(x),
nargs="*",
help="Valid data path for tasks not uploaded to huggingface datasets",
)
parser.add_argument("--run_all_templates", action="store_true")
parser.add_argument("--template_names", type=str, nargs="*", required=False)
parser.add_argument(
"--fix_demon_samples", action="store_true", help="Fix few-shot demonstrations for all test iterations."
)
parser.add_argument("--num_valid_samples", type=int, help="Num test samples for evaluation")
parser.add_argument("--num_valid_ratio", type=int, help="Test samples ratio for evaluation")
parser.add_argument("--use_sentinel", action="store_true", help="Use sentinel token just as in pretrain.")
parser.add_argument("--test_data_to_decoder", action="store_true", help="Put test input to decoder or not")
parser.add_argument(
"--add_eos_loss", action="store_true", help="If true, loss for eos token be included to total loss comparison."
)
parser.add_argument(
"--reduction",
type=str,
default="sum",
choices=["sum", "mean"],
help="Loss reduction method for classification tasks",
)
parser.add_argument("--generation_hp_path", type=str, help="Yaml path for generation eval hyperparameters")
args = parser.parse_args()
return args
def get_generation_kwargs(generation_config_path: str) -> Dict[str, Any]:
nested_args = OmegaConf.load(generation_config_path)
generation_kwargs = nested_args.generation
return generation_kwargs
def main():
args = get_args()
assert (
len(args.dataset_name) == len(args.subtask_name)
and len(args.dataset_name) == len(args.output_file)
and len(args.train_path) == len(args.valid_path)
and len(args.train_path) == len(args.dataset_name)
), "Number of dataset_name, subtask_name, output_file must all be matched"
model_kwargs = {
"torch_dtype": TORCH_DTYPE_MAPPING[args.torch_dtype],
"revision": args.model_revision,
"low_cpu_mem_usage": True,
"cache_dir": args.model_cache_dir,
}
# check tokenizer include sentinel token
print("loading tokenizer")
tokenizer = AutoTokenizer.from_pretrained(args.model_path)
if args.use_sentinel:
assert args.first_sentinel_token in tokenizer.get_vocab(), "Put proper extra token"
# load model for proper k-shot type
print("loading model")
if args.type == "fid-k-shot":
# We recommend saving model after 'load_t5' using 'model.save_pretrained(save_path)'
# and then loading the model using 'FiDT5.from_pretrained(save_path, **model_kwargs)'
# from next time because 'load_t5' function is too slow
model_base = AutoModelForSeq2SeqLM.from_pretrained(args.model_path, **model_kwargs)
model = FiDModel(model_base.config)
model.load_t5(model_base.state_dict())
elif args.type.startswith("rag"):
rag_mapper = {"rag-token-k-shot": RagTokenModel, "rag-sequence-k-shot": RagSequenceModel}
model = rag_mapper[args.type].from_pretrained(args.model_path, **model_kwargs)
else:
model = AutoModelForSeq2SeqLM.from_pretrained(args.model_path, **model_kwargs)
print("model weights loading to gpu")
if args.num_gpus > 1:
from parallelformers import parallelize
parallelize(model, num_gpus=args.num_gpus, fp16=False, verbose="simple")
else:
model.to("cuda")
# set decoder start token id
decoder_start_id = (
model.config.decoder_start_token_id if tokenizer.pad_token_id is None else tokenizer.pad_token_id
)
first_sentinel_id = tokenizer.get_vocab()[args.first_sentinel_token]
# Iterate many k-shot evals
for (dataset_name, subtask_name, output_file, train_path, valid_path) in zip(
args.dataset_name, args.subtask_name, args.output_file, args.train_path, args.valid_path
):
for num_k in args.num_k:
for seed in args.seed:
fix_seed(seed)
template_mappings, prompts = get_template_mappings(dataset_name, subtask_name)
if args.run_all_templates:
template_list = template_mappings.values()
else:
template_list = [template_mappings[template] for template in args.template_names]
for template in template_list:
try:
template = prompts.templates[template]
print(template.get_name())
train_dataset, valid_dataset = get_train_valid_dataset(
dataset_name, subtask_name, train_path, valid_path
)
# Sample valid(test) dataset if given sampling args
if args.num_valid_ratio and not args.num_valid_samples:
args.num_valid_samples = int(len(valid_dataset) * args.num_valid_ratio)
if args.num_valid_ratio or args.num_valid_samples:
if args.num_valid_samples >= len(valid_dataset):
args.num_valid_samples = len(valid_dataset)
print(
"Number of valid set entered is greater than total set, so all the valid set will be used"
)
else:
valid_random_indices = random.sample(range(len(valid_dataset)), args.num_valid_samples)
valid_dataset = valid_dataset.select(valid_random_indices)
if args.fix_demon_samples:
train_random_indices = random.sample(range(len(train_dataset)), num_k)
else:
num_demon_examples = min(num_k * len(valid_dataset), len(train_dataset))
train_random_indices = random.sample(range(len(train_dataset)), num_demon_examples)
train_dataset = train_dataset.select(train_random_indices)
print(f"Num Train Samples for k-shot: {num_k}")
print(f"Num Test Samples: {len(valid_dataset)}")
# create proper dataset for k-shot method and task type
task_key = "-".join([dataset_name, subtask_name]) if subtask_name else dataset_name
metric_name: List[str] = METRIC_MAPPER[task_key]
mapped_ds = METHOD_MAPPING[metric_name[-1]][args.type](
tokenizer=tokenizer,
decoder_start_id=decoder_start_id,
first_sentinel_id=first_sentinel_id,
train_dataset=train_dataset,
valid_dataset=valid_dataset,
template=template,
num_k=num_k,
denoiser_prefix=args.denoiser_prefix,
use_sentinel=args.use_sentinel,
test_data_to_decoder=args.test_data_to_decoder,
add_eos_loss=args.add_eos_loss,
num_proc=os.cpu_count() // 2,
)
logging_message = deepcopy(args.type).upper()
logging_message = re.sub("K-SHOT", f"{num_k}-SHOT", logging_message)
run_fn_ = RUN_MAPPER[METHOD_MAPPING[metric_name[-1]][args.type]]
if "generation_kwargs" not in inspect.signature(run_fn_).parameters:
generation_kwargs = None
print(f"{logging_message} classification")
result = run_fn_(
model,
tokenizer,
mapped_ds,
metric_name,
args.reduction,
args.num_gpus != 1,
args.logging_samples,
)
else:
print(f"{logging_message} generation")
generation_kwargs = get_generation_kwargs(args.generation_hp_path)
print(generation_kwargs)
result = run_fn_(
model,
tokenizer,
mapped_ds,
metric_name,
generation_kwargs,
args.num_gpus != 1,
args.logging_samples,
args.use_sentinel,
args.first_sentinel_token,
)
iter_args = deepcopy(args)
iter_args.seed = seed
iter_args.num_k = num_k
iter_args.dataset_name = dataset_name
iter_args.subtask_name = subtask_name
iter_args.output_file = output_file
iter_args.train_path = train_path
iter_args.valid_path = valid_path
print(f"Results with [{template.get_name()}]")
print(iter_args)
print(result)
if output_file:
write_results(
results=result,
template_name=template.get_name(),
args=iter_args,
generation_kwargs=generation_kwargs,
)
except:
print(f"[{template.get_name()}] raised error!")
if __name__ == "__main__":
main()