-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy patheval_ppl_utils.py
286 lines (262 loc) · 12 KB
/
eval_ppl_utils.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
from datautils import get_loaders
from lm_eval import evaluator
from pprint import pprint
import torch
import torch.nn as nn
from tqdm import tqdm
import numpy as np
import os
import copy
from categories import subcategories, categories
import pdb
@torch.no_grad()
def evaluate(lm, args, logger):
results = {}
hf_device_map = lm.model.hf_device_map
hf_device = f"cuda:{hf_device_map[f'model.layers.{0}']}"
print("hf_device_map: ", hf_device_map, "hf_device: ", hf_device)
if args.eval_ppl:
for dataset in ["wikitext2", "c4", "ptb"]:
cache_testloader = f'{args.cache_dir}/testloader_{args.model_family}_{dataset}_all.cache'
if os.path.exists(cache_testloader):
testloader = torch.load(cache_testloader)
logger.info(f"load calibration from {cache_testloader}")
else:
dataloader, testloader = get_loaders(
dataset,
seed=args.seed,
model=args.model,
seqlen=lm.seqlen,
)
torch.save(testloader, cache_testloader)
if "c4" in dataset:
testenc = testloader
else:
testenc = testloader.input_ids
nsamples = testenc.numel() // lm.seqlen
use_cache = lm.model.config.use_cache
lm.model.config.use_cache = False
lm.model.eval()
nlls = []
for i in tqdm(range(nsamples)):
batch = testenc[:, (i * lm.seqlen) : ((i + 1) * lm.seqlen)].to(hf_device)
if "opt" in args.net.lower():
logs = lm.model.model.decoder(batch.to(hf_device))
elif "llama" in args.net.lower() or "mixtral" in args.net.lower():
logs = lm.model.model(batch.to(hf_device))
elif "falcon" in args.model:
logs = lm.model.transformer(batch.to(hf_device))
hidden_states = logs[0]
logits = lm.model.lm_head(hidden_states)
shift_logits = logits[:, :-1, :]
shift_labels = testenc[:, (i * lm.seqlen) : ((i + 1) * lm.seqlen)][:, 1:].to(lm.model.lm_head.weight.device)
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
)
neg_log_likelihood = loss.float() * lm.seqlen
nlls.append(neg_log_likelihood)
if i == args.limit:
break
ppl = torch.exp(torch.stack(nlls).sum() / (nsamples * lm.seqlen))
logger.info(f'{dataset} : {ppl.item()}')
lm.model.config.use_cache = use_cache
results[dataset] = ppl.item()
if args.tasks != "":
t_results = evaluator.simple_evaluate(
lm,
tasks=args.tasks,
num_fewshot=args.num_fewshot,
limit=None if args.limit == -1 else args.limit,
)
results.update(t_results)
logger.info(results)
pprint(results)
# for test of MMLU
if 'hendrycksTest' in args.tasks:
all_cors = []
all_cors_norm = []
subcat_cors = {subcat: [] for subcat_lists in subcategories.values() for subcat in subcat_lists}
cat_cors = {cat: [] for cat in categories}
cat_cors_norm = {cat: [] for cat in categories}
for key in t_results['results'].keys():
if not 'hendrycksTest' in key:
continue
subject = key.split('-')[-1]
cors = t_results['results'][key]['acc']
cors_norm = t_results['results'][key]['acc_norm']
subcats = subcategories[subject]
for subcat in subcats:
subcat_cors[subcat].append(cors)
for key in categories.keys():
if subcat in categories[key]:
cat_cors[key].append(cors)
cat_cors_norm[key].append(cors_norm)
all_cors.append(cors)
all_cors_norm.append(cors_norm)
for cat in cat_cors:
cat_acc = np.mean(cat_cors[cat])
logger.info("Average accuracy {:.4f} - {}".format(cat_acc, cat))
weighted_acc = np.mean(all_cors)
logger.info("Average accuracy: {:.4f}".format(weighted_acc))
return results
@torch.no_grad()
def evaluate_lrquant(lm, args, logger, fp_lm):
results = {}
# torch.save(lm.model.state_dict(),os.path.join(args.output_dir, f"current.pth"))
if args.multigpu:
if "opt" in args.net.lower():
map_layers_to_multi_gpus(lm.model.model.decoder.layers)
input_device = lm.model.model.decoder.layers[0].device
output_device = lm.model.model.decoder.layers[-1].device
lm._device = input_device
assert input_device == output_device
lm.model.model.decoder.embed_positions.to(input_device)
lm.model.model.decoder.embed_tokens.to(input_device)
lm.model.model.decoder.final_layer_norm.to(output_device)
lm.model.lm_head.to(output_device)
elif "llama" in args.net.lower():
map_layers_to_multi_gpus(lm.model.model.layers)
input_device = lm.model.model.layers[0].device
output_device = lm.model.model.layers[-1].device
assert input_device == output_device
lm._device = input_device
lm.model.model.embed_tokens.to(input_device)
lm.model.model.norm.to(output_device)
lm.model.lm_head.to(output_device)
elif "falcon" in args.net.lower():
map_layers_to_multi_gpus(lm.model.transformer.h)
input_device = lm.model.transformer.h[0].device
output_device = lm.model.transformer.h[-1].device
assert input_device == output_device
lm._device = input_device
lm.model.transformer.word_embeddings.to(input_device)
lm.model.transformer.ln_f.to(output_device)
lm.model.lm_head.to(output_device)
else:
if "opt" in args.net.lower():
lm.model.model.decoder = lm.model.model.decoder.to(lm.device)
elif "llama" in args.net.lower():
lm.model = lm.model.to(lm.device)
elif "falcon" in args.net.lower():
lm.model.transformer = lm.model.transformer.to(lm.device)
if args.eval_ppl:
for dataset in ["wikitext2","ptb","c4","ptb-new",'c4-new']:
cache_testloader = f'{args.cache_dir}/testloader_{args.model_family}_{dataset}_all.cache'
if os.path.exists(cache_testloader):
testloader = torch.load(cache_testloader)
logger.info(f"load calibration from {cache_testloader}")
else:
_, testloader = get_loaders(
dataset,
seed=args.seed,
model=args.model,
seqlen=lm.seqlen,
)
torch.save(testloader, cache_testloader)
if "c4" in dataset:
testenc = testloader
else:
testenc = testloader.input_ids
lm.model.load_state_dict(torch.load(os.path.join(args.output_dir, f"current.pth")))
lm.model.eval()
nsamples = testenc.numel() // lm.seqlen
use_cache = lm.model.config.use_cache
lm.model.config.use_cache = False
fp_lm.model.config.use_cache = False
fp_lm.model.eval()
lm2 = copy.deepcopy(lm)
lm2.model = lm2.model.cpu()
lm2.model.config.use_cache = False
lm2.model.eval()
if dataset != args.calib_dataset and args.tta:
lm.model = lm.model.cpu()
lm2.model = lm2.model.to(lm2.device)
lm2.model.eval()
tta_loader = []
for i in range(nsamples):
tta_loader.append(testenc[:, (i * lm.seqlen) : ((i + 1) * lm.seqlen)])
with torch.enable_grad():#torch.enable_grad():
tta(
lm2,
args,
tta_loader, #dataloader_test
fp_lm,
logger
)
lm2.model = lm2.model.to(lm.device)
lm2.model.eval()
tmp_lm = lm2
else:
lm.model = lm.model.to(lm.device)
tmp_lm = lm
nlls = []
for i in tqdm(range(nsamples)):
batch = testenc[:, (i * tmp_lm.seqlen) : ((i + 1) * tmp_lm.seqlen)].to(tmp_lm.device) #1*2048
#c4 testenc:524288 = 2048*256
# x = batch[0] #2048
if "opt" in args.net.lower():
outputs = tmp_lm.model.model.decoder(batch)
elif "llama" in args.net.lower():
outputs = tmp_lm.model.model(batch) # 1*2048*4096
hidden_states = outputs[0] #1*2048*4096
logits = tmp_lm.model.lm_head(hidden_states) #1*2048*32000
shift_logits = logits[:, :-1, :] #1*2047*32000
shift_labels = testenc[:, (i * tmp_lm.seqlen) : ((i + 1) * tmp_lm.seqlen)][
:, 1:
].to(tmp_lm.model.lm_head.weight.device) #1*2047
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
)
neg_log_likelihood = loss.float() * tmp_lm.seqlen
nlls.append(neg_log_likelihood)
if i == args.limit:
break
ppl = torch.exp(torch.stack(nlls).sum() / (nsamples * tmp_lm.seqlen))
logger.info(f'{dataset} : {ppl.item()}')
tmp_lm.model.config.use_cache = use_cache
results[dataset] = ppl.item()
tmp_lm.model = tmp_lm.model.cpu()
tmp_lm.model.config.use_cache = False
tmp_lm.model.eval()
if args.tasks != "":
t_results = evaluator.simple_evaluate(
lm,
tasks=args.tasks,
num_fewshot=args.num_fewshot,
limit=None if args.limit == -1 else args.limit,
)
results.update(t_results)
logger.info(results)
pprint(results)
# for test of MMLU
if 'hendrycksTest' in args.tasks:
all_cors = []
all_cors_norm = []
subcat_cors = {subcat: [] for subcat_lists in subcategories.values() for subcat in subcat_lists}
cat_cors = {cat: [] for cat in categories}
cat_cors_norm = {cat: [] for cat in categories}
for key in t_results['results'].keys():
if not 'hendrycksTest' in key:
continue
subject = key.split('-')[-1]
cors = t_results['results'][key]['acc']
cors_norm = t_results['results'][key]['acc_norm']
subcats = subcategories[subject]
for subcat in subcats:
subcat_cors[subcat].append(cors)
for key in categories.keys():
if subcat in categories[key]:
cat_cors[key].append(cors)
cat_cors_norm[key].append(cors_norm)
all_cors.append(cors)
all_cors_norm.append(cors_norm)
for cat in cat_cors:
cat_acc = np.mean(cat_cors[cat])
logger.info("Average accuracy {:.4f} - {}".format(cat_acc, cat))
weighted_acc = np.mean(all_cors)
logger.info("Average accuracy: {:.4f}".format(weighted_acc))
return results