run_bertology.py 18.5 KB
Newer Older
1
#!/usr/bin/env python3
thomwolf's avatar
thomwolf committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
# Copyright 2018 CMU and The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Bertology: this script shows how you can explore the internals of the models in the library to:
    - compute the entropy of the head attentions
    - compute the importance of each head
    - prune (remove) the low importance head.
    Some parts of this script are adapted from the code of Michel et al. (http://arxiv.org/abs/1905.10650)
    which is available at https://github.com/pmichel31415/are-16-heads-really-better-than-1
"""
thomwolf's avatar
thomwolf committed
22
import os
23
24
import argparse
import logging
thomwolf's avatar
thomwolf committed
25
from datetime import timedelta, datetime
thomwolf's avatar
thomwolf committed
26
from tqdm import tqdm
27
28
29

import numpy as np

thomwolf's avatar
thomwolf committed
30
31
32
33
34
import torch
from torch.utils.data import DataLoader, SequentialSampler, TensorDataset, Subset
from torch.utils.data.distributed import DistributedSampler
from torch.nn import CrossEntropyLoss, MSELoss

35
from transformers import (WEIGHTS_NAME,
thomwolf's avatar
thomwolf committed
36
37
38
                                  BertConfig, BertForSequenceClassification, BertTokenizer,
                                  XLMConfig, XLMForSequenceClassification, XLMTokenizer,
                                  XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer)
39

thomwolf's avatar
thomwolf committed
40
from run_glue import set_seed, load_and_cache_examples, ALL_MODELS, MODEL_CLASSES
thomwolf's avatar
thomwolf committed
41

Adrian Bauer's avatar
Adrian Bauer committed
42
43
44
from transformers import glue_compute_metrics as compute_metrics
from transformers import glue_output_modes as output_modes
from transformers import glue_processors as processors
thomwolf's avatar
thomwolf committed
45

46
47
logger = logging.getLogger(__name__)

thomwolf's avatar
thomwolf committed
48
49

def entropy(p):
thomwolf's avatar
thomwolf committed
50
    """ Compute the entropy of a probability distribution """
thomwolf's avatar
thomwolf committed
51
52
53
54
    plogp = p * torch.log(p)
    plogp[p == 0] = 0
    return -plogp.sum(dim=-1)

thomwolf's avatar
thomwolf committed
55

thomwolf's avatar
thomwolf committed
56
def print_2d_tensor(tensor):
thomwolf's avatar
thomwolf committed
57
    """ Print a 2D tensor """
thomwolf's avatar
thomwolf committed
58
59
    logger.info("lv, h >\t" + "\t".join(f"{x + 1}" for x in range(len(tensor))))
    for row in range(len(tensor)):
thomwolf's avatar
thomwolf committed
60
61
62
63
        if tensor.dtype != torch.long:
            logger.info(f"layer {row + 1}:\t" + "\t".join(f"{x:.5f}" for x in tensor[row].cpu().data))
        else:
            logger.info(f"layer {row + 1}:\t" + "\t".join(f"{x:d}" for x in tensor[row].cpu().data))
thomwolf's avatar
thomwolf committed
64

thomwolf's avatar
thomwolf committed
65

thomwolf's avatar
thomwolf committed
66
def compute_heads_importance(args, model, eval_dataloader, compute_entropy=True, compute_importance=True, head_mask=None):
thomwolf's avatar
thomwolf committed
67
68
    """ This method shows how to compute:
        - head attention entropy
thomwolf's avatar
thomwolf committed
69
70
        - head importance scores according to http://arxiv.org/abs/1905.10650
    """
thomwolf's avatar
thomwolf committed
71
72
73
74
    # Prepare our tensors
    n_layers, n_heads = model.bert.config.num_hidden_layers, model.bert.config.num_attention_heads
    head_importance = torch.zeros(n_layers, n_heads).to(args.device)
    attn_entropy = torch.zeros(n_layers, n_heads).to(args.device)
thomwolf's avatar
thomwolf committed
75
76
77
78

    if head_mask is None:
        head_mask = torch.ones(n_layers, n_heads).to(args.device)
    head_mask.requires_grad_(requires_grad=True)
thomwolf's avatar
thomwolf committed
79
80
81
82
    preds = None
    labels = None
    tot_tokens = 0.0

thomwolf's avatar
thomwolf committed
83
84
85
86
    for step, batch in enumerate(tqdm(eval_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])):
        batch = tuple(t.to(args.device) for t in batch)
        input_ids, input_mask, segment_ids, label_ids = batch

thomwolf's avatar
thomwolf committed
87
        # Do a forward pass (not with torch.no_grad() since we need gradients for importance score - see below)
thomwolf's avatar
thomwolf committed
88
89
90
        outputs = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids, head_mask=head_mask)
        loss, logits, all_attentions = outputs[0], outputs[1], outputs[-1]  # Loss and logits are the first, attention the last
        loss.backward()  # Backpropagate to populate the gradients in the head mask
thomwolf's avatar
thomwolf committed
91
92
93
94
95
96
97

        if compute_entropy:
            for layer, attn in enumerate(all_attentions):
                masked_entropy = entropy(attn.detach()) * input_mask.float().unsqueeze(1)
                attn_entropy[layer] += masked_entropy.sum(-1).sum(0).detach()

        if compute_importance:
thomwolf's avatar
thomwolf committed
98
            head_importance += head_mask.grad.abs().detach()
thomwolf's avatar
thomwolf committed
99
100
101
102
103
104
105
106

        # Also store our logits/labels if we want to compute metrics afterwards
        if preds is None:
            preds = logits.detach().cpu().numpy()
            labels = label_ids.detach().cpu().numpy()
        else:
            preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
            labels = np.append(labels, label_ids.detach().cpu().numpy(), axis=0)
thomwolf's avatar
thomwolf committed
107
108
109
110
111
112

        tot_tokens += input_mask.float().detach().sum().data

    # Normalize
    attn_entropy /= tot_tokens
    head_importance /= tot_tokens
thomwolf's avatar
thomwolf committed
113
114
115
116
117
118
119
    # Layerwise importance normalization
    if not args.dont_normalize_importance_by_layer:
        exponent = 2
        norm_by_layer = torch.pow(torch.pow(head_importance, exponent).sum(-1), 1/exponent)
        head_importance /= norm_by_layer.unsqueeze(-1) + 1e-20

    if not args.dont_normalize_global_importance:
thomwolf's avatar
thomwolf committed
120
121
        head_importance = (head_importance - head_importance.min()) / (head_importance.max() - head_importance.min())

thomwolf's avatar
thomwolf committed
122
123
124
125
126
127
128
129
130
131
132
133
134
135
    # Print/save matrices
    np.save(os.path.join(args.output_dir, 'attn_entropy.npy'), attn_entropy.detach().cpu().numpy())
    np.save(os.path.join(args.output_dir, 'head_importance.npy'), head_importance.detach().cpu().numpy())

    logger.info("Attention entropies")
    print_2d_tensor(attn_entropy)
    logger.info("Head importance scores")
    print_2d_tensor(head_importance)
    logger.info("Head ranked by importance scores")
    head_ranks = torch.zeros(head_importance.numel(), dtype=torch.long, device=args.device)
    head_ranks[head_importance.view(-1).sort(descending=True)[1]] = torch.arange(head_importance.numel(), device=args.device)
    head_ranks = head_ranks.view_as(head_importance)
    print_2d_tensor(head_ranks)

thomwolf's avatar
thomwolf committed
136
    return attn_entropy, head_importance, preds, labels
thomwolf's avatar
thomwolf committed
137

thomwolf's avatar
thomwolf committed
138

thomwolf's avatar
thomwolf committed
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
def mask_heads(args, model, eval_dataloader):
    """ This method shows how to mask head (set some heads to zero), to test the effect on the network,
        based on the head importance scores, as described in Michel et al. (http://arxiv.org/abs/1905.10650)
    """
    _, head_importance, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False)
    preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
    original_score = compute_metrics(args.task_name, preds, labels)[args.metric_name]
    logger.info("Pruning: original score: %f, threshold: %f", original_score, original_score * args.masking_threshold)

    new_head_mask = torch.ones_like(head_importance)
    num_to_mask = max(1, int(new_head_mask.numel() * args.masking_amount))

    current_score = original_score
    while current_score >= original_score * args.masking_threshold:
        head_mask = new_head_mask.clone() # save current head mask
        # heads from least important to most - keep only not-masked heads
        head_importance[head_mask == 0.0] = float('Inf')
        current_heads_to_mask = head_importance.view(-1).sort()[1]

        if len(current_heads_to_mask) <= num_to_mask:
            break

        # mask heads
        current_heads_to_mask = current_heads_to_mask[:num_to_mask]
        logger.info("Heads to mask: %s", str(current_heads_to_mask.tolist()))
        new_head_mask = new_head_mask.view(-1)
        new_head_mask[current_heads_to_mask] = 0.0
        new_head_mask = new_head_mask.view_as(head_mask)
        print_2d_tensor(new_head_mask)

        # Compute metric and head importance again
        _, head_importance, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False, head_mask=new_head_mask)
        preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
        current_score = compute_metrics(args.task_name, preds, labels)[args.metric_name]
        logger.info("Masking: current score: %f, remaning heads %d (%.1f percents)", current_score, new_head_mask.sum(), new_head_mask.sum()/new_head_mask.numel() * 100)

    logger.info("Final head mask")
    print_2d_tensor(head_mask)
    np.save(os.path.join(args.output_dir, 'head_mask.npy'), head_mask.detach().cpu().numpy())

    return head_mask


def prune_heads(args, model, eval_dataloader, head_mask):
    """ This method shows how to prune head (remove heads weights) based on
        the head importance scores as described in Michel et al. (http://arxiv.org/abs/1905.10650)
    """
    # Try pruning and test time speedup
    # Pruning is like masking but we actually remove the masked weights
    before_time = datetime.now()
    _, _, preds, labels = compute_heads_importance(args, model, eval_dataloader,
                                                   compute_entropy=False, compute_importance=False, head_mask=head_mask)
    preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
    score_masking = compute_metrics(args.task_name, preds, labels)[args.metric_name]
    original_time = datetime.now() - before_time

    original_num_params = sum(p.numel() for p in model.parameters())
    heads_to_prune = dict((layer, (1 - head_mask[layer].long()).nonzero().tolist()) for layer in range(len(head_mask)))
    assert sum(len(h) for h in heads_to_prune.values()) == (1 - head_mask.long()).sum().item()
    model.prune_heads(heads_to_prune)
    pruned_num_params = sum(p.numel() for p in model.parameters())

    before_time = datetime.now()
    _, _, preds, labels = compute_heads_importance(args, model, eval_dataloader,
                                                    compute_entropy=False, compute_importance=False, head_mask=None)
    preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
    score_pruning = compute_metrics(args.task_name, preds, labels)[args.metric_name]
    new_time = datetime.now() - before_time

    logger.info("Pruning: original num of params: %.2e, after pruning %.2e (%.1f percents)", original_num_params, pruned_num_params, pruned_num_params/original_num_params * 100)
    logger.info("Pruning: score with masking: %f score with pruning: %f", score_masking, score_pruning)
    logger.info("Pruning: speed ratio (new timing / original timing): %f percents", original_time/new_time * 100)


def main():
214
    parser = argparse.ArgumentParser()
tuvuumass's avatar
tuvuumass committed
215
    ## Required parameters
thomwolf's avatar
thomwolf committed
216
217
    parser.add_argument("--data_dir", default=None, type=str, required=True,
                        help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
tuvuumass's avatar
tuvuumass committed
218
219
220
    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(
                            ALL_MODELS))
thomwolf's avatar
thomwolf committed
221
222
223
224
225
226
227
    parser.add_argument("--task_name", default=None, type=str, required=True,
                        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
    parser.add_argument("--output_dir", default=None, type=str, required=True,
                        help="The output directory where the model predictions and checkpoints will be written.")

    ## Other parameters
    parser.add_argument("--config_name", default="", type=str,
tuvuumass's avatar
tuvuumass committed
228
                        help="Pretrained config name or path if not the same as model_name_or_path")
thomwolf's avatar
thomwolf committed
229
    parser.add_argument("--tokenizer_name", default="", type=str,
tuvuumass's avatar
tuvuumass committed
230
                        help="Pretrained tokenizer name or path if not the same as model_name_or_path")
thomwolf's avatar
thomwolf committed
231
232
233
234
235
236
    parser.add_argument("--cache_dir", default="", type=str,
                        help="Where do you want to store the pre-trained models downloaded from s3")
    parser.add_argument("--data_subset", type=int, default=-1,
                        help="If > 0: limit the data to a subset of data_subset instances.")
    parser.add_argument("--overwrite_output_dir", action='store_true',
                        help="Whether to overwrite data in output directory")
Adrian Bauer's avatar
Adrian Bauer committed
237
238
    parser.add_argument('--overwrite_cache', action='store_true',
                        help="Overwrite the cached training and evaluation sets")
thomwolf's avatar
thomwolf committed
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256

    parser.add_argument("--dont_normalize_importance_by_layer", action='store_true',
                        help="Don't normalize importance score by layers")
    parser.add_argument("--dont_normalize_global_importance", action='store_true',
                        help="Don't normalize all importance scores between 0 and 1")

    parser.add_argument("--try_masking", action='store_true',
                        help="Whether to try to mask head until a threshold of accuracy.")
    parser.add_argument("--masking_threshold", default=0.9, type=float,
                        help="masking threshold in term of metrics (stop masking when metric < threshold * original metric value).")
    parser.add_argument("--masking_amount", default=0.1, type=float,
                        help="Amount to heads to masking at each masking step.")
    parser.add_argument("--metric_name", default="acc", type=str,
                        help="Metric to use for head masking.")

    parser.add_argument("--max_seq_length", default=128, type=int,
                        help="The maximum total input sequence length after WordPiece tokenization. \n"
                             "Sequences longer than this will be truncated, sequences shorter padded.")
thomwolf's avatar
thomwolf committed
257
258
    parser.add_argument("--batch_size", default=1, type=int, help="Batch size.")

259
    parser.add_argument("--seed", type=int, default=42)
thomwolf's avatar
thomwolf committed
260
261
    parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
    parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available")
thomwolf's avatar
thomwolf committed
262
263
    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
264
265
    args = parser.parse_args()

thomwolf's avatar
thomwolf committed
266
267
268
269
270
271
272
    if args.server_ip and args.server_port:
        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
        import ptvsd
        print("Waiting for debugger attach")
        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
        ptvsd.wait_for_attach()

thomwolf's avatar
thomwolf committed
273
    # Setup devices and distributed training
thomwolf's avatar
thomwolf committed
274
275
    if args.local_rank == -1 or args.no_cuda:
        args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
thomwolf's avatar
thomwolf committed
276
        args.n_gpu = torch.cuda.device_count()
thomwolf's avatar
thomwolf committed
277
278
279
    else:
        torch.cuda.set_device(args.local_rank)
        args.device = torch.device("cuda", args.local_rank)
thomwolf's avatar
thomwolf committed
280
        args.n_gpu = 1
thomwolf's avatar
thomwolf committed
281
        torch.distributed.init_process_group(backend='nccl')  # Initializes the distributed backend
thomwolf's avatar
thomwolf committed
282

thomwolf's avatar
thomwolf committed
283
    # Setup logging
thomwolf's avatar
thomwolf committed
284
    logging.basicConfig(level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
thomwolf's avatar
thomwolf committed
285
    logger.info("device: {} n_gpu: {}, distributed: {}".format(args.device, args.n_gpu, bool(args.local_rank != -1)))
286

thomwolf's avatar
thomwolf committed
287
    # Set seeds
thomwolf's avatar
thomwolf committed
288
    set_seed(args)
thomwolf's avatar
thomwolf committed
289
290

    # Prepare GLUE task
thomwolf's avatar
thomwolf committed
291
292
293
294
295
    args.task_name = args.task_name.lower()
    if args.task_name not in processors:
        raise ValueError("Task not found: %s" % (args.task_name))
    processor = processors[args.task_name]()
    args.output_mode = output_modes[args.task_name]
thomwolf's avatar
thomwolf committed
296
    label_list = processor.get_labels()
thomwolf's avatar
thomwolf committed
297
    num_labels = len(label_list)
thomwolf's avatar
thomwolf committed
298

thomwolf's avatar
thomwolf committed
299
    # Load pretrained model and tokenizer
thomwolf's avatar
thomwolf committed
300
    if args.local_rank not in [-1, 0]:
thomwolf's avatar
thomwolf committed
301
302
303
304
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab

    args.model_type = ""
    for key in MODEL_CLASSES:
tuvuumass's avatar
tuvuumass committed
305
        if key in args.model_name_or_path.lower():
thomwolf's avatar
thomwolf committed
306
307
308
            args.model_type = key  # take the first match in model types
            break
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
tuvuumass's avatar
tuvuumass committed
309
    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
thomwolf's avatar
thomwolf committed
310
311
312
313
314
315
316
317
318
319
                                          num_labels=num_labels,
                                          finetuning_task=args.task_name,
                                          output_attentions=True,
                                          cache_dir=args.cache_dir if args.cache_dir else None)
    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
                                                cache_dir=args.cache_dir if args.cache_dir else None)
    model = model_class.from_pretrained(args.model_name_or_path,
                                        from_tf=bool('.ckpt' in args.model_name_or_path),
                                        config=config,
                                        cache_dir=args.cache_dir if args.cache_dir else None)
thomwolf's avatar
thomwolf committed
320

thomwolf's avatar
thomwolf committed
321
    if args.local_rank == 0:
thomwolf's avatar
thomwolf committed
322
323
324
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab

    # Distributed and parallel training
thomwolf's avatar
thomwolf committed
325
    model.to(args.device)
thomwolf's avatar
thomwolf committed
326
    if args.local_rank != -1:
thomwolf's avatar
thomwolf committed
327
328
329
330
331
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
                                                          output_device=args.local_rank,
                                                          find_unused_parameters=True)
    elif args.n_gpu > 1:
        model = torch.nn.DataParallel(model)
332

thomwolf's avatar
thomwolf committed
333
334
335
    # Print/save training arguments
    torch.save(args, os.path.join(args.output_dir, 'run_args.bin'))
    logger.info("Training/evaluation parameters %s", args)
thomwolf's avatar
thomwolf committed
336

thomwolf's avatar
thomwolf committed
337
338
    # Prepare dataset for the GLUE task
    eval_data = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=True)
thomwolf's avatar
thomwolf committed
339
    if args.data_subset > 0:
thomwolf's avatar
thomwolf committed
340
        eval_data = Subset(eval_data, list(range(min(args.data_subset, len(eval_data)))))
thomwolf's avatar
thomwolf committed
341
342
343
344
345
    eval_sampler = SequentialSampler(eval_data) if args.local_rank == -1 else DistributedSampler(eval_data)
    eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.batch_size)


    # Compute head entropy and importance score
thomwolf's avatar
thomwolf committed
346
    compute_heads_importance(args, model, eval_dataloader)
thomwolf's avatar
thomwolf committed
347

thomwolf's avatar
thomwolf committed
348

thomwolf's avatar
thomwolf committed
349
350
    # Try head masking (set heads to zero until the score goes under a threshole)
    # and head pruning (remove masked heads and see the effect on the network)
thomwolf's avatar
thomwolf committed
351
    if args.try_masking and args.masking_threshold > 0.0 and args.masking_threshold < 1.0:
thomwolf's avatar
thomwolf committed
352
353
        head_mask = mask_heads(args, model, eval_dataloader)
        prune_heads(args, model, eval_dataloader, head_mask)
thomwolf's avatar
thomwolf committed
354

thomwolf's avatar
thomwolf committed
355
356

if __name__ == '__main__':
thomwolf's avatar
thomwolf committed
357
    main()