README.md 7.96 KB
Newer Older
Sylvain Gugger's avatar
Sylvain Gugger committed
1
2
<!---
Copyright 2020 The HuggingFace Team. All rights reserved.
3

Sylvain Gugger's avatar
Sylvain Gugger committed
4
5
6
7
8
9
10
11
12
13
14
15
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.
-->
16
17
18

## SQuAD

Sylvain Gugger's avatar
Sylvain Gugger committed
19
20
21
22
23
24
25
Based on the script [`run_qa.py`](https://github.com/huggingface/transformers/blob/master/examples/question-answering/run_qa.py).

**Note:** This script only works with models that have a fast tokenizer (backed by the 馃 Tokenizers library) as it
uses special features of those tokenizers. You can check if your favorite model has a fast tokenizer in
[this table](https://huggingface.co/transformers/index.html#bigtable), if it doesn't you can still use the old version
of the script.

26
The old version of this script can be found [here](https://github.com/huggingface/transformers/tree/master/examples/legacy/question-answering).
27
28
29
#### Fine-tuning BERT on SQuAD1.0

This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large)
Sylvain Gugger's avatar
Sylvain Gugger committed
30
on a single tesla V100 16GB.
31
32

```bash
Sylvain Gugger's avatar
Sylvain Gugger committed
33
python run_qa.py \
34
  --model_name_or_path bert-base-uncased \
Sylvain Gugger's avatar
Sylvain Gugger committed
35
  --dataset_name squad \
36
37
  --do_train \
  --do_eval \
Sylvain Gugger's avatar
Sylvain Gugger committed
38
  --per_device_train_batch_size 12 \
39
  --learning_rate 3e-5 \
Sylvain Gugger's avatar
Sylvain Gugger committed
40
  --num_train_epochs 2 \
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
  --max_seq_length 384 \
  --doc_stride 128 \
  --output_dir /tmp/debug_squad/
```

Training with the previously defined hyper-parameters yields the following results:

```bash
f1 = 88.52
exact_match = 81.22
```

#### Distributed training


Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD1.1:

```bash
59
python -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answering/run_squad.py \
60
    --model_name_or_path bert-large-uncased-whole-word-masking \
Sylvain Gugger's avatar
Sylvain Gugger committed
61
    --dataset_name squad \
62
63
64
65
66
67
68
    --do_train \
    --do_eval \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./examples/models/wwm_uncased_finetuned_squad/ \
Sylvain Gugger's avatar
Sylvain Gugger committed
69
70
    --per_device_eval_batch_size=3   \
    --per_device_train_batch_size=3   \
71
72
73
74
75
76
77
78
79
80
```

Training with the previously defined hyper-parameters yields the following results:

```bash
f1 = 93.15
exact_match = 86.91
```

This fine-tuned model is available as a checkpoint under the reference
81
[`bert-large-uncased-whole-word-masking-finetuned-squad`](https://huggingface.co/bert-large-uncased-whole-word-masking-finetuned-squad).
82

Sylvain Gugger's avatar
Sylvain Gugger committed
83
#### Fine-tuning XLNet with beam search on SQuAD
84

Sylvain Gugger's avatar
Sylvain Gugger committed
85
This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset.
86
87
88
89

##### Command for SQuAD1.0:

```bash
Sylvain Gugger's avatar
Sylvain Gugger committed
90
python run_qa_beam_search.py \
91
    --model_name_or_path xlnet-large-cased \
Sylvain Gugger's avatar
Sylvain Gugger committed
92
    --dataset_name squad \
93
94
95
96
97
98
99
    --do_train \
    --do_eval \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./wwm_cased_finetuned_squad/ \
Sylvain Gugger's avatar
Sylvain Gugger committed
100
101
    --per_device_eval_batch_size=4  \
    --per_device_train_batch_size=4   \
102
103
104
105
106
107
108
109
    --save_steps 5000
```

##### Command for SQuAD2.0:

```bash
export SQUAD_DIR=/path/to/SQUAD

Sylvain Gugger's avatar
Sylvain Gugger committed
110
python run_qa_beam_search.py \
111
    --model_name_or_path xlnet-large-cased \
Sylvain Gugger's avatar
Sylvain Gugger committed
112
    --dataset_name squad_v2 \
113
114
115
116
117
118
119
120
    --do_train \
    --do_eval \
    --version_2_with_negative \
    --learning_rate 3e-5 \
    --num_train_epochs 4 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./wwm_cased_finetuned_squad/ \
Sylvain Gugger's avatar
Sylvain Gugger committed
121
122
    --per_device_eval_batch_size=2  \
    --per_device_train_batch_size=2   \
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
    --save_steps 5000
```

Larger batch size may improve the performance while costing more memory.

##### Results for SQuAD1.0 with the previously defined hyper-parameters:

```python
{
"exact": 85.45884578997162,
"f1": 92.5974600601065,
"total": 10570,
"HasAns_exact": 85.45884578997162,
"HasAns_f1": 92.59746006010651,
"HasAns_total": 10570
}
```

##### Results for SQuAD2.0 with the previously defined hyper-parameters:

```python
{
"exact": 80.4177545691906,
"f1": 84.07154997729623,
"total": 11873,
"HasAns_exact": 76.73751686909581,
"HasAns_f1": 84.05558584352873,
"HasAns_total": 5928,
"NoAns_exact": 84.0874684608915,
"NoAns_f1": 84.0874684608915,
"NoAns_total": 5945
}
```

157
158
159
#### Fine-tuning BERT on SQuAD1.0 with relative position embeddings

The following examples show how to fine-tune BERT models with different relative position embeddings. The BERT model 
Sylvain Gugger's avatar
Sylvain Gugger committed
160
`bert-base-uncased` was pretrained with default absolute position embeddings. We provide the following pretrained 
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
models which were pre-trained on the same training data (BooksCorpus and English Wikipedia) as in the BERT model 
training, but with different relative position embeddings. 

* `zhiheng-huang/bert-base-uncased-embedding-relative-key`, trained from scratch with relative embedding proposed by 
Shaw et al., [Self-Attention with Relative Position Representations](https://arxiv.org/abs/1803.02155)
* `zhiheng-huang/bert-base-uncased-embedding-relative-key-query`, trained from scratch with relative embedding method 4 
in Huang et al. [Improve Transformer Models with Better Relative Position Embeddings](https://arxiv.org/abs/2009.13658)
* `zhiheng-huang/bert-large-uncased-whole-word-masking-embedding-relative-key-query`, fine-tuned from model 
`bert-large-uncased-whole-word-masking` with 3 additional epochs with relative embedding method 4 in Huang et al. 
[Improve Transformer Models with Better Relative Position Embeddings](https://arxiv.org/abs/2009.13658)


##### Base models fine-tuning

```bash
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
python -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answering/run_squad.py \
    --model_name_or_path zhiheng-huang/bert-base-uncased-embedding-relative-key-query \
Sylvain Gugger's avatar
Sylvain Gugger committed
179
    --dataset_name squad \
180
181
182
183
184
185
    --do_train \
    --do_eval \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 512 \
    --doc_stride 128 \
Sylvain Gugger's avatar
Sylvain Gugger committed
186
187
188
    --output_dir relative_squad \
    --per_device_eval_batch_size=60 \
    --per_device_train_batch_size=6
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
```
Training with the above command leads to the following results. It boosts the BERT default from f1 score of 88.52 to 90.54.

```bash
'exact': 83.6802270577105, 'f1': 90.54772098174814
```

The change of `max_seq_length` from 512 to 384 in the above command leads to the f1 score of 90.34. Replacing the above 
model `zhiheng-huang/bert-base-uncased-embedding-relative-key-query` with 
`zhiheng-huang/bert-base-uncased-embedding-relative-key` leads to the f1 score of 89.51. The changing of 8 gpus to one 
gpu training leads to the f1 score of 90.71.

##### Large models fine-tuning

```bash
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
python -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answering/run_squad.py \
    --model_name_or_path zhiheng-huang/bert-large-uncased-whole-word-masking-embedding-relative-key-query \
Sylvain Gugger's avatar
Sylvain Gugger committed
207
    --dataset_name squad \
208
209
210
211
212
213
    --do_train \
    --do_eval \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 512 \
    --doc_stride 128 \
Sylvain Gugger's avatar
Sylvain Gugger committed
214
    --output_dir relative_squad \
215
216
217
218
219
220
221
    --per_gpu_eval_batch_size=6 \
    --per_gpu_train_batch_size=2 \
    --gradient_accumulation_steps 3
```
Training with the above command leads to the f1 score of 93.52, which is slightly better than the f1 score of 93.15 for 
`bert-large-uncased-whole-word-masking`.

222
223
224
225
226
227
## SQuAD with the Tensorflow Trainer

```bash
python run_tf_squad.py \
    --model_name_or_path bert-base-uncased \
    --output_dir model \
228
    --max_seq_length 384 \
229
230
231
232
    --num_train_epochs 2 \
    --per_gpu_train_batch_size 8 \
    --per_gpu_eval_batch_size 16 \
    --do_train \
233
    --logging_dir logs \    
234
235
    --logging_steps 10 \
    --learning_rate 3e-5 \
236
    --doc_stride 128    
237
238
```

239
For the moment evaluation is not available in the Tensorflow Trainer only the training.