yoshitomo-matsubara commited on
Commit
fd2e5db
1 Parent(s): b1c53f2

initial commit

Browse files
README.md ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: en
3
+ tags:
4
+ - bert
5
+ - rte
6
+ - glue
7
+ - kd
8
+ - torchdistill
9
+ license: apache-2.0
10
+ datasets:
11
+ - rte
12
+ metrics:
13
+ - accuracy
14
+ ---
15
+
16
+ `bert-base-uncased` fine-tuned on RTE dataset, using fine-tuned `bert-large-uncased` as a teacher model, [***torchdistill***](https://github.com/yoshitomo-matsubara/torchdistill) and [Google Colab](https://colab.research.google.com/github/yoshitomo-matsubara/torchdistill/blob/master/demo/glue_kd_and_submission.ipynb) for knowledge distillation.
17
+ The training configuration (including hyperparameters) is available [here](https://github.com/yoshitomo-matsubara/torchdistill/blob/main/configs/sample/glue/rte/kd/bert_base_uncased_from_bert_large_uncased.yaml).
18
+ I submitted prediction files to [the GLUE leaderboard](https://gluebenchmark.com/leaderboard), and the overall GLUE score was **78.9**.
config.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "bert-base-uncased",
3
+ "architectures": [
4
+ "BertForSequenceClassification"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "finetuning_task": "rte",
8
+ "gradient_checkpointing": false,
9
+ "hidden_act": "gelu",
10
+ "hidden_dropout_prob": 0.1,
11
+ "hidden_size": 768,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 3072,
14
+ "layer_norm_eps": 1e-12,
15
+ "max_position_embeddings": 512,
16
+ "model_type": "bert",
17
+ "num_attention_heads": 12,
18
+ "num_hidden_layers": 12,
19
+ "pad_token_id": 0,
20
+ "position_embedding_type": "absolute",
21
+ "problem_type": "single_label_classification",
22
+ "transformers_version": "4.6.1",
23
+ "type_vocab_size": 2,
24
+ "use_cache": true,
25
+ "vocab_size": 30522
26
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a8d70e71e4e304fc905099faa9090838b449b426749817ec6ac6ad141f5bf340
3
+ size 438024457
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
1
+ {"unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]"}
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
1
+ {"do_lower_case": true, "unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]", "tokenize_chinese_chars": true, "strip_accents": null, "do_lower": true, "model_max_length": 512, "special_tokens_map_file": null, "name_or_path": "bert-base-uncased"}
training.log ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2021-05-31 17:50:29,582 INFO __main__ Namespace(adjust_lr=False, config='torchdistill/configs/sample/glue/rte/kd/bert_base_uncased_from_bert_large_uncased.yaml', log='log/glue/rte/kd/bert_base_uncased_from_bert_large_uncased.txt', private_output='leaderboard/glue/kd/bert_base_uncased_from_bert_large_uncased/', seed=None, student_only=False, task_name='rte', test_only=False, world_size=1)
2
+ 2021-05-31 17:50:29,613 INFO __main__ Distributed environment: NO
3
+ Num processes: 1
4
+ Process index: 0
5
+ Local process index: 0
6
+ Device: cuda
7
+ Use FP16 precision: True
8
+
9
+ 2021-05-31 17:50:39,943 WARNING datasets.builder Reusing dataset glue (/root/.cache/huggingface/datasets/glue/rte/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad)
10
+ 2021-05-31 17:50:41,587 INFO __main__ Start training
11
+ 2021-05-31 17:50:41,587 INFO torchdistill.models.util [teacher model]
12
+ 2021-05-31 17:50:41,587 INFO torchdistill.models.util Using the original teacher model
13
+ 2021-05-31 17:50:41,588 INFO torchdistill.models.util [student model]
14
+ 2021-05-31 17:50:41,588 INFO torchdistill.models.util Using the original student model
15
+ 2021-05-31 17:50:41,588 INFO torchdistill.core.distillation Loss = 1.0 * OrgLoss
16
+ 2021-05-31 17:50:41,588 INFO torchdistill.core.distillation Freezing the whole teacher model
17
+ 2021-05-31 17:50:45,473 INFO torchdistill.misc.log Epoch: [0] [ 0/78] eta: 0:01:02 lr: 9.957264957264958e-05 sample/s: 5.016064534395792 loss: 0.0097 (0.0097) time: 0.8039 data: 0.0065 max mem: 3804
18
+ 2021-05-31 17:51:25,664 INFO torchdistill.misc.log Epoch: [0] [50/78] eta: 0:00:22 lr: 7.820512820512821e-05 sample/s: 5.000885583822104 loss: 0.0065 (0.0081) time: 0.8035 data: 0.0049 max mem: 5115
19
+ 2021-05-31 17:51:47,044 INFO torchdistill.misc.log Epoch: [0] Total time: 0:01:02
20
+ 2021-05-31 17:51:48,268 INFO /usr/local/lib/python3.7/dist-packages/datasets/metric.py Removing /root/.cache/huggingface/metrics/glue/rte/default_experiment-1-0.arrow
21
+ 2021-05-31 17:51:48,268 INFO __main__ Validation: accuracy = 0.6570397111913358
22
+ 2021-05-31 17:51:48,268 INFO __main__ Updating ckpt at ./resource/ckpt/glue/rte/kd/rte-bert-base-uncased_from_bert-large-uncased
23
+ 2021-05-31 17:51:50,012 INFO torchdistill.misc.log Epoch: [1] [ 0/78] eta: 0:01:02 lr: 6.623931623931624e-05 sample/s: 5.007037577218317 loss: 0.0042 (0.0042) time: 0.8055 data: 0.0066 max mem: 5115
24
+ 2021-05-31 17:52:30,204 INFO torchdistill.misc.log Epoch: [1] [50/78] eta: 0:00:22 lr: 4.4871794871794874e-05 sample/s: 4.994946466605415 loss: 0.0033 (0.0032) time: 0.8037 data: 0.0050 max mem: 5115
25
+ 2021-05-31 17:52:51,720 INFO torchdistill.misc.log Epoch: [1] Total time: 0:01:02
26
+ 2021-05-31 17:52:52,943 INFO /usr/local/lib/python3.7/dist-packages/datasets/metric.py Removing /root/.cache/huggingface/metrics/glue/rte/default_experiment-1-0.arrow
27
+ 2021-05-31 17:52:52,943 INFO __main__ Validation: accuracy = 0.6859205776173285
28
+ 2021-05-31 17:52:52,943 INFO __main__ Updating ckpt at ./resource/ckpt/glue/rte/kd/rte-bert-base-uncased_from_bert-large-uncased
29
+ 2021-05-31 17:52:54,985 INFO torchdistill.misc.log Epoch: [2] [ 0/78] eta: 0:01:03 lr: 3.290598290598291e-05 sample/s: 5.004806067380402 loss: 0.0014 (0.0014) time: 0.8145 data: 0.0153 max mem: 5115
30
+ 2021-05-31 17:53:35,181 INFO torchdistill.misc.log Epoch: [2] [50/78] eta: 0:00:22 lr: 1.153846153846154e-05 sample/s: 5.00735140278351 loss: 0.0012 (0.0013) time: 0.8039 data: 0.0051 max mem: 5115
31
+ 2021-05-31 17:53:56,746 INFO torchdistill.misc.log Epoch: [2] Total time: 0:01:02
32
+ 2021-05-31 17:53:57,970 INFO /usr/local/lib/python3.7/dist-packages/datasets/metric.py Removing /root/.cache/huggingface/metrics/glue/rte/default_experiment-1-0.arrow
33
+ 2021-05-31 17:53:57,970 INFO __main__ Validation: accuracy = 0.6823104693140795
34
+ 2021-05-31 17:53:58,006 INFO __main__ [Teacher: bert-large-uncased]
35
+ 2021-05-31 17:54:01,530 INFO /usr/local/lib/python3.7/dist-packages/datasets/metric.py Removing /root/.cache/huggingface/metrics/glue/rte/default_experiment-1-0.arrow
36
+ 2021-05-31 17:54:01,530 INFO __main__ Test: accuracy = 0.740072202166065
37
+ 2021-05-31 17:54:04,735 INFO __main__ [Student: bert-base-uncased]
38
+ 2021-05-31 17:54:05,975 INFO /usr/local/lib/python3.7/dist-packages/datasets/metric.py Removing /root/.cache/huggingface/metrics/glue/rte/default_experiment-1-0.arrow
39
+ 2021-05-31 17:54:05,976 INFO __main__ Test: accuracy = 0.6859205776173285
40
+ 2021-05-31 17:54:05,976 INFO __main__ Start prediction for private dataset(s)
41
+ 2021-05-31 17:54:05,977 INFO __main__ rte/test: 3000 samples
vocab.txt ADDED
The diff for this file is too large to render. See raw diff