Hugging Face's logo

BluebrainAI
/
duo-predict-gpt2-medium-wikitext

Feature Extraction
Transformers
Safetensors
duo-predict-gpt2
Generated from Trainer
custom_code
Model card Files
xet
Community
shivanandmn commited on
Commit
8f4638b
·
verified ·
1 Parent(s): 3c358cd

Model save

Browse files
Files changed (3) hide show
  1. README.md +39 -39
  2. model.safetensors +1 -1
  3. modeling_duo_predict_gpt2.py +23 -15
README.md CHANGED
@@ -17,9 +17,9 @@ should probably proofread and complete it, then remove this comment. -->
17
 
18
  This model is a fine-tuned version of [](https://huggingface.co/) on an unknown dataset.
19
  It achieves the following results on the evaluation set:
20
- - Loss: 1.5713
21
  - Accuracy: 0.0073
22
- - Perplexity: 4.8128
23
  - Bleu: 1.0
24
 
25
  ## Model description
@@ -50,43 +50,43 @@ The following hyperparameters were used during training:
50
 
51
  ### Training results
52
 
53
- | Training Loss | Epoch | Step | Validation Loss | Accuracy | Perplexity | Bleu |
54
- |:-------------:|:------:|:-----:|:---------------:|:--------:|:----------:|:------:|
55
- | 7.6316 | 0.1403 | 500 | 3.7041 | 0.0073 | 40.6115 | 1.0 |
56
- | 6.4879 | 0.2807 | 1000 | 3.1196 | 0.0073 | 22.6384 | 0.9995 |
57
- | 5.3189 | 0.4210 | 1500 | 2.5976 | 0.0073 | 13.4321 | 1.0 |
58
- | 4.7557 | 0.5613 | 2000 | 2.3369 | 0.0073 | 10.3487 | 1.0 |
59
- | 4.351 | 0.7017 | 2500 | 2.1234 | 0.0073 | 8.3592 | 1.0 |
60
- | 4.0701 | 0.8420 | 3000 | 2.0021 | 0.0073 | 7.4045 | 1.0 |
61
- | 3.9112 | 0.9823 | 3500 | 1.9280 | 0.0073 | 6.8756 | 1.0 |
62
- | 3.7775 | 1.1226 | 4000 | 1.8769 | 0.0073 | 6.5331 | 1.0 |
63
- | 3.703 | 1.2630 | 4500 | 1.8369 | 0.0073 | 6.2768 | 1.0 |
64
- | 3.6443 | 1.4033 | 5000 | 1.8061 | 0.0073 | 6.0866 | 1.0 |
65
- | 3.5659 | 1.5436 | 5500 | 1.7789 | 0.0073 | 5.9232 | 1.0 |
66
- | 3.5303 | 1.6840 | 6000 | 1.7552 | 0.0073 | 5.7848 | 1.0 |
67
- | 3.4926 | 1.8243 | 6500 | 1.7348 | 0.0073 | 5.6676 | 1.0 |
68
- | 3.464 | 1.9646 | 7000 | 1.7167 | 0.0073 | 5.5660 | 1.0 |
69
- | 3.3432 | 2.1050 | 7500 | 1.7016 | 0.0073 | 5.4826 | 1.0 |
70
- | 3.3215 | 2.2453 | 8000 | 1.6883 | 0.0073 | 5.4104 | 1.0 |
71
- | 3.3122 | 2.3856 | 8500 | 1.6768 | 0.0073 | 5.3483 | 1.0 |
72
- | 3.2836 | 2.5260 | 9000 | 1.6651 | 0.0073 | 5.2860 | 1.0 |
73
- | 3.2582 | 2.6663 | 9500 | 1.6541 | 0.0073 | 5.2281 | 1.0 |
74
- | 3.2387 | 2.8066 | 10000 | 1.6434 | 0.0073 | 5.1726 | 1.0 |
75
- | 3.223 | 2.9470 | 10500 | 1.6338 | 0.0073 | 5.1232 | 1.0 |
76
- | 3.126 | 3.0873 | 11000 | 1.6289 | 0.0073 | 5.0984 | 1.0 |
77
- | 3.1149 | 3.2276 | 11500 | 1.6212 | 0.0073 | 5.0590 | 1.0 |
78
- | 3.1048 | 3.3679 | 12000 | 1.6149 | 0.0073 | 5.0276 | 1.0 |
79
- | 3.0966 | 3.5083 | 12500 | 1.6074 | 0.0073 | 4.9897 | 1.0 |
80
- | 3.0904 | 3.6486 | 13000 | 1.6013 | 0.0073 | 4.9597 | 1.0 |
81
- | 3.0979 | 3.7889 | 13500 | 1.5959 | 0.0073 | 4.9326 | 1.0 |
82
- | 3.0772 | 3.9293 | 14000 | 1.5907 | 0.0073 | 4.9074 | 1.0 |
83
- | 2.9795 | 4.0696 | 14500 | 1.5887 | 0.0073 | 4.8975 | 1.0 |
84
- | 2.9807 | 4.2099 | 15000 | 1.5847 | 0.0073 | 4.8779 | 1.0 |
85
- | 2.976 | 4.3503 | 15500 | 1.5817 | 0.0073 | 4.8634 | 1.0 |
86
- | 2.9795 | 4.4906 | 16000 | 1.5778 | 0.0073 | 4.8444 | 1.0 |
87
- | 2.9594 | 4.6309 | 16500 | 1.5754 | 0.0073 | 4.8328 | 1.0 |
88
- | 2.9698 | 4.7713 | 17000 | 1.5731 | 0.0073 | 4.8217 | 1.0 |
89
- | 2.9604 | 4.9116 | 17500 | 1.5713 | 0.0073 | 4.8128 | 1.0 |
90
 
91
 
92
  ### Framework versions
 
17
 
18
  This model is a fine-tuned version of [](https://huggingface.co/) on an unknown dataset.
19
  It achieves the following results on the evaluation set:
20
+ - Loss: 2.2546
21
  - Accuracy: 0.0073
22
+ - Perplexity: 9.5311
23
  - Bleu: 1.0
24
 
25
  ## Model description
 
50
 
51
  ### Training results
52
 
53
+ | Training Loss | Epoch | Step | Validation Loss | Accuracy | Perplexity | Bleu |
54
+ |:-------------:|:------:|:-----:|:---------------:|:--------:|:----------:|:----:|
55
+ | 7.6654 | 0.1403 | 500 | 3.7315 | 0.0073 | 41.7396 | 1.0 |
56
+ | 7.0276 | 0.2807 | 1000 | 3.4735 | 0.0073 | 32.2490 | 1.0 |
57
+ | 6.4629 | 0.4210 | 1500 | 3.1863 | 0.0073 | 24.1987 | 1.0 |
58
+ | 5.9671 | 0.5613 | 2000 | 2.9542 | 0.0073 | 19.1873 | 1.0 |
59
+ | 5.6969 | 0.7017 | 2500 | 2.8233 | 0.0073 | 16.8331 | 1.0 |
60
+ | 5.5077 | 0.8420 | 3000 | 2.7351 | 0.0073 | 15.4112 | 1.0 |
61
+ | 5.3536 | 0.9823 | 3500 | 2.6607 | 0.0073 | 14.3059 | 1.0 |
62
+ | 5.2099 | 1.1226 | 4000 | 2.6000 | 0.0073 | 13.4641 | 1.0 |
63
+ | 5.1158 | 1.2630 | 4500 | 2.5493 | 0.0073 | 12.7980 | 1.0 |
64
+ | 5.0453 | 1.4033 | 5000 | 2.5125 | 0.0073 | 12.3362 | 1.0 |
65
+ | 4.955 | 1.5436 | 5500 | 2.4806 | 0.0073 | 11.9489 | 1.0 |
66
+ | 4.9157 | 1.6840 | 6000 | 2.4537 | 0.0073 | 11.6310 | 1.0 |
67
+ | 4.8756 | 1.8243 | 6500 | 2.4300 | 0.0073 | 11.3584 | 1.0 |
68
+ | 4.844 | 1.9646 | 7000 | 2.4100 | 0.0073 | 11.1342 | 1.0 |
69
+ | 4.7136 | 2.1050 | 7500 | 2.3948 | 0.0073 | 10.9657 | 1.0 |
70
+ | 4.6911 | 2.2453 | 8000 | 2.3805 | 0.0073 | 10.8105 | 1.0 |
71
+ | 4.6741 | 2.3856 | 8500 | 2.3668 | 0.0073 | 10.6637 | 1.0 |
72
+ | 4.6485 | 2.5260 | 9000 | 2.3538 | 0.0073 | 10.5257 | 1.0 |
73
+ | 4.623 | 2.6663 | 9500 | 2.3416 | 0.0073 | 10.3976 | 1.0 |
74
+ | 4.6016 | 2.8066 | 10000 | 2.3303 | 0.0073 | 10.2806 | 1.0 |
75
+ | 4.5823 | 2.9470 | 10500 | 2.3202 | 0.0073 | 10.1776 | 1.0 |
76
+ | 4.4802 | 3.0873 | 11000 | 2.3143 | 0.0073 | 10.1182 | 1.0 |
77
+ | 4.4671 | 3.2276 | 11500 | 2.3073 | 0.0073 | 10.0469 | 1.0 |
78
+ | 4.4557 | 3.3679 | 12000 | 2.3006 | 0.0073 | 9.9800 | 1.0 |
79
+ | 4.4437 | 3.5083 | 12500 | 2.2928 | 0.0073 | 9.9023 | 1.0 |
80
+ | 4.4402 | 3.6486 | 13000 | 2.2862 | 0.0073 | 9.8375 | 1.0 |
81
+ | 4.4482 | 3.7889 | 13500 | 2.2800 | 0.0073 | 9.7763 | 1.0 |
82
+ | 4.4279 | 3.9293 | 14000 | 2.2752 | 0.0073 | 9.7303 | 1.0 |
83
+ | 4.3188 | 4.0696 | 14500 | 2.2730 | 0.0073 | 9.7087 | 1.0 |
84
+ | 4.3193 | 4.2099 | 15000 | 2.2691 | 0.0073 | 9.6704 | 1.0 |
85
+ | 4.3158 | 4.3503 | 15500 | 2.2652 | 0.0073 | 9.6329 | 1.0 |
86
+ | 4.3196 | 4.4906 | 16000 | 2.2619 | 0.0073 | 9.6012 | 1.0 |
87
+ | 4.2946 | 4.6309 | 16500 | 2.2589 | 0.0073 | 9.5722 | 1.0 |
88
+ | 4.3078 | 4.7713 | 17000 | 2.2564 | 0.0073 | 9.5487 | 1.0 |
89
+ | 4.2974 | 4.9116 | 17500 | 2.2546 | 0.0073 | 9.5311 | 1.0 |
90
 
91
 
92
  ### Framework versions
model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9837d5fdb80b6c053423ca0c71bdf5d0f982483681680c4175aed3873c4947ed
3
  size 1417229824
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9839d5f4cd171dea6e15ba583be6e1cce05dea4e8444d950b8ccd0ca73da4483
3
  size 1417229824
modeling_duo_predict_gpt2.py CHANGED
@@ -77,7 +77,7 @@ def scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.
77
  attn_bias = torch.zeros(L, S, dtype=query.dtype, device=query.device)
78
  if is_causal:
79
  assert attn_mask is None
80
- temp_mask = torch.ones(L, S, dtype=torch.bool).tril(diagonal=0)
81
  attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
82
  attn_bias.to(query.dtype)
83
 
@@ -129,10 +129,10 @@ def sdpa_attention_forward(
129
  query,
130
  key,
131
  value,
132
- attn_mask=create_attention_mask_matrix(query.shape[-2]).to(query.device),
133
  dropout_p=dropout,
134
  scale=scaling,
135
- is_causal=is_causal,
136
  )
137
  attn_output = attn_output.transpose(1, 2).contiguous()
138
 
@@ -582,9 +582,12 @@ class DuoPredictGPT2Model(DuoPredictGPT2PretrainedModel):
582
  inputs_embeds = self.wte(input_ids)
583
  position_embeds = self.wpe(position_ids)
584
  ###TODO: correctly initialized
585
- hidden_states = torch.empty((batch_size, input_shape[-1], self.embed_dim), device=device)
586
- hidden_states[:, ::2] = inputs_embeds[:, ::2] + position_embeds.to(inputs_embeds.device)
587
- hidden_states[:, 1::2] = inputs_embeds[:, 1::2] + position_embeds[:, :self.config.max_position_embeddings-1].to(inputs_embeds.device)
 
 
 
588
 
589
  # Attention mask.
590
  _use_sdpa = self._attn_implementation == "sdpa" and output_attentions is False and head_mask is None
@@ -836,17 +839,21 @@ class DuoPredictGPT2LMHeadModel(DuoPredictGPT2PretrainedModel, GenerationMixin):
836
  lm_logits = self.lm_head(hidden_states)
837
 
838
  loss = None
 
839
  if labels is not None:
840
- # Flatten the tokens
841
- total_labels = torch.full((lm_logits.shape[:2]), -100, dtype=input_ids.dtype, device=input_ids.device)
842
- total_labels[:, :-1:2] = labels[:, 1: ]
843
- total_labels[:, 1:-1:2] = labels[:, :-1]
 
 
 
844
  loss = self.loss_function(
845
- lm_logits,
846
- total_labels,
847
- vocab_size=self.config.vocab_size,
848
- **kwargs,
849
- )
850
 
851
  if not return_dict:
852
  output = (lm_logits,) + transformer_outputs[1:]
@@ -897,5 +904,6 @@ if __name__ == "__main__":
897
  model = DuoPredictGPT2LMHeadModel(cg)
898
  from src.utils.model_utlis import print_trainable_parameters
899
  print_trainable_parameters(model)
 
900
  model(torch.randint(0, 10000, (1, 100)))
901
  print()
 
77
  attn_bias = torch.zeros(L, S, dtype=query.dtype, device=query.device)
78
  if is_causal:
79
  assert attn_mask is None
80
+ temp_mask = torch.ones(L, S, dtype=torch.bool, device=query.device).tril(diagonal=0)
81
  attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
82
  attn_bias.to(query.dtype)
83
 
 
129
  query,
130
  key,
131
  value,
132
+ attn_mask=create_attention_mask_matrix(query.shape[-2]).to(query.device) if query.shape[1]>module.config.max_position_embeddings else None,
133
  dropout_p=dropout,
134
  scale=scaling,
135
+ is_causal=False if query.shape[1]>module.config.max_position_embeddings else True,
136
  )
137
  attn_output = attn_output.transpose(1, 2).contiguous()
138
 
 
582
  inputs_embeds = self.wte(input_ids)
583
  position_embeds = self.wpe(position_ids)
584
  ###TODO: correctly initialized
585
+ if inputs_embeds.shape[1] != position_embeds.shape[1]:
586
+ hidden_states = torch.empty((batch_size, input_shape[-1], self.embed_dim), device=device)
587
+ hidden_states[:, ::2] = inputs_embeds[:, ::2] + position_embeds.to(inputs_embeds.device)
588
+ hidden_states[:, 1::2] = inputs_embeds[:, 1::2] + position_embeds[:, :self.config.max_position_embeddings-1].to(inputs_embeds.device)
589
+ else:
590
+ hidden_states = inputs_embeds + position_embeds
591
 
592
  # Attention mask.
593
  _use_sdpa = self._attn_implementation == "sdpa" and output_attentions is False and head_mask is None
 
839
  lm_logits = self.lm_head(hidden_states)
840
 
841
  loss = None
842
+ bs, seq = lm_logits.shape[:2]
843
  if labels is not None:
844
+ if seq>labels.shape[1]:
845
+ # Flatten the tokens
846
+ total_labels = torch.full((bs, seq-1), -100, dtype=input_ids.dtype, device=input_ids.device)
847
+ total_labels[:, :-1:2] = labels[:, 1: ]
848
+ total_labels[:, 1::2] = labels[:, :-1]
849
+ else:
850
+ total_labels = labels[:, 1:]
851
  loss = self.loss_function(
852
+ lm_logits[:, :-1],
853
+ total_labels,
854
+ vocab_size=self.config.vocab_size,
855
+ **kwargs,
856
+ )
857
 
858
  if not return_dict:
859
  output = (lm_logits,) + transformer_outputs[1:]
 
904
  model = DuoPredictGPT2LMHeadModel(cg)
905
  from src.utils.model_utlis import print_trainable_parameters
906
  print_trainable_parameters(model)
907
+ model.eval()
908
  model(torch.randint(0, 10000, (1, 100)))
909
  print()