Update modeling_arlow_gpt.py
Browse files- modeling_arlow_gpt.py +26 -54
modeling_arlow_gpt.py
CHANGED
@@ -1,15 +1,14 @@
|
|
1 |
# modeling_arlow_gpt.py
|
2 |
-
import
|
3 |
-
|
4 |
-
from transformers import PreTrainedModel, CLIPModel, GPT2Model
|
5 |
from typing import Optional, Union, Dict, Tuple
|
6 |
from .configuration_arlow_gpt import ArlowGPTConfig
|
7 |
|
8 |
class ArlowGPTPreTrainedModel(PreTrainedModel):
|
9 |
-
"""Base class for ArlowGPT model."""
|
10 |
config_class = ArlowGPTConfig
|
11 |
base_model_prefix = "arlow_gpt"
|
12 |
supports_gradient_checkpointing = True
|
|
|
13 |
|
14 |
def _init_weights(self, module):
|
15 |
if isinstance(module, nn.Linear):
|
@@ -18,59 +17,32 @@ class ArlowGPTPreTrainedModel(PreTrainedModel):
|
|
18 |
module.bias.data.zero_()
|
19 |
|
20 |
class ArlowGPTModel(ArlowGPTPreTrainedModel):
|
|
|
|
|
|
|
21 |
def __init__(self, config: ArlowGPTConfig):
|
22 |
super().__init__(config)
|
23 |
-
|
24 |
-
self.
|
25 |
-
self.gpt2 = GPT2Model.from_pretrained(config.gpt2_model_name)
|
26 |
-
|
27 |
-
self.feature_projection = nn.Linear(
|
28 |
-
self.clip.vision_model.config.hidden_size + self.gpt2.config.hidden_size,
|
29 |
-
config.projection_dim
|
30 |
-
)
|
31 |
|
32 |
# Initialize weights and apply final processing
|
33 |
self.post_init()
|
34 |
|
35 |
-
def
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
pixel_values: torch.Tensor,
|
40 |
-
return_dict: bool = True,
|
41 |
-
) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
|
42 |
-
vision_outputs = self.clip.get_image_features(pixel_values=pixel_values)
|
43 |
-
text_outputs = self.gpt2(
|
44 |
-
input_ids=input_ids,
|
45 |
-
attention_mask=attention_mask
|
46 |
-
).last_hidden_state
|
47 |
-
|
48 |
-
batch_size = text_outputs.shape[0]
|
49 |
-
seq_length = text_outputs.shape[1]
|
50 |
-
|
51 |
-
vision_features = vision_outputs.unsqueeze(1).expand(
|
52 |
-
batch_size, seq_length, -1
|
53 |
-
)
|
54 |
-
|
55 |
-
combined_features = torch.cat(
|
56 |
-
[vision_features, text_outputs],
|
57 |
-
dim=-1
|
58 |
-
)
|
59 |
-
|
60 |
-
hidden_states = self.feature_projection(combined_features)
|
61 |
|
62 |
-
|
63 |
-
|
64 |
-
|
|
|
|
|
|
|
|
|
65 |
|
66 |
-
|
67 |
-
|
68 |
-
super().__init__(config)
|
69 |
-
self.arlow_gpt = ArlowGPTModel(config)
|
70 |
-
self.output_projection = nn.Linear(config.projection_dim, config.vocab_size)
|
71 |
-
|
72 |
-
# Initialize weights and apply final processing
|
73 |
-
self.post_init()
|
74 |
|
75 |
def forward(
|
76 |
self,
|
@@ -88,7 +60,7 @@ class ArlowGPTForCausalLM(ArlowGPTPreTrainedModel):
|
|
88 |
)
|
89 |
|
90 |
hidden_states = outputs["hidden_states"]
|
91 |
-
logits = self.
|
92 |
|
93 |
loss = None
|
94 |
if labels is not None:
|
@@ -96,10 +68,10 @@ class ArlowGPTForCausalLM(ArlowGPTPreTrainedModel):
|
|
96 |
loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
|
97 |
|
98 |
if return_dict:
|
99 |
-
return
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
return (loss, logits) if loss is not None else logits
|
104 |
|
105 |
def prepare_inputs_for_generation(
|
|
|
1 |
# modeling_arlow_gpt.py
|
2 |
+
from transformers import PreTrainedModel, PreTrainedModel, CLIPModel, GPT2Model
|
3 |
+
from transformers.modeling_outputs import Seq2SeqLMOutput
|
|
|
4 |
from typing import Optional, Union, Dict, Tuple
|
5 |
from .configuration_arlow_gpt import ArlowGPTConfig
|
6 |
|
7 |
class ArlowGPTPreTrainedModel(PreTrainedModel):
|
|
|
8 |
config_class = ArlowGPTConfig
|
9 |
base_model_prefix = "arlow_gpt"
|
10 |
supports_gradient_checkpointing = True
|
11 |
+
_keys_to_ignore_on_load_missing = [r"clip", r"gpt2"]
|
12 |
|
13 |
def _init_weights(self, module):
|
14 |
if isinstance(module, nn.Linear):
|
|
|
17 |
module.bias.data.zero_()
|
18 |
|
19 |
class ArlowGPTModel(ArlowGPTPreTrainedModel):
|
20 |
+
# Same as before
|
21 |
+
|
22 |
+
class ArlowGPTForImageTextToText(ArlowGPTPreTrainedModel):
|
23 |
def __init__(self, config: ArlowGPTConfig):
|
24 |
super().__init__(config)
|
25 |
+
self.arlow_gpt = ArlowGPTModel(config)
|
26 |
+
self.lm_head = nn.Linear(config.projection_dim, config.vocab_size)
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
# Initialize weights and apply final processing
|
29 |
self.post_init()
|
30 |
|
31 |
+
def save_pretrained(self, save_directory, **kwargs):
|
32 |
+
"""Override save_pretrained to save all components"""
|
33 |
+
super().save_pretrained(save_directory, **kwargs)
|
34 |
+
self.arlow_gpt.save_pretrained(save_directory)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
|
36 |
+
@classmethod
|
37 |
+
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
|
38 |
+
"""Override from_pretrained to handle custom loading logic"""
|
39 |
+
config = kwargs.get("config", None)
|
40 |
+
if config is None:
|
41 |
+
config = ArlowGPTConfig.from_pretrained(pretrained_model_name_or_path)
|
42 |
+
kwargs["config"] = config
|
43 |
|
44 |
+
config._name_or_path = pretrained_model_name_or_path
|
45 |
+
return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
def forward(
|
48 |
self,
|
|
|
60 |
)
|
61 |
|
62 |
hidden_states = outputs["hidden_states"]
|
63 |
+
logits = self.lm_head(hidden_states)
|
64 |
|
65 |
loss = None
|
66 |
if labels is not None:
|
|
|
68 |
loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
|
69 |
|
70 |
if return_dict:
|
71 |
+
return Seq2SeqLMOutput(
|
72 |
+
loss=loss,
|
73 |
+
logits=logits,
|
74 |
+
)
|
75 |
return (loss, logits) if loss is not None else logits
|
76 |
|
77 |
def prepare_inputs_for_generation(
|