This repository has been archived by the owner on Dec 16, 2022. It is now read-only.
/
vision_text_model.py
229 lines (193 loc) · 7.32 KB
/
vision_text_model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
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
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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
import logging
from typing import Dict, List, Optional
import numpy as np
import torch
from allennlp.data.fields.text_field import TextFieldTensors
from allennlp.data.vocabulary import Vocabulary
from allennlp.models.model import Model
from allennlp.modules.transformer import (
TransformerEmbeddings,
ImageFeatureEmbeddings,
BiModalEncoder,
)
logger = logging.getLogger(__name__)
@Model.register("vision_model")
class VisionTextModel(Model):
"""
`VisionTextModel` takes as input a single text input and a single image input
to produce some output. Example tasks include visual question-answering, visual
entailment, etc.
# Parameters
vocab : `Vocabulary`
text_embeddings : `TransformerEmbeddings`
image_embeddings : `ImageFeatureEmbeddings`
encoder : `BiModalEncoder`
pooled_output_dim : `int`
fusion_method : `str`, optional (default = `"sum"`)
dropout : `float`, optional (default = `0.1`)
label_namespace : `str`, optional (default = `"labels"`)
is_multilabel: `bool`, optional (default = `False`)
Whether the output classification is multilabel.
(i.e., can have multiple correct answers)
"""
def __init__(
self,
vocab: Vocabulary,
text_embeddings: TransformerEmbeddings,
image_embeddings: ImageFeatureEmbeddings,
encoder: BiModalEncoder,
pooled_output_dim: int,
fusion_method: str = "sum",
dropout: float = 0.1,
label_namespace: str = "labels",
is_multilabel: bool = False,
*,
ignore_text: bool = False,
ignore_image: bool = False,
) -> None:
super().__init__(vocab)
from allennlp.modules.backbones import VilbertBackbone
self.backbone = VilbertBackbone(
vocab,
text_embeddings,
image_embeddings,
encoder,
pooled_output_dim,
fusion_method,
dropout,
)
num_labels = vocab.get_vocab_size(label_namespace)
self.label_namespace = label_namespace
self.classifier = torch.nn.Linear(pooled_output_dim, num_labels)
self.dropout = torch.nn.Dropout(dropout)
self.is_multilabel = is_multilabel
self.ignore_text = ignore_text
self.ignore_images = ignore_image
@classmethod
def from_huggingface_model_name(
cls,
vocab: Vocabulary,
model_name: str,
image_feature_dim: int,
image_num_hidden_layers: int,
image_hidden_size: int,
image_num_attention_heads: int,
combined_hidden_size: int,
combined_num_attention_heads: int,
pooled_output_dim: int,
image_intermediate_size: int,
image_attention_dropout: float,
image_hidden_dropout: float,
image_biattention_id: List[int],
text_biattention_id: List[int],
text_fixed_layer: int,
image_fixed_layer: int,
pooled_dropout: float = 0.1,
fusion_method: str = "sum",
*,
ignore_text: bool = False,
ignore_image: bool = False,
):
text_embeddings = TransformerEmbeddings.from_pretrained_module(model_name)
image_embeddings = ImageFeatureEmbeddings(
feature_size=image_feature_dim,
embedding_size=image_hidden_size,
dropout=image_hidden_dropout,
)
encoder = BiModalEncoder.from_pretrained_module(
model_name,
num_hidden_layers2=image_num_hidden_layers,
hidden_size2=image_hidden_size,
num_attention_heads2=image_num_attention_heads,
combined_hidden_size=combined_hidden_size,
combined_num_attention_heads=combined_num_attention_heads,
intermediate_size2=image_intermediate_size,
attention_dropout2=image_attention_dropout,
hidden_dropout2=image_hidden_dropout,
biattention_id1=text_biattention_id,
biattention_id2=image_biattention_id,
fixed_layer1=text_fixed_layer,
fixed_layer2=image_fixed_layer,
)
return cls(
vocab=vocab,
text_embeddings=text_embeddings,
image_embeddings=image_embeddings,
encoder=encoder,
pooled_output_dim=pooled_output_dim,
fusion_method=fusion_method,
dropout=pooled_dropout,
ignore_text=ignore_text,
ignore_image=ignore_image,
)
def forward(
self, # type: ignore
box_features: torch.Tensor,
box_coordinates: torch.Tensor,
box_mask: torch.Tensor,
text: TextFieldTensors,
labels: Optional[torch.Tensor] = None,
label_weights: Optional[torch.Tensor] = None,
) -> Dict[str, torch.Tensor]:
"""
# Parameters
box_features : `Tensor`
Shape: `(batch_size, num_boxes, feature_size)`
box_coordinates : `Tensor`
Shape: `(batch_size, num_boxes, 4)`
box_mask : `Tensor`
A bool and 0-1 tensor of shape `(batch_size, num_boxes)`.
text : `TextFieldTensors`
label : `Optional[Tensor]`
label_weights : `Optional[Tensor]`
"""
batch_size = box_features.size(0)
if self.ignore_images:
box_features = torch.zeros_like(box_features)
box_coordinates = torch.zeros_like(box_coordinates)
box_coordinates[..., 2] = 1
box_coordinates[..., 3] = 1
box_mask = torch.ones_like(box_mask)
if self.ignore_text:
dummy_text = {}
for embedder_name, tensor_dict in text.items():
dummy_tensor_dict = {}
for tensor_name, tensor in tensor_dict.items():
if "mask" in tensor_name:
tensor = torch.ones_like(tensor)
else:
tensor = torch.zeros_like(tensor)
dummy_tensor_dict[tensor_name] = tensor
dummy_text[embedder_name] = dummy_tensor_dict
text = dummy_text
backbone_outputs = self.backbone(box_features, box_coordinates, box_mask, text)
# Shape: (batch_size, num_labels)
logits = self.classifier(backbone_outputs["pooled_boxes_and_text"])
# Shape: (batch_size, num_labels)
if self.is_multilabel:
probs = torch.sigmoid(logits)
else:
probs = torch.softmax(logits, dim=-1)
outputs = {"logits": logits, "probs": probs}
outputs = self._compute_loss_and_metrics(batch_size, outputs, labels, label_weights)
return outputs
def _compute_loss_and_metrics(
self,
batch_size: int,
outputs: Dict[str, torch.Tensor],
label: torch.Tensor,
label_weights: Optional[torch.Tensor] = None,
) -> Dict[str, torch.Tensor]:
return outputs
def get_metrics(self, reset: bool = False) -> Dict[str, float]:
result = self.accuracy.get_metric(reset)
return {"accuracy": result}
def make_output_human_readable(
self, output_dict: Dict[str, torch.Tensor]
) -> Dict[str, torch.Tensor]:
batch_labels = []
for batch_index, batch in enumerate(output_dict["probs"]):
labels = np.argmax(batch, axis=-1)
batch_labels.append(labels)
output_dict["labels"] = batch_labels
return output_dict