File size: 1,008 Bytes
38c803c
 
 
 
99407d3
38c803c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2a5de8e
20d44a7
38c803c
 
 
 
13e6ee3
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
import torch.nn as nn
from transformers import VisionEncoderDecoderModel, PreTrainedModel, AutoConfig

class LegibilityModel(PreTrainedModel):
    def __init__(self, config):
        config = AutoConfig.from_pretrained("microsoft/trocr-base-handwritten")
        super(LegibilityModel, self).__init__(config=config)

        # base model architecture
        self.model = VisionEncoderDecoderModel(config).encoder

        # change dropout during training
        self.stack = nn.Sequential(
            nn.Dropout(0),
            nn.Linear(768, 768),
            nn.ReLU(),
            nn.Dropout(0),
            nn.Linear(768, 1)
        )

    # choice, img0, img1 are not used by the model, but are passed by the trainer
    def forward(self, img_batch, choice=None, img0=None, img1=None):
        output = self.model(img_batch)
        # average the output of the last hidden layer
        output = output.last_hidden_state.mean(dim=1)
        scores = self.stack(output)
        return scores.squeeze()