modeling_resnet.py

import torch
from torch import nn

from transformers import ResNetPreTrainedModel
from transformers.modeling_outputs import ImageClassifierOutputWithNoAttention
from transformers.image_processing_utils import BaseImageProcessor
from transformers import ResNetConfig, ResNetModel
from typing import Optional

class ResNetForZeroBitWatermarkDetection(ResNetPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.resnet = ResNetModel(config)
        self.classifier = nn.Sequential(
            nn.Flatten(),
            nn.Linear(config.hidden_sizes[-1], 128),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(128, 1))
        self.register_buffer('beta', torch.tensor([1.0]))
        
        # initialize weights and apply final processing
        self.post_init()


    # TODO docstring
    def forward(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> ImageClassifierOutputWithNoAttention:

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.resnet(pixel_values, output_hidden_states=output_hidden_states, return_dict=return_dict)

        pooled_output = outputs.pooler_output if return_dict else outputs[1]

        x = self.classifier(pooled_output)

        # generalized-Gaussian recalibration, centering and scaling is already included in last linear layer
        x =  0.5 + torch.sign(x) * 0.5 * torch.special.gammainc(1 / self.beta, torch.abs(x)**self.beta)

        # Laplacian calibration, centering and scaling is already included in last linear layer
        # if beta==1
        #x =  0.5 + torch.sign(x) * 0.5 * (1 - torch.exp(-torch.abs(x))) # laplacian

        logits = torch.log(x) - torch.log1p(-x)

        loss = None
       
        if not return_dict:
            output = (logits,) + outputs[2:]
            return (loss,) + output if loss is not None else output

        return ImageClassifierOutputWithNoAttention(loss=loss, logits=logits, hidden_states=outputs.hidden_states)