Teklia

/

doc-ufcn-generic-historical-line

like 17

Follow

Teklia 21

Hello! There is no dedicated function to do this directly. However, you can retrieve the encoder representations very easily:

• I suggest you clone the repository directly: https://gitlab.teklia.com/dla/doc-ufcn
• Install the environment using pip install -e .
• To keep only the encoder part, you can simply comment/remove the last part of the forward method in the model.py file:

    def forward(self, input_tensor):
        """
        Define the forward step of the network.
        It consists in 4 successive dilated blocks followed by 3
        convolutional blocks, a final convolution and a softmax layer.
        :param input_tensor: The input tensor.
        :return: The output tensor.
        """
        with autocast(enabled=self.amp):
            tensor = self.dilated_block1(input_tensor)
            out_block1 = tensor
            tensor = self.dilated_block2(self.pool(tensor))
            out_block2 = tensor
            tensor = self.dilated_block3(self.pool(tensor))
            out_block3 = tensor
            tensor = self.dilated_block4(self.pool(tensor))
            return tensor

You can then follow the example given in the documentation and apply the model as follows:

import cv2
from doc_ufcn import models
from doc_ufcn.main import DocUFCN

image = cv2.cvtColor(cv2.imread(IMAGE_PATH), cv2.COLOR_BGR2RGB)

model_path, parameters = models.download_model('generic-historical-line')
model = DocUFCN(len(parameters['classes']), parameters['input_size'], 'cpu')
model.load(model_path, parameters['mean'], parameters['std'], mode="eval")

_, raw_output, _, _ = model.predict(image, raw_output=True)
print(raw_output.shape)

Got it. I implemented a similar workaround. Thank you!