mkocabas / coordconv-pytorch Goto Github PK

There is a bug in the original implementation as they using x_dim to tile the x_range tensor.
It will result in xx_channel with shape [x_dim, x_dim] and yy_channel with shape [y_dim, y_dim]
and then the concatenate is not allowed for non-square case.

Please refer to correct implementation as
https://gist.github.com/leVirve/0377a8fbac455bfd44e374e5cf8b1260

you realize another channel,but i found it's wrong or i don't understand it ?
L104-you want add a new channel with r-->radius
but this xx_channel is not range 0-1;
so i change it to below and i change some place to reduce the transpose:

class AddCoords(nn.Module):

    def __init__(self, with_r=False):
        super().__init__()
        self.with_r = with_r

    def forward(self, input_tensor):
        """
        Args:
            input_tensor: shape(batch, channel, x_dim, y_dim)
        """
        batch_size, _, x_dim, y_dim = input_tensor.size()

        xx_channel = torch.arange(x_dim).repeat(1, y_dim, 1).transpose(1, 2)
        yy_channel = torch.arange(y_dim).repeat(1, x_dim, 1)
        print(xx_channel)

        xx_channel = xx_channel.repeat(batch_size, 1, 1, 1)
        yy_channel = yy_channel.repeat(batch_size, 1, 1, 1)

        xx_channel_01 = xx_channel.float() / (x_dim - 1)
        yy_channel_01 = yy_channel.float() / (y_dim - 1)

        xx_channel = xx_channel_01 * 2 - 1
        yy_channel = yy_channel_01 * 2 - 1

        ret = torch.cat([
            input_tensor,
            xx_channel.type_as(input_tensor),
            yy_channel.type_as(input_tensor)], dim=1)

        if self.with_r:
            rr = torch.sqrt(torch.pow(xx_channel_01.type_as(input_tensor) - 0.5, 2) + torch.pow(yy_channel_01.type_as(input_tensor) - 0.5, 2))
            ret = torch.cat([ret, rr], dim=1)

        return ret

another ways to realize it is to use "torch.meshgrid"

Hi @mkocabas,

There is a slight mismatch in the notation used to denote the extracted dimensions from the shape of the input tensor. here

batch_size, _, x_dim, y_dim = input_tensor.size()

You extract x_dim from dimension number 2 and y_dim from dimension number 3. As per the image based 4d tensors of Pytorch, it is the height that comes before width. Please refer to the documentation of Pytorch's Conv2d layer. Notice the input and output tensor shapes.

The code would run fine since the extracted dimensions end up in right place when constructing the output tensor, but for a code reader like me (implementation checking), it seems like a mistake at first glance. Perhaps, the notation can be changed as per Pytorch.
Thanks

Cheers!
@akanimax

1

Currently, I have been using this implementation for one of my object detection models.
As the model is training, batch processing time increases linearly. I still haven't figured out why.

Does anyone see a similar behavior?

Is there an example code integrating this with FasterRCNN as mentioned in the paper?

It seems to me like this solution has potential impact on localization tasks applied to OCR. I'd be curious to see what impact it would have on the ICDAR 2013 Table Competition dataset.

See https://arxiv.org/pdf/1804.06236.pdf for Reference

Hello,
At the code,

if with_r == True,this should be in_channels + 3 instead of in_channels + 2

Does anyone know where should we add the coord conv layers if we do regression task?

As I can see in numpy example it appends maps with values in range [-1,1], but as I understand from paper they suggest i,j int coordinates of pixels.

I went through the paper on archivX (this one https://arxiv.org/pdf/1807.03247.pdf)

but could not find the reason of why the center of the radius channel is shifted by 0.5 in both x and y direction? (Is there any particular reason? My intuition is it won't matter. Please let me know, if I'm missing something.)