walsvid / coordconv

Pytorch implementation of "An intriguing failing of convolutional neural networks and the CoordConv solution" - https://arxiv.org/abs/1807.03247

CoordConv

Pytorch implementation of CoordConv for N-D ConvLayers, and the experiments.

Reference from the paper: An intriguing failing of convolutional neural networks and the CoordConv solution

Extends the CoordinateChannel concatenation from 2D to 1D and 3D tensors.

Requirements

• pytorch 0.4.0
• torchvision 0.2.1
• torchsummary 1.3
• sklearn 0.19.1

Usage

from coordconv import CoordConv1d, CoordConv2d, CoordConv3d

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.coordconv = CoordConv2d(2, 32, 1, with_r=True)
        self.conv1 = nn.Conv2d(32, 64, 1)
        self.conv2 = nn.Conv2d(64, 64, 1)
        self.conv3 = nn.Conv2d(64,  1, 1)
        self.conv4 = nn.Conv2d( 1,  1, 1)

    def forward(self, x):
        x = self.coordconv(x)
        x = F.relu(self.conv1(x))
        x = F.relu(self.conv2(x))
        x = F.relu(self.conv3(x))
        x = self.conv4(x)
        x = x.view(-1, 64*64)
        return x

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
net = Net().to(device)

Experiments

Implement experiments from origin paper.

Coordinate Classification

Use experiments/generate_data.py to generate Uniform and Quadrant datasets for Coordinate Classification task.

Use experiments/train_and_test.py to train and test neural network model.

Uniform Datasets

Train	Test	Predictions

Quadrant Datasets

Train	Test	Predictions

coordconv's Issues

doesnt seem to work on quadrant data

Hi, thanks for sharing the code. The jupyter notebook is very helpful indeed, but it doesn't seem to work on the quadrant data whereas the same code works fine for the uniform data. Anybody else facing this problem?

Implemenation for rank==3 incomplete

CoordConv/coordconv.py

Line 71 in 3f86264

batch_size_shape, channel_in_shape, dim_z, dim_y, dim_x = input_tensor.shape

For example, batch_size_shape not used, leading to batch_size mismatch.

Question about generate_data.py

Although it does not affect the final results,
please consider this minor change in your code to reflect the original paper contents

It will be helpful for debugging

paper :

## bias is basically neglected in tensorflow
onehots = np.pad(np.eye(3136).reshape((3136, 56, 56, 1)), ((0,0), (4,4), (4,4), (0,0)), "constant");  
images = tf.nn.conv2d(onehots, np.ones((9, 9, 1, 1)), [1]*4, "SAME")

code as is :

CoordConv/experiments/generate_data.py

Line 25 in 647bcfe

conv_layer = nn.Conv2d(in_channels=1, out_channels=1, kernel_size=(9, 9), padding=4, stride=1)

code to be:

conv_layer = nn.Conv2d(in_channels=1, out_channels=1, kernel_size=(9, 9), padding=4, stride=1, bias=False)

Training time increases linearly for each epoch.

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?

CoordConv regression task

did you try coordinate regression task in the original paper?