Conv1d throws an error when using 'channels_first' data format

[MoayedHajiAli]

System information.

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Amazon EC2
• TensorFlow installed from (source or binary): source
• TensorFlow version (use command below): 2.5.0
• Python version: 3.6.13
• tested on CPU

Describe the problem.
calling a Conv1D layer with the channels first data format throws an error of

The Conv2D op currently only supports the NHWC tensor format on the CPU. The op was given the format: NCHW [Op:Conv2D]

This error is not present when using 'channels_last' data format and produce the expected behavior. The issue is also present in the Conv2d layers.

Standalone code to reproduce the issue.
K.layers.Conv1D(10, 3, data_format='channels_first')(tf.random.normal((10, 3, 10), dtype=tf.float32))

[collinzrj]

The convolution layer of keras use the tensorflow op conv1d. From line 2002-2026 in tensorflow/python/ops/nn_ops

if value.shape.ndims in (4, 3, 2, 1, 0, None):
  result = gen_nn_ops.conv2d(
      value,
      filters,
      strides,
      padding,
      use_cudnn_on_gpu=use_cudnn_on_gpu,
      data_format=data_format,
      dilations=dilations,
      name=name)
else:
  result = squeeze_batch_dims(
      value,
      functools.partial(
          gen_nn_ops.conv2d,
          filter=filters,
          strides=strides,
          padding=padding,
          use_cudnn_on_gpu=use_cudnn_on_gpu,
          data_format=data_format,
          dilations=dilations,
      ),
      inner_rank=3,
      name=name)
return array_ops.squeeze(result, [spatial_start_dim])

It seems that the implementation of conv1d in tensorflow uses conv2d, so the error says "Conv2D". Changing to GPU solves this problem, or use 'channels_last' as you said.

[tilakrayal]

@MoayedHajiAli ,

Please check the above comment and also please take a look at this issue link which is similar.It helps.Thanks!

[MoayedHajiAli]

Thank you a lot for your help. I am closing the issue.