Why can I set both dilation_rate and strides '>1' in Conv1DTranspose?

[Zhaopudark]

System information.

• Have I written custom code (as opposed to using a stock example script provided in Keras): no
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Windows 11
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): 2.8.0
• Python version: 3.9.7
• Bazel version (if compiling from source): no
• GPU model and memory: RTX3090 24GB
• Exact command to reproduce:

>>> import tensorflow as tf
>>> x = tf.random.normal(shape=[1,5,1])
>>> y = tf.keras.layers.Conv1DTranspose(filters=3,kernel_size=[3,],strides=(2,),dilation_rate=(2,),padding="same",)
>>> print(y(x).shape)
(1, 10, 3)
>>> print(y.strides)
(2,)
>>> print(y.dilation_rate)
(2,)

Describe the problem.

When using Conv1DTranspose y = tf.keras.layers.Conv1DTranspose(filters=3,kernel_size=[3,],strides=(2,),dilation_rate=(2,),padding="same",)
havn't thrown an exception when stride and dilation_rate are both ">1".

It's different from the description in https://github.com/keras-team/keras/blob/v2.8.0/keras/layers/convolutional.py#L880-L884 and AIP docs in https://tensorflow.google.cn/api_docs/python/tf/keras/layers/Conv1DTranspose.

I have no idea whever this is a docs's bug or a code's bug.

[pokecheater]

have a similar issue https://github.com/keras-team/tf-keras/issues/79 seems like I am not alone with it :)

[old-school-kid]

@Zhaopudark Dilation basically means to leave out features (tokens or pixels) in between two features which will be convolved together. If you keep both dilation_rate and strides greater than 1 then some features will nevr be taken into account which is illegal. Hope this helps!

[Zhaopudark]

@old-school-kid Thanks, I totally agree with you. That answers my doubts. I really felt confused about the code of Conv1DTranspose that different from its' documentation description before. But now, I think it just lacks mandatory restrictions, but does not affect correct use if I understand the effects on dilation_rate and strides.

[chunduriv]

When we run the code on the CPU, we get InvalidArgumentError as shown below

import tensorflow as tf
x = tf.random.normal(shape=[1,5,1])
y = tf.keras.layers.Conv1DTranspose(filters=3,kernel_size=[3,],strides=(2,),dilation_rate=(2,),padding="same",)

print(y(x).shape)

Output:

---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
[<ipython-input-2-155ce2f8a75e>](https://localhost:8080/#) in <module>()
----> 1 print(y(x).shape)

1 frames
[/usr/local/lib/python3.7/dist-packages/tensorflow/python/framework/ops.py](https://localhost:8080/#) in raise_from_not_ok_status(e, name)
   7184 def raise_from_not_ok_status(e, name):
   7185   e.message += (" name: " + name if name is not None else "")
-> 7186   raise core._status_to_exception(e) from None  # pylint: disable=protected-access
   7187 
   7188 

InvalidArgumentError: Exception encountered when calling layer "conv1d_transpose" (type Conv1DTranspose).

Current libxsmm and customized CPU implementations do not yet support dilation rates larger than 1. [Op:Conv2DBackpropInput]

Call arguments received:
  • inputs=tf.Tensor(shape=(1, 5, 1), dtype=float32)
 

Where as on the GPU, it is not throwing any exception as shown below

print(y(x).shape) (1, 10, 3)

Please find the gist here for reference. Thanks!

[old-school-kid]

@chunduriv In the documentation it clearly mentions under dilation

Currently, specifying a dilation_rate value != 1 is incompatible with specifying a stride value != 1.

And this should also be the ideal case. But while executing with GPU it accepts value which I think is a bug. Can you please look into it? TIA

[chunduriv]

@old-school-kid,

Yes, it is working as expected on CPU, but it is a bug on GPU.

[gowthamkpr]

This is working as expected in the latest version of tf-nightly. Please find the gist here. Thanks!

[old-school-kid]

@gowthamkpr On CPU it throws Invalid Argument Error which is right imo. But on GPU it throws Unimplemented Error which is confusing IMO.

[tilakrayal]

@Zhaopudark, I tried to execute the code on tensorflow v2.13(both cpu and gpu), and it was working as expected by throwing the error strides > 1 not supported in conjunction with dilation_rate > 1. Received: strides=(2,) and dilation_rate=(2,) which was informative for the user. Kindly find the gist of it here.

CPU and GPU. Thank you!

[Zhaopudark]

@tilakrayal Thanks. Since it has been solved in the latest version of TensorFlow, should I close this issue?

[tilakrayal]

@Zhaopudark, As the issue got resolved, Could you please feel free to move this issue to the closed status. Thank you!

[Zhaopudark]

@tilakrayal OK! Thanks a lot

[google-ml-butler[bot]]

Are you satisfied with the resolution of your issue? Yes No