How to create `tf.keras.layers.Input` without batch_size dimension?

[baoachun]

Click to expand!

Issue Type

Support

Have you reproduced the bug with TF nightly?

No

Source

source

Tensorflow Version

2.9.1

Custom Code

No

OS Platform and Distribution

centos7

Mobile device

No response

Python version

3.8.14

Bazel version

5.0

GCC/Compiler version

10.2

CUDA/cuDNN version

11.4

GPU model and memory

A30

Current Behaviour?

One input of my model has nothing to do with batch_size, for example, its shape is [1,2,3], how to avoid automatically adding 1 dimension when creating tf.keras.layers.Input ? If I manually slice it, the slice operator will be introduced, resulting in a decrease in inference performance.

Standalone code to reproduce the issue

import tensorflow as tf

x = tf.keras.layers.Input(shape=(32,64))
# x.shape: (None, 32, 64)
x = x[0, :, :]
# x.shape: (32, 64)

[Pouyanpi]

Hi @baoachun ,

The first dimension is reserved for batch size. Please note that the shape argument must be the shape of each sample, not the shape of one batch.

In case you have fixed batch size you can use the batch_size parameter. Also, it is not clear what you exactly mean by "One input of my model has nothing to do with batch_size".

Probably it would be useful if you provide a more concrete and complete the description of your use case.

Best,
PR

[baoachun]

Hi @baoachun ,

The first dimension is reserved for batch size. Please note that the shape argument must be the shape of each sample, not the shape of one batch.

In case you have fixed batch size you can use the batch_size parameter. Also, it is not clear what you exactly mean by "One input of my model has nothing to do with batch_size".

Probably it would be useful if you provide a more concrete and complete the description of your use case.

Best, PR

Some inputs of my model are weights, such as the second input of matmul, so this input has nothing to do with the batch dimension. The reason for declaring weights as inputs is that the weights are saved in multiple servers under parameter server framework.

[sachinprasadhs]

Thank you for reporting the issue.
We're now prioritizing Keras 3 development, a multi-backend framework that offers significant advantages by supporting JAX, TensorFlow, PyTorch and OpenVINO.
This means we won't be adding new features to TF-Keras (Keras 2).

TensorFlow users can benefit directly from Keras 3, which is included by default starting with version 2.16.1.
We encourage you to migrate to Keras 3 and experience its enhanced capabilities. If you feel this specific feature is crucial for the Keras framework itself, please open a new issue at: https://github.com/keras-team/keras/issues

For more details and usage guide about Keras 3, refer: https://keras.io/getting_started/about/