[FIXED] Can tf.reshape() be implemented when defining/ compiling a neural network model with 'None' in input shape (eg, (None, 128, 128, 128, 1))?

Issue

I am building a deep learning model on tensorflow and require to reshape the input in some layer during runtime. Part of the code is defined below:

``` init = RandomNormal(stddev=0.02, seed=42)
``` shape_= [layer_in.shape[0], layer_in.shape[1], layer_in.shape[2], layer_in.shape[3]*layer_in.shape[4]]
``` shape_t = tf.convert_to_tensor(shape_)
``` layer_in = tf.reshape(layer_in, shape=shape_t)
``` g = Conv3D(n_filters, (4,4,4), strides=(2,2,2), padding='same', kernel_initializer=init)(layer_in)
``` g = LeakyReLU(alpha=0.2)(g)
``` return g

However, after running the code, it shows an error: 
ValueError: Attempt to convert a value (None) with an unsupported type (<class 'NoneType'>) to a Tensor.

When the error occurs, the shape of layer_in is (None, 128, 128, 128, 1) 
Also checked tf.reshape(), the size parameter has to be a tensor so it seems it doesn't work to run tf.reshape() in this case.
How can it be resolved? 

Solution

The error arises because layer_in.shape[0] is None, indicating an unknown batch size. TensorFlow expects a tensor for the shape argument in tf.reshape, but None cannot be converted. There are different options you can consider to resolve this issue:

Solution: Use -1 to infer the batch dimension during runtime:

shape_ = [-1, layer_in.shape[1], layer_in.shape[2], layer_in.shape[3] * layer_in.shape[4]]

Use tf.shape for Dynamic Dimensions:

For fully dynamic shapes, use tf.shape(layer_in) to retrieve the dimensions at runtime:

shape_ = tf.shape(layer_in)
shape_ = tf.concat([shape_[:1], shape_[1:3], [shape_[3] * shape_[4]]], axis=0)

Now you can safely reshape layer_in using the constructed shape tensor:

layer_in = tf.reshape(layer_in, shape=shape_)

The final version should be like:

init = RandomNormal(stddev=0.02, seed=42)
shape_ = [-1, layer_in.shape[1], layer_in.shape[2], layer_in.shape[3] * layer_in.shape[4]]  # Use -1 for batch dimension

layer_in = tf.reshape(layer_in, shape=shape_)
g = Conv3D(n_filters, (4, 4, 4), strides=(2, 2, 2), padding='same', kernel_initializer=init)(layer_in)
g = LeakyReLU(alpha=0.2)(g)

return g

If you're working with eager execution, you might not need to explicitly convert shapes to tensors. For more complex dynamic reshaping, explore tf.keras.layers.Reshape layer for clarity and potential performance benefits.

Answered By - Mohammad Hossein Zardary