I have this tensor of multidimensional array:

# read from memory-mapped files

# File paths for memory-mapped files
input_file_path = 'tensor_input.dat'
output_file_path = 'tensor_output.dat'

# Define the shapes of the arrays
tensor_input_shape = (12, 5, 120, 2, 600, 600)  # num of ID, num of channels, timeline, timesteps, height, width
tensor_output_shape = (12,)  # num of ID

tensor_input = np.memmap(input_file_path, dtype='uint8', mode='r', shape=tensor_input_shape)

# Debugging outputs
dbg(tensor_input)

Outputting:

tensor type is numpy array
shape: (12, 5, 120, 2, 600, 600)
dtype: uint8

The tensor input size is so large, that’s why I use mmap.

The ConvLSTM2D keras accept 5D (batch, timestep, height, width, channel). Therefore I expect I can reshape the mmap into (batch*timeline, timestep, height, width, channel).

I can swap or change axis with np.transpose. But I’m afraid if I messed up the memory structure, mainly there is flatenned axis at batch and timeline axis.