1
$\begingroup$

time_distributed_24 (TimeDis (None, 16, 64, 64, 512) 0


conv_lst_m2d_2 (ConvLSTM2D) (None, 16, 64, 64, 128) 2949632


time_distributed_25 (TimeDis (None, 16, 64, 64, 128) 512


time_distributed_26 (TimeDis (None, 16, 128, 128, 128) 0

For example, why does this ConvLSTM2D layer have 2949632 parameters? A standard LSTM layer has 4(n*m + n^2 + n) parameters where m = input dim, n = output dim.

A standard 3x3 conv layer with 128 kernels over a 64x64x512 tensor would have 3x3x512x128 parameters

What would be the n and m for the LSTM part?

$\endgroup$

1 Answer 1

1
$\begingroup$

i=512, input dimension or input channel

h=128, output dimension or output channel

k=3, kernel size

number of parameters = 4 * h * ( k**2 * (i+h) + 1 ) = 2949632

if k == 1, ConvLSTM is just a LSTM.

You can check the original paper for detail. https://arxiv.org/abs/1506.04214

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.