search

jacobgil / pytorch-tensor-decompositions

PyTorch implementation of [1412.6553] and [1511.06530] tensor decomposition methods for convolutional layers.
https://jacobgil.github.io/deeplearning/tensor-decompositions-deep-learning
280 stars 63 forks source link

Channel mismatches #3

Closed chowkamlee81 closed 6 years ago

chowkamlee81 commented 6 years ago

Hi ,

Gone thru your code. Great work and encouraging for the rest of the people to work on tensor decomposition. I had a doubt

Since this algorithm works layer by layer by decomposition of channels. Every layer number of channel input and output will be reduced. For next successive layer, If Prev_OUTPUT_Channel != Curr_INPUT_Channel, how coupling is done out? Kindly help me to understand

jacobgil commented 6 years ago

Hi,

Each layer will be decomposed into a block of several layers. The number of input channels to the new block, will be the number of input channels to the original layer. The number of output channels from the new block, will be the number of output channels from the original layer.

The reduction is done in the body of the block, but the input/output does not change (unlike in pruning for example).

Does this answer the question?

chowkamlee81 commented 6 years ago

Sounds good for me . Thanks for making me to understand