Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
Description
The dimension of the weight matrix of the Dense layer in V1.0 is (num_input, num_output), whereas it is (num_output, num_input) in Caffe and Torch. To make it easy to convert caffe models, we can change the dimension order.
To be compatible with previous checkpoint files which were generated with (num_input, num_output), we need to change the version ID to (1.0.1) and add it into the checkpoint file via SINGA-261. The net load() function should be updated to convert the weight matrix according to the version ID: transpose the weight matrix if it was dumped from v1.0.0; do nothing for >=v1.0.1.