[SINGA-278] Convert trained caffe parameters to singa

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Assignee: Xiangrui
• Reporter: Xiangrui
• Created: 03/Dec/16 02:48
• Updated: 11/Jan/17 06:46
• Resolved: 11/Jan/17 06:46

Description

Convert trained parameters of caffe model to singa.
Run vgg as an example. Some tricks should be noticed:
1. The order of image axes in caffe is height, width and channels due to opencv implementation, while it is width, height, channels in singa if you use python PIL.
2. Another problem caused by these two libraries is the order of channels, BGR(caffe, opencv) v.s. RGB(singa, PIL).
3. It needs to transpose the weight tensor in InnerProduct(Dense) layer.