Improved caffe converter #6822
Improved caffe converter #6822
Conversation
|
Please revert the submodule change. |
|
Done. |
|
build still fails, @piiswrong can you check the log to see if this is something I can fix? |
|
Error is due to you changed submodule: https://github.com/dmlc/mxnet/pull/6822/files |
|
But I reverted this change, or so I thought. |
|
Try checkout upstream master and cherry pick your changes. |
- added support for networks which uses batch normalization without a scale layer following the batch norm, i.e. gamma is fixed to 1 - extended naming convention used when implementing batch normalization in caffe - added support for old caffe versions where dilation didn't exist. This is needed to convert models which depends on old caffe - added support for deconvolution layer - added support for older version of caffe where kernel_size, pad and stride parameters were not iterable - fixed crash happening when a bottom layer doesn't exist in the internal top_to_layers dictionary, this can happen if the name of the input is not 'data' - added ignore-by-design support for converting 'Crop' layers - fixed batch norm layer comparison to take into account the rescaling factor - added careful condition in tester to swap (RGB,BGR) input channels only if they are of size 3 or 4, which is the same check the conversion does - allow comparing layers of models with no mean file - added support for comparing the parameters of deconvolution layers
arikpoz
force-pushed
the
improved_caffe_converter
branch
from
bb5af5c to
c671c9e
Compare
|
submodules issue seem resolved. |
|
Is caffe converter tested? Maybe we should setup CI? @mli Could you have a look? |
|
The only error is due to test timeout, can someone check this, maybe update the timeout? |
|
the caffe converter is tested in the CI, but we only tried to convert a few commonly cnns, such as vgg/resnet. Each test is expensive. @arikpoz do you have a better idea how to test the converter? |
|
The current test seems good to me, checking both performance and layer by layer outputs. What takes times is the performance test since we run inference on many images. The layer by layer test is done on a single image comparing all the outputs (and network parameters) of the caffe network and the mxnet network. Since we run inference on one image only, it is very fast to execute. my 2c. |
…he#6822) - added support for networks which uses batch normalization without a scale layer following the batch norm, i.e. gamma is fixed to 1 - extended naming convention used when implementing batch normalization in caffe - added support for old caffe versions where dilation didn't exist. This is needed to convert models which depends on old caffe - added support for deconvolution layer - added support for older version of caffe where kernel_size, pad and stride parameters were not iterable - fixed crash happening when a bottom layer doesn't exist in the internal top_to_layers dictionary, this can happen if the name of the input is not 'data' - added ignore-by-design support for converting 'Crop' layers - fixed batch norm layer comparison to take into account the rescaling factor - added careful condition in tester to swap (RGB,BGR) input channels only if they are of size 3 or 4, which is the same check the conversion does - allow comparing layers of models with no mean file - added support for comparing the parameters of deconvolution layers
Extended caffe-to-mxnet converter and improved converter test