Caffe turn mxnet model --mmdnn mentioned mmdnn using the model conversion was not very clear command conversion, recently re-checked some, finishing at mmdnn conversion command.
Step conversion
mmdnn step conversion command is mmconvert, from mmdnn / conversion / _script / convert.py, whose parameters are:
| command | meaning |
|---|---|
| -sf | Enter the model type |
| -in | Enter the model of the network (if the network file independent) |
| -iw | Enter the weights file |
| -df | Output model type |
| -if | Output Model Name |
During generated .py, .npy intermediate structure model as a model of the target code form. In addition, like TensorFlow model need to specify the output node, such as -dstNode MMdnn_Output, but -dstNode I did not find the key value in the convert.py do not know running will not be a problem.
Example:
mmconvert -sf tensorflow -in imagenet_resnet_v2_152.ckpt.meta -iw imagenet_resnet_v2_152.ckpt --dstNode MMdnn_Output -df pytorch -om tf_to_pytorch_resnet_152.pth
python -m mmdnn.conversion._script.convertToIR -sf tensorflow -in imagenet_resnet_v2_152.ckpt.meta -iw imagenet_resnet_v2_152.ckpt --dstNode MMdnn_Output -df pytorch -om tf_to_pytorch_resnet_152.pth
Step conversion
1, turn IR
Other models to convert mmdnn IR command format is mmtoir, from mmdnn / conversion / _script / convertToIR.py, which parameters are:
| command | meaning |
|---|---|
| -f | Enter the model type |
| -n | Enter the model of the network (if the network file independent) |
| -w | Enter the weights file |
| –inputShape | Input Size (optional) |
| -node | An output node (TensorFlow) |
| -O | Intermediate file name .json, .pb, .npy |
Example:
mmtoir -f tensorflow -n imagenet_resnet_v2_152.ckpt.meta -w imagenet_resnet_v2_152.ckpt --inputShape 3,112,112 -node MMdnn_Output -o converted
python -m mmdnn.conversion._script.convertToIR -f tensorflow -n imagenet_resnet_v2_152.ckpt.meta -w imagenet_resnet_v2_152.ckpt --inputShape 3,112,112 -node MMdnn_Output -o converted
2, IR turn output model Code
mmdnn transform IR code format to another form of the model command mmtocode, from mmdnn / conversion / _script / IRToCode.py, which parameters are:
| command | meaning |
|---|---|
| -f | Output model type |
| -n | IR model structure |
| -w | IR model weights |
| -O | Output model structure |
| -dw | Output model weights (optional) |
Example:
mmtocode -f pytorch -n converted.pb -w converted.npy -o converted_pytorch.py -dw converted_pytorch.npy
python -m mmdnn.conversion._script. IRToCode -f pytorch -n converted.pb -w converted.npy -o converted_pytorch.py -dw converted_pytorch.npy
3. Export final model
mmdnn Export command the ultimate goal of the model is mmtomodel, from mmdnn / conversion / _script / IRToModel.py, whose parameters are:
| command | meaning |
|---|---|
| -f | Output model type |
| -in | Output model structure |
| -iw | IR model weights |
| -O | Output Model Name |
Example:
mmtomodel -f pytorch -in converted_pytorch.py -iw converted_pytorch.npy -o converted_pytorch.pth
python -m mmdnn.conversion._script. IRToModel -f pytorch -in converted_pytorch.py -iw converted_pytorch.npy -o converted_pytorch.pth
