Skip to content

Latest commit

 

History

History
119 lines (78 loc) · 4.97 KB

README.md

File metadata and controls

119 lines (78 loc) · 4.97 KB
Raw

Caffe README

Caffe pre-trained model

We tested some Caffe pre-trained models to others, get more detail from this file

Models Caffe Keras Tensorflow CNTK MXNet PyTorch CoreML ONNX
Vgg19 √ √ √ √ √ √ √ √
AlexNet √ √ √ √ √ √ √
Inception_v1 √ √ √ √ √ √ √ √
Inception_v4 √ √ o √ √ √
ResNet 152 √ √ √ √ √ √ √ √
SqueezeNet √ √ √ √ √ √ √ √
Xception √ √ √ √ √ √
voc-FCN8s/FCN16s/FCN32s √ √

√ - Correctness tested

o - Some difference after conversion

space - not tested

Usage

Download Caffe pre-trained model

$ mmdownload -f caffe

Supported models: ['resnet50', 'resnet101', 'voc-fcn8s', 'inception_v4', 'voc-fcn16s', 'squeezenet', 'alexnet', 'inception_v1', 'resnet152', 'xception', 'vgg16', 'vgg19', 'voc-fcn32s']

$ mmdownload -f caffe -n resnet152 -o ./

Downloading file [./resnet152-deploy.prototxt] from [http://data.mxnet.io/models/imagenet/test/caffe/ResNet-152-deploy.prototxt]
100% [..............................................................................] 98034 / 98034
Downloading file [./resnet152.caffemodel] from [http://data.mxnet.io/models/imagenet/test/caffe/ResNet-152-model.caffemodel]
100% [......................................................................] 241444171 / 241444171
Caffe Model resnet152 saved as [./resnet152-deploy.prototxt] and [./resnet152.caffemodel].

One-step conversion

Above MMdnn@0.1.4, we provide one command to achieve the conversion

$ mmconvert -sf caffe -in resnet152-deploy.prototxt -iw resnet152.caffemodel -df cntk -om caffe_resnet152.dnn
.
.
.
CNTK model file is saved as [caffe_resnet152.dnn], generated by [069867aa7f674b73aebaf9de34cf2507.py] and [069867aa7f674b73aebaf9de34cf2507.npy].

Then you get the CNTK original model caffe_resnet152.dnn converted from Caffe. Temporal files are removed automatically.

If you want to assume a fixed inputshape, you can use "--inputShape 224,224,3"

Step-by-step conversion (for debugging)

Convert models from Caffe to IR (Caffe -> IR)

You can use following bash command to convert the network architecture [resnet152-deploy.prototxt] with weights [resnet152.caffemodel] to IR architecture file [caffe_resnet_IR.pb], [caffe_resnet_IR.json] and IR weights file [caffe_resnet_IR.npy]

$ mmtoir -f caffe -n resnet152-deploy.prototxt -w resnet152.caffemodel -o caffe_resnet_IR
.
.
.
IR network structure is saved as [caffe_resnet_IR.json].
IR network structure is saved as [caffe_resnet_IR.pb].
IR weights are saved as [caffe_resnet_IR.npy].

Convert models from IR to Caffe code (IR -> Caffe)

You can use following bash command to convert the IR architecture file [caffe_resnet_IR.pb] and weights file [caffe_resnet_IR.npy] to Caffe Python code file[caffe_converted.py] and IR weights file suit for caffe model[caffe_converted.npy]

$ mmtocode -f caffe -n caffe_resnet_IR.pb -w caffe_resnet_IR.npy -d caffe_converted.py -dw caffe_converted.npy

Parse file [caffe_resnet_IR.pb] with binary format successfully.
Target network code snippet is saved as [caffe_converted.py].
Target weights are saved as [caffe_converted.npy].

Generate Caffe model from code snippet file and weight file

You can use following bash command to generate caffe architecture file [caffe_target.prototxt] and weights file [caffe_target.caffemodel] from python code [caffe_converted.py] and weights file [caffe_converted.npy] for further usage.

$ mmtomodel -f caffe -in caffe_converted.py -iw caffe_converted.npy -o caffe_target

Caffe model files are saved as [caffe_target.prototxt] and [caffe_target.caffemodel], generated by [caffe_converted.py] and [caffe_converted.npy].

Then you can use Caffe to load the caffe_target.prototxt and caffe_target.caffemodel files for future use.

Develop version

Ubuntu 16.04 with

  • Caffe 1.0.0

@ 05/03/2018

Limitation

  • Currently no RNN related operations support

Acknowledgement

Thanks to Saumitro Dasgupta, the initial code of caffe-to-tensorflow references to his project caffe-tensorflow.