Gluon Converter

Convert MXNet-Gluon model to Caffe.

Requirements

• python >= 3.6 (f-string is used in codes)
• mxnet
• gluoncv
• numpy
• caffe(optional)

Usage

1. Construct your gluon model, for example

   from gluoncv import resnet18_v1
   net = resnet18_v1(pretrained=True)
2. Convert it to caffe NetParamert, for example

   from convert import convert_model
   text_net, binary_weights = convert_model(net, input_shape=(1,3,224,224), softmax=False, to_bgr=True, merge_bn=True)
   For caffe, the order of inputs' channels is often BGR but not RGB.
   if you want to convert a ssd-like model in gluoncv, please use convert_ssd_model API but not convert_model.
3. Save to files, for example

   from convert import save_model
   save_model(text_net, binary_weights, prefix="tmp/resnet18_v1")

How do I convert a ssd-like model?

1. To fetch attributes needed by PriorBox and DetectionOutput layers, convert_ssd_model will extract them from gluon-net and anchors, box_decoder, class_decoder in it.
2. Before convert symbols to caffemodel, fake symbols for priorbox and detction_output are added into the origin symbols.
3. Since step could not be extract from anchors in gluon-net, it will be setted by default in caffe (step=img_size/layer_size, refer to caffe-ssd/prior_box_layer.cpp).
   You can create an instance of gluoncv.model_zoo.SSD and train it as gluoncv/scripts/detection/ssd/train_ssd.py, for example, ssd300_mobilenetv2 --

   from gluoncv.model_zoo import SSD
   image_size = 300
   layer_size = (19, 10, 5, 3, 2, 1)
   net = SSD(network="mobilenetv2_1.0", 
         base_size=image_size, 
         features=['features_linearbottleneck12_elemwise_add0_output',     # FeatureMap: 19x19
                   'features_linearbottleneck16_batchnorm2_fwd_output'],   # FeatureMap: 10x10
         num_filters=[256, 256, 128, 128],    # Expand feature extractor with FeatureMaps: 5x5, 3x3, 2x2, 1x1 (stride=2)
         sizes=[21, 45, 99, 153, 207, 261, 315],
         ratios=[[1, 2, 0.5]] + [[1, 2, 0.5, 3, 1.0/3]] * 3 + [[1, 2, 0.5]] * 2,
         steps=[image_size/layer_size for layer_size in layer_size],   # Default setting in DetectionOutput caffe-layer
         classes=['A', 'B', 'C'],
         pretrained=True)
   # ...train as train_ssd.py

I've tested the ssd models converted from gluoncv on caffe-ssd and ncnn and they works well.

How to convert MobileNetv2?

ReLU6 is one of components in MobileNetv2, which is implemented with a clip symbol with range [0,6]. But caffe does not support clip. Therefore, to convert MobileNetv2, converter will replace clip symbol with range [0,6] with Activation(relu). And of course, some errors will be introduced especially for quantized-models.
However, as I know, some branches of caffe and some platform(such as ncnn) support ReLU6, please reset the type of activation layers manually if you want to deploy it to such branches or platforms.

Support Layers

• Convolution -> Convolution
• BatchNorm -> BatchNorm & Scale
• Activation (relu only) -> ReLU
• Pooling -> Pooling (MAX/AVG)
• elemwise_add -> Eltwise (ADD)
• FullyConnected -> InnerProduct
• Flatten -> Flatten
• Concat -> Concat
• Dropout -> Dropout
• softmax -> Softmax
• transpose -> Permute (caffe-ssd)
• Reshape -> Reshape (caffe-ssd)
• ReLU6 -> ReLU