shape inference error in depthwise_conv

[xianyuggg]

Platform: ubuntu

MMdnn version: 0.3.1

Source framework with version: Tensorflow 1.13.1

Pre-trained model: mobilenet_v1 from https://github.com/tensorflow/models/tree/master/research/slim/nets

Running scripts: mmtoir

Description: shape inference error in parsing depthwise_conv (converting from tf-slim model to IR file), the output shape's channel won't match with depthwise_conv's kernel channel

[JiahaoYao]

Have u ever tried this? https://github.com/microsoft/MMdnn/tree/master/mmdnn/conversion/tensorflow#frozen-graph-conversion

[xianyuggg]

Have u ever tried this? https://github.com/microsoft/MMdnn/tree/master/mmdnn/conversion/tensorflow#frozen-graph-conversion

Sorry I haven't tried this but it just seems the inferenced output shape for depthwise-conv got wrong mmtoir --srcFramework tensorflow -in ${ckpt_meta} -iw ${ckpt} --dstNodeName ${dst_node} -d ${dst_file_path}