Closed biycyr closed 5 years ago
It seems that tensorflow model has merge batchnorm of Conv with convolution of Conv. FeatureExtractor/MobilenetV2/Conv/BatchNorm/batchnorm/mul_1 and FeatureExtractor/MobilenetV2/Conv/BatchNorm/batchnorm/add_1 are now Conv2D op and addtion op. Does this mean the load_caffe_weights.py needs a rewrite? @chuanqi305
How to solve the problem?
I didn't. I choose to use Mobilenetv2_SSD(not ssdlite). In which, batchnorm layers are not merged yet. Although you may still need to change the way of naming .dat in dump_tensorflow_weights.py, at least it is easier to change the network structure.
My fault. Just add --nobn true while using gen_model.py
I have this problem while using load_caffe_weights.py. I have checked ssdlite_mobilenet_v2_coco_2018_05_09/frozen_inference_graph.pb and there is no tensor with 'Conv' and 'moving_mean' in the name. Does any one have similar problem?