Convert bactchnorm weights from caffe to pytorch

1

I am trying to convert caffemodel weights to pytorch and i am wondering what is num_batches_tracked in pytorch that is similar to caffe.
Here is my caffe BN prototxt

layer {
  name: "conv5_3_1x1_increase/bn"
  type: "BatchNorm"
  bottom: "conv5_3_1x1_increase"
  top: "conv5_3_1x1_increase"
  batch_norm_param {
    use_global_stats: true
  }
}
layer {
  name: "conv5_3_1x1_increase/bn/scale"
  type: "Scale"
  bottom: "conv5_3_1x1_increase"
  top: "conv5_3_1x1_increase"
  scale_param {
    bias_term: true
  }
}

I can convert caffemodel like this

for layer_name, param in caffe_model.items():
    if '/bn' in layer_name and '/scale' not in layer_name:
        factor = param[2].data[0]
        mean = np.array(param[0].data, dtype=np.float32) / factor
        variance = np.array(param[1].data, dtype=np.float32) / factor
    if '/scale' in layer_name:
        gamma = np.array(param[0].data, dtype=np.float32)
        beta = np.array(param[1].data, dtype=np.float32)
        params[layer_name + '.weight'] = gamma
        params[layer_name + '.bias'] = beta
        params[layer_name + '.running_mean'] = mean
        params[layer_name + '.running_var'] = variance

but pytorch has a variable called num_batches_tracked and how do i find this in caffe model. ?

2

I don’t think it’s in the caffe2 model. You probably don’t need it. It’s only used when computing running_mean/var with a simple moving average. By default we use an exponential moving average.

3

Interesting! Could you share to me know do you copy weight of convolution and concat layer in caffe to pytorch? I am working on it also. Thanks