How to convert standard densenet121 pretrained model to efficient version?

[ZhengRui]

I tried to convert pretrained densenet121 model provided in https://github.com/shicai/DenseNet-Caffe
to efficient version obeying your DenseBlock naming convention. I have the following prototxt (efficient_densenet121.prototxt) and script to copy params (from standard DenseNet_121.prototxt and corresponding model) until the end of first transition layer:

name: "densenet121_efficient"

input: "data"
input_dim: 1
input_dim: 3
input_dim: 224
input_dim: 224

layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  convolution_param {
    num_output: 64
    bias_term: false
    pad: 3
    kernel_size: 7
    stride: 2
  }
}
layer {
  name: "conv1/bn"
  type: "BatchNorm"
  bottom: "conv1"
  top: "conv1/bn"
  batch_norm_param {
    eps: 1e-5
  }
}
layer {
  name: "conv1/scale"
  type: "Scale"
  bottom: "conv1/bn"
  top: "conv1/bn"
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu1"
  type: "ReLU"
  bottom: "conv1/bn"
  top: "conv1/bn"
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1/bn"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
    pad: 1
    ceil_mode: false
  }
}

# dense block 1
layer {
  name: "denseblock1"
  type: "DenseBlock"
  bottom: "pool1"
  top: "concat_2_6"
  denseblock_param {
    numTransition: 6
    initChannel: 64
    growthRate: 32
    use_dropout: false
    use_BC: true
    BC_ultra_space_efficient: true
  }
}

# transition layer 1
layer {
  name: "conv2_blk/bn"
  type: "BatchNorm"
  bottom: "concat_2_6"
  top: "conv2_blk/bn"
  batch_norm_param {
    eps: 1e-5
  }
}
layer {
  name: "conv2_blk/scale"
  type: "Scale"
  bottom: "conv2_blk/bn"
  top: "conv2_blk/bn"
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu2_blk"
  type: "ReLU"
  bottom: "conv2_blk/bn"
  top: "conv2_blk/bn"
}
layer {
  name: "conv2_blk"
  type: "Convolution"
  bottom: "conv2_blk/bn"
  top: "conv2_blk"
  convolution_param {
    num_output: 128
    bias_term: false
    kernel_size: 1
  }
}
layer {
  name: "pool2"
  type: "Pooling"
  bottom: "conv2_blk"
  top: "pool2"
  pooling_param {
    pool: AVE
    kernel_size: 2
    stride: 2
  }
}

script to test copying params:

import sys
sys.path.insert(0, './efficient-caffe/python')
import caffe
import numpy as np
import google.protobuf as pb2

import ipdb

##
# @brief
#
# @param i ith denseblock, 1st denseblock prefix: conv2_, 2nd denseblock prefix: conv3_
# @param j jth subblock, prefix: convi_j/
# @param k which param in efficient denseblock
#           0: 3x3 filter, 1/2: scale/bias of scale layer before 1x1 conv
#           3/4: global mean/var of bn layer before 1x1 conv
#           5: 1x1 filter, 6/7: scale/bias of scale layer before 3x3 conv
#           8/9: global mean/var of bn layer before 3x3 conv
#           10: moving average factor of bn layers
#
# @return layer_name in standard source net
def find_src_layer(i, j, k):
    layer_name = 'conv{}_{}/'.format(i,j)
    idx = 0
    assert k in np.arange(11)
    if k == 0:
        layer_name += 'x2'
    elif k == 1:
        layer_name += 'x1/scale'
    elif k == 2:
        layer_name += 'x1/scale'
        idx = 1
    elif k == 3:
        layer_name += 'x1/bn'
    elif k == 4:
        layer_name += 'x1/bn'
        idx = 1
    elif k == 5:
        layer_name += 'x1'
    elif k == 6:
        layer_name += 'x2/scale'
    elif k == 7:
        layer_name += 'x2/scale'
        idx = 1
    elif k == 8:
        layer_name += 'x2/bn'
    elif k == 9:
        layer_name += 'x2/bn'
        idx = 1
    else:
        layer_name += 'x1/bn'
        idx = 2

    return layer_name, idx

caffe.set_mode_gpu()
net_dst = caffe.Net('./efficient_densenet121.prototxt', './DenseNet_121.caffemodel', caffe.TEST)
net_src = caffe.Net('./DenseNet_121.prototxt', './DenseNet_121.caffemodel', caffe.TEST)

net_dst_proto = caffe.proto.caffe_pb2.NetParameter()
with open('./efficient_densenet121.prototxt', 'rb') as fd:
    pb2.text_format.Merge(fd.read(), net_dst_proto)

j = 1
# copy denseblock params
for i, layer in enumerate(net_dst.layers):
    if layer.type == 'DenseBlock':
        # params are saved in layer.blobs
        layer_proto = net_dst_proto.layer[i-1]
        repeat = layer_proto.denseblock_param.numTransition
        j += 1

        if layer_proto.denseblock_param.use_BC:
            for k, param in enumerate(layer.blobs):
                nth_param = k / repeat
                nth_repeat = k % repeat + 1
                src_layer, nth = find_src_layer(j, nth_repeat, nth_param)
                # print src_layer, nth, nth_repeat, nth_param
                assert param.data.shape == net_src.params[src_layer][nth].data.shape or \
                        param.data.size == net_src.params[src_layer][nth].data.size
                param.data[:] = net_src.params[src_layer][nth].data.copy().reshape(param.data.shape)
        else:
            pass

        # sanity check
        inp = np.ones((1,3,224,224))
        o_dst = net_dst.forward(data=inp)
        o_src = net_src.forward(data=inp)
        top_blob = net_dst.top_names[net_dst._layer_names[i]][0] # concat_6_2 layer for first denseblock
        print top_blob, net_dst.blobs[top_blob].data.mean(), net_src.blobs[top_blob].data.mean(), \
                net_dst.blobs[top_blob].data.std(), net_src.blobs[top_blob].data.std()

However, when I run the above script, the mean and std of 'concat_2_6' layer does not match, not very big difference but obviously some issues exist when copying parameters, especially I found the result of efficient version seems irrelevant to the value of final parameter of DenseBlock layer. In https://github.com/Tongcheng/caffe/blob/master/src/caffe/layers/DenseBlock_layer.cpp#L159 it says it is related to batch norm layer's moving average factor, I am wondering is it because in standard densenet all batch norm layers's last parameter is the same, so in this efficient implementation it only has one parameter? Other than that, I still don't know why its value doesn't affect output and what will be the correct mapping of parameters from standard model to efficient model?

[szk495732563]

Hi, have you solved this problem yet? I need the pretrained model of efficient version too, can you help me? thanks @ZhengRui

[ZhengRui]

Finally I decided to use the Pytorch pre-trained model, https://github.com/gpleiss/efficient_densenet_pytorch/issues/13