pytorch2caffe工具使用教程及出现KeyError: 'AdaptiveAvgPool2dBackward'时解决方案
工具使用教程：
工具下载地址：https://github.com/longcw/pytorch2caffe
依赖包安装：
python环境：python2.7
配置caffe：具体方法请参考https://blog.csdn.net/xunan003/article/details/74905718
安装pytorch&torchvision（版本0.2.0）：pip install torch==0.2.0 torchvision
安装其他依赖工具：
sudo apt-get install graphziv
pip install pydot
pip install graphziv使用方法：
该工具中最主要的两个文件：prototxt.py和pytorch2caffe.py，前者不用理会，关键看后者，后者代码粘贴并解释如下：
使用指令：python pytorch2caffe.py
import sys
sys.path.append('/home/xn/caffe-ssd/python')
import caffe
from collections import OrderedDict
import torch.nn as nn
import torch.nn.functional as F
import torch
import numpy as np
from torch.autograd import Variable
from prototxt import *
import pydot
#以下为layer类型
layer_dict = {'ConvNdBackward': 'Convolution',
'ThresholdBackward': 'ReLU',
'MaxPool2dBackward': 'Pooling',
'AvgPool2dBackward': 'Pooling',
'AdaptiveAvgPool2dBackward': 'Pooling', #caffe中没有的自适应池化，我自己添加否则转化不成功。
'DropoutBackward': 'Dropout',
'AddmmBackward': 'InnerProduct',
'BatchNormBackward': 'BatchNorm',
'AddBackward': 'Eltwise',
'ViewBackward': 'Reshape',
'ConcatBackward': 'Concat',
'UpsamplingNearest2d': 'Deconvolution',
'UpsamplingBilinear2d': 'Deconvolution',
'SigmoidBackward': 'Sigmoid',
'LeakyReLUBackward': 'ReLU',
'NegateBackward': 'Power',
'MulBackward': 'Eltwise',
'SpatialCrossMapLRNFunc': 'LRN'}
layer_id = 0
def pytorch2caffe(input_var, output_var, protofile, caffemodel):
global layer_id
net_info = pytorch2prototxt(input_var, output_var)
print_prototxt(net_info)
save_prototxt(net_info, protofile)
if caffemodel is None:
return
net = caffe.Net(protofile, caffe.TEST)
params = net.params
layer_id = 1
seen = set()
def convert_layer(func):
if True:
global layer_id
parent_type = str(type(func).__name__)
if hasattr(func, 'next_functions'):
for u in func.next_functions:
if u[0] is not None:
child_type = str(type(u[0]).__name__)
child_name = child_type + str(layer_id)
if child_type != 'AccumulateGrad' and (
parent_type != 'AddmmBackward' or child_type != 'TransposeBackward'):
if u[0] not in seen:
convert_layer(u[0])
seen.add(u[0])
if child_type != 'ViewBackward':
layer_id = layer_id + 1
parent_name = parent_type + str(layer_id)
print('converting %s' % parent_name)
if parent_type == 'ConvNdBackward':
if func.next_functions[1][0] is not None:
weights = func.next_functions[1][0].variable.data
if func.next_functions[2][0]:
biases = func.next_functions[2][0].variable.data
else:
biases = None
save_conv2caffe(weights, biases, params[parent_name])
elif parent_type == 'BatchNormBackward':
running_mean = func.running_mean
running_var = func.running_var
bn_name = parent_name + "_bn"
save_bn2caffe(running_mean, running_var, params[bn_name])
affine = func.next_functions[1][0] is not None
if affine:
scale_weights = func.next_functions[1][0].variable.data
scale_biases = func.next_functions[2][0].variable.data
scale_name = parent_name + "_scale"
save_scale2caffe(scale_weights, scale_biases, params[scale_name])
elif parent_type == 'AddmmBackward':
biases = func.next_functions[0][0].variable.data
weights = func.next_functions[2][0].next_functions[0][0].variable.data
save_fc2caffe(weights, biases, params[parent_name])
elif parent_type == 'UpsamplingNearest2d':
print('UpsamplingNearest2d')
convert_layer(output_var.grad_fn)
print('save caffemodel to %s' % caffemodel)
net.save(caffemodel)
def save_conv2caffe(weights, biases, conv_param):
if biases is not None:
conv_param[1].data[...] = biases.numpy()
conv_param[0].data[...] = weights.numpy()
def save_fc2caffe(weights, biases, fc_param):
print(biases.size(), weights.size())
print(fc_param[1].data.shape)
print(fc_param[0].data.shape)
fc_param[1].data[...] = biases.numpy()
fc_param[0].data[...] = weights.numpy()
def save_bn2caffe(running_mean, running_var, bn_param):
bn_param[0].data[...] = running_mean.numpy()
bn_param[1].data[...] = running_var.numpy()
bn_param[2].data[...] = np.array([1.0])
def save_scale2caffe(weights, biases, scale_param):
scale_param[1].data[...] = biases.numpy()
scale_param[0].data[...] = weights.numpy()
def pytorch2prototxt(input_var, output_var):
global layer_id
net_info = OrderedDict()
props = OrderedDict()
props['name'] = 'pytorch'
props['input'] = 'data'
props['input_dim'] = input_var.size()
layers = []
layer_id = 1
seen = set()
top_names = dict()
def add_layer(func):
global layer_id
parent_type = str(type(func).__name__)
parent_bottoms = []
if hasattr(func, 'next_functions'):
for u in func.next_functions:
if u[0] is not None:
child_type = str(type(u[0]).__name__)
child_name = child_type + str(layer_id)
if child_type != 'AccumulateGrad' and (
parent_type != 'AddmmBackward' or child_type != 'TransposeBackward'):
if u[0] not in seen:
top_name = add_layer(u[0])
parent_bottoms.append(top_name)
seen.add(u[0])
else:
top_name = top_names[u[0]]
parent_bottoms.append(top_name)
if child_type != 'ViewBackward':
layer_id = layer_id + 1
parent_name = parent_type + str(layer_id)
layer = OrderedDict()
layer['name'] = parent_name
layer['type'] = layer_dict[parent_type]
parent_top = parent_name
if len(parent_bottoms) > 0:
layer['bottom'] = parent_bottoms
else:
layer['bottom'] = ['data']
layer['top'] = parent_top
if parent_type == 'MulBackward':
eltwise_param = {
'operation': 'PROD',
}
layer['eltwise_param'] = eltwise_param
elif parent_type == 'NegateBackward':
power_param = {
'power': 1,
'scale': -1.,
'shift': 0
}
layer['power_param'] = power_param
elif parent_type == 'LeakyReLUBackward':
negative_slope = func.additional_args[0]
layer['relu_param'] = {'negative_slope': negative_slope}
elif parent_type == 'UpsamplingNearest2d':
conv_param = OrderedDict()
factor = func.scale_factor
conv_param['num_output'] = func.saved_tensors[0].size(1)
conv_param['group'] = conv_param['num_output']
conv_param['kernel_size'] = (2 * factor - factor % 2)
conv_param['stride'] = factor
conv_param['pad'] = int(np.ceil((factor - 1) / 2.))
conv_param['weight_filler'] = {'type': 'bilinear'}
conv_param['bias_term'] = 'false'
layer['convolution_param'] = conv_param
layer['param'] = {'lr_mult': 0, 'decay_mult': 0}
elif parent_type == 'UpsamplingBilinear2d':
conv_param = OrderedDict()
factor = func.scale_factor[0]
conv_param['num_output'] = func.input_size[1]
conv_param['group'] = conv_param['num_output']
conv_param['kernel_size'] = (2 * factor - factor % 2)
conv_param['stride'] = factor
conv_param['pad'] = int(np.ceil((factor - 1) / 2.))
conv_param['weight_filler'] = {'type': 'bilinear'}
conv_param['bias_term'] = 'false'
layer['convolution_param'] = conv_param
layer['param'] = {'lr_mult': 0, 'decay_mult': 0}
elif parent_type == 'ConcatBackward':
concat_param = OrderedDict()
concat_param['axis'] = func.dim
layer['concat_param'] = concat_param
elif parent_type == 'ConvNdBackward':
# Only for UpsamplingCaffe
if func.transposed is True and func.next_functions[1][0] is None:
layer['type'] = layer_dict['UpsamplingBilinear2d']
conv_param = OrderedDict()
factor = func.stride[0]
conv_param['num_output'] = func.next_functions[0][0].saved_tensors[0].size(1)
conv_param['group'] = conv_param['num_output']
conv_param['kernel_size'] = (2 * factor - factor % 2)
conv_param['stride'] = factor
conv_param['pad'] = int(np.ceil((factor - 1) / 2.))
conv_param['weight_filler'] = {'type': 'bilinear'}
conv_param['bias_term'] = 'false'
layer['convolution_param'] = conv_param
layer['param'] = {'lr_mult': 0, 'decay_mult': 0}
else:
weights = func.next_functions[1][0].variable
conv_param = OrderedDict()
conv_param['num_output'] = weights.size(0)
conv_param['pad_h'] = func.padding[0]
conv_param['pad_w'] = func.padding[1]
conv_param['kernel_h'] = weights.size(2)
conv_param['kernel_w'] = weights.size(3)
conv_param['stride'] = func.stride[0]
conv_param['dilation'] = func.dilation[0]
if func.next_functions[2][0] == None:
conv_param['bias_term'] = 'false'
layer['convolution_param'] = conv_param
elif parent_type == 'BatchNormBackward':
bn_layer = OrderedDict()
bn_layer['name'] = parent_name + "_bn"
bn_layer['type'] = 'BatchNorm'
bn_layer['bottom'] = parent_bottoms
bn_layer['top'] = parent_top
batch_norm_param = OrderedDict()
batch_norm_param['use_global_stats'] = 'true'
batch_norm_param['eps'] = func.eps
bn_layer['batch_norm_param'] = batch_norm_param
affine = func.next_functions[1][0] is not None
# func.next_functions[1][0].variable.data
if affine:
scale_layer = OrderedDict()
scale_layer['name'] = parent_name + "_scale"
scale_layer['type'] = 'Scale'
scale_layer['bottom'] = parent_top
scale_layer['top'] = parent_top
scale_param = OrderedDict()
scale_param['bias_term'] = 'true'
scale_layer['scale_param'] = scale_param
else:
scale_layer = None
elif parent_type == 'ThresholdBackward':
parent_top = parent_bottoms[0]
elif parent_type == 'MaxPool2dBackward':
pooling_param = OrderedDict()
pooling_param['pool'] = 'MAX'
pooling_param['kernel_size'] = func.kernel_size[0]
pooling_param['stride'] = func.stride[0]
# http://netaz.blogspot.com/2016/08/confused-about-caffes-pooling-layer.html
padding = func.padding[0]
# padding = 0 if func.padding[0] in {0, 1} else func.padding[0]
pooling_param['pad'] = padding
layer['pooling_param'] = pooling_param
elif parent_type == 'AvgPool2dBackward':
pooling_param = OrderedDict()
pooling_param['pool'] = 'AVE'
pooling_param['kernel_size'] = func.kernel_size[0]
pooling_param['stride'] = func.stride[0]
pooling_param['pad'] = func.padding[0]
layer['pooling_param'] = pooling_param
#自己添加的AdaptiveAvgPool模块，vgg16_bn model中有该层参数，但原始code没有，不添加转换不成功。
elif parent_type == 'AdaptiveAvgPool2dBackward':
pooling_param = OrderedDict()
pooling_param['pool'] = 'MAX'
pooling_param['kernel_size'] = 1
layer['pooling_param'] = pooling_param
elif parent_type == 'DropoutBackward':
parent_top = parent_bottoms[0]
dropout_param = OrderedDict()
dropout_param['dropout_ratio'] = func.p
layer['dropout_param'] = dropout_param
elif parent_type == 'AddmmBackward':
inner_product_param = OrderedDict()
inner_product_param['num_output'] = func.next_functions[0][0].variable.size(0)
layer['inner_product_param'] = inner_product_param
elif parent_type == 'ViewBackward':
parent_top = parent_bottoms[0]
elif parent_type == 'AddBackward':
eltwise_param = OrderedDict()
eltwise_param['operation'] = 'SUM'
layer['eltwise_param'] = eltwise_param
elif parent_type == 'SpatialCrossMapLRNFunc':
layer['lrn_param'] = {
'local_size': func.size,
'alpha': func.alpha,
'beta': func.beta,
}
layer['top'] = parent_top # reset layer['top'] as parent_top may change
if parent_type != 'ViewBackward':
if parent_type == "BatchNormBackward":
layers.append(bn_layer)
if scale_layer is not None:
layers.append(scale_layer)
else:
layers.append(layer)
# layer_id = layer_id + 1
top_names[func] = parent_top
return parent_top
add_layer(output_var.grad_fn)
net_info['props'] = props
net_info['layers'] = layers
return net_info
def plot_graph(top_var, fname, params=None):
"""
This method don't support release v0.1.12 caused by a bug fixed in: https://github.com/pytorch/pytorch/pull/1016
So if you want to use `plot_graph`, you have to build from master branch or wait for next release.
Plot the graph. Make sure that require_grad=True and volatile=False
:param top_var: network output Varibale
:param fname: file name
:param params: dict of (name, Variable) to add names to node that
:return: png filename
"""
from graphviz import Digraph
import pydot
dot = Digraph(comment='LRP',
node_attr={'style': 'filled', 'shape': 'box'})
# , 'fillcolor': 'lightblue'})
seen = set()
if params is not None:
assert isinstance(params.values()[0], Variable)
param_map = {id(v): k for k, v in params.items()}
def size_to_str(size):
return '(' + (', ').join(['%d' % v for v in size]) + ')'
def add_nodes(var):
if var not in seen:
if torch.is_tensor(var):
dot.node(str(id(var)), size_to_str(var.size()), fillcolor='orange')
elif hasattr(var, 'variable'):
u = var.variable
name = '{}\n '.format(param_map[id(u)]) if params is not None else ''
node_name = '{}{}'.format(name, size_to_str(u.size()))
dot.node(str(id(var)), node_name, fillcolor='lightblue')
else:
dot.node(str(id(var)), str(type(var).__name__))
seen.add(var)
if hasattr(var, 'next_functions'):
for u in var.next_functions:
if u[0] is not None:
dot.edge(str(id(u[0])), str(id(var)))
add_nodes(u[0])
if hasattr(var, 'saved_tensors'):
for t in var.saved_tensors:
dot.edge(str(id(t)), str(id(var)))
add_nodes(t)
add_nodes(top_var.grad_fn)
dot.save(fname)
(graph,) = pydot.graph_from_dot_file(fname)
im_name = '{}.png'.format(fname)
graph.write_png(im_name)
print(im_name)
return im_name
#该部分为主函数
if __name__ == '__main__':
import torchvision
import os
#m = '/home/xn/pytorch2caffe/vgg16_bn-6c64b313.pth'
#m为从torchvision.models中下载的pytorch pth文件，你可在/usr/local/lib/python2.7/dist-packages/torchvision/models文件夹（你安装的torchvision路径）中找到相关的网络，这里vgg16_bn是我要加载的model名称，/models文件夹下没有的则无法加载，需要自己定义并添加相关code。括号中的参数（pretrained=True）在models/vgg.py中vgg16_bn函数中有定义，若此处碰到缺少参数或者多出参数的情况请根据models文件夹下具体网络中的定义修改。
m = torchvision.models.vgg16_bn(pretrained=True)
#m = torchvision.models.inception_v3(pretrained=True, transform_input=False)
m.eval()
print(m)
#网络输入的大小，需提前掌握，否则会转换不成功。
input_var = Variable(torch.rand(1, 3, 224, 224))
output_var = m(input_var)
# plot graph to png
#可视化图片的具体路径
output_dir = '/home/xn/pytorch2caffe/demo/'
plot_graph(output_var, os.path.join(output_dir, 'vgg16_bn.dot'))
pytorch2caffe(input_var, output_var, os.path.join(output_dir, 'vgg16_bn-pytorch2caffe.prototxt'),
os.path.join(output_dir, 'vgg16_bn-pytorch2caffe.caffemodel'))KeyError: 'AdaptiveAvgPool2dBackward'解决方案：
1. 屏蔽掉上述代码中def pytorch2caffe函数中：
net_info = pytorch2prototxt(input_var, output_var)
print_prototxt(net_info)
save_prototxt(net_info, protofile)即不写对应的prototxt文件，只转换出caffemodel。此方法无法获得model中每层的具体参数或名称。
2. 手动添加AdaptiveAvgPool2dBackward相关内容。
在第14行layer_dict中添加参数
'AdaptiveAvgPool2dBackward': 'Pooling',在第292行add_layer函数中添加内容：
elif parent_type == 'AdaptiveAvgPool2dBackward':
pooling_param = OrderedDict()
pooling_param['ouput_size'] = func.output_size[0]
layer['pooling_param'] = pooling_param因为在pytorch中自适应池化AdaptiveAvgPool参数只有output_size。更改之后执行程序出错：
Error parsing text-format caffe.NetParameter: 594:20: Message type "caffe.PoolingParameter" has no field named "output_size".大概的原因是：由于caffe中关于pooling类型的层只有max和ave两种类型，并不支持AdaptiveAvgPool，所以会提示没有名为output_size的参数。
我们将292行修改的部分中 pooling_param['ouput_size'] = func.output_size[0] 删除，则报如下错误
pool_param.has_kernel_size() != !(pool_param.has_kernel_h() && pool_param.has_kernel_w()) Filter size is kernel_size OR kernel_h and kernel_w; not both大概的原因是：由于caffe中max和ave两种类型必须要有池化核kernel才能计算，所以会提示需要参数kernel_size。
然后我们又修改如下：
elif parent_type == 'AdaptiveAvgPool2dBackward':
pooling_param = OrderedDict()
pooling_param['type'] = 'MAX'
pooling_param['kernel_size'] = func.kernel_size[0]
layer['pooling_param'] = pooling_param
继续出错，大概的原因是从model中获取不到kernel_size和type参数，因为AdaptiveAvgPool并没有这两个参数。
被逼无奈，查看输出日志，发现:
I0424 15:33:15.656667 3208 net.cpp:408] MaxPool2dBackward44 -> MaxPool2dBackward44
I0424 15:33:15.656678 3208 net.cpp:150] Setting up MaxPool2dBackward44
I0424 15:33:15.656684 3208 net.cpp:157] Top shape: 1 512 7 7 (25088)
I0424 15:33:15.656688 3208 net.cpp:165] Memory required for data: 223483904
I0424 15:33:15.656692 3208 layer_factory.hpp:77] Creating layer AdaptiveAvgPool2dBackward45
I0424 15:33:15.656698 3208 net.cpp:100] Creating Layer AdaptiveAvgPool2dBackward45
I0424 15:33:15.656703 3208 net.cpp:434] AdaptiveAvgPool2dBackward45 <- MaxPool2dBackward44
I0424 15:33:15.656710 3208 net.cpp:408] AdaptiveAvgPool2dBackward45 -> AdaptiveAvgPool2dBackward45
I0424 15:33:15.656725 3208 net.cpp:150] Setting up AdaptiveAvgPool2dBackward45
I0424 15:33:15.656733 3208 net.cpp:157] Top shape: 1 512 7 7 (25088)发现AdaptiveAvgPool2dBackward层和其前一层MaxPool2dBackward输出的特征图size大小相同，所以我用MAX池化将其kernel_size设置为1，可代替AdaptiveAvgPool层，所以最终修改为：
elif parent_type == 'AdaptiveAvgPool2dBackward':
pooling_param = OrderedDict()
pooling_param['pool'] = 'MAX'
pooling_param['kernel_size'] = 1
layer['pooling_param'] = pooling_param注意与上一次修改的不同之处，上一次修改kernel_size参数从pytorch model中抓取，而本次为恒定赋值为1。
至此pytorch2caffe成功转换，可视化图如下：
其他错误：
ValueError: could not broadcast input array from shape (4096,25088) into shape (4096,51200)
数据输入错误，具体设置
input_var = Variable(torch.rand(1, 3, 224, 224))