Pytorch官方目前无法像tensorflow,咖啡那样直接给出形状信息,详见
https://github.com/pytorch/pytorch/pull/3043
以下代码算一种解决方案。由于CNN, RNN等模块实现不一样,添加其他模块支持可能需要改代码。
例如RNN中偏见是bool类型,其权重也不是存于体重属性中,不过我们只关注形状够用了。
该方法必须构造一个输入调用向前后(模型(x)调用)才可获取形状
utf - 8编码:
从进口OrderedDict集合
进口火炬
从火炬。autograd导入变量
进口火炬。神经网络是神经网络
导入模型。crnn作为crnn
进口json
def get_output_size (summary_dict、输出):
如果isinstance(输出元组):
因为我在xrange (len(输出)):
summary_dict[我]=OrderedDict ()
summary_dict[我]=get_output_size (summary_dict[我],[我])输出
其他:
summary_dict [' output_shape ']=列表(output.size ())
返回summary_dict
def总结(input_size、模型):
def register_hook(模块):
def钩(模块、输入、输出):
class_name=str (module.__class__) .split (“。”) [1] .split(“的”)[0]
module_idx=len(总结)
m_key=' % s - %我' % (class_name module_idx + 1)
摘要[m_key]=OrderedDict ()
摘要[m_key] [' input_shape ']=列表(输入[0].size ())
摘要[m_key]=get_output_size(总结(m_key),输出)
params=0
“体重”如果hasattr(模块):
params +=torch.prod (torch.LongTensor(列表(module.weight.size ())))
如果module.weight.requires_grad:
摘要[m_key]['学习']=True
其他:
摘要[m_key]['学习']=False
#如果hasattr(模块,“偏见”):
#参数+=torch.prod (torch.LongTensor(列表(module.bias.size ())))
摘要[m_key] [' nb_params ']=参数
如果不是isinstance(模块,nn.Sequential)和\
不是isinstance(模块,nn.ModuleList)和\==模型(模块):
hooks.append (module.register_forward_hook(钩))
#检查如果有多个输入到网络
如果isinstance (input_size[0],(列表、元组)):
x=[变量(torch.rand (1 * in_size))在input_size in_size]
其他:
x=变量(torch.rand (1 * input_size))
#创建属性
摘要=OrderedDict ()
钩子=[]
#注册钩
model.apply (register_hook)
#传球前进
模型(x)
#删除这些钩子
钩子的h:
h.remove ()
返回的总结
crnn=crnn。CRNN (32, 3755, 256, (1)
x=总结([32128],crnn)
打印json.dumps (x)
之前
以pytorch版CRNN为例,输出形状如下
{
" Conv2d-1 ": {
“input_shape”: 1, 1, 128,
“output_shape”: (64, 128),
“学”:没错,
“nb_params”: 576
},
" ReLU-2 ": {
“input_shape”: (64, 128),
“output_shape”: (64, 128),
“nb_params”: 0
},
" MaxPool2d-3 ": {
“input_shape”: (64, 128),
“output_shape”: (64, 64),
“nb_params”: 0
},
" Conv2d-4 ": {
“input_shape”: (64, 64),
“output_shape”: (128, 64),
“学”:没错,
“nb_params”: 73728
},
" ReLU-5 ": {
“input_shape”: (128, 64),
“output_shape”: (128, 64),
“nb_params”: 0
},
" MaxPool2d-6 ": {
“input_shape”: (128, 64),
“output_shape”: (1 128 8, 32),
“nb_params”: 0
},
" Conv2d-7 ": {
“input_shape”: (1 128 8, 32),
“output_shape”: (1 256 8, 32),
“学”:没错,
“nb_params”: 294912
},
" BatchNorm2d-8 ": {
“input_shape”: (1 256 8, 32),
“output_shape”: (1 256 8, 32),
“学”:没错,
“nb_params”: 256
},
" ReLU-9 ": {
“input_shape”: (1 256 8, 32),
“output_shape”: (1 256 8, 32),
“nb_params”: 0
},
" Conv2d-10 ": {
“input_shape”: (1 256 8, 32),
“output_shape”: (1 256 8, 32),
“学”:没错,
“nb_params”: 589824
},
" ReLU-11 ": {
“input_shape”: (1 256 8, 32),
“output_shape”: (1 256 8, 32),
“nb_params”: 0
},
" MaxPool2d-12 ": {
“input_shape”: (1 256 8, 32),
“output_shape”: 1 256 4, 33岁,
“nb_params”: 0
},
" Conv2d-13 ": {
“input_shape”: 1 256 4, 33岁,
“output_shape”: 1 512 4, 33岁,
“学”:没错,
“nb_params”: 1179648
},
" BatchNorm2d-14 ": {
“input_shape”: 1 512 4, 33岁,
“output_shape”: 1 512 4, 33岁,
“学”:没错,
“nb_params”: 512
},
" ReLU-15 ": {
“input_shape”: 1 512 4, 33岁,
“output_shape”: 1 512 4, 33岁,
“nb_params”: 0
},
" Conv2d-16 ": {
“input_shape”: 1 512 4, 33岁,
“output_shape”: 1 512 4, 33岁,
“学”:没错,
“nb_params”: 2359296
},
" ReLU-17 ": {
“input_shape”: 1 512 4, 33岁,
“output_shape”: 1 512 4, 33岁,
“nb_params”: 0
},
" MaxPool2d-18 ": {
“input_shape”: 1 512 4, 33岁,
“output_shape”: 1, 512, 2, 34,
“nb_params”: 0
},
" Conv2d-19 ": {
“input_shape”: 1, 512, 2, 34,
“output_shape”: 1 512 1, 33岁,
“学”:没错,
“nb_params”: 1048576
},
" BatchNorm2d-20 ": {
“input_shape”: 1 512 1, 33岁,
“output_shape”: 1 512 1, 33岁,
“学”:没错,
“nb_params”: 512
},
" ReLU-21 ": {
“input_shape”: 1 512 1, 33岁,
“output_shape”: 1 512 1, 33岁,
“nb_params”: 0
},
" LSTM-22 ": {
“input_shape”: (33, 512),
" 0 ":{
“output_shape”: (33, 512)
},
" 1 ":{
" 0 ":{
“output_shape”: [2, 256)
},
" 1 ":{
“output_shape”: [2, 256)
}
},
“nb_params”: 0
},
" Linear-23 ": {
“input_shape”: [512],
“output_shape”: [256],
“学”:没错,
“nb_params”: 131072
},
" BidirectionalLSTM-24 ": {
“input_shape”: (33, 512),
“output_shape”: (33, 256),
“nb_params”: 0
},
" LSTM-25 ": {
“input_shape”: (33, 256),
" 0 ":{
“output_shape”: (33, 512)
},
" 1 ":{
" 0 ":{
“output_shape”: [2, 256)
},
" 1 ":{
“output_shape”: [2, 256)
}
},
“nb_params”: 0
},
" Linear-26 ": {
“input_shape”: [512],
“output_shape”: [3755],
“学”:没错,
“nb_params”: 1922560
},
" BidirectionalLSTM-27 ": {
“input_shape”: (33, 256),
null
null
null
null
null
null
