pytorch 查看模型参数，查看模型特定层输入输出，模型结构图绘制总结

1 参考链接
大杂烩 
https://zhuanlan.zhihu.com/p/33992733

绘制图形时候记得安装graphviz插件 
https://stackoverflow.com/questions/35064304/runtimeerror-make-sure-the-graphviz-executables-are-on-your-systems-path-aft

画模型 
https://blog.csdn.net/qq_18293213/article/details/79047742

2 总结代码
#coding:utf-8
from collections import OrderedDict
import pandas as pd
import numpy as np
import torch
from torch.autograd import Variable
import torch.nn as nn
from graphviz import Digraph

class CNN(nn.Module):
    def __init__(self):
        super(CNN, self).__init__()
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels=1, out_channels=16, kernel_size=5, stride=1, padding=2),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(in_channels=16, out_channels=32, kernel_size=5, stride=1, padding=2),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2)
        )
        self.out = nn.Linear(32 * 7 * 7, 10)

    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = x.view(x.size(0), -1)  # (batch, 32*7*7)
        out = self.out(x)
        return out

def make_dot(var, params=None):
    """ Produces Graphviz representation of PyTorch autograd graph
    Blue nodes are the Variables that require grad, orange are Tensors
    saved for backward in torch.autograd.Function
    Args:
        var: output Variable
        params: dict of (name, Variable) to add names to node that
            require grad (TODO: make optional)
    """
    if params is not None:
        assert isinstance(params.values()[0], Variable)
        param_map = {id(v): k for k, v in params.items()}

    node_attr = dict(style='filled',
                     shape='box',
                     align='left',
                     fontsize='12',
                     ranksep='0.1',
                     height='0.2')
    dot = Digraph(node_attr=node_attr, graph_attr=dict(size="12,12"))
    seen = set()

    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 = param_map[id(u)] if params is not None else ''
                node_name = '%s\n %s' % (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(var.grad_fn)
    return dot

def get_names_dict(model):
    """
    Recursive walk to get names including path
    """
    names = {}

    def _get_names(module, parent_name=''):
        for key, module in module.named_children():
            name = parent_name + '.' + key if parent_name else key
            names[name] = module
            if isinstance(module, torch.nn.Module):
                _get_names(module, parent_name=name)

    _get_names(model)
    return names

def torch_summarize_df(input_size, model, weights=False, input_shape=True, nb_trainable=False):
    """
    Summarizes torch model by showing trainable parameters and weights.

    author: wassname
    url: https://gist.github.com/wassname/0fb8f95e4272e6bdd27bd7df386716b7
    license: MIT

    Modified from:
    - https://github.com/pytorch/pytorch/issues/2001#issuecomment-313735757
    - https://gist.github.com/wassname/0fb8f95e4272e6bdd27bd7df386716b7/

    Usage:
        import torchvision.models as models
        model = models.alexnet()
        df = torch_summarize_df(input_size=(3, 224,224), model=model)
        print(df)

        #              name class_name        input_shape       output_shape  nb_params
        # 1     features=>0     Conv2d  (-1, 3, 224, 224)   (-1, 64, 55, 55)      23296#(3*11*11+1)*64
        # 2     features=>1       ReLU   (-1, 64, 55, 55)   (-1, 64, 55, 55)          0
        # ...
    """

    def register_hook(module):
        def hook(module, input, output):
            name = ''
            for key, item in names.items():
                if item == module:
                    name = key
            #
            class_name = str(module.__class__).split('.')[-1].split("'")[0]
            module_idx = len(summary)

            m_key = module_idx + 1

            summary[m_key] = OrderedDict()
            summary[m_key]['name'] = name
            summary[m_key]['class_name'] = class_name
            if input_shape:
                summary[m_key][
                    'input_shape'] = (-1,) + tuple(input[0].size())[1:]
            summary[m_key]['output_shape'] = (-1,) + tuple(output.size())[1:]
            if weights:
                summary[m_key]['weights'] = list(
                    [tuple(p.size()) for p in module.parameters()])

            #             summary[m_key]['trainable'] = any([p.requires_grad for p in module.parameters()])
            if nb_trainable:
                params_trainable = sum(
                    [torch.LongTensor(list(p.size())).prod() for p in module.parameters() if p.requires_grad])
                summary[m_key]['nb_trainable'] = params_trainable
            params = sum([torch.LongTensor(list(p.size())).prod() for p in module.parameters()])
            summary[m_key]['nb_params'] = params

        if not isinstance(module, nn.Sequential) and \
                not isinstance(module, nn.ModuleList) and \
                not (module == model):
            hooks.append(module.register_forward_hook(hook))

    # Names are stored in parent and path+name is unique not the name
    names = get_names_dict(model)

    # check if there are multiple inputs to the network
    if isinstance(input_size[0], (list, tuple)):
        x = [Variable(torch.rand(1, *in_size)) for in_size in input_size]
    else:
        x = Variable(torch.rand(1, *input_size))

    if next(model.parameters()).is_cuda:
        x = x.cuda()

    # create properties
    summary = OrderedDict()
    hooks = []

    # register hook
    model.apply(register_hook)

    # make a forward pass
    model(x)

    # remove these hooks
    for h in hooks:
        h.remove()

    # make dataframe
    df_summary = pd.DataFrame.from_dict(summary, orient='index')

    return df_summary

def torch_summarize(model, show_weights=True, show_parameters=True):
    """Summarizes torch model by showing trainable parameters and weights."""
    from torch.nn.modules.module import _addindent

    tmpstr = model.__class__.__name__ + ' (\n'
    for key, module in model._modules.items():
        # if it contains layers let call it recursively to get params and weights
        if type(module) in [
            torch.nn.modules.container.Container,
            torch.nn.modules.container.Sequential
        ]:
            modstr = torch_summarize(module)
        else:
            modstr = module.__repr__()
        modstr = _addindent(modstr, 2)

        params = sum([np.prod(p.size()) for p in module.parameters()])
        weights = tuple([tuple(p.size()) for p in module.parameters()])

        tmpstr += '  (' + key + '): ' + modstr
        if show_weights:
            tmpstr += ', weights={}'.format(weights)
        if show_parameters:
            tmpstr += ', parameters={}'.format(params)
        tmpstr += '\n'

    tmpstr = tmpstr + ')'
    return tmpstr

def analysis():
    # summarize model
    '''
    (features): Sequential (
    (0): Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2)), weights=((64L, 3L, 11L, 11L), (64L,)), parameters=23296
    (1): ReLU(inplace), weights=(), parameters=0
    (2): MaxPool2d(kernel_size=(3, 3), stride=(2, 2), dilation=(1, 1), ceil_mode=False), weights=(), parameters=0
    (3): Conv2d(64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2)), weights=((192L, 64L, 5L, 5L), (192L,)), parameters=307392
    (4): ReLU(inplace), weights=(), parameters=0
    '''
    import torchvision.models as models

    model = models.alexnet()
    print(torch_summarize(model))

    #summarize input and output
    #              name class_name        input_shape       output_shape  nb_params
    # 1     features=>0     Conv2d  (-1, 3, 224, 224)   (-1, 64, 55, 55)      23296#(3*11*11+1)*64
    # 2     features=>1       ReLU   (-1, 64, 55, 55)   (-1, 64, 55, 55)          0
    # ...
    df = torch_summarize_df(input_size=(3, 224, 224), model=model)
    print(df)

    # make dot

    x = Variable(torch.randn(1, 3, 224, 224))
    y = model(x)
    g = make_dot(y)
    g.view()

if __name__ == '__main__':

    analysis()

————————————————
版权声明：本文为CSDN博主「HxShine」的原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/qq_16949707/article/details/81981507