pytorch 实用工具总结 - 极市社区

知乎博主Big Fish从各种资料加上自己实践整理出来的pytorch工具汇总，非常实用，欢迎收藏~
作者：Big Fish
https://zhuanlan.zhihu.com/p/33992733

以下工具主要包括：

模型层数：print_layers_num

模型参数总量：print_model_parm_nums

模型的计算图：def print_autograd_graph():或者参见tensorboad

模型滤波器可视化：show_save_tensor

模型在具体的输入下的尺寸信息summary以及参数量：show_summary

模型计算量：print_model_parm_flops

格式较混乱，但上述代码均可用，后续会继续整理。

#coding:utf8
import torch
import torchvision

import torch.nn as nn
from torch.autograd import Variable
import torchvision.models as models

import numpy as np

def test():
    model = models.resnet18()
    print model.layer1[0].conv1.weight.data

    print model.layer1[0].conv1.__class__#<class 'torch.nn.modules.conv.Conv2d'>
    print model.layer1[0].conv1.kernel_size

    input = torch.autograd.Variable(torch.randn(20, 16, 50, 100))
    print input.size()
    print np.prod(input.size())

def print_model_parm_nums():
    model = models.alexnet()
    total = sum([param.nelement() for param in model.parameters()])
    print('  + Number of params: %.2fM' % (total / 1e6))

def print_model_parm_flops():

    # prods = {}
    # def save_prods(self, input, output):
        # print 'flops:{}'.format(self.__class__.__name__)
        # print 'input:{}'.format(input)
        # print '_dim:{}'.format(input[0].dim())
        # print 'input_shape:{}'.format(np.prod(input[0].shape))
        # grads.append(np.prod(input[0].shape))

    prods = {}
    def save_hook(name):
        def hook_per(self, input, output):
            # print 'flops:{}'.format(self.__class__.__name__)
            # print 'input:{}'.format(input)
            # print '_dim:{}'.format(input[0].dim())
            # print 'input_shape:{}'.format(np.prod(input[0].shape))
            # prods.append(np.prod(input[0].shape))
            prods[name] = np.prod(input[0].shape)
            # prods.append(np.prod(input[0].shape))
        return hook_per

    list_1=[]
    def simple_hook(self, input, output):
        list_1.append(np.prod(input[0].shape))
    list_2={}
    def simple_hook2(self, input, output):
        list_2['names'] = np.prod(input[0].shape)

    multiply_adds = False
    list_conv=[]
    def conv_hook(self, input, output):
        batch_size, input_channels, input_height, input_width = input[0].size()
        output_channels, output_height, output_width = output[0].size()

        kernel_ops = self.kernel_size[0] * self.kernel_size[1] * (self.in_channels / self.groups) * (2 if multiply_adds else 1)
        bias_ops = 1 if self.bias is not None else 0

        params = output_channels * (kernel_ops + bias_ops)
        flops = batch_size * params * output_height * output_width

        list_conv.append(flops)

    list_linear=[] 
    def linear_hook(self, input, output):
        batch_size = input[0].size(0) if input[0].dim() == 2 else 1

        weight_ops = self.weight.nelement() * (2 if multiply_adds else 1)
        bias_ops = self.bias.nelement()

        flops = batch_size * (weight_ops + bias_ops)
        list_linear.append(flops)

    list_bn=[] 
    def bn_hook(self, input, output):
        list_bn.append(input[0].nelement())

    list_relu=[] 
    def relu_hook(self, input, output):
        list_relu.append(input[0].nelement())

    list_pooling=[]
    def pooling_hook(self, input, output):
        batch_size, input_channels, input_height, input_width = input[0].size()
        output_channels, output_height, output_width = output[0].size()

        kernel_ops = self.kernel_size * self.kernel_size
        bias_ops = 0
        params = output_channels * (kernel_ops + bias_ops)
        flops = batch_size * params * output_height * output_width

        list_pooling.append(flops)

    def foo(net):
        childrens = list(net.children())
        if not childrens:
            if isinstance(net, torch.nn.Conv2d):
                # net.register_forward_hook(save_hook(net.__class__.__name__))
                # net.register_forward_hook(simple_hook)
                # net.register_forward_hook(simple_hook2)
                net.register_forward_hook(conv_hook)
            if isinstance(net, torch.nn.Linear):
                net.register_forward_hook(linear_hook)
            if isinstance(net, torch.nn.BatchNorm2d):
                net.register_forward_hook(bn_hook)
            if isinstance(net, torch.nn.ReLU):
                net.register_forward_hook(relu_hook)
            if isinstance(net, torch.nn.MaxPool2d) or isinstance(net, torch.nn.AvgPool2d):
                net.register_forward_hook(pooling_hook)
            return
        for c in childrens:
                foo(c)

    resnet = models.alexnet()
    foo(resnet)
    input = Variable(torch.rand(3,224,224).unsqueeze(0), requires_grad = True)
    out = resnet(input)

    total_flops = (sum(list_conv) + sum(list_linear) + sum(list_bn) + sum(list_relu) + sum(list_pooling))

    print('  + Number of FLOPs: %.2fG' % (total_flops / 1e9))

    # print list_bn

    # print 'prods:{}'.format(prods)
    # print 'list_1:{}'.format(list_1)
    # print 'list_2:{}'.format(list_2)
    # print 'list_final:{}'.format(list_final)

def print_forward():
    model = torchvision.models.resnet18()
    select_layer = model.layer1[0].conv1

    grads={}
    def save_grad(name):
        def hook(self, input, output):
            grads[name] = input
        return hook

    select_layer.register_forward_hook(save_grad('select_layer'))

    input = Variable(torch.rand(3,224,224).unsqueeze(0), requires_grad = True)
    out = model(input)
    # print grads['select_layer']
    print grads

def print_value():
    grads = {}
    def save_grad(name):
        def hook(grad):
            grads[name] = grad
        return hook

    x = Variable(torch.randn(1,1), requires_grad=True)
    y = 3*x
    z = y**2

    # In here, save_grad('y') returns a hook (a function) that keeps 'y' as name
    y.register_hook(save_grad('y'))
    z.register_hook(save_grad('z'))
    z.backward()
    print 'HW'
    print("grads['y']: {}".format(grads['y']))
    print(grads['z'])

def print_layers_num():
    # resnet = models.resnet18()
    resnet = models.resnet18()
    def foo(net):
        childrens = list(net.children())
        if not childrens:
            if isinstance(net, torch.nn.Conv2d):
                print ' '
                #可以用来统计不同层的个数
                # net.register_backward_hook(print)
            return 1
        count = 0
        for c in childrens:
                count += foo(c)
        return count
    print(foo(resnet))

def check_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

    # Test
    import torchvision.models as models
    model = models.alexnet()
    print(torch_summarize(model))

#https://gist.github.com/wassname/0fb8f95e4272e6bdd27bd7df386716b7
#summarize a torch model like in keras, showing parameters and output shape
def show_summary():
    from collections import OrderedDict
    import pandas as pd
    import numpy as np

    import torch
    from torch.autograd import Variable
    import torch.nn.functional as F
    from torch import nn

    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 'torch.nn.modules.conv.Conv2d'>
                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

    # Test on alexnet
    import torchvision.models as models
    model = models.alexnet()
    df = torch_summarize_df(input_size=(3, 224, 224), model=model)
    print(df)

    # # Output
    #              name class_name        input_shape       output_shape  nb_params
    # 1     features=>0     Conv2d  (-1, 3, 224, 224)   (-1, 64, 55, 55)      23296#nn.Conv2d(3, 64, kernel_size=11, stride=4, padding=2),
    # 2     features=>1       ReLU   (-1, 64, 55, 55)   (-1, 64, 55, 55)          0
    # 3     features=>2  MaxPool2d   (-1, 64, 55, 55)   (-1, 64, 27, 27)          0
    # 4     features=>3     Conv2d   (-1, 64, 27, 27)  (-1, 192, 27, 27)     307392
    # 5     features=>4       ReLU  (-1, 192, 27, 27)  (-1, 192, 27, 27)          0
    # 6     features=>5  MaxPool2d  (-1, 192, 27, 27)  (-1, 192, 13, 13)          0
    # 7     features=>6     Conv2d  (-1, 192, 13, 13)  (-1, 384, 13, 13)     663936
    # 8     features=>7       ReLU  (-1, 384, 13, 13)  (-1, 384, 13, 13)          0
    # 9     features=>8     Conv2d  (-1, 384, 13, 13)  (-1, 256, 13, 13)     884992
    # 10    features=>9       ReLU  (-1, 256, 13, 13)  (-1, 256, 13, 13)          0
    # 11   features=>10     Conv2d  (-1, 256, 13, 13)  (-1, 256, 13, 13)     590080
    # 12   features=>11       ReLU  (-1, 256, 13, 13)  (-1, 256, 13, 13)          0
    # 13   features=>12  MaxPool2d  (-1, 256, 13, 13)    (-1, 256, 6, 6)          0
    # 14  classifier=>0    Dropout         (-1, 9216)         (-1, 9216)          0
    # 15  classifier=>1     Linear         (-1, 9216)         (-1, 4096)   37752832
    # 16  classifier=>2       ReLU         (-1, 4096)         (-1, 4096)          0
    # 17  classifier=>3    Dropout         (-1, 4096)         (-1, 4096)          0
    # 18  classifier=>4     Linear         (-1, 4096)         (-1, 4096)   16781312
    # 19  classifier=>5       ReLU         (-1, 4096)         (-1, 4096)          0
    # 20  classifier=>6     Linear         (-1, 4096)         (-1, 1000)    4097000

def show_save_tensor():
    import torch
    from torchvision import utils
    import torchvision.models as models
    from matplotlib import pyplot as plt

    def vis_tensor(tensor, ch = 0, all_kernels=False, nrow=8, padding = 2):
        '''
        ch: channel for visualization
        allkernels: all kernels for visualization
        '''
        n,c,h,w = tensor.shape
        if all_kernels:
            tensor = tensor.view(n*c ,-1, w, h)
        elif c != 3:
            tensor = tensor[:, ch,:,:].unsqueeze(dim=1)

        rows = np.min((tensor.shape[0]//nrow + 1, 64 ))  
        grid = utils.make_grid(tensor, nrow=nrow, normalize=True, padding=padding)
        # plt.figure(figsize=(nrow,rows))
        plt.imshow(grid.numpy().transpose((1, 2, 0)))#CHW HWC

    def save_tensor(tensor, filename, ch=0, all_kernels=False, nrow=8, padding=2):
        n,c,h,w = tensor.shape
        if all_kernels:
            tensor = tensor.view(n*c ,-1, w, h)
        elif c != 3:
            tensor = tensor[:, ch,:,:].unsqueeze(dim=1)
        utils.save_image(tensor, filename, nrow = nrow,normalize=True, padding=padding)

    vgg = models.resnet18(pretrained=True)
    mm = vgg.double()
    filters = mm.modules
    body_model = [i for i in mm.children()][0]
    # layer1 = body_model[0]
    layer1 = body_model
    tensor = layer1.weight.data.clone()
    vis_tensor(tensor)
    save_tensor(tensor,'test.png')

    plt.axis('off')
    plt.ioff()
    plt.show()

def print_autograd_graph():
    from graphviz import Digraph
    import torch
    from torch.autograd import Variable

    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 all(isinstance(p, Variable) for p in params.values())        
            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()))
                    node_name = '%s\n %s' % (param_map.get(id(u.data)), 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

    from torchvision import models

    torch.manual_seed(1)
    inputs = torch.randn(1,3,224,224)
    model = models.resnet18(pretrained=False)
    y = model(Variable(inputs))
    #print(y)

    g = make_dot(y, params=model.state_dict())
    g.view()
    #g

if __name__=='__main__':
    import fire
    fire. Fire()

• 3
• 0
• 6420

有用