PyTorch下的可視化工具

來(lái)源：知乎—錦恢

01

import torch import torch.nn as nn  class ConvNet(nn.Module):     def __init__(self):         super(ConvNet, self).__init__()          self.conv1 = nn.Sequential(             nn.Conv2d(1, 16, 3, 1, 1),             nn.ReLU(),             nn.AvgPool2d(2, 2)         )          self.conv2 = nn.Sequential(             nn.Conv2d(16, 32, 3, 1, 1),             nn.ReLU(),             nn.MaxPool2d(2, 2)         )          self.fc = nn.Sequential(             nn.Linear(32 * 7 * 7, 128),             nn.ReLU(),             nn.Linear(128, 64),             nn.ReLU()         )          self.out = nn.Linear(64, 10)      def forward(self, x):         x = self.conv1(x)         x = self.conv2(x)         x = x.view(x.size(0), -1)         x = self.fc(x)         output = self.out(x)         return output

 MyConvNet = ConvNet() print(MyConvNet)

 ConvNet(   (conv1): Sequential(     (0): Conv2d(1, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))     (1): ReLU()     (2): AvgPool2d(kernel_size=2, stride=2, padding=0)   )   (conv2): Sequential(     (0): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))     (1): ReLU()     (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)   )   (fc): Sequential(     (0): Linear(in_features=1568, out_features=128, bias=True)     (1): ReLU()     (2): Linear(in_features=128, out_features=64, bias=True)     (3): ReLU()   )   (out): Linear(in_features=64, out_features=10, bias=True) )

需要說(shuō)明的是，這兩個(gè)庫(kù)都是基于Graphviz開(kāi)發(fā)的，因此倘若你的電腦上沒(méi)有安裝并且沒(méi)有添加環(huán)境變量，請(qǐng)自行安裝Graphviz工具，https://blog.csdn.net/lizzy05/article/details/88529483

1.1 通過(guò)HiddenLayer可視化網(wǎng)絡(luò)

 pip install hiddenlayer

import hiddenlayer as h vis_graph = h.build_graph(MyConvNet, torch.zeros([1 ,1, 28, 28]))   # 獲取繪制圖像的對(duì)象 vis_graph.theme = h.graph.THEMES["blue"].copy()     # 指定主題顏色 vis_graph.save("./demo1.png")   # 保存圖像的路徑

1.2 通過(guò)PyTorchViz可視化網(wǎng)絡(luò)

 pip install torchviz

from torchviz import make_dot x = torch.randn(1, 1, 28, 28).requires_grad_(True)  # 定義一個(gè)網(wǎng)絡(luò)的輸入值 y = MyConvNet(x)    # 獲取網(wǎng)絡(luò)的預(yù)測(cè)值  MyConvNetVis = make_dot(y, params=dict(list(MyConvNet.named_parameters()) + [('x', x)])) MyConvNetVis.format = "png" # 指定文件生成的文件夾 MyConvNetVis.directory = "data" # 生成文件 MyConvNetVis.view()

默認(rèn)情況下，上述程序運(yùn)行后會(huì)自動(dòng)打開(kāi).png文件

02

import torchvision import torch.utils.data as Data # 準(zhǔn)備訓(xùn)練用的MNIST數(shù)據(jù)集 train_data = torchvision.datasets.MNIST(     root = "./data/MNIST",  # 提取數(shù)據(jù)的路徑     train=True, # 使用MNIST內(nèi)的訓(xùn)練數(shù)據(jù)     transform=torchvision.transforms.ToTensor(),    # 轉(zhuǎn)換成torch.tensor     download=False   # 如果是第一次運(yùn)行的話，置為T(mén)rue，表示下載數(shù)據(jù)集到root目錄 )  # 定義loader train_loader = Data.DataLoader(     dataset=train_data,     batch_size=128,     shuffle=True,     num_workers=0 )  test_data = torchvision.datasets.MNIST(     root="./data/MNIST",     train=False,    # 使用測(cè)試數(shù)據(jù)     download=False )  # 將測(cè)試數(shù)據(jù)壓縮到0-1 test_data_x = test_data.data.type(torch.FloatTensor) / 255.0 test_data_x = torch.unsqueeze(test_data_x, dim=1) test_data_y = test_data.targets  # 打印一下測(cè)試數(shù)據(jù)和訓(xùn)練數(shù)據(jù)的shape print("test_data_x.shape:", test_data_x.shape) print("test_data_y.shape:", test_data_y.shape)  for x, y in train_loader:     print(x.shape)     print(y.shape)     break

 test_data_x.shape: torch.Size([10000, 1, 28, 28]) test_data_y.shape: torch.Size([10000]) torch.Size([128, 1, 28, 28]) torch.Size([128])

2.1 通過(guò)tensorboardX可視化訓(xùn)練過(guò)程

 pip install tensorboardX pip install tensorboard

from tensorboardX import SummaryWriter logger = SummaryWriter(log_dir="data/log")  # 獲取優(yōu)化器和損失函數(shù) optimizer = torch.optim.Adam(MyConvNet.parameters(), lr=3e-4) loss_func = nn.CrossEntropyLoss() log_step_interval = 100      # 記錄的步數(shù)間隔  for epoch in range(5):     print("epoch:", epoch)     # 每一輪都遍歷一遍數(shù)據(jù)加載器     for step, (x, y) in enumerate(train_loader):         # 前向計(jì)算->計(jì)算損失函數(shù)->(從損失函數(shù))反向傳播->更新網(wǎng)絡(luò)         predict = MyConvNet(x)         loss = loss_func(predict, y)         optimizer.zero_grad()   # 清空梯度（可以不寫(xiě)）         loss.backward()     # 反向傳播計(jì)算梯度         optimizer.step()    # 更新網(wǎng)絡(luò)         global_iter_num = epoch * len(train_loader) + step + 1  # 計(jì)算當(dāng)前是從訓(xùn)練開(kāi)始時(shí)的第幾步(全局迭代次數(shù))         if global_iter_num % log_step_interval == 0:             # 控制臺(tái)輸出一下             print("global_step:{}, loss:{:.2}".format(global_iter_num, loss.item()))             # 添加的第一條日志：損失函數(shù)-全局迭代次數(shù)             logger.add_scalar("train loss", loss.item() ,global_step=global_iter_num)             # 在測(cè)試集上預(yù)測(cè)并計(jì)算正確率             test_predict = MyConvNet(test_data_x)             _, predict_idx = torch.max(test_predict, 1)     # 計(jì)算softmax后的最大值的索引，即預(yù)測(cè)結(jié)果             acc = accuracy_score(test_data_y, predict_idx)             # 添加第二條日志：正確率-全局迭代次數(shù)             logger.add_scalar("test accuary", acc.item(), global_step=global_iter_num)             # 添加第三條日志：這個(gè)batch下的128張圖像             img = vutils.make_grid(x, nrow=12)             logger.add_image("train image sample", img, global_step=global_iter_num)             # 添加第三條日志：網(wǎng)絡(luò)中的參數(shù)分布直方圖             for name, param in MyConvNet.named_parameters():                 logger.add_histogram(name, param.data.numpy(), global_step=global_iter_num)

logdir后面的參數(shù)是日志文件的文件夾的路徑

以下是筆者安裝使用tensorboard時(shí)遇到的一些錯(cuò)誤

2.2 HiddenLayer可視化訓(xùn)練過(guò)程

import hiddenlayer as hl import time  # 記錄訓(xùn)練過(guò)程的指標(biāo) history = hl.History() # 使用canvas進(jìn)行可視化 canvas = hl.Canvas()  # 獲取優(yōu)化器和損失函數(shù) optimizer = torch.optim.Adam(MyConvNet.parameters(), lr=3e-4) loss_func = nn.CrossEntropyLoss() log_step_interval = 100      # 記錄的步數(shù)間隔  for epoch in range(5):     print("epoch:", epoch)     # 每一輪都遍歷一遍數(shù)據(jù)加載器     for step, (x, y) in enumerate(train_loader):         # 前向計(jì)算->計(jì)算損失函數(shù)->(從損失函數(shù))反向傳播->更新網(wǎng)絡(luò)         predict = MyConvNet(x)         loss = loss_func(predict, y)         optimizer.zero_grad()   # 清空梯度（可以不寫(xiě)）         loss.backward()     # 反向傳播計(jì)算梯度         optimizer.step()    # 更新網(wǎng)絡(luò)         global_iter_num = epoch * len(train_loader) + step + 1  # 計(jì)算當(dāng)前是從訓(xùn)練開(kāi)始時(shí)的第幾步(全局迭代次數(shù))         if global_iter_num % log_step_interval == 0:             # 控制臺(tái)輸出一下             print("global_step:{}, loss:{:.2}".format(global_iter_num, loss.item()))             # 在測(cè)試集上預(yù)測(cè)并計(jì)算正確率             test_predict = MyConvNet(test_data_x)             _, predict_idx = torch.max(test_predict, 1)  # 計(jì)算softmax后的最大值的索引，即預(yù)測(cè)結(jié)果             acc = accuracy_score(test_data_y, predict_idx)              # 以epoch和step為索引，創(chuàng)建日志字典             history.log((epoch, step),                         train_loss=loss,                         test_acc=acc,                         hidden_weight=MyConvNet.fc[2].weight)              # 可視化             with canvas:                 canvas.draw_plot(history["train_loss"])                 canvas.draw_plot(history["test_acc"])                 canvas.draw_image(history["hidden_weight"])

不同于tensorboard，hiddenlayer會(huì)在程序運(yùn)行的過(guò)程中動(dòng)態(tài)生成圖像，而不是模型訓(xùn)練完后

03

from visdom import Visdom from sklearn.datasets import  load_iris import torch import numpy as np from PIL import Image

# 繪制圖像需要的數(shù)據(jù) iris_x, iris_y = load_iris(return_X_y=True)  # 獲取繪圖對(duì)象，相當(dāng)于plt vis = Visdom()  # 添加折線圖 x = torch.linspace(-6, 6, 100).view([-1, 1]) sigmoid = torch.nn.Sigmoid() sigmoid_y = sigmoid(x) tanh = torch.nn.Tanh() tanh_y = tanh(x) relu = torch.nn.ReLU() relu_y = relu(x) # 連接三個(gè)張量 plot_x = torch.cat([x, x, x], dim=1) plot_y = torch.cat([sigmoid_y, tanh_y, relu_y], dim=1) # 繪制線性圖 vis.line(X=plot_x, Y=plot_y, win="line plot", env="main",          opts={              "dash" : np.array(["solid", "dash", "dashdot"]),              "legend" : ["Sigmoid", "Tanh", "ReLU"]          })

# 繪制2D和3D散點(diǎn)圖 # 參數(shù)Y用來(lái)指定點(diǎn)的分布，win指定圖像的窗口名稱，env指定圖像所在的環(huán)境，opts通過(guò)字典來(lái)指定一些樣式 vis.scatter(iris_x[ : , 0 : 2], Y=iris_y+1, win="windows1", env="main") vis.scatter(iris_x[ : , 0 : 3], Y=iris_y+1, win="3D scatter", env="main",             opts={                 "markersize" : 4,   # 點(diǎn)的大小                 "xlabel" : "特征1",                 "ylabel" : "特征2"             })

# 添加莖葉圖 x = torch.linspace(-6, 6, 100).view([-1, 1]) y1 = torch.sin(x) y2 = torch.cos(x)  # 連接張量 plot_x = torch.cat([x, x], dim=1) plot_y = torch.cat([y1, y2], dim=1) # 繪制莖葉圖 vis.stem(X=plot_x, Y=plot_y, win="stem plot", env="main",          opts={              "legend" : ["sin", "cos"],              "title" : "莖葉圖"          })

# 計(jì)算鳶尾花數(shù)據(jù)集特征向量的相關(guān)系數(shù)矩陣 iris_corr = torch.from_numpy(np.corrcoef(iris_x, rowvar=False)) # 繪制熱力圖 vis.heatmap(iris_corr, win="heatmap", env="main",             opts={                 "rownames" : ["x1", "x2", "x3", "x4"],                 "columnnames" : ["x1", "x2", "x3", "x4"],                 "title" : "熱力圖"             })

# 可視化圖片 img_Image = Image.open("./example.jpg") img_array = np.array(img_Image.convert("L"), dtype=np.float32) img_tensor = torch.from_numpy(img_array) print(img_tensor.shape)  # 這次env自定義 vis.image(img_tensor, win="one image", env="MyPlotEnv",           opts={               "title" : "一張圖像"           })

# 可視化文本 text = "hello world" vis.text(text=text, win="text plot", env="MyPlotEnv",          opts={              "title" : "可視化文本"          })

在Environment中輸入不同的env參數(shù)可以看到我們?cè)诓煌h(huán)境下繪制的圖片。對(duì)于分類(lèi)圖集特別有用

04

import jsonwith open(r"...\.visdom\normal.json", "r", encoding="utf-8") as f:    dataset : dict = json.load(f)jsons : dict = dataset["jsons"]      # 這里存著你想要恢復(fù)的數(shù)據(jù)reload : dict = dataset["reload"]    # 這里存著有關(guān)窗口尺寸的數(shù)據(jù) print(jsons.keys())     # 查看所有的win

dict_keys(['jsons', 'reload'])dict_keys(['1.wav', '2.wav', '3.wav', '4.wav', '5.wav', '6.wav', '7.wav', '8.wav', '9.wav', '10.wav', '11.wav', '12.wav', '13.wav', '14.wav'])

from visdom import Visdomvis = Visdom()print(vis.get_env_list())

Setting up a new session...['main', 'normal']

from visdom import Visdomimport jsonvis = Visdom()window = vis.get_window_data(win="1.wav", env="normal")    window = json.loads(window)         # window 是 str，需要解析為字典content = window["content"]data = content["data"][0]print(data.keys())

Setting up a new session...dict_keys(['z', 'x', 'y', 'zmin', 'zmax', 'type', 'colorscale'])

猜您喜歡：
 戳我，查看GAN的系列專(zhuān)輯~！
一頓午飯外賣(mài)，成為CV視覺(jué)前沿弄潮兒！
CVPR 2022 | 25+方向、最新50篇GAN論文
 ICCV 2021 | 35個(gè)主題GAN論文匯總
超110篇！CVPR 2021最全GAN論文梳理
超100篇！CVPR 2020最全GAN論文梳理

拆解組新的GAN：解耦表征MixNMatch
StarGAN第2版：多域多樣性圖像生成

附下載 | 《可解釋的機(jī)器學(xué)習(xí)》中文版
附下載 |《TensorFlow 2.0 深度學(xué)習(xí)算法實(shí)戰(zhàn)》
附下載 |《計(jì)算機(jī)視覺(jué)中的數(shù)學(xué)方法》分享

《基于深度學(xué)習(xí)的表面缺陷檢測(cè)方法綜述》
《零樣本圖像分類(lèi)綜述: 十年進(jìn)展》
《基于深度神經(jīng)網(wǎng)絡(luò)的少樣本學(xué)習(xí)綜述》