Kaggle竞赛：Histopathologic Cancer Detection（VGG加RNN）

文章目录

划分训练集和测试集训练及评估函数加载函数训练函数挺不错的vgg不太好的RNN模型的保存和加载计算最终的结果

划分训练集和测试集

import pandas as pd import os import shutil import random !cd train && ls -l |grep "^-"|wc -l 220025 Train_label = pd.read_csv('train_labels.csv') name_0 = list(Train_label[Train_label['label'] == 0]['id']) name_0 = ['0'+name + '.tif' for name in name_0] len(name_0) 130908 name_1 = list(Train_label[Train_label['label'] == 1]['id']) name_1 = ['1'+name + '.tif' for name in name_1] len(name_1) 89117 for i in range(len(name_0)): if i/len(name_0) < 0.1: shutil.move('./train/'+name_0[i], './test/0') else: shutil.move('./train/'+name_0[i], './train/0') for i in range(len(name_1)): if i/len(name_1) < 0.1: shutil.move('./train/'+name_1[i], './test/1') else: shutil.move('./train/'+name_1[i], './train/1') !cd train/1 && ls -l |grep "^-"|wc -l 82593 !cd train/0 && ls -l |grep "^-"|wc -l 116929 !cd test/0 && ls -l |grep "^-"|wc -l 13979 !cd test/1 && ls -l |grep "^-"|wc -l 6524

训练及评估函数

%%time def train(model, train_loader, loss_func, optimizer, device): total_loss = 0 for i, (images, targets) in enumerate(train_loader): images = images.to(device) targets = targets.to(device) outputs = model(images) loss = loss_func(outputs, targets) optimizer.zero_grad() loss.backward() optimizer.step() total_loss += loss.item() if (i + 1) % 100 == 0: print ("Step [{}/{}] Train Loss: {:.4f}" .format(i+1, len(train_loader), loss.item())) return total_loss / len(train_loader) CPU times: user 8 µs, sys: 1 µs, total: 9 µs Wall time: 14.8 µs def evaluate(model, val_loader, train_loader, device): """ model: CNN networks val_loader: a Dataloader object with validation data device: evaluate on cpu or gpu device return classification accuracy of the model on val dataset """ # evaluate the model model.eval() # context-manager that disabled gradient computation with torch.no_grad(): correct = 0 total = 0 for i, (images, targets) in enumerate(train_loader): images = images.to(device) targets = targets.to(device) outputs = model(images) _, predicted = torch.max(outputs.data, dim=1) correct += (predicted == targets).sum().item() total += targets.size(0) accuracy = correct / total print('Accuracy on Train Set: {:.4f} %'.format(100 * accuracy)) correct = 0 total = 0 for i, (images, targets) in enumerate(val_loader): images = images.to(device) targets = targets.to(device) outputs = model(images) _, predicted = torch.max(outputs.data, dim=1) correct += (predicted == targets).sum().item() total += targets.size(0) accuracy = correct / total print('Accuracy on Test Set: {:.4f} %'.format(100 * accuracy)) return accuracy import matplotlib.pyplot as plt # 画单条变化曲线 def show_curve(ys, title): x = np.array(range(len(ys))) y = np.array(ys) plt.plot(x, y, c='b') plt.axis() plt.title('{} curve'.format(title)) plt.xlabel('epoch') plt.ylabel('{}'.format(title)) plt.show() # 画两条对比变化曲线 def show_double_curve(y1, y2, title, ylabel, label1, label2, locc = 'upper right'): x = np.array(range(len(y1))) y1 = np.array(y1) y2 = np.array(y2) plt.plot(x, y1, c='b', label = label1) plt.plot(x, y2, c='r', label = label2) plt.axis() plt.title('{} curve'.format(title)) plt.xlabel('epoch') plt.ylabel('{}'.format(ylabel)) plt.legend(bbox_to_anchor=(1.0, 1), borderaxespad=0, loc = locc) plt.show() # 画多条对比曲线 def show_multiple_curve(ys, title, ylabel, labels, locc = 'upper right'): x = np.array(range(len(ys[0]))) for i in range(len(ys)): plt.plot(x, np.array(ys[i]), label = labels[i]) plt.axis() plt.title('{} curve'.format(title)) plt.xlabel('epoch') plt.ylabel('{}'.format(ylabel)) plt.legend(bbox_to_anchor=(1.0, 1), borderaxespad=0, loc = locc) plt.show()

加载函数

import torch import torch.nn as nn import torchvision import torchvision.transforms as transforms import numpy as np train_transform = transforms.Compose([ transforms.RandomHorizontalFlip(0.6), transforms.RandomVerticalFlip(0.3), transforms.Pad(4), transforms.RandomCrop(96), transforms.ToTensor(), transforms.Normalize(mean = (0.70541453, 0.55419785, 0.69922107), std = (0.24204749, 0.28683773, 0.21985182)) ]) test_transform = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(mean = (0.70541453, 0.55419785, 0.69922107), std = (0.24204749, 0.28683773, 0.21985182)) ]) train_dataset = torchvision.datasets.ImageFolder(root='./train', transform=train_transform) test_dataset = torchvision.datasets.ImageFolder(root='./test', transform=test_transform) train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=100, shuffle=True) test_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=100, shuffle=False) def get_mean_std(dataset): """Get mean and std by sample ratio """ dataloader = torch.utils.data.DataLoader(dataset, batch_size=int(len(dataset)), shuffle=True) train = iter(dataloader).next()[0] mean = np.mean(train.numpy(), axis=(0,2,3)) std = np.std(train.numpy(), axis=(0,2,3)) return mean, std train_mean, train_std = get_mean_std(train_dataset) test_mean, test_std = get_mean_std(test_dataset) print(train_mean, train_std) print(test_mean,test_std) [ 0.70215523 0.54543036 0.69617653] [ 0.238572 0.28159615 0.21588241] [ 0.70541453 0.55419785 0.69922107] [ 0.24204749 0.28683773 0.21985182]

训练函数

def fit(model, num_epochs, optimizer, device): # loss and optimizer loss_func = nn.CrossEntropyLoss() model.to(device) loss_func.to(device) # log train loss and test accuracy losses = [] accs = [] for epoch in range(num_epochs): print('Epoch {}/{}:'.format(epoch + 1, num_epochs)) # train step loss = train(model, train_loader, loss_func, optimizer, device) losses.append(loss) # evaluate step accuracy = evaluate(model, test_loader, train_loader, device) accs.append(accuracy) # 优化学习率的调整 if epoch%10 == 0: for param_group in optimizer.param_groups: param_group['lr'] = param_group['lr']*0.6 save_path = './vgg'+str(epoch)+'.pt' torch.save(model.state_dict(), save_path) # show curve show_curve(losses, "train loss") show_curve(accs, "test accuracy") return losses, accs

挺不错的vgg

import math import torch import torch.nn as nn import torch.utils.data as Data import torchvision class VGG(nn.Module): def __init__(self, cfg): super(VGG, self).__init__() self.features = self._make_layers(cfg) # linear layer self.fc = nn.Linear(512, 2) def forward(self, x): out = self.features(x) out = out.view(out.size(0), -1) out = self.fc(out) return out def _make_layers(self, cfg): """ cfg: a list define layers this layer contains 'M': MaxPool, number: Conv2d(out_channels=number) -> BN -> ReLU """ layers = [] in_channels = 3 for x in cfg: if x == 'M': layers += [nn.MaxPool2d(kernel_size=2, stride=2)] elif x == 'MM': layers += [nn.MaxPool2d(kernel_size=3, stride=3)] elif x == 'D': layers += [nn.Dropout(p=0.1)] else: layers += [nn.Conv2d(in_channels, x, kernel_size=3, padding=1), nn.BatchNorm2d(x), nn.ReLU(inplace=True)] in_channels = x layers += [nn.AvgPool2d(kernel_size=1, stride=1)] return nn.Sequential(*layers) vggnet = VGG([64, 'M', 128, 'M', 128, 256, 'M', 256, 512, 'MM', 512, 512, 'M', 512, 'M']) print(vggnet) VGG( (features): Sequential( (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): ReLU(inplace) (3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (4): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (5): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (6): ReLU(inplace) (7): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (8): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (9): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (10): ReLU(inplace) (11): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (12): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (13): ReLU(inplace) (14): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (15): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (16): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (17): ReLU(inplace) (18): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (19): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (20): ReLU(inplace) (21): MaxPool2d(kernel_size=3, stride=3, padding=0, dilation=1, ceil_mode=False) (22): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (23): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (24): ReLU(inplace) (25): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (26): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (27): ReLU(inplace) (28): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (29): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (30): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (31): ReLU(inplace) (32): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (33): AvgPool2d(kernel_size=1, stride=1, padding=0) ) (fc): Linear(in_features=512, out_features=2, bias=True) ) from torchsummaryX import summary import torch summary(vggnet, torch.zeros((128, 3, 96, 96))) =========================================================================================== Kernel Shape Output Shape Params (K) \ Layer 0_features.Conv2d_0 [3, 64, 3, 3] [128, 64, 96, 96] 1.792 1_features.BatchNorm2d_1 [64] [128, 64, 96, 96] 0.128 2_features.ReLU_2 - [128, 64, 96, 96] - 3_features.MaxPool2d_3 - [128, 64, 48, 48] - 4_features.Conv2d_4 [64, 128, 3, 3] [128, 128, 48, 48] 73.856 5_features.BatchNorm2d_5 [128] [128, 128, 48, 48] 0.256 6_features.ReLU_6 - [128, 128, 48, 48] - 7_features.MaxPool2d_7 - [128, 128, 24, 24] - 8_features.Conv2d_8 [128, 128, 3, 3] [128, 128, 24, 24] 147.584 9_features.BatchNorm2d_9 [128] [128, 128, 24, 24] 0.256 10_features.ReLU_10 - [128, 128, 24, 24] - 11_features.Conv2d_11 [128, 256, 3, 3] [128, 256, 24, 24] 295.168 12_features.BatchNorm2d_12 [256] [128, 256, 24, 24] 0.512 13_features.ReLU_13 - [128, 256, 24, 24] - 14_features.MaxPool2d_14 - [128, 256, 12, 12] - 15_features.Conv2d_15 [256, 256, 3, 3] [128, 256, 12, 12] 590.08 16_features.BatchNorm2d_16 [256] [128, 256, 12, 12] 0.512 17_features.ReLU_17 - [128, 256, 12, 12] - 18_features.Conv2d_18 [256, 512, 3, 3] [128, 512, 12, 12] 1180.16 19_features.BatchNorm2d_19 [512] [128, 512, 12, 12] 1.024 20_features.ReLU_20 - [128, 512, 12, 12] - 21_features.MaxPool2d_21 - [128, 512, 4, 4] - 22_features.Conv2d_22 [512, 512, 3, 3] [128, 512, 4, 4] 2359.81 23_features.BatchNorm2d_23 [512] [128, 512, 4, 4] 1.024 24_features.ReLU_24 - [128, 512, 4, 4] - 25_features.Conv2d_25 [512, 512, 3, 3] [128, 512, 4, 4] 2359.81 26_features.BatchNorm2d_26 [512] [128, 512, 4, 4] 1.024 27_features.ReLU_27 - [128, 512, 4, 4] - 28_features.MaxPool2d_28 - [128, 512, 2, 2] - 29_features.Conv2d_29 [512, 512, 3, 3] [128, 512, 2, 2] 2359.81 30_features.BatchNorm2d_30 [512] [128, 512, 2, 2] 1.024 31_features.ReLU_31 - [128, 512, 2, 2] - 32_features.MaxPool2d_32 - [128, 512, 1, 1] - 33_features.AvgPool2d_33 - [128, 512, 1, 1] - 34_fc [512, 2] [128, 2] 1.026 Mult-Adds (M) Layer 0_features.Conv2d_0 15.9252 1_features.BatchNorm2d_1 6.4e-05 2_features.ReLU_2 - 3_features.MaxPool2d_3 - 4_features.Conv2d_4 169.869 5_features.BatchNorm2d_5 0.000128 6_features.ReLU_6 - 7_features.MaxPool2d_7 - 8_features.Conv2d_8 84.9347 9_features.BatchNorm2d_9 0.000128 10_features.ReLU_10 - 11_features.Conv2d_11 169.869 12_features.BatchNorm2d_12 0.000256 13_features.ReLU_13 - 14_features.MaxPool2d_14 - 15_features.Conv2d_15 84.9347 16_features.BatchNorm2d_16 0.000256 17_features.ReLU_17 - 18_features.Conv2d_18 169.869 19_features.BatchNorm2d_19 0.000512 20_features.ReLU_20 - 21_features.MaxPool2d_21 - 22_features.Conv2d_22 37.7487 23_features.BatchNorm2d_23 0.000512 24_features.ReLU_24 - 25_features.Conv2d_25 37.7487 26_features.BatchNorm2d_26 0.000512 27_features.ReLU_27 - 28_features.MaxPool2d_28 - 29_features.Conv2d_29 9.43718 30_features.BatchNorm2d_30 0.000512 31_features.ReLU_31 - 32_features.MaxPool2d_32 - 33_features.AvgPool2d_33 - 34_fc 0.001024 ------------------------------------------------------------------------------------------- Params (K): 9374.85 Mult-Adds (M): 780.3410559999999 =========================================================================================== Kernel ShapeOutput ShapeParams (K)Mult-Adds (M)Layer0_features.Conv2d_0[3, 64, 3, 3][128, 64, 96, 96]1.79215.9252481_features.BatchNorm2d_1[64][128, 64, 96, 96]0.1280.0000642_features.ReLU_2-[128, 64, 96, 96]NaNNaN3_features.MaxPool2d_3-[128, 64, 48, 48]NaNNaN4_features.Conv2d_4[64, 128, 3, 3][128, 128, 48, 48]73.856169.8693125_features.BatchNorm2d_5[128][128, 128, 48, 48]0.2560.0001286_features.ReLU_6-[128, 128, 48, 48]NaNNaN7_features.MaxPool2d_7-[128, 128, 24, 24]NaNNaN8_features.Conv2d_8[128, 128, 3, 3][128, 128, 24, 24]147.58484.9346569_features.BatchNorm2d_9[128][128, 128, 24, 24]0.2560.00012810_features.ReLU_10-[128, 128, 24, 24]NaNNaN11_features.Conv2d_11[128, 256, 3, 3][128, 256, 24, 24]295.168169.86931212_features.BatchNorm2d_12[256][128, 256, 24, 24]0.5120.00025613_features.ReLU_13-[128, 256, 24, 24]NaNNaN14_features.MaxPool2d_14-[128, 256, 12, 12]NaNNaN15_features.Conv2d_15[256, 256, 3, 3][128, 256, 12, 12]590.08084.93465616_features.BatchNorm2d_16[256][128, 256, 12, 12]0.5120.00025617_features.ReLU_17-[128, 256, 12, 12]NaNNaN18_features.Conv2d_18[256, 512, 3, 3][128, 512, 12, 12]1180.160169.86931219_features.BatchNorm2d_19[512][128, 512, 12, 12]1.0240.00051220_features.ReLU_20-[128, 512, 12, 12]NaNNaN21_features.MaxPool2d_21-[128, 512, 4, 4]NaNNaN22_features.Conv2d_22[512, 512, 3, 3][128, 512, 4, 4]2359.80837.74873623_features.BatchNorm2d_23[512][128, 512, 4, 4]1.0240.00051224_features.ReLU_24-[128, 512, 4, 4]NaNNaN25_features.Conv2d_25[512, 512, 3, 3][128, 512, 4, 4]2359.80837.74873626_features.BatchNorm2d_26[512][128, 512, 4, 4]1.0240.00051227_features.ReLU_27-[128, 512, 4, 4]NaNNaN28_features.MaxPool2d_28-[128, 512, 2, 2]NaNNaN29_features.Conv2d_29[512, 512, 3, 3][128, 512, 2, 2]2359.8089.43718430_features.BatchNorm2d_30[512][128, 512, 2, 2]1.0240.00051231_features.ReLU_31-[128, 512, 2, 2]NaNNaN32_features.MaxPool2d_32-[128, 512, 1, 1]NaNNaN33_features.AvgPool2d_33-[128, 512, 1, 1]NaNNaN34_fc[512, 2][128, 2]1.0260.001024 # Hyper-parameters num_epochs = 32 lr = 1e-5 # Device configuration device = torch.device('cuda:1') # optimizer optimizer = torch.optim.Adam(vggnet.parameters(), lr=lr) fit(vggnet, num_epochs, optimizer, device) Epoch 1/32: Step [100/1996] Train Loss: 0.6435 Step [200/1996] Train Loss: 0.3753 Step [300/1996] Train Loss: 0.3293 Step [400/1996] Train Loss: 0.3352 Step [500/1996] Train Loss: 0.4041 Step [600/1996] Train Loss: 0.3956 Step [700/1996] Train Loss: 0.4544 Step [800/1996] Train Loss: 0.3475 Step [900/1996] Train Loss: 0.3589 Step [1000/1996] Train Loss: 0.2784 Step [1100/1996] Train Loss: 0.3917 Step [1200/1996] Train Loss: 0.2907 Step [1300/1996] Train Loss: 0.2534 Step [1400/1996] Train Loss: 0.2708 Step [1500/1996] Train Loss: 0.2601 Step [1600/1996] Train Loss: 0.2590 Step [1700/1996] Train Loss: 0.2357 Step [1800/1996] Train Loss: 0.3057 Step [1900/1996] Train Loss: 0.2773 Accuracy on Train Set: 85.5044 % Accuracy on Test Set: 84.6803 % Epoch 2/32: Step [100/1996] Train Loss: 0.3205 Step [200/1996] Train Loss: 0.2809 Step [300/1996] Train Loss: 0.2924 Step [400/1996] Train Loss: 0.3045 Step [500/1996] Train Loss: 0.3805 Step [600/1996] Train Loss: 0.3427 Step [700/1996] Train Loss: 0.2957 Step [800/1996] Train Loss: 0.2942 Step [900/1996] Train Loss: 0.3697 Step [1000/1996] Train Loss: 0.2627 Step [1100/1996] Train Loss: 0.3467 Step [1200/1996] Train Loss: 0.1966 Step [1300/1996] Train Loss: 0.3518 Step [1400/1996] Train Loss: 0.3212 Step [1500/1996] Train Loss: 0.3209 Step [1600/1996] Train Loss: 0.2841 Step [1700/1996] Train Loss: 0.3483 Step [1800/1996] Train Loss: 0.3784 Step [1900/1996] Train Loss: 0.2279 Accuracy on Train Set: 88.2840 % Accuracy on Test Set: 88.3871 % ... Epoch 30/32: Step [100/1996] Train Loss: 0.1237 Step [200/1996] Train Loss: 0.1236 Step [300/1996] Train Loss: 0.1687 Step [400/1996] Train Loss: 0.1455 Step [500/1996] Train Loss: 0.0666 Step [600/1996] Train Loss: 0.1796 Step [700/1996] Train Loss: 0.2382 Step [800/1996] Train Loss: 0.0954 Step [900/1996] Train Loss: 0.1450 Step [1000/1996] Train Loss: 0.0748 Step [1100/1996] Train Loss: 0.1481 Step [1200/1996] Train Loss: 0.1721 Step [1300/1996] Train Loss: 0.0872 Step [1400/1996] Train Loss: 0.1785 Step [1500/1996] Train Loss: 0.1604 Step [1600/1996] Train Loss: 0.2249 Step [1700/1996] Train Loss: 0.1070 Step [1800/1996] Train Loss: 0.1704 Step [1900/1996] Train Loss: 0.2214 Accuracy on Train Set: 94.4222 % Accuracy on Test Set: 94.2350 % Epoch 31/32: Step [100/1996] Train Loss: 0.1209 Step [200/1996] Train Loss: 0.1470 Step [300/1996] Train Loss: 0.1198 Step [400/1996] Train Loss: 0.0985 Step [500/1996] Train Loss: 0.1920 Step [600/1996] Train Loss: 0.0988 Step [700/1996] Train Loss: 0.1128 Step [800/1996] Train Loss: 0.1519 Step [900/1996] Train Loss: 0.1162 Step [1000/1996] Train Loss: 0.1357 Step [1100/1996] Train Loss: 0.0848 Step [1200/1996] Train Loss: 0.1087 Step [1300/1996] Train Loss: 0.1963 Step [1400/1996] Train Loss: 0.1025 Step [1500/1996] Train Loss: 0.1235 Step [1600/1996] Train Loss: 0.1427 Step [1700/1996] Train Loss: 0.1792 Step [1800/1996] Train Loss: 0.1654 Step [1900/1996] Train Loss: 0.1518 Accuracy on Train Set: 94.4452 % Accuracy on Test Set: 94.7130 % Epoch 32/32: Step [100/1996] Train Loss: 0.1606 Step [200/1996] Train Loss: 0.0931 Step [300/1996] Train Loss: 0.0738 Step [400/1996] Train Loss: 0.1140 Step [500/1996] Train Loss: 0.1462 Step [600/1996] Train Loss: 0.1682 Step [700/1996] Train Loss: 0.0993 Step [800/1996] Train Loss: 0.2225 Step [900/1996] Train Loss: 0.0880 Step [1000/1996] Train Loss: 0.1019 Step [1100/1996] Train Loss: 0.1696 Step [1200/1996] Train Loss: 0.1691 Step [1300/1996] Train Loss: 0.1612 Step [1400/1996] Train Loss: 0.1052 Step [1500/1996] Train Loss: 0.1834 Step [1600/1996] Train Loss: 0.1454 Step [1700/1996] Train Loss: 0.1393 Step [1800/1996] Train Loss: 0.1126 Step [1900/1996] Train Loss: 0.1546 Accuracy on Train Set: 94.6497 % Accuracy on Test Set: 94.7910 %

不太好的RNN

import math import torch import torch.nn as nn import torch.utils.data as Data import torchvision class MyRnn(nn.Module): def __init__(self, in_dim, hidden_dim, n_layer, n_class): super(MyRnn, self).__init__() self.n_layer = n_layer self.hidden_dim = hidden_dim self.lstm = nn.LSTM(in_dim, hidden_dim, n_layer,bidirectional=True) self.fc1 = nn.Linear(hidden_dim*4, 512) self.fc2 = nn.Linear(512, 64) self.fc3 = nn.Linear(64, 2) self.relu = nn.ReLU() self.dropout = nn.Dropout(0.2) def forward(self, x): x = x.view(-1,96,96*3) y = x.permute([1, 0, 2]) out, _ = self.lstm(y) encoding = torch.cat([out[0], out[-1]], dim=1) result = self.relu(encoding) result = self.dropout(result) result = self.fc1(result) result = self.relu(result) result = self.dropout(result) result = self.fc2(result) result = self.relu(result) result = self.dropout(result) result = self.fc3(result) return result from torchsummaryX import summary import torch summary(model, torch.zeros((128, 3, 96, 96))) ================================================================== Kernel Shape Output Shape Params (K) Mult-Adds (M) Layer 0_lstm - [96, 128, 2048] 35946.5 35.9137 1_relu - [128, 4096] - - 2_dropout - [128, 4096] - - 3_fc1 [4096, 512] [128, 512] 2097.66 2.09715 4_relu - [128, 512] - - 5_dropout - [128, 512] - - 6_fc2 [512, 64] [128, 64] 32.832 0.032768 7_relu - [128, 64] - - 8_dropout - [128, 64] - - 9_fc3 [64, 2] [128, 2] 0.13 0.000128 ------------------------------------------------------------------ Params (K): 38077.121999999996 Mult-Adds (M): 38.043775999999994 ================================================================== Kernel ShapeOutput ShapeParams (K)Mult-Adds (M)Layer0_lstm-[96, 128, 2048]35946.49635.9137281_relu-[128, 4096]NaNNaN2_dropout-[128, 4096]NaNNaN3_fc1[4096, 512][128, 512]2097.6642.0971524_relu-[128, 512]NaNNaN5_dropout-[128, 512]NaNNaN6_fc2[512, 64][128, 64]32.8320.0327687_relu-[128, 64]NaNNaN8_dropout-[128, 64]NaNNaN9_fc3[64, 2][128, 2]0.1300.000128 model = MyRnn(288, 1024, 2, 2) model MyRnn( (lstm): LSTM(288, 1024, num_layers=2, bidirectional=True) (fc1): Linear(in_features=4096, out_features=512, bias=True) (fc2): Linear(in_features=512, out_features=64, bias=True) (fc3): Linear(in_features=64, out_features=2, bias=True) (relu): ReLU() (dropout): Dropout(p=0.2) ) # Hyper-parameters num_epochs = 70 lr = 1e-4 # Device configuration device = torch.device('cuda:1') # optimizer optimizer = torch.optim.Adam(model.parameters(), lr=lr) fit(model, num_epochs, optimizer, device) Epoch 1/70: Step [100/1559] Train Loss: 0.6516 Step [200/1559] Train Loss: 0.5623 Step [300/1559] Train Loss: 0.5309 Step [400/1559] Train Loss: 0.5903 Step [500/1559] Train Loss: 0.5542 Step [600/1559] Train Loss: 0.5293 Step [700/1559] Train Loss: 0.5013 Step [800/1559] Train Loss: 0.5523 Step [900/1559] Train Loss: 0.5209 Step [1000/1559] Train Loss: 0.5653 Step [1100/1559] Train Loss: 0.4562 Step [1200/1559] Train Loss: 0.4115 Step [1300/1559] Train Loss: 0.4277 Step [1400/1559] Train Loss: 0.4983 Step [1500/1559] Train Loss: 0.4299 Accuracy on Train Set: 79.8193 % Accuracy on Test Set: 80.5492 % Epoch 2/70: Step [100/1559] Train Loss: 0.5156 Step [200/1559] Train Loss: 0.5145 Step [300/1559] Train Loss: 0.5255 Step [400/1559] Train Loss: 0.5319 Step [500/1559] Train Loss: 0.4295 Step [600/1559] Train Loss: 0.4307 Step [700/1559] Train Loss: 0.4712 Step [800/1559] Train Loss: 0.4979 Step [900/1559] Train Loss: 0.3949 Step [1000/1559] Train Loss: 0.5165 Step [1100/1559] Train Loss: 0.3943 Step [1200/1559] Train Loss: 0.4350 Step [1300/1559] Train Loss: 0.5146 Step [1400/1559] Train Loss: 0.2779 Step [1500/1559] Train Loss: 0.4307 Accuracy on Train Set: 81.4396 % Accuracy on Test Set: 83.0171 % Epoch 3/70: Step [100/1559] Train Loss: 0.4699 Step [200/1559] Train Loss: 0.4977 Step [300/1559] Train Loss: 0.4993 Step [400/1559] Train Loss: 0.4876 Step [500/1559] Train Loss: 0.4446 Step [600/1559] Train Loss: 0.5243 Step [700/1559] Train Loss: 0.4131 Step [800/1559] Train Loss: 0.4891 Step [900/1559] Train Loss: 0.5236 Step [1000/1559] Train Loss: 0.3616 Step [1100/1559] Train Loss: 0.5291 Step [1200/1559] Train Loss: 0.4133 Step [1300/1559] Train Loss: 0.3593 Step [1400/1559] Train Loss: 0.4399 Step [1500/1559] Train Loss: 0.3548 Accuracy on Train Set: 82.4125 % Accuracy on Test Set: 83.7097 % ... Epoch 69/70: Step [100/1559] Train Loss: 0.1127 Step [200/1559] Train Loss: 0.1356 Step [300/1559] Train Loss: 0.0965 Step [400/1559] Train Loss: 0.0732 Step [500/1559] Train Loss: 0.0706 Step [600/1559] Train Loss: 0.1228 Step [700/1559] Train Loss: 0.0812 Step [800/1559] Train Loss: 0.1042 Step [900/1559] Train Loss: 0.1054 Step [1000/1559] Train Loss: 0.1461 Step [1100/1559] Train Loss: 0.1589 Step [1200/1559] Train Loss: 0.0724 Step [1300/1559] Train Loss: 0.1406 Step [1400/1559] Train Loss: 0.0955 Step [1500/1559] Train Loss: 0.1438 Accuracy on Train Set: 96.0907 % Accuracy on Test Set: 89.5966 % Epoch 70/70: Step [100/1559] Train Loss: 0.1222 Step [200/1559] Train Loss: 0.0991 Step [300/1559] Train Loss: 0.1290 Step [400/1559] Train Loss: 0.0402 Step [600/1559] Train Loss: 0.0958 Step [800/1559] Train Loss: 0.0545 Step [900/1559] Train Loss: 0.1083 Step [1000/1559] Train Loss: 0.1351 Step [1100/1559] Train Loss: 0.0810 Step [1200/1559] Train Loss: 0.0882 Step [1300/1559] Train Loss: 0.1177 Step [1400/1559] Train Loss: 0.0832 Step [1500/1559] Train Loss: 0.0738 Accuracy on Train Set: 96.1648 % Accuracy on Test Set: 89.8259 %

模型的保存和加载

save_path = './vgg14.pt' device = torch.device('cuda:0') saved_parametes = torch.load(save_path) new_vgg = VGG([64, 'M', 128, 'M', 128, 256, 'M', 256, 512, 'MM', 512, 512, 'M', 512, 'M']).to(device) new_vgg.load_state_dict(saved_parametes)

计算最终的结果

import os from PIL import Image def get_result(model, path): imgs = os.listdir(path) correct = 0 model.eval() with torch.no_grad(): for name in imgs: img = Image.open(path+name) img = test_transform(img) s = model(img.unsqueeze(0).to(device)) _, pre = torch.max(s.data, dim=1) if pre.item() == ord(name[0])-48: correct += 1 print('在'+path+'上的准确率为：', correct/len(imgs)) get_result(vggnet, "./train/1/") 在./train/1/上的准确率为： 0.9282869008269466 get_result(vggnet, "./train/0/") 在./train/0/上的准确率为： 0.9639695883826938 get_result(vggnet, "./test/1/") 在./test/1/上的准确率为： 0.9216738197424893 get_result(vggnet, "./test/0/") 在./test/0/上的准确率为： 0.9601545174905215 def get_resultcsv(model): path = './predict/' # df = pd.read_csv('sample_submission.csv') imgs = os.listdir(path) model.eval() label = [] i = 0 with torch.no_grad(): for name in imgs: i += 1 img = Image.open(path+name) img = test_transform(img) s = model(img.unsqueeze(0).to(device)) _, pre = torch.max(s.data, dim=1) label.append(pre.item()) if i%1000 == 0: print ("Step [{}/{}] prediction!".format(i, len(imgs))) df = pd.DataFrame({'id': [x[:-4] for x in imgs],'label':label}) df.to_csv("result.csv",index=False,sep=',') %%time import pandas as pd get_resultcsv(vggnet) Step [1000/57458] prediction! ... Step [56000/57458] prediction! Step [57000/57458] prediction! CPU times: user 5min 12s, sys: 59.5 s, total: 6min 11s Wall time: 18min 16s