PyTorchとscikit-learnでより見やすい性能指標を作る

はじめに

自分が普段分類問題を解いてる中でこれ見やすいんじゃねと思った指標をコード以外でも残しときたいので書きました.

code only

PyTorchの転移学習チュートリアル(1)にあるコードを改良してます．
importはあんまりなので怒らないで・・・

import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
import numpy as np
import torchvision
from torchvision import datasets, models, transforms
import matplotlib.pyplot as plt
import time
import os
import copy
from PIL import Image
from sklearn.metrics import *
import pandas as pd
from torch.utils.tensorboard import SummaryWriter
import datetime


def train_model(model, criterion, optimizer, scheduler, num_epochs=25, save_model_name="vgg16_transferlearning"):
    writer = SummaryWriter()
    save_model_dir="H:\model"
    os.makedirs(save_model_dir, exist_ok=True)
    d = datetime.datetime.now()
    save_day = "{}_{}{}_{}-{}".format(d.year, d.month, d.day, d.hour, d.minute)
    since = time.time()

    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0
    best_precision = 0.0

    for epoch in range(num_epochs):
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)

        # Each epoch has a training and validation phase
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()  # Set model to training mode
            else:
                model.eval()   # Set model to evaluate mode

            # 損失関数と正答率？
            running_loss = 0.0
            running_corrects = 0

            # Iterate over data.
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                # zero the parameter gradients
                optimizer.zero_grad()

                # forward
                # track history if only in train
                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs)
                    # row
                    axis = 1
                    _, preds = torch.max(outputs, axis)
                    # 損失関数を用いて損失(loss)を計算
                    loss = criterion(outputs, labels) 

                    # backward + optimize only if in training phase
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                # statistics 学習の評価＆統計
                running_loss += loss.item() * inputs.size(0) # inputs.size(0) == batchsize
                running_corrects += torch.sum(preds == labels.data)
            if phase == 'train':
                scheduler.step()
                if epoch%10 == 0:
                    torch.save(model_ft.state_dict(), os.path.join(save_model_dir, save_model_name+"_{}_{}.pkl".format(epoch, save_day)))
                    print("saving model epoch :{}".format(epoch))

            # 評価項目 (loss, accracy, recall, precision)
            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects.double() / dataset_sizes[phase]
            epoch_recall = recall_score(y_true=labels.cpu(), y_pred=preds.cpu(), pos_label=0)
            epoch_precision = precision_score(y_true=labels.cpu(), y_pred=preds.cpu(), pos_label=0)

            writer.add_scalar('Loss/{}'.format(phase), epoch_loss, epoch)
            writer.add_scalar('Accuracy/{}'.format(phase), epoch_acc, epoch)
            writer.add_scalar('Recall/{}'.format(phase), epoch_recall, epoch)
            writer.add_scalar('Precision/{}'.format(phase), epoch_precision, epoch)

            print('{} Loss: {:.4f} Acc: {:.4f} Recall: {:.4f} Precision: {:.4f}'.format(
                phase, epoch_loss, epoch_acc, epoch_recall, epoch_precision))

            # deep copy the model
            if phase == 'val' and epoch_acc > best_acc:
                if epoch_recall==1 and epoch_precision > best_precision:
                    torch.save(model_ft.state_dict(), 
                               os.path.join(save_model_dir, save_model_name+"_{}_{}_recall_1.0.pkl".format(epoch, save_day)))
                    print("saving model recall=1.0 epoch :{}".format(epoch))
                    recall_1_precision = epoch_precision
                best_precision = epoch_precision
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())
        print()

    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(
        time_elapsed // 60, time_elapsed % 60))
    print('Best val Acc: {:4f}, Precision: {:.4f}'.format(best_acc, best_precision))

    # load best model weights
    model.load_state_dict(best_model_wts)
    torch.save(model_ft.state_dict(), 
               os.path.join(save_model_dir, save_model_name+"_{}_{}_best.pkl".format(epoch, save_day)))
    writer.close()

    return model

これで実行

model_ft = model_ft.to(device)
criterion = nn.CrossEntropyLoss()
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1) # わからん


# Writer will output to ./runs/ directory by default
writer = SummaryWriter()
dummy_iamge = torch.rand(inputs.shape[0:])
print(dummy_iamge.shape)
dummy_iamge = dummy_iamge.to(device)
writer.add_graph(model_ft, dummy_iamge)
writer.close()
model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=25)

出力

Epoch 0/24
----------
saving model epoch :0
train Loss: 0.6785 Acc: 0.5913 Recall: 1.0000 Precision: 1.0000
val Loss: 0.6839 Acc: 0.4138 Recall: 0.3012 Precision: 1.0000

Epoch 1/24
----------
train Loss: 0.5544 Acc: 0.7340 Recall: 1.0000 Precision: 1.0000
val Loss: 0.2682 Acc: 0.9475 Recall: 1.0000 Precision: 0.9765

.....

Epoch 24/24
----------
train Loss: 0.0956 Acc: 0.9738 Recall: 1.0000 Precision: 1.0000
val Loss: 0.0232 Acc: 1.0000 Recall: 1.0000 Precision: 1.0000

Training complete in 6m 10s
Best val Acc: 1.000000, Precision: 1.0000

使った性能指標とツール

• PyTorch
• scikit-learn
• Tensorboard

使った理由

Pytorchに移行した最大の理由がTensorboardをそのまま使えることだったのです．
今回は分類問題をメインに考えているので混同行列を使って再現率と適合率を出したいと考えてました．
最終的にaccuracy，loss，recall，precisionをTensorboard内で確認することができたのでめで，たしですね．

さいごに

やっぱ性能評価時は可視化できると良いぞ〜
1次元より2次元の方が情報が詰まってるしね．
ただ，情報の詰めすぎは情報が複雑になっちゃって読み取れなくなるから気をつけてね〜

もっと良い書き方あんねんとかあればご指摘お願いします．

参考

(1) TRANSFER LEARNING FOR COMPUTER VISION TUTORIAL