版權聲明：本文為博主原創文章，歡迎轉載，並請注明出處。聯系方式：460356155@qq.com
Fastai是在pytorch上封裝的深度學習框架，效果出眾，以下是訓練CIFAR10的過程。
導入庫
from fastai import * from fastai.vision import * from fastai.callbacks import CSVLogger, SaveModelCallback
驗證集上訓練結果計算和顯示
def show_result(learn):
# 得到驗證集上的准確度
probs, val_labels = learn.get_preds(ds_type=DatasetType.Valid)
print('Accuracy', accuracy(probs, val_labels)),
print('Error Rate', error_rate(probs, val_labels))
訓練結果混淆矩陣及預測錯誤最多的類型顯示
def show_matrix(learn):
# 畫訓練結果的混合矩陣
interp = ClassificationInterpretation.from_learner(learn)
interp.confusion_matrix()
interp.plot_confusion_matrix(dpi=120)
# 顯示判斷錯誤最多的類型，min_val指定錯誤次數，默認1
# 打印順序為actual, predicted, number of occurrences.
interp.most_confused(min_val=5)
# 模型預測最困難的9個樣本顯示
# 顯示順序為預測值、實際值、損失值、預測對的概率
interp.plot_top_losses(9, figsize=(10, 10))
下載數據集，因調用linux的tar進行解壓，在windows下會出錯，可手動解壓，解壓后的目錄：
# 下載數據集 untar_data(URLs.CIFAR)
# 訓練數據目錄
path = Path(r'C:\Users\Administrator\.fastai\data\cifar10')
定義數據及數據在線增強方式
# 數據在線增強方式定義
tfms = get_transforms(do_flip=False)
data = (ImageList.from_folder(path) # Where to find the data? -> in path and its subfolders
.split_by_rand_pct() # How to split in train/valid? -> use the folders
.label_from_folder() # How to label? -> depending on the folder of the filenames
.add_test_folder() # Optionally add a test set (here default name is test)
.transform(tfms, size=(32, 32)) # Data augmentation? -> use tfms with a size of 164
.databunch(bs=128) # Finally? -> use the defaults for conversion to ImageDataBunch
.normalize(imagenet_stats))
查看數據
# 查看數據信息 data.classes, data.c, data
(['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck'], 10, ImageDataBunch; Train: LabelList (39072 items) x: ImageList Image (3, 32, 32),Image (3, 32, 32),Image (3, 32, 32),Image (3, 32, 32),Image (3, 32, 32) y: CategoryList airplane,airplane,airplane,airplane,airplane Path: C:\Users\Administrator\.fastai\data\cifar10; Valid: LabelList (9767 items) x: ImageList Image (3, 32, 32),Image (3, 32, 32),Image (3, 32, 32),Image (3, 32, 32),Image (3, 32, 32) y: CategoryList airplane,deer,deer,deer,automobile Path: C:\Users\Administrator\.fastai\data\cifar10; Test: LabelList (10000 items) x: ImageList Image (3, 32, 32),Image (3, 32, 32),Image (3, 32, 32),Image (3, 32, 32),Image (3, 32, 32) y: EmptyLabelList ,,,, Path: C:\Users\Administrator\.fastai\data\cifar10)
創建訓練器
# 創建learn learn = cnn_learner(data, models.resnet50, metrics=[accuracy, error_rate], callback_fns=[ShowGraph, SaveModelCallback])
第一階段訓練
# 最佳學習率尋找 learn.lr_find(end_lr=1)
LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.
# 畫出學習率尋找曲線，給出建議學習率 learn.recorder.plot(suggestion=True)
# 根據學習率曲線得到max_lr，開始訓練 learn.fit_one_cycle(cyc_len=15, max_lr=1.78e-2)
訓練結果
# 計算和顯示訓練結果 show_result(learn)
Accuracy tensor(0.8407)
Error Rate tensor(0.1593)
# 保存訓練模型
learn.save('stg1')
第二階段訓練
learn.load('stg1')
learn.unfreeze()
learn.lr_find(end_lr=1)
LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.
learn.recorder.plot(suggestion=True)
learn.fit_one_cycle(15, slice(1e-6, 5e-5))
Better model found at epoch 0 with val_loss value: 0.5218283534049988. Better model found at epoch 1 with val_loss value: 0.5134344696998596. Better model found at epoch 4 with val_loss value: 0.5069490671157837. Better model found at epoch 6 with val_loss value: 0.491120308637619.
訓練結果
# 計算和顯示訓練結果 show_result(learn)
Accuracy tensor(0.8509) Error Rate tensor(0.1491)
保存模型
learn.save('stg2')
# 畫訓練結果的混合矩陣 interp = ClassificationInterpretation.from_learner(learn) interp.confusion_matrix() interp.plot_confusion_matrix(dpi=120)
顯示預測錯誤次數最多的類型，錯誤次數大於5，輸出順序為actual, predicted, number of occurrences.
interp.most_confused(5)
[('bird', 'frog', 86),
('truck', 'automobile', 71),
('deer', 'frog', 66),
('dog', 'bird', 59),
('airplane', 'ship', 57),
('bird', 'airplane', 54),
('dog', 'frog', 54),
('bird', 'deer', 53),
('dog', 'deer', 50),
('cat', 'dog', 47),
('deer', 'bird', 47),
('automobile', 'truck', 45),
('ship', 'airplane', 45),
('cat', 'frog', 44),
('bird', 'dog', 37),
('ship', 'automobile', 34),
('ship', 'truck', 32),
('airplane', 'bird', 31),
('deer', 'dog', 26),
('frog', 'bird', 25),
('dog', 'cat', 24),
('dog', 'horse', 24),
('airplane', 'automobile', 23),
('horse', 'deer', 23),
('airplane', 'truck', 22),
('airplane', 'deer', 20),
('frog', 'deer', 17),
('cat', 'deer', 16),
('horse', 'dog', 14),
('automobile', 'ship', 13),
('deer', 'horse', 13),
('truck', 'ship', 13),
('bird', 'ship', 12),
('cat', 'bird', 12),
('deer', 'airplane', 12),
('dog', 'truck', 12),
('truck', 'airplane', 12),
('frog', 'dog', 11),
('airplane', 'frog', 10),
('deer', 'ship', 10),
('dog', 'airplane', 9),
('frog', 'automobile', 8),
('horse', 'frog', 8),
('ship', 'bird', 8),
('cat', 'truck', 7),
('horse', 'airplane', 7),
('horse', 'bird', 7),
('ship', 'deer', 7),
('dog', 'automobile', 6),
('truck', 'frog', 6),
('automobile', 'frog', 5),
('bird', 'cat', 5),
('bird', 'truck', 5),
('cat', 'ship', 5),
('dog', 'ship', 5),
('frog', 'airplane', 5)]
預測最困難的9個樣本
# 模型預測最困難的9個樣本顯示 # 顯示順序為預測值、實際值、損失值、預測對的概率 interp.plot_top_losses(9, figsize=(10, 10))
