Upload 20221206.py

20221206.py

@@ -0,0 +1,241 @@
+# -*- coding: utf-8 -*-
+"""20221206.ipynb
+
+Automatically generated by Colaboratory.
+
+Original file is located at
+    https://colab.research.google.com/drive/1QJCfDvr9ofVasBT4JZbaZHr44QYCMuB0
+
+###1.1 安裝套件(若在colab訓練每次都需要執行)
+"""
+
+!pip install fastbook -q
+
+"""###1.2 讀取套件"""
+
+from fastbook import *
+from fastai.vision.widgets import *
+divice = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+divice
+
+import fastai;
+print('fastai version:',fastai.__version__)
+print('torch version:',torch.__version__)
+
+"""###1.3 準備資料集"""
+
+from google.colab import drive
+drive.mount('/content/drive')
+
+path = Path('/content/drive/MyDrive/dataset/mycat2') 
+keywords = {'Lion': 'Lion','Tiger':'Tiger','Snow leopard':'Snow leopard'}
+array = keywords.items()
+if not path.exists():
+  !mkdir -p /content/drive/MyDrive/dataset/mycat2
+  for key,value in array:
+    print(key,value)
+    dest = (path/key)
+    dest.mkdir(exist_ok=True)
+    urls = search_images_ddg(f'{value}',max_images=300)
+    download_images (dest,urls=urls)
+
+"""###1.4 清洗資料"""
+
+fns = get_image_files(path)
+failed = verify_images(fns)
+failed.map(Path.unlink) #unlink broken images
+
+"""###2.1. 設定訓練資料路徑"""
+
+path = Path('/content/drive/MyDrive/dataset/mycat2')
+path.ls()
+
+#建立模型銓重儲存路徑
+myPath = '/content/drive/MyDrive/dataset/models'
+!mkdir -p $myPath
+
+"""###2.2. 資料讀取框架"""
+
+dataset = DataBlock(
+    blocks=(ImageBlock,CategoryBlock),
+    get_items = get_image_files,
+    splitter = RandomSplitter(valid_pct=0.2,seed=42),
+    item_tfms = Resize(224),
+    get_y = parent_label
+)
+
+#利用框架正式讀取資料
+dls = dataset.dataloaders(path,bs=16,num_workers=16)
+
+#讀取結果
+print(dls.c,dls.vocab,len(dls.train_ds),len(dls.valid_ds))
+
+print('訓練資料')
+dls.show_batch(max_n=5, nrows=1,unique=True)
+
+dls.show_batch(max_n=5,nrows=1)
+
+"""###3.1. 選擇模型架構以及對應的預訓練權重
+
+###Note: metrics是模型訓練人員觀察的指標, 可設定多個
+"""
+
+learn = vision_learner(dls, resnet34, metrics=[accuracy, error_rate], pretrained=True)
+learn.fit_one_cycle(3, 1e-3)
+
+"""###3.2 儲存第一次訓練好的權重"""
+
+myModel=myPath+'/resnet34_stage-1.pkl'
+learn.export(myModel)
+
+"""###3.3 解凍權重再次訓練"""
+
+learn.unfreeze()
+lr_min,lr_steep = learn.lr_find(suggest_funcs=(minimum, steep))
+print(f"Minimum/10: {lr_min:.2e}, steepest point: {lr_steep:.2e}")
+
+learn.fit_one_cycle(6, lr_max= 2.75e-04)
+
+"""###3.4 儲存新的權重"""
+
+myModel=myPath+'/resnet34_stage-2.pkl'
+learn.export(myModel)
+
+"""###4.1. 結果檢核(Confusion Matrix)"""
+
+# Contains interpretation methods for classification models
+interp = ClassificationInterpretation.from_learner(learn)
+# Plot the confusion matrix
+interp.plot_confusion_matrix()
+
+interp.plot_top_losses(5, nrows=1)
+
+"""###4.2 ROC Curve and AUC"""
+
+preds,y, loss = learn.get_preds(with_loss=True)
+# get accuracy
+acc = accuracy(preds, y)
+print('The accuracy is {0} %.'.format(acc))
+
+from sklearn.metrics import roc_curve, auc
+# probs from log preds
+probs = np.exp(preds[:,1])
+# Compute ROC curve
+fpr, tpr, thresholds = roc_curve(y, probs, pos_label=1)
+
+# Compute ROC area
+roc_auc = auc(fpr, tpr)
+print('ROC area is {0}'.format(roc_auc))
+
+plt.figure()
+plt.plot(fpr, tpr, color='darkorange', label='ROC curve (area = %0.2f)' % roc_auc)
+plt.plot([0, 1], [0, 1], color='navy', linestyle='--')
+plt.xlim([-0.01, 1.01])
+plt.ylim([-0.01, 1.01])
+plt.axis('square')
+plt.xlabel('False Positive Rate')
+plt.ylabel('True Positive Rate')
+plt.title('Receiver operating characteristic')
+plt.legend(loc="lower right")
+
+"""###5. 預測"""
+
+## 模型位置
+from fastbook import *
+from fastai.vision.widgets import *
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+device
+
+"""###5.1 讀取先前訓練好的權重"""
+
+myPath='/content/drive/MyDrive/dataset/models'
+myModel=myPath+'/resnet34_stage-2.pkl'
+learn = load_learner(myModel)
+
+"""###5.2 讀取檔案並送入模型預測"""
+
+## 執行預測 - method I
+fnames_Lion = get_image_files('/content/drive/MyDrive/dataset/mycat2/Lion')
+fnames_Tiger = get_image_files('/content/drive/MyDrive/dataset/mycat2/Tiger')
+fnames_Snow_leopard = get_image_files('/content/drive/MyDrive/dataset/mycat2/Snow leopard')
+
+fnames_Lion
+
+pred_class,pred_idx,outputs = learn.predict(fnames_Lion[3])
+print("Actual: Lion, Predicted = {}".format(pred_class))
+
+pred_class,pred_idx,outputs = learn.predict(fnames_Tiger[7])
+print("Actual: Tiger, Predicted = {}".format(pred_class))
+
+"""###6. Visualization with Grad-CAM"""
+
+class Hook():
+  def __init__(self, m):
+      self.hook = m.register_forward_hook(self.hook_func)   
+  def hook_func(self, m, i, o): self.stored = o.detach().clone()
+  def __enter__(self, *args): return self
+  def __exit__(self, *args): self.hook.remove()
+
+class HookBwd():
+  def __init__(self, m):
+      self.hook = m.register_backward_hook(self.hook_func)   
+  def hook_func(self, m, gi, go): self.stored = go[0].detach().clone()
+  def __enter__(self, *args): return self
+  def __exit__(self, *args): self.hook.remove()
+
+"""###讀取要繪製的影像"""
+
+fnames_Australian = get_image_files('/content/drive/MyDrive/dataset/mycat2/Tiger')
+test_dl = learn.dls.test_dl(fnames_Tiger, with_label=True)
+print(len(test_dl.get_idxs()))
+# pred_probas, _, pred_classes = learn.get_preds(dl=test_dl, with_decoded=True)
+
+test_dl.show_batch()
+
+from torchvision.transforms.functional import to_tensor
+fn = test_dl.items[0]
+x_dec = PILImage.create(fn);
+
+#Resize: 224  填充黑邊
+rsz = Resize(224, method=ResizeMethod.Pad, pad_mode=PadMode.Zeros)
+x_dec = rsz(x_dec)
+x = to_tensor(x_dec)
+x.unsqueeze_(0)
+x.shape,type(x)
+
+"""###繪製最後一層的feature map的Grad-CAM"""
+
+cls = 1
+with HookBwd(learn.model[0]) as hookg:
+  with Hook(learn.model[0]) as hook:
+      # output = learn.model.eval()(x.cuda())
+      output = learn.model.eval()(x.cpu())
+      act = hook.stored
+  output[0,cls].backward()
+  grad = hookg.stored
+
+w = grad[0].mean(dim=[1,2], keepdim=True)
+cam_map = (w * act[0]).sum(0)
+
+_,ax = plt.subplots()
+x_dec.show(ctx=ax)
+ax.imshow(cam_map.detach().cpu(), alpha=0.6, extent=(0,224,224,0),
+              interpolation='bilinear', cmap='magma');
+
+"""###看其他feature map的grad-cam"""
+
+with HookBwd(learn.model[0][-2]) as hookg:
+    with Hook(learn.model[0][-2]) as hook:
+        # output = learn.model.eval()(x.cuda())
+        output = learn.model.eval()(x.cpu())
+        act = hook.stored
+    output[0,cls].backward()
+    grad = hookg.stored
+
+w = grad[0].mean(dim=[1,2], keepdim=True)
+cam_map = (w * act[0]).sum(0)
+
+_,ax = plt.subplots()
+x_dec.show(ctx=ax)
+ax.imshow(cam_map.detach().cpu(), alpha=0.6, extent=(0,224,224,0),
+              interpolation='bilinear', cmap='magma');