Update app.py

app.py

@@ -8,6 +8,9 @@ import torchvision.transforms as transforms
 import pandas as pd
 import cv2
 import numpy as np
+from PIL import Image
+from get_data import get_augmentation
+from configs.dataset_params import normalize_params
 
 def overlapping_features_on_input(model,output, feature_maps, input, target):
     W=model.linear.layer.weight
@@ -20,14 +23,16 @@ def overlapping_features_on_input(model,output, feature_maps, input, target):
      label=np.argmax(output)+1
 
     Interpretable_Selection= W[label,:]
-    
+    print("W",Interpretable_Selection)
     input_np=np.array(input)
     h,w= input.shape[:2]
-    
+    print("h,w:",h,w)
     Interpretable_Features=[]
     Feature_image_list=[]
+    
     for S in range(len(Interpretable_Selection)):
         if Interpretable_Selection[S] > 0:
+
                Interpretable_Features.append(feature_maps[S])
                Feature_image=cv2.resize(feature_maps[S],(w,h))
                Feature_image=((Feature_image-np.min(Feature_image))/(np.max(Feature_image)-np.min(Feature_image)))*255
@@ -47,63 +52,71 @@ def genreate_intepriable_output(input,dataset="CUB2011", arch="resnet50",seed=12
     n_classes = dataset_constants[dataset]["num_classes"]
   
     model = get_model(arch, n_classes, reduced_strides)
-    tr=transforms.ToTensor()
-    device = torch.device("cpu")
+    tr=transforms.Compose([
+            transforms.Resize(256),
+            transforms.CenterCrop(224),
+            transforms.ToTensor(),
+        ])
+    TR=get_augmentation(0.1, img_size, False, False, True, True, normalize_params["CUB2011"])
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     if folder is None:
-        folder = Path(f"tmp/{arch}/{dataset}/{seed}/")
+        folder = Path.home() / f"tmp/{arch}/{dataset}/{seed}/"
     
-    state_dict = torch.load(folder / f"{model_type}_{n_features}_{n_per_class}_FinetunedModel.pth", map_location=torch.device('cpu'))
-    selection= torch.load(folder / f"SlDD_Selection_50.pt", map_location=torch.device('cpu'))
+    state_dict = torch.load(folder / f"{model_type}_{n_features}_{n_per_class}_FinetunedModel.pth")
+    selection= torch.load(folder / f"SlDD_Selection_50.pt")
     state_dict['linear.selection']=selection
     
     feature_sel, sparse_layer, current_mean, current_std, bias_sparse = extract_sel_mean_std_bias_assignemnt(state_dict)
     model.set_model_sldd(feature_sel, sparse_layer, current_mean, current_std, bias_sparse)
     model.load_state_dict(state_dict)
+    input=Image.fromarray(input)
 
     input = tr(input)
     input= input.unsqueeze(0)
     input= input.to(device)
     model = model.to(device)
-    output, feature_maps, final_features = model(input, with_feature_maps=True, with_final_features=True)
-    print("final features:",final_features)
-    output=output.detach().cpu().numpy()
-    output= np.argmax(output)+1
-    
-
-    print("outputclass:",output)
-    data_dir=Path("tmp/Datasets/CUB200/CUB_200_2011/")
-    labels = pd.read_csv(data_dir/"image_class_labels.txt", sep=' ', names=['img_id', 'target'])
-    namelist=pd.read_csv(data_dir/"images.txt",sep=' ',names=['img_id','file_name'])
-    classlist=pd.read_csv(data_dir/"classes.txt",sep=' ',names=['cl_id','class_name'])
-    options_output=labels[labels['target']==output]
-    options_output=options_output.sample(1)
-    others=labels[labels['target']!=output]
-    options_others=others.sample(3)
-    options = pd.concat([options_others, options_output], ignore_index=True)
-    shuffled_options = options.sample(frac=1).reset_index(drop=True)
-    print("shuffled:",shuffled_options)
-    op=[]
-
-    for i in shuffled_options['img_id']:
-        print(i)
-        filenames=namelist.loc[namelist['img_id']==i,'file_name'].values[0]
-        targets=shuffled_options.loc[shuffled_options['img_id']==i,'target'].values[0]
-        
+    model.eval()
+    with torch.no_grad():
+        output, feature_maps, final_features = model(input, with_feature_maps=True, with_final_features=True)
+        print("final features:",final_features)
+        output=output.detach().cpu().numpy()
+        output= np.argmax(output)+1
         
 
-        classes=classlist.loc[classlist['cl_id']==targets, 'class_name'].values[0]
-        
-
-        op_img=cv2.imread(data_dir/f"images/{filenames}")
-
-        op_images=tr(op_img)
-        op_images=op_images.unsqueeze(0)
-        op_images=op_images.to(device)
-        OP, feature_maps_op =model(op_images,with_feature_maps=True,with_final_features=False)
-        print("OP:",OP,
-              "feature_maps_op:",feature_maps_op.shape)
-        opt= overlapping_features_on_input(model,OP, feature_maps_op,op_img,targets)
-        op+=opt
+        print("outputclass:",output)
+        data_dir=Path.home()/"tmp/Datasets/CUB200/CUB_200_2011/"
+        labels = pd.read_csv(data_dir/"image_class_labels.txt", sep=' ', names=['img_id', 'target'])
+        namelist=pd.read_csv(data_dir/"images.txt",sep=' ',names=['img_id','file_name'])
+        classlist=pd.read_csv(data_dir/"classes.txt",sep=' ',names=['cl_id','class_name'])
+        options_output=labels[labels['target']==output]
+        options_output=options_output.sample(1)
+        others=labels[labels['target']!=output]
+        options_others=others.sample(3)
+        options = pd.concat([options_others, options_output], ignore_index=True)
+        shuffled_options = options.sample(frac=1).reset_index(drop=True)
+        print("shuffled:",shuffled_options)
+        op=[]
+
+
+        for i in shuffled_options['img_id']:
+
+            filenames=namelist.loc[namelist['img_id']==i,'file_name'].values[0]
+            targets=shuffled_options.loc[shuffled_options['img_id']==i,'target'].values[0]
+
+
+            classes=classlist.loc[classlist['cl_id']==targets, 'class_name'].values[0]
+            
+
+            op_img=cv2.imread(data_dir/f"images/{filenames}")
+            
+            op_imag=Image.fromarray(op_img)
+            op_images=TR(op_imag)
+            op_images=op_images.unsqueeze(0)
+            op_images=op_images.to(device)
+            OP, feature_maps_op =model(op_images,with_feature_maps=True,with_final_features=False)
+      
+            opt= overlapping_features_on_input(model,OP, feature_maps_op,op_img,targets)
+            op+=opt
     
     return op