convnext

app.py

@@ -13,15 +13,7 @@ for key in list(model_weights):
 
 
 def get_model():
-    model = timm.create_model('tf_efficientnet_b1', pretrained=True, num_classes=2, global_pool='catavgmax')
-    num_in_features = model.get_classifier().in_features
-    from torch import nn
-
-    model.fc = nn.Sequential(
-        nn.Linear(in_features=num_in_features, out_features=1024, bias=False),
-        nn.ReLU(),
-        nn.Linear(in_features=1024, out_features=2, bias=False),
-    )
+    model = timm.create_model('convnext_base.fb_in22k_ft_in1k', pretrained=True, num_classes=2)
 
     return model
 
@@ -33,15 +25,26 @@ model.eval()
 import requests
 from PIL import Image
 from torchvision import transforms
+import albumentations as A
+
+CROP = 224
+SIZE = CROP + CROP//8
+
+ho_trans_center = A.Compose([
+    A.Resize(SIZE,SIZE, interpolation=cv2.INTER_AREA),
+    A.CenterCrop(height=CROP, width=CROP, always_apply=True),                                        
+    ])
+topt = A.Compose([
+    A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
+    ToTensorV2(),
+    ])
 
 # Download human-readable labels for ImageNet.
 labels = ['good', 'ill']
-CROP=384
 
 def predict(inp):
-  img = torchvision.transforms.ToTensor()(inp)
-  img = torchvision.transforms.Resize((800, 800))(img)
-  img = torchvision.transforms.CenterCrop(CROP)(img)
+  img = ho_trans_center(image = inp)['image']
+  img = topt(image = img)['image']
   img = img.unsqueeze(0)
   with torch.no_grad():
     prediction = model(img).softmax(1).numpy()
@@ -51,7 +54,7 @@ def predict(inp):
 import gradio as gr
 
 gr.Interface(fn=predict, 
-             inputs=gr.Image(type="pil"),
+             inputs=gr.Image(),
              outputs=gr.Label(num_top_classes=1),
              ).launch()
 

v5-epoch=19-val_loss=0.1464-val_accuracy=0.9514.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f34c9f1ef5bf747a84a52eff907ffafb9f37c7a023a4ea9e5b736fbc6e4156be
+size 1051254575