Initial Commit for SkinGuard (#1)

- Initial Commit for SkinGuard (1fd6d2dc419c10e463f950d2ab45e9891b7ab6ce)

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/ISIC_0000381.png filter=lfs diff=lfs merge=lfs -text

Final_ViT_Model_50_Epochs.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e11954779371fb0c29ac8bbe5b5a504442ee7892b0a4b463b7a8e9c0517b00b6
+size 343261946

app.py

@@ -0,0 +1,61 @@
+### 1. Imports and class names setup ###
+import gradio as gr
+import os
+import torch
+
+from model import create_ViT
+from timeit import default_timer as timer
+
+# Setup class names
+class_names = ['Normal', 'Malignant']
+
+### 2. Model and transforms preparation ###
+ViT, manual_transforms = create_ViT()
+
+# Load saved weights
+ViT.load_state_dict(
+  torch.load(f='Final_ViT_Model_50_Epochs.pth',
+  map_location=torch.device('cpu') # Load model to cpu
+  )
+)
+
+### 3. Predict function ###
+def predict(img):
+  start_time = timer()
+
+  img = manual_transforms(img).unsqueeze(0)
+
+  ViT.eval()
+  with torch.inference_mode():
+      pred_prob = torch.sigmoid(ViT(img).squeeze())
+      pred_label_idx = int(torch.round(pred_prob))
+  if pred_label_idx:
+    pred_labels_and_preds = {class_names[pred_label_idx]: float(pred_prob.item())}
+  else:
+    pred_labels_and_preds = {class_names[pred_label_idx]: 1-float(pred_prob.item())}
+
+  end_time = timer()
+  pred_time = round(end_time-start_time, 4)
+
+  return pred_labels_and_preds, pred_time
+
+### 4. Gradio App ###
+# Create interface for gradio
+title = 'SkinGuard'
+description = 'An AI model that can predict whether a CLOSE-UP picture of skin is normal or shows a sign of Skin Cancer. \n IMPORTANT INSTRUCTION: When taking a picture of your skin, take the picture AS CLOSE AS POSSIBLE to the target area.'
+article = 'This is our demo for the DEDA Entrepreneurship Competition 2024. Created by Imran, Brian, Lukas, and Rohit, all in Baez CSE Period 5. \n IMPORTANT NOTE: If you have followed the important instruction in the description of this demo and STILL get malignant results constantly, do NOT take it as an official diagnosis. Contact a professional healthcare professional for more information if you are concerned.'
+# Create examples list
+example_list = [['examples/' + example] for example in os.listdir('examples')]
+
+# Create the gradio demo
+demo = gr.Interface(fn=predict,# maps intputs to outputs
+                    inputs=gr.Image(type='pil'),
+                    outputs=[gr.Label(num_top_classes=1, label='Predictions'),
+                             gr.Label(label='Prediction Time (s)')],
+                    examples=example_list,
+                    title=title,
+                    description=description,
+                    article=article)
+# Launch the model!!
+demo.launch(debug=False,
+            share=True)

model.py

@@ -0,0 +1,39 @@
+import torch
+import torchvision
+
+from torch import nn
+
+def create_ViT():
+  ViT_weights = torchvision.models.ViT_B_16_Weights.DEFAULT
+  ViT_model = torchvision.models.vit_b_16(weights=ViT_weights)
+
+  # Freeze pre-trained weights
+  for param in ViT_model.parameters():
+    param.requires_grad = False
+
+  # Find the encoder module and its layers
+  encoder = ViT_model.encoder
+  encoder_layers = encoder.layers
+  # Modify each Encoder layer to include dropout
+  for layer in encoder_layers:
+    # Access the Multi-head Self-Attention module (might be named differently)
+    attn_module = layer.self_attention  # Replace with the actual module name in your model
+    # Add dropout layer after the attention module
+    attn_module.add_module('my_dropout', nn.Dropout(p=0.4))
+  # Add your new head for classification (same as before)
+  ViT_model.heads = nn.Sequential(
+      nn.Dropout(p=0.5),
+      nn.Linear(in_features=768, out_features=1, bias=True)
+  )
+
+  manual_transforms = torchvision.transforms.Compose([
+    torchvision.transforms.RandomRotation(25),
+    torchvision.transforms.RandomAffine(degrees=0, translate=(0.15, 0.15), shear=15),
+    torchvision.transforms.RandomHorizontalFlip(),
+    torchvision.transforms.RandomVerticalFlip(),
+    torchvision.transforms.ColorJitter(brightness=(0.9, 1.5)),
+    torchvision.transforms.Resize((224, 224)),
+    torchvision.transforms.ToTensor()
+  ])
+
+  return ViT_model, manual_transforms

requirements.txt

@@ -0,0 +1,3 @@
+torch==2.2.1
+torchvision==0.17.1
+gradio==4.26.0