First commit

.gitattributes

@@ -33,3 +33,9 @@ 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
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,19 @@
+# Model files
+*.pth
+*.pt
+*.bin
+
+# OS files
+.DS_Store
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Virtual environment
+venv/
+env/
+.env/
+
+# IDE files
+.vscode/
+.idea/ 

README.md

@@ -1,13 +1,42 @@
 ---
-title: Scene Classification Swin L
-emoji: 📚
-colorFrom: green
-colorTo: yellow
+title: "AML 16"
+version: "1.0.0"
+emoji: "🤗"
+colorFrom: indigo
+colorTo: pink
 sdk: gradio
-sdk_version: 5.34.2
+sdk_version: "5.29.0"
 app_file: app.py
 pinned: false
-short_description: scene_classification_swin_l
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# AML 16
+
+This is a demo application for the best-performing model (Swin-Large) created for the AML 16 project.  
+The app uses Gradio to provide an interactive interface where users can upload an image, view the top-1 predicted scene category, see a reference image from the predicted class, and explore the top-5 prediction probabilities in a bar chart.
+
+The model was trained for scene classification and deployed using Hugging Face Spaces.
+
+- predict.py
+This file handles loading the trained Swin-Large model and making predictions.
+It loads the model weights from Hugging Face Hub, applies the correct image preprocessing, and outputs:
+
+The uploaded image,
+
+A reference image from the predicted class,
+
+The Top-5 prediction probabilities.
+
+The model was customized with an updated classifier head, and class labels are loaded from a labels.json file. A random sample image from the predicted class folder is also shown for better visualization.
+
+- app.py
+This file builds the Gradio interface.
+It lets users upload an image, runs the prediction using predict.py, and displays:
+
+The uploaded image,
+
+An image for the top-1 predicted class,
+
+The predicted class label,
+
+A bar chart showing the Top-5 prediction probabilities.

app.py

@@ -0,0 +1,16 @@
+import gradio as gr
+from predict import predict
+
+demo = gr.Interface(
+    fn=predict,
+    inputs=gr.Image(type="filepath", label="Upload Image"),
+    outputs=[
+        gr.Image(label="Uploaded Image"),
+        gr.Image(label="Top-1 Class Example"),
+        gr.Label(label="Top-5 Probabilities")
+    ],
+    title="Scene Classification with Reference Image Testing using Large SWIN version 1.1.0",
+    description="Upload an image to get the predicted class with a sample image and top-5 prediction chart."
+)
+
+demo.launch()

labels.json

@@ -0,0 +1,42 @@
+[
+    "airport_terminal",
+    "amphitheatre",
+    "amusement_park",
+    "art_gallery",
+    "bakery_shop",
+    "bar",
+    "bookstore",
+    "botanical_garden",
+    "bridge",
+    "bus_interior",
+    "butchers_shop",
+    "campsite",
+    "classroom",
+    "coffee_shop",
+    "construction_site",
+    "courtyard",
+    "driveway",
+    "fire_station",
+    "fountain",
+    "gas_station",
+    "harbour",
+    "highway",
+    "kindergarten_classroom",
+    "lobby",
+    "market_outdoor",
+    "museum",
+    "office",
+    "parking_lot",
+    "phone_booth",
+    "playground",
+    "railroad_track",
+    "restaurant",
+    "river",
+    "shed",
+    "staircase",
+    "supermarket",
+    "swimming_pool_outdoor",
+    "track",
+    "valley",
+    "yard"
+]

predict.py

@@ -0,0 +1,91 @@
+import torch
+import torch.nn as nn
+import torch.nn.init as init
+from transformers import SwinForImageClassification
+from huggingface_hub import hf_hub_download
+from PIL import Image
+import json
+import os
+import random
+from torchvision import transforms
+
+# Load labels
+with open("labels.json", "r") as f:
+    class_names = json.load(f)
+print("class_names:", class_names)
+
+MODEL_NAME = "microsoft/swin-large-patch4-window7-224"  
+
+class SwinCustom(nn.Module):
+    def __init__(self, model_name=MODEL_NAME, num_classes=40):
+        super(SwinCustom, self).__init__()
+        self.model = SwinForImageClassification.from_pretrained(model_name, num_labels=num_classes, ignore_mismatched_sizes=True)
+        in_features = self.model.classifier.in_features
+        self.model.classifier = nn.Sequential(
+            nn.Linear(in_features, in_features),
+            nn.LeakyReLU(),
+            nn.Dropout(0.3),
+            nn.Linear(in_features, num_classes)
+        )
+        # Weight initialization
+        for m in self.model.classifier:
+            if isinstance(m, nn.Linear):
+                init.kaiming_uniform_(m.weight, a=0, mode='fan_in', nonlinearity='leaky_relu')
+
+    def forward(self, images):
+        outputs = self.model(images)
+        return outputs.logits
+
+model_path = hf_hub_download(repo_id="Noha90/AML_16", filename="large_swin_best_model.pth")
+print("Model path:", model_path)
+model = SwinCustom(model_name=MODEL_NAME, num_classes=40)
+state_dict = torch.load(model_path, map_location="cpu")
+if "model_state_dict" in state_dict:
+    state_dict = state_dict["model_state_dict"]
+model.load_state_dict(state_dict, strict=False)
+model.eval()
+
+# Preprocessing
+transform = transforms.Compose([
+    transforms.Resize((224, 224)),
+    transforms.ToTensor(),
+    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+])
+
+def predict(image_path):
+    image = Image.open(image_path).convert("RGB")
+    x = transform(image).unsqueeze(0)
+    with torch.no_grad():
+        outputs = model(x)
+        print("Logits:", outputs)
+        probs = torch.nn.functional.softmax(outputs, dim=1)[0]
+        print("Probs:", probs)
+        print("Sum of probs:", probs.sum())
+        top5 = torch.topk(probs, k=5)
+
+    top1_idx = int(top5.indices[0])
+    top1_label = class_names[top1_idx]
+
+    # Select a random image from the class subfolder
+    class_folder = f"sample_images/{str(top1_label).replace(' ', '_')}"
+    reference_image = None
+    if os.path.isdir(class_folder):
+        image_files = [f for f in os.listdir(class_folder) if f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp", ".gif", ".webp"))]
+        if image_files:
+            chosen_file = random.choice(image_files)
+            ref_path = os.path.join(class_folder, chosen_file)
+            print(f"[DEBUG] Randomly selected reference image: {ref_path}")
+            reference_image = Image.open(ref_path).convert("RGB")
+        else:
+            print(f"[DEBUG] No images found in {class_folder}")
+    else:
+        print(f"[DEBUG] Class folder does not exist: {class_folder}")
+
+    top5_probs = {class_names[int(idx)]: float(score) for idx, score in zip(top5.indices, top5.values)}
+    print(f"image path: {image_path}")
+    print(f"top1_label: {top1_label}")
+    print(f"[DEBUG] Top-5 indices: {top5.indices}")
+    print(f"[DEBUG] Top-5 labels: {[class_names[int(idx)] for idx in top5.indices]}")
+    print(f"[DEBUG] Top-5 probs: {top5_probs}")
+
+    return image, reference_image, top5_probs

requirements.txt

@@ -0,0 +1,7 @@
+torch
+torchvision
+transformers
+Pillow
+gradio
+numpy
+huggingface_hub