initial commit

.gitignore

@@ -0,0 +1,18 @@
+# Python
+__pycache__/
+*.pyc
+.venv/
+venv/
+env/
+
+# Gradio local run artifacts
+flagged/
+.gradio/
+
+# OS / editor
+.DS_Store
+.vscode/
+.idea/
+
+# Env/secrets
+.env

README.md

@@ -1,15 +1,28 @@
 ---
-title: Instant Gram
-emoji: 🏢
-colorFrom: pink
-colorTo: gray
+title: Instant-Gram
+emoji: 🫘
+colorFrom: yellow
+colorTo: green
 sdk: gradio
-sdk_version: 6.28.0
-python_version: '3.12'
+sdk_version: 4.44.0
 app_file: app.py
 pinned: false
 license: mit
 short_description: Image Classifier for Indian Lentils
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Instant-Gram 🫘
+
+A ResNet18 image classifier fine-tuned on the [Indian Lentils dataset](https://data.mendeley.com/datasets/r4yhfnzc5y/1)
+— 20 varieties of lentils, beans, chickpeas, and grams (Kidney Bean, Horse Gram,
+Chana, Black-Eyed Pea, and more).
+
+Upload a photo, or use your camera, to classify a lentil/bean/chickpea variety
+in real time.
+
+**Model:** ResNet18 (ImageNet-pretrained, fine-tuned), 99.9% held-out accuracy
+— matches an EfficientNetB0 baseline (the architecture a [published benchmark
+study](https://www.sciencedirect.com/science/article/pii/S2666154323004507#cited-by) found best for this dataset) while training 1.55x faster (T4 GPU), via per-architecture learning-rate tuning.
+
+Full methodology, GPU performance benchmarking (mixed precision, `torch.compile`,
+cross-architecture comparison), and findings: see the project write-up.

app.py

@@ -0,0 +1,59 @@
+import spaces
+import gradio as gr
+import torch
+import torch.nn as nn
+from torchvision import models, transforms
+from PIL import Image
+
+# Order matches full_train.classes from the training notebook (ImageFolder's
+# alphabetical folder sort) — do not reorder without re-checking the notebook.
+CLASS_NAMES = [
+    "Black_Beans", "Black_EyedPea", "Brown_Chickpea", "Desi_chickpea",
+    "Horse_Gram", "Husked_BlackGram", "Husked_RedLentil", "Kidney_Bean",
+    "Red_Beans", "Red_KidneyBean", "Red_chickpea", "Soya_Bean",
+    "White_Chickpea", "White_beans", "White_peabeans", "Whole_BlackGram",
+    "Whole_GreenGram", "Whole_TurkishGram", "chana1", "chana2",
+]
+
+# ZeroGPU only grants GPU access inside a function decorated with
+# @spaces.GPU — no CUDA context exists at import time, so the model is
+# loaded on CPU here and moved to CUDA per-call inside predict().
+model = models.resnet18(weights=None)
+model.fc = nn.Linear(model.fc.in_features, len(CLASS_NAMES))
+model.load_state_dict(torch.load("resnet_lentils.pth", map_location="cpu"))
+model.eval()
+
+# Same eval_transform used during evaluation in the training notebook.
+transform = transforms.Compose([
+    transforms.Resize(256),
+    transforms.CenterCrop(224),
+    transforms.ToTensor(),
+    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+])
+
+
+@spaces.GPU
+def predict(image: Image.Image):
+    if image is None:
+        return {}
+    device = torch.device("cuda")
+    model.to(device)
+    x = transform(image.convert("RGB")).unsqueeze(0).to(device)
+    with torch.no_grad():
+        probs = torch.softmax(model(x), dim=1)[0]
+    return {CLASS_NAMES[i]: float(probs[i]) for i in range(len(CLASS_NAMES))}
+
+
+demo = gr.Interface(
+    fn=predict,
+    inputs=gr.Image(type="pil", sources=["upload", "webcam"], label="Photo of a lentil / bean / chickpea"),
+    outputs=gr.Label(num_top_classes=5, label="Prediction"),
+    title="Indian Lentils Classifier",
+    description=(
+        "ResNet18, fine-tuned on the Indian Lentils dataset (20 classes). "
+        "Upload a photo or use your camera to classify a lentil, bean, or chickpea variety."
+    ),
+)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,5 @@
+torch
+torchvision
+gradio
+pillow
+spaces