Create README.md

README.md

@@ -0,0 +1,100 @@
+---
+license: apache-2.0
+datasets:
+- uoft-cs/cifar10
+metrics:
+- accuracy
+base_model:
+- microsoft/resnet-18
+---
+# ResNet18 Fine-Tuned on CIFAR-10
+
+This model is a fine-tuned version of **ResNet18** (originally pretrained on ImageNet) on the [CIFAR-10 dataset](https://www.cs.toronto.edu/~kriz/cifar.html). It achieves the following results on the validation/test set:
+
+- **Validation Accuracy**: `88.60%`  
+
+---
+
+## Model description
+
+- **Architecture**: ResNet18 with the final fully-connected layer replaced by a 10-class output layer for CIFAR-10 (airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck).  
+- **Pretrained Weights**: ImageNet1K  
+- **Fine-Tuning**: The model was fine-tuned on CIFAR-10 images resized to 128×128 pixels.  
+- **Data Augmentation**: Random horizontal flip, random rotation, normalization to mean=0.5 and std=0.5.  
+
+> **Intended uses & limitations**  
+> - **Intended use**: Educational/demo purposes or as a starting point for further fine-tuning on similar image classification tasks.  
+> - **Not intended for**: Production-critical tasks without further evaluation, as CIFAR-10 is relatively small-scale, and the model may not generalize to non-CIFAR data without additional fine-tuning.
+
+---
+
+## Training procedure
+
+**Hyperparameters** (approximate):
+- **optimizer**: Adam  
+- **learning_rate**: 1e-3  
+- **batch_size**: 32  
+- **num_epochs**: 15  
+
+**GPU/CPU**:  
+- This model was trained on a single GPU (`torch.device("cuda")`) if available, otherwise CPU.
+
+**Training logs** (for each epoch on the training set):
+
+| Epoch | Training Loss | Training Accuracy | Validation Accuracy |
+|------:|--------------:|------------------:|--------------------:|
+| 1     | 0.7013        | 76.52%           | -                   |
+| 2     | 0.4248        | 85.64%           | -                   |
+| 3     | 0.3185        | 89.07%           | -                   |
+| 4     | 0.2341        | 92.06%           | -                   |
+| 5     | 0.1762        | 93.86%           | -                   |
+| 6     | 0.1302        | 95.55%           | -                   |
+| 7     | 0.1085        | 96.31%           | -                   |
+| 8     | 0.0925        | 96.82%           | -                   |
+| 9     | 0.0765        | 97.37%           | -                   |
+| 10    | 0.0683        | 97.68%           | -                   |
+| 11    | 0.0655        | 97.83%           | -                   |
+| 12    | 0.0548        | 98.18%           | -                   |
+| 13    | 0.0513        | 98.27%           | -                   |
+| 14    | 0.0461        | 98.49%           | -                   |
+| 15    | 0.0470        | 98.41%           | **88.60%**          |
+
+> **Note**: Validation accuracy was computed at the end of training (final epoch).
+
+---
+
+## Usage
+
+Below is a sample usage snippet in Python. **Replace** `username/model_repo_name` with the actual model repo id on Hugging Face.
+
+```python
+import torch
+import torch.nn as nn
+from torchvision import models, transforms
+from huggingface_hub import hf_hub_download
+from PIL import Image
+
+# Download the weights from the Hugging Face Hub
+ckpt_path = hf_hub_download(repo_id="username/model_repo_name", filename="cnn_model.pth")
+
+# Define the same model architecture
+model = models.resnet18(pretrained=False)
+model.fc = nn.Linear(model.fc.in_features, 10)  # for CIFAR-10
+model.load_state_dict(torch.load(ckpt_path, map_location="cpu"))
+model.eval()
+
+# Define transforms
+transform = transforms.Compose([
+    transforms.Resize((128, 128)),
+    transforms.ToTensor(),
+    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
+])
+
+# Example inference
+image = Image.open("your_image.jpg").convert("RGB")
+input_tensor = transform(image).unsqueeze(0)  # add batch dimension
+with torch.no_grad():
+    logits = model(input_tensor)
+    predicted_class = logits.argmax(dim=1).item()
+
+print("Predicted class ID:", predicted_class)