.
🧠 Model Card: Reactor Stage Classifier (Fine-tuned ResNet-50)
Model Description
This model is a fine-tuned version of Microsoft’s ResNet-50 designed to classify images of a plasma reactor into three operational stages:
l (low), h (high), and elm (edge-localized mode).
The dataset consists of labeled reactor images, preprocessed and augmented to improve robustness.
The fine-tuning was performed using the Hugging Face transformers and datasets libraries on Google Colab, with integration to the Hugging Face Hub for model saving and version control.
🧩 Training Details
Base model: microsoft/resnet-50
Frameworks: transformers, datasets, torchvision, evaluate
Dataset format: ImageFolder (ZIP file from Google Drive, already labeled)
Dataset split: 85% training / 15% testing
Training samples: ~850 (train) / ~150 (test)
Training parameters:
Learning rate: 2e-4
Batch size: 16
Epochs: 10
Mixed precision (fp16): enabled
⚙️ Data Preprocessing & Augmentation
The preprocessing pipeline used several data augmentations to improve generalization, applied with torchvision.transforms:
Random resized crop (224×224)
Random horizontal flip
Color jitter (brightness, contrast, saturation, hue)
Random grayscale (10%)
Normalization with pretrained ResNet mean and std
Validation data was resized and center-cropped to the same input size.
