DhruvParth/ddpm-celebahq-256-fineTuned-skin_rash_v2_12epochs

Fine-tuned Skin Rash Detection Model v2

Model Overview

• Base Model: google/ddpm-celebahq-256
• Fine-tuning Dataset: Skin Rash Dataset
• Output Image Size: 256 x 256 pixels

Updates and Changes

1. Parameter Increase:

   • Previous Model: 18.54 Million Parameters
   • Current Model: 113.67 Million Parameters

2. Image Output Size:

   • Previous: 32 x 32 pixels
   • Current: 256 x 256 pixels

3. Batch Size:

   • Previous: 64 samples per batch
   • Current: 6 samples per batch

4. Gradient Accumulation:

   • Implemented gradient accumulation to simulate a larger batch size without exceeding GPU memory limits.

5. Timesteps:

   • Reduced timesteps from 1000 to 40 for faster training and inference.

Training Details

• Epochs: 12
• Number of Training Images: 656
• Optimizer: AdamW with learning rate adjusted for gradient accumulation
• Learning Rate: Initially set at (1 imes 10^-5), adjusted for gradient accumulation
• Gradient Accumulation Steps: 2
• Loss Function: Mean Squared Error (MSE) Loss

Performance

• Training Loss: Monitored and recorded over 12 epochs, with periodic visualization of generated images to ensure the model learns effectively.

Training Environment

• Hardware: GPU with 15GB RAM
• Software: PyTorch, torchvision, diffusers library

Usage

• The fine-tuned model can generate high-resolution images of skin rashes, which can be useful for medical analysis and diagnosis.
• Due to the increased parameter count and higher resolution, the model is more accurate in capturing fine details in the images.

Example Usage Code

from diffusers import DDPMPipeline

pipeline = DDPMPipeline.from_pretrained('DhruvParth/ddpm-celebahq-256-fineTuned-skin_rash_v2_12epochs')
image = pipeline().images[0]
image

#### Considerations
- **Batch Size**: Due to GPU memory constraints, batch size is limited to 6 samples.
- **Training Time**: Reduced timesteps to 40, allowing faster training and inference without significantly compromising on quality.
- **Gradient Accumulation**: Enabled to manage large model parameter count and small batch size effectively.