大纲

In this notebook you will:

• See a powerful custom diffusion model pipeline in action (with information on how to make your own version)
• Create your own mini pipeline by:
  • Recapping the core ideas behind diffusion models
  • Loading in data from the Hub for training
  • Exploring how we add noise to this data with a scheduler
  • Creating and training the UNet model
  • Putting the pieces together into a working pipeline
• Edit and run a script for initializing longer training runs, that will handle
  • Multi-GPU training via 🤗 Accelerate
  • Experiment logging to track critical stats
  • Uploading the final model to the Hugging Face Hub

步骤

Step 2: Download a training dataset

Step 3: Define the Scheduler

Step 4: Define the Model

Step 5: Create a Training Loop

Step 6: Generate Images