In order to build a machine learning system, it is advised to start with a new small project and improve it progressively:
- Find a similar problem to yours and download code (and test the model to check results)
- Find ways to scale your computation if needed (namely, AWS/Google Cloud)
- Start with smaller datasets to avoid losing time just waiting for a single epoch
- Start with a simple architecture
- Use visualization/debugging (for instance, TensorBoard)
- Fine-tune the model, fine-tune hyperparameters, depth, architecture, layers, and the loss function
- Expand your dataset and ensure that it is as clean as possible
- Split your dataset into training, development, and testing sets
- Evaluate your model