Debugging learning

List of factors that might impact the learning phrase:

1. Batch size

2. Learning rate & Optimizer & warm up state

3. Number of demonstrations  (the demonstrations are not enough to sufficiently being trained.)

4. Loss is wrong

 

Number of demonstration depending

- How much randomized (e.g., domain, object position, pose)

- Multi-modal solutions.

 

Memory reduction

1. fp32 -> fp16

 

How to compare computation complexity

- gFLOPs

- fp32 thoughout (?, not sure)