More fixes

[jonashaag]

• Improve memory usage during training by allowing for garbage collection
• Fix model loading from checkpoint
• Remove unused imports
• Add feature: skip initial N frames of each audio sample