mel_roformer_experiments.md

January 19, 2025 ยท View on GitHub

Mel Roformer models

All experiments were made using MUSDB18HQ dataset. All metrics were measured using 'test' set. Training was made using 'train' set.

Experiments table

Average SDR ScoreChunk sizeDepthDimmlp expansion factorSkip connectionHop sizeFFT SizeDropoutBatch SizeDL CheckpointComment
5.1235882002641No44120480/032 (48 GB)Config / Weights
6.46988820041281No44120480.1/0.128 (80 GB)Config / Weights
6.70228820041281No88240960/020 (80 GB)Config / Weights
7.81278820062561Yes44120480.1/0.116 (80 GB)Config / Weights
6.490817640041281Yes44120480.1/0.18 (48 GB)Config / Weights
6.522417640041282Yes44120480.1/0.18 (48 GB)Config / Weights
7.041235280041281No88240960/05 (80 GB)Config / Weights
8.217535280042561No44120480/05 (80 GB)Config / WeightsTrained longer on different strategies. Looks like it a bit overfit in the end
1.055735280041281No88220480/06 (48 GB)---Looks like big hop size is not great
6.865248510041281No44120480.1/0.15 (48 GB)Config / Weights
8.9400*48510083844Yes88240960/02 (80 GB)Config / Weights (part 1, part2)Very big file with weights > 3GB. Continue to increase metrics
  • Note 1: Some models probably undertrained