Open Saurabhbhati opened 1 year ago
Based on the README and the paper I am using the following hyper-parameters for training the variable-rate CPC
python cpc/train.py --pathDB /path_datasets/LibriSpeech/train-clean-100 --file_extension '.flac' --pathCheckpoint ./hcpc --normMode layerNorm --dropout --n_process_loader 1 --batchSizeGPU 32 --CPCCTC --nPredicts 12 --CPCCTCNumMatched 12 --limitNegsInBatch 8 --nEpoch 50 --nGPU 1 --nLevelsGRU 2 --schedulerRamp 10 --multiLevel --segmentationMode boundaryPredictor --nPredictsSegment 2 --CPCCTCNumMatchedSegment 2 --adjacentNegatives --targetQuantizerSegment robustKmeans
However, the phone segmentation results (R-value 73.23) are lower than the ones in the paper (R-value 81.98) for Librispeech dataset. I seem to be missing something. Could you please take a look and share the optimal parameters?
Thank you.
Based on the README and the paper I am using the following hyper-parameters for training the variable-rate CPC
python cpc/train.py --pathDB /path_datasets/LibriSpeech/train-clean-100 --file_extension '.flac' --pathCheckpoint ./hcpc --normMode layerNorm --dropout --n_process_loader 1 --batchSizeGPU 32 --CPCCTC --nPredicts 12 --CPCCTCNumMatched 12 --limitNegsInBatch 8 --nEpoch 50 --nGPU 1 --nLevelsGRU 2 --schedulerRamp 10 --multiLevel --segmentationMode boundaryPredictor --nPredictsSegment 2 --CPCCTCNumMatchedSegment 2 --adjacentNegatives --targetQuantizerSegment robustKmeansHowever, the phone segmentation results (R-value 73.23) are lower than the ones in the paper (R-value 81.98) for Librispeech dataset. I seem to be missing something. Could you please take a look and share the optimal parameters?
Thank you.