Speech Enhancement Generative Adversarial Network in PyTorch

Train SEGAN replica:

Execute training with default parameters (kwidth=31, skip connections with concat merge, etc.), CUDA, a batch_size of 100:

python train.py --save_path seganv1_ckpt --cuda --batch_size 100 \
		--clean_trainset data/clean_trainset \
		--noisy_trainset data/noisy_trainset \
		--clean_valset data/clean_valset \
		--noisy_valset data/noisy_valset \
		--cache_dir data/cache

TODO: Write clean script

python clean.py --pretrained_ckpt seganv1_ckpt/weights_SEGAN-52546.ckpt \
		--cfg_file seganv1_ckpt/train.opts --synthesis_path enhanced_results \
		--test_files data/noisy_testset

Disclaimer:

• Multi-GPU is not supported yet in this framework.
• Virtual Batch Norm is not included, and similar results to those of original paper can be obtained with regular BatchNorm in D (ONLY D).