lstm.json (1388B)
1 { 2 "_comments": [ 3 "Reminders and tips:", 4 " * For your data, use a long ny like 32768.", 5 " * For this model, it really helps if you have the delay in your data set", 6 " correctly. I've seen improvements fixing a delay that was off by 10", 7 " samples.", 8 " * gamma below is picked so that we end up with a learning rate of about", 9 " 1e-4 after 1000 epochs. I've found LSTMs to work with a pretty aggressive", 10 " learning rate that would be out of the question for other architectures.", 11 " * Number of units between 8 and 96, layers from 1 to 5 all seem to be ok", 12 " depending on the dataset, though bigger models might not make real-time.", 13 "", 14 "Dev note: Ensure that tests/test_bin/test_train/test_main.py's data is ", 15 "representative of this!" 16 ], 17 "net": { 18 "name": "LSTM", 19 "config": { 20 "num_layers": 3, 21 "hidden_size": 18, 22 "train_burn_in": 8192, 23 "train_truncate": null 24 } 25 }, 26 "loss": { 27 "val_loss": "esr", 28 "mask_first": 8192, 29 "pre_emph_mrstft_weight": 0.002, 30 "pre_emph_mrstft_coef": 0.85 31 }, 32 "optimizer": { 33 "lr": 0.008 34 }, 35 "lr_scheduler": { 36 "class": "ExponentialLR", 37 "kwargs": { 38 "gamma": 0.995 39 } 40 } 41 }