neural-amp-modeler

Neural network emulator for guitar amplifiers
Log | Files | Refs | README | LICENSE

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 }