Train with my own data

[thanostriantafyllou3]

Hello,

Could you please provide some code to train GlaucusAE with my own data?

I have been trying to do that but without any success.

I am storing my data as np.complex64 in a .npy file.

Thank you in advance!

[thanostriantafyllou3]

Specifically I was getting an issue about mat-mul different shapes at the fc_encoder. I removed fc_encoder and fc_decoder and I am now able to train a model with data: signal_data = torch.randn(11, 11264, dtype=torch.complex64) instead of torchsig (notice that at the step() function I just remove the metadata variable for that to work).

However, is it possible to fix fc_encoder too? I assume it's either a bug or the way the data is represented?

[Teque5]

Training

Training is often quite complex and for this release it was beyond the scope of our work. I'm not sure what your training loop looks like but I highly recommend using the lightning's Trainer class with any kind of data (including TorchSig). The training should look something like:

import lightning as L
from lightning.pytorch.callbacks import (EarlyStopping, ModelCheckpoint,
                                         RichProgressBar)

checkpoint = ModelCheckpoint(monitor="val_rf_loss", filename="check-{epoch:03d}-{val_spec_loss:05f}", save_top_k=3, mode='min')
early_stopping = EarlyStopping(monitor="val_rf_loss", mode="min", patience=10) # whatever patience you can tolerate
progress = RichProgressBar()

model = GlaucusRVQVAE() # or any model
datamodule = SomeDataModule() # should be L.LightningDataModule with train_dataloader, val_dataloader, test_dataloader methods

# train model
trainer = L.Trainer(max_epochs=-1, callbacks=[checkpoint, early_stopping, progress])
trainer.fit(model, datamodule=datamodule)

# rewind to best checkpoint
model.load_from_checkpoint(trainer.checkpoint_callback.best_model_path)
# freeze weights
model.freeze()
model.eval()
# evaluate model
trainer.test(datamodule=datamodule)

Issue with x, metadata = batch

I believe your confusion arises from the fact that we expect our dataloader to produce batches that contain both complex RF and metadata describing it. You can do that or make the modification you suggest. We expect something akin to this if you are using data_format="nl":

>>> rf_tensor, rf_metadata = next(iter(datamodule.train_dataloader())) # first iteration only
>>> rf_tensor.shape, rf_tensor.dtype
(torch.Size([256, 11264]), torch.complex64)
>>> type(rf_metadata)
dict

This work was completed before torchsig existed and we have our own suite of generators, recordings, and RF impairments implemented on-the-fly.

Suggestions

• We recommend using as much data as possible along with a curriculum training strategy where you use high SNRs with easy RF impairments and progressively ramp to lower SNRs with harder impairments.
• If you have a small corpus of data I suggest you use transfer learning and load one of our model checkpoints before you begin training.