def main():
# Set parameters
vae_epoch = 2
can_epoch = 1000
batch_size = 64
latent_dim = 10
beta_eeg = 5.0
train = True
# Read data sets
data_root = "/home/zainkhan/bci-representation-learning"
eeg_train, eeg_test, pupil_train, pupil_test, sub_cond = utils.read_single_trial_datasets(
data_root)
if train:
# Train VAE
vae = VAE(beta=beta_eeg, latent_dim=latent_dim)
vae.compile(optimizer=keras.optimizers.Adam())
vae.fit(eeg_train, epochs=vae_epoch, batch_size=batch_size)
# Save VAE
#vae.encoder.save("vae_encoder")
#vae.decoder.save("vae_decoder")
# Train CAN
can = CAN(
vae=vae,
can_data=pupil_train,
vae_data=eeg_train,
latent_dim=latent_dim,
epochs=can_epoch,
batch_size=batch_size,
)
can.compile(optimizer=keras.optimizers.Adam(), run_eagerly=True)
can.fit(pupil_train,
epochs=can_epoch,
batch_size=batch_size,
shuffle=False)
# Save CAN
can.encoder.save("can_encoder")
can.decoder.save("can_decoder")
else:
# Load all encoders/decoders
vae = VAE(beta=beta_eeg, latent_dim=latent_dim)
vae.encoder = keras.models.load_model("vae_encoder")
vae.decoder = keras.models.load_model("vae_decoder")
can = CAN(vae=vae, vae_data=eeg_train, latent_dim=latent_dim)
can.encoder = keras.models.load_model("can_encoder")
can.decoder = keras.models.load_model("can_decoder")
# VAE predictions
encoded_data = vae.encoder.predict(eeg_test)
decoded_data = vae.decoder.predict(encoded_data)
fn = utils.get_filename("predictions/", "test-eeg")
while real_imgs is None or real_imgs.size(0) < num_samples:
imgs = next(real_dl)
if real_imgs is None:
real_imgs = imgs[0]
else:
real_imgs = torch.cat((real_imgs, imgs[0]), 0)
real_imgs = real_imgs[:num_samples].expand(-1, 3, -1, -1)
with torch.no_grad():
samples = None
while samples is None or samples.size(0) < num_samples:
imgs = model.forward(
z=torch.randn(args.batch_size, model.latent_dim).to(device))
if samples is None:
samples = imgs
else:
samples = torch.cat((samples, imgs), 0)
samples = samples[:num_samples].expand(-1, 3, -1, -1)
samples = samples.cpu()
fid = fid_score.calculate_fid_given_images(real_imgs, samples,
args.batch_size, device)
tb_writer.add_scalar('fid', fid)
print("FID score: {:.3f}".format(fid), flush=True)
imgs = next(real_dl)[0][0:4].to(device)
noise = torch.randn(4, args.latent_dim).to(device)
mu, log_var = model.encoder(imgs).chunk(2, dim=-1)
sampled_z = model.reparameterize(mu, log_var).unsqueeze(1)
interpolate(model, sqrt_K=14, noises=sampled_z)
