# for dt_str, sample_idx in zip(['train', 'test'], [train_idx, test_idx]):
sample_idx = np.intersect1d(multi_dt.get_comb_idx(omic_combn),
sample_idx)
dl = DataLoader(
dataset=Subset(multi_dt, sample_idx),
batch_size=128,
collate_fn=lambda x: collate_wrapper(x, omic_combn),
shuffle=False)
latent_test = []
for i, x in enumerate(dl):
x = [[
x_i.to(device) if x_i is not None else None
for x_i in y
] for y in x]
latent_test += [model.get_latent(x, elbo_bool=omic_combn)]
res = np.concatenate(latent_test)
np.savetxt(output_dir + file_sub +
"_".join([str(x) for x in omic_combn]) + "_" +
dt_str + "_latent.csv",
res,
delimiter=",")
np.savetxt(output_dir + file_sub +
"_".join([str(x) for x in omic_combn]) + "_" +
dt_str + "_barcode.csv",
barcode[sample_idx],
fmt="%s")
plt.close()
# === save feature
for i, x in enumerate(model.get_beta()):
loss = annealing_factor * latent_loss + recon_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
latent_loss_all += [latent_loss.item()]
recon_loss_all += [recon_loss.item()]
history['loss'] += [
np.array(latent_loss_all) + np.array(recon_loss_all)
]
if np.isnan(history['loss'][-1].mean()):
raise ValueError
# print([x.mean() / 128 for x in history['loss']])
with torch.no_grad():
for test_elbo_comb in [[True, False], [False, True], [True, True]]:
for dt_str, dl in zip(['train'], [train_loader, test_loader]):
topic_prop_test = []
for i, x in enumerate(dl):
topic_prop_test += [
model.get_latent([y.to(device) for y in x],
elbo_bool=test_elbo_comb)
]
res = np.concatenate(topic_prop_test)
np.savetxt(os.path.join(
output_dir,
file_sub + "_".join([str(x) for x in test_elbo_comb]) + \
"_" + dt_str + "_latent.csv"),
res, delimiter=",")