loss_indices = np.where(~np.isnan(targets))
if use_gpu:
targets = targets.cuda()
inputs = inputs[:, begin_loss_ind:, :]
depths = depths[:, begin_loss_ind:]
mse = mse_criterion(pred[loss_indices], targets[loss_indices])
avg_mse += mse
if mse > 0: #obsolete i think
ct += 1
avg_mse = avg_mse / ct
#save model
(outputm_npy, labelm_npy) = parseMatricesFromSeqs(pred.cpu().numpy(), targets.cpu().numpy(), depths, tmp_dates, n_depths,
n_test_dates_target, u_depths_target,
unique_tst_dates_target)
#store output
output_mats[i,:,:] = outputm_npy
if i == 0:
#store label
label_mats = labelm_npy
loss_output = outputm_npy[~np.isnan(labelm_npy)]
loss_label = labelm_npy[~np.isnan(labelm_npy)]
mat_rmse = np.sqrt(((loss_output - loss_label) ** 2).mean())
#save model
total_output_npy = np.average(output_mats,axis=0)
loss_output = total_output_npy[~np.isnan(label_mats)]
h_state = None
lstm_net.hidden = lstm_net.init_hidden(batch_size=inputs.size()[0])
pred, h_state = lstm_net(inputs, h_state)
pred = pred.view(pred.size()[0], -1)
pred = pred[:, begin_loss_ind:]
#calculate error
loss_indices = np.where(~np.isnan(targets))
inputs = inputs[:, begin_loss_ind:, :]
depths = depths[:, begin_loss_ind:]
mse = mse_criterion(pred[loss_indices], targets[loss_indices])
avg_mse += mse
#fill in data structs to save model outputs
(output_npy, label_npy) = parseMatricesFromSeqs(pred, targets, depths, \
tmp_dates, n_depths, \
n_test_dates, u_depths, \
unique_tst_dates)
#save model
saveModel(lstm_net.state_dict(), optimizer.state_dict(), save_path)
print("training finished in "+ str(epoch) + " epochs")
print("finished trial", trial)
print("rmse=", np.sqrt(avg_mse))
loss_output = output_npy[~np.isnan(label_npy)]
loss_label = label_npy[~np.isnan(label_npy)]
mat_rmse = np.sqrt(((loss_output - loss_label) ** 2).mean())
print("Total rmse=", mat_rmse)
saveFeatherFullData(output_npy, label_npy, unique_tst_dates, args.lake_name, trial)
err_per_trial[trial] = mat_rmse