outputs_cpu = outputs.data.cpu().numpy()
lossL1 = criterionL1(outputs, targets)
L1val_accum += lossL1.item()
if i == 0:
input_ndarray = inputs_cpu.cpu().numpy()[0]
v_norm = (np.max(np.abs(input_ndarray[0, :, :]))**2 +
np.max(np.abs(input_ndarray[1, :, :]))**2)**0.5
outputs_denormalized = data.denormalize(outputs_cpu[0], v_norm)
targets_denormalized = data.denormalize(
targets_cpu.cpu().numpy()[0], v_norm)
utils.makeDirs(["results_train"])
utils.imageOut("results_train/epoch{}_{}".format(epoch, i),
outputs_denormalized,
targets_denormalized,
saveTargets=True)
# data for graph plotting
L1_accum /= len(trainLoader)
L1val_accum /= len(valiLoader)
if saveL1:
if epoch == 0:
utils.resetLog(prefix + "L1.txt")
utils.resetLog(prefix + "L1val.txt")
utils.log(prefix + "L1.txt", "{} ".format(L1_accum), False)
utils.log(prefix + "L1val.txt", "{} ".format(L1val_accum), False)
torch.save(netG.state_dict(), prefix + "modelG")
targets_denormalized_comp, outputs_denormalized_comp = targets_denormalized_comp.float(
).cuda(), outputs_denormalized_comp.float().cuda()
outputs_dn.data.resize_as_(outputs_denormalized_comp).copy_(
outputs_denormalized_comp)
targets_dn.data.resize_as_(targets_denormalized_comp).copy_(
targets_denormalized_comp)
loss_dn = criterionLoss(outputs_dn, targets_dn)
Lossval_dn_accum += loss_dn.item()
# write output image, note - this is currently overwritten for multiple models
os.chdir("./results_test/")
utils.imageOut("%04d" % (i),
outputs_cpu,
targets_cpu,
normalize=False,
saveMontage=True) # write normalized with error
os.chdir("../")
log(lf, "\n")
Lossval_accum /= len(testLoader)
lossPer_p_accum /= len(testLoader)
lossPer_v_accum /= len(testLoader)
lossPer_accum /= len(testLoader)
Lossval_dn_accum /= len(testLoader)
log(
lf, "Loss percentage (p, v, combined): %f %% %f %% %f %% " %
(lossPer_p_accum * 100, lossPer_v_accum * 100, lossPer_accum * 100))
log(lf, "Loss error: %f" % (Lossval_accum))
log(lf, "Denormalized error: %f" % (Lossval_dn_accum))
