def test_psnr_batch():
shape4d = [4, 6, 32, 32]
input, input2 = create_input_pair(shape4d, sigma=0.1)
input_numpy = utils.torch_to_numpy(input)
input2_numpy = utils.torch_to_numpy(input2)
err = loss.psnr(input, input2, batch=True).item()
err_numpy = 0
for i in range(shape4d[0]):
err_curr = psnr(input_numpy[i], input2_numpy[i])
err_numpy += err_curr
err_numpy /= shape4d[0]
assert np.allclose(err, err_numpy)
shape5d = [4, 6, 1, 32, 32]
input, input2 = create_input_pair(shape5d, sigma=0.1)
input_numpy = utils.torch_to_numpy(input)
input2_numpy = utils.torch_to_numpy(input2)
err = loss.psnr(input, input2)
err_numpy = 0
for i in range(shape5d[0]):
err_numpy += psnr(input_numpy[i][:,0], input2_numpy[i][:,0])
err_numpy /= shape5d[0]
assert np.allclose(err, err_numpy)
def _evaluate(self, val_logs):
losses = []
outputs = defaultdict(list)
targets = defaultdict(list)
for log in val_logs:
losses.append(log['val_loss'].cpu().numpy())
for i, (fname, slice) in enumerate(zip(log['fname'],
log['slice'])):
outputs[fname].append((slice, log['output'][i]))
targets[fname].append((slice, log['target'][i]))
metrics = dict(val_loss=losses, nmse=[], ssim=[], psnr=[])
for fname in outputs:
output = np.stack([out for _, out in sorted(outputs[fname])])
target = np.stack([tgt for _, tgt in sorted(targets[fname])])
metrics['nmse'].append(evaluate.nmse(target, output))
metrics['ssim'].append(evaluate.ssim(target, output))
metrics['psnr'].append(evaluate.psnr(target, output))
metrics = {
metric: np.mean(values)
for metric, values in metrics.items()
}
print(metrics, '\n')
# save the metrics data
metric_file_path = Path(
self.hparams.exp_dir) / self.hparams.exp / "validation_metrics"
metric_file_path.mkdir(parents=True, exist_ok=True)
metric_file_path = metric_file_path / "metrics.csv"
df = pd.DataFrame([metrics])
if metric_file_path.exists():
df.to_csv(metric_file_path, mode="a", header=False, index=False)
else:
df.to_csv(metric_file_path, mode="w", header=True, index=False)
return dict(log=metrics, **metrics)
def test_psnr(shape):
input, input2 = create_input_pair(shape, sigma=0.1)
input_numpy = utils.torch_to_numpy(input)
input2_numpy = utils.torch_to_numpy(input2)
err = loss.psnr(input, input2, batch=False).item()
err_numpy = psnr(input_numpy, input2_numpy)
assert np.allclose(err, err_numpy)
def eval(args, model, data_loader):
model.eval()
psnrl = []
with torch.no_grad():
for (input, target, mean, std) in data_loader:
input = input.to(args.device)
recons = model(input).to('cpu').squeeze(1)
target = target.cpu()
psnrl.append(psnr(target.numpy(), recons.numpy()))
print(f'PSNR: {np.mean(psnrl):.2f}')
return
def eval(args, model, data_loader):
model.eval()
psnr_l = []
ssim_l = []
with torch.no_grad():
for (input, target, mean, std, norm) in data_loader:
input = input.to(args.device)
recons = model(input).to('cpu').squeeze(1)
# recons = transforms.complex_abs(recons) # complex to real
recons = recons.squeeze()
target=target.to('cpu')
psnr_l.append(psnr(target.numpy(), recons.numpy()))
ssim_l.append(ssim(target.numpy(), recons.numpy()))
print(f'PSNR: {np.mean(psnr_l):.2f} +- {np.std(psnr_l):.2f}, SSIM: {np.mean(ssim_l):.4f} +- {np.std(ssim_l):.4f}')
return
def _evaluate(self, val_logs):
losses = []
outputs = defaultdict(list)
targets = defaultdict(list)
for log in val_logs:
losses.append(log['val_loss'].cpu().numpy())
for i, (fname, slice) in enumerate(zip(log['fname'], log['slice'])):
outputs[fname].append((slice, log['output'][i]))
targets[fname].append((slice, log['target'][i]))
metrics = dict(val_loss=losses, nmse=[], ssim=[], psnr=[])
for fname in outputs:
output = np.stack([out for _, out in sorted(outputs[fname])])
target = np.stack([tgt for _, tgt in sorted(targets[fname])])
metrics['nmse'].append(evaluate.nmse(target, output))
metrics['ssim'].append(evaluate.ssim(target, output))
metrics['psnr'].append(evaluate.psnr(target, output))
metrics = {metric: np.mean(values) for metric, values in metrics.items()}
print(metrics, '\n')
return dict(log=metrics, **metrics)