def _intermediate_eval(self, session, batch):
if session.params['eval_space'] == 'original':
batch.original_poses = self.normalizer.denormalize(
batch.poses, batch.normalization_params)
train_results = {
'DEFAULT': Evaluator.to_batch(batch, session.params['eval_space'])
}
train_mean_results = Evaluator.means_per_metric(train_results)
Evaluator.results_to_cpu(train_mean_results)
val_mean_results = session.test_model(self.val_loader)
Evaluator.results_to_cpu(val_mean_results)
return train_mean_results, val_mean_results
def test_to_batch():
batch_size = 42
poses = torch.rand(batch_size, 21, 3)
labels = poses + torch.ones(batch_size, 21, 3)
batch = datasets.PoseCorrectionBatch(poses, labels, poses, labels)
true_results = {
'coord_diff': torch.ones(batch_size),
'distance': torch.sqrt(3.0 * torch.ones(batch_size)),
'bone_length': torch.zeros(batch_size),
'proportion': torch.zeros(batch_size),
}
results_norm = Evaluator.to_batch(batch)
results_orig = Evaluator.to_batch(batch, space='original')
for metric_name in Evaluator.metric_names:
assert torch.allclose(results_norm[metric_name],
true_results[metric_name],
atol=1e-6)
assert torch.allclose(results_orig[metric_name],
true_results[metric_name],
atol=1e-6)
