def evaluate_config(config):
data_model = PlusShapeModel()
try:
results = {'IoU': [], 'DICE': [], 'max_h': [], 'avg_h': []}
sample = np.load(f'results/{config.replace(".", "-")}_0_sample.npy')
# print(config, sample.shape)
# return
curves = data_model.trace_fourier_curves(sample)
curves_dense = data_model.trace_fourier_curves(sample, n_points=1000)
for j in range(len(curves)):
points = torch.tensor(curves[j]).float().cpu()
params = fit_plus_shape_to_points(points)
iou, dice = iou_and_dice_plus_shape(params, points)
max_h, avg_h = max_and_avg_hausdorff_distance_plus_shape(
params, curves_dense[j])
print(config, j, iou, dice, max_h, avg_h, flush=True)
results['IoU'].append(iou)
results['DICE'].append(dice)
results['max_h'].append(max_h)
results['avg_h'].append(avg_h)
iou = np.mean(results['IoU'])
dice = np.mean(results['DICE'])
max_h = np.mean(results['max_h'])
avg_h = np.mean(results['avg_h'])
print(iou, dice, max_h, avg_h)
np.save(f'results/{config.replace(".", "-")}_0',
np.stack([iou, dice, max_h, avg_h]))
except Exception as e:
print(f'ERROR with config "{config}"')
print(e)
traceback.print_exc()
'n_reflections': c['ndim_x']
},
name=f'perm_{i+1}'))
nodes.append(
Node(nodes[-1],
AffineCoupling, {
'F_class': F_fully_connected,
'F_args': {
'internal_size': c['hidden_layer_sizes']
}
},
name=f'ac_{i+1}'))
nodes.append(OutputNode(nodes[-1], name='z'))
model = ReversibleGraphNet(nodes, verbose=False)
model.to(c['device'])
model.params_trainable = list(
filter(lambda p: p.requires_grad, model.parameters()))
def model_inverse(test_z):
return model([test_z], rev=True)
c['model'] = model
c['model_inverse'] = model_inverse
# create namedtuple from config dictionary
c = namedtuple("Configuration", c.keys())(*c.values())
for i in range(c['n_blocks']):
if i > 0:
x_lane.append(Node(x_lane[-1],
HouseholderPerm,
{'fixed': False, 'n_reflections': c['ndim_x']},
name=f'perm_{i}'))
x_lane.append(Node(x_lane[-1],
HierarchicalAffineCouplingBlock,
{'c_internal': [c['hidden_layer_sizes'], c['hidden_layer_sizes']//2, c['hidden_layer_sizes']//4, c['hidden_layer_sizes']//8],
'max_splits': c['recursion_depth']},
name=f'hac_{i+1}'))
x_lane.append(OutputNode(x_lane[-1], name='z'))
model = ReversibleGraphNet(x_lane, verbose=False)
model.to(c['device'])
model.params_trainable = list(filter(lambda p: p.requires_grad, model.parameters()))
def model_inverse(test_z):
return model(test_z, rev=True)
c['model'] = model
c['model_inverse'] = model_inverse
# create namedtuple from config dictionary
c = namedtuple("Configuration",c.keys())(*c.values())
# create namedtuple from config dictionary
c = namedtuple("Configuration",c.keys())(*c.values())
##############################
### MODEL ARCHITECTURE ###
##############################
nodes = [InputNode(c.ndim_x, name='x')]
for i in range(c.n_blocks):
nodes.append(Node(nodes[-1],
HouseholderPerm,
{'fixed': False, 'n_reflections': c.ndim_x},
name=f'perm_{i+1}'))
nodes.append(Node(nodes[-1],
AffineCoupling,
{'F_class': F_fully_connected,
'F_args': {'internal_size': c.hidden_layer_sizes}},
name=f'ac_{i+1}'))
nodes.append(OutputNode(nodes[-1], name='z'))
model = ReversibleGraphNet(nodes, verbose=False)
model.to(c.device)
def model_inverse(test_z):
return model([test_z], rev=True)