def build_net(steps=steps_inference, mode="inference"):
unet = unet_class.UNet(
feature_widths_down,
feature_widths_up,
downsampling_factors,
kernel_sizes_down,
kernel_sizes_up,
padding=padding,
constant_upsample=constant_upsample,
trans_equivariant=trans_equivariant,
input_voxel_size=voxel_size,
input_fov=voxel_size,
)
if voxel_size == voxel_size_input:
net, input_shape, output_shape = make_net(unet, labels, steps, loss_name=loss_name, mode=mode)
else:
net, input_shape, output_shape = make_net_upsample(unet, labels, steps, upsample_factor,
final_kernel_size, final_feature_width,
loss_name=loss_name, mode=mode)
logging.info(
"Built {0:} with input shape {1:} and output_shape {2:}".format(
net, input_shape, output_shape
)
)
return net, input_shape, output_shape
def build_net(mode="inference"):
if mode != "training" and not os.path.exists(
"{0:}_io_names.json".format(network_name)):
logging.info("Building mode training first to generate io names")
build_net(mode="training")
tf.reset_default_graph()
if mode == "inference" or mode == "forward":
padding = padding_inference
add_context = add_context_inference
elif mode == "training":
padding = padding_train
add_context = add_context_training
else:
raise ValueError("Unkown mode: {0:}".format(mode))
unet = unet_class.UNet(
feature_widths_down,
feature_widths_up,
downsampling_factors,
kernel_sizes_down,
kernel_sizes_up,
skip_connections=skip_connections,
padding=padding,
constant_upsample=constant_upsample,
trans_equivariant=trans_equivariant,
enforce_even_context=enforce_even_context,
input_voxel_size=voxel_size,
input_fov=voxel_size,
)
net, input_shape, output_shape = make_net(network_name,
unet,
n_out,
add_context,
input_name=input_name,
output_names=output_names,
loss_name=loss_name,
mode=mode)
logging.info(
"Built {0:} with input shape {1:} and output_shape {2:}".format(
net, input_shape, output_shape))
return net, input_shape, output_shape
