def main():
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
set_determinism(12345)
device = torch.device("cuda:0")
# load generator
network_filepath = "./network_final.pth"
data = torch.load(network_filepath)
latent_size = 64
gen_net = Generator(latent_shape=latent_size,
start_shape=(latent_size, 8, 8),
channels=[32, 16, 8, 1],
strides=[2, 2, 2, 1])
gen_net.conv.add_module("activation", torch.nn.Sigmoid())
gen_net.load_state_dict(data["g_net"])
gen_net = gen_net.to(device)
# create fakes
output_dir = "./generated_images"
if not os.path.isdir(output_dir):
os.mkdir(output_dir)
num_fakes = 10
print("Generating %d fakes and saving in %s" % (num_fakes, output_dir))
fake_latents = make_latent(num_fakes, latent_size).to(device)
save_generator_fakes(output_dir, gen_net(fake_latents))
def test_type_shape(self, input_data, expected_type, expected_count,
expected_shape):
results = []
for p in TEST_NDARRAYS + (None, ):
input_data = deepcopy(input_data)
if p is not None:
input_data["indices"] = p(input_data["indices"])
set_determinism(0)
result = generate_label_classes_crop_centers(**input_data)
self.assertIsInstance(result, expected_type)
self.assertEqual(len(result), expected_count)
self.assertEqual(len(result[0]), expected_shape)
# check for consistency between numpy, torch and torch.cuda
results.append(result)
if len(results) > 1:
for x, y in zip(result[0], result[-1]):
assert_allclose(x, y, type_test=False)
def tearDown(self):
set_determinism(None)
def setUp(self):
set_determinism(0)
super().setUp()
def test_invert(self):
set_determinism(seed=0)
im_fname, seg_fname = [
make_nifti_image(i)
for i in create_test_image_3d(101, 100, 107, noise_max=100)
]
transform = Compose([
LoadImaged(KEYS),
AddChanneld(KEYS),
Orientationd(KEYS, "RPS"),
Spacingd(KEYS,
pixdim=(1.2, 1.01, 0.9),
mode=["bilinear", "nearest"],
dtype=np.float32),
ScaleIntensityd("image", minv=1, maxv=10),
RandFlipd(KEYS, prob=0.5, spatial_axis=[1, 2]),
RandAxisFlipd(KEYS, prob=0.5),
RandRotate90d(KEYS, spatial_axes=(1, 2)),
RandZoomd(KEYS,
prob=0.5,
min_zoom=0.5,
max_zoom=1.1,
keep_size=True),
RandRotated(KEYS,
prob=0.5,
range_x=np.pi,
mode="bilinear",
align_corners=True),
RandAffined(KEYS, prob=0.5, rotate_range=np.pi, mode="nearest"),
ResizeWithPadOrCropd(KEYS, 100),
ToTensord(
"image"
), # test to support both Tensor and Numpy array when inverting
CastToTyped(KEYS, dtype=[torch.uint8, np.uint8]),
])
data = [{"image": im_fname, "label": seg_fname} for _ in range(12)]
# num workers = 0 for mac or gpu transforms
num_workers = 0 if sys.platform == "darwin" or torch.cuda.is_available(
) else 2
dataset = CacheDataset(data, transform=transform, progress=False)
loader = DataLoader(dataset, num_workers=num_workers, batch_size=5)
# set up engine
def _train_func(engine, batch):
self.assertTupleEqual(batch["image"].shape[1:], (1, 100, 100, 100))
engine.state.output = batch
engine.fire_event(IterationEvents.MODEL_COMPLETED)
return engine.state.output
engine = Engine(_train_func)
engine.register_events(*IterationEvents)
# set up testing handler
TransformInverter(
transform=transform,
loader=loader,
output_keys=["image", "label"],
batch_keys="label",
nearest_interp=True,
postfix="inverted1",
to_tensor=[True, False],
device="cpu",
num_workers=0
if sys.platform == "darwin" or torch.cuda.is_available() else 2,
).attach(engine)
# test different nearest interpolation values
TransformInverter(
transform=transform,
loader=loader,
output_keys=["image", "label"],
batch_keys="image",
nearest_interp=[True, False],
post_func=[lambda x: x + 10, lambda x: x],
postfix="inverted2",
num_workers=0
if sys.platform == "darwin" or torch.cuda.is_available() else 2,
).attach(engine)
engine.run(loader, max_epochs=1)
set_determinism(seed=None)
self.assertTupleEqual(engine.state.output["image"].shape,
(2, 1, 100, 100, 100))
self.assertTupleEqual(engine.state.output["label"].shape,
(2, 1, 100, 100, 100))
# check the nearest inerpolation mode
for i in engine.state.output["image_inverted1"]:
torch.testing.assert_allclose(
i.to(torch.uint8).to(torch.float), i.to(torch.float))
self.assertTupleEqual(i.shape, (1, 100, 101, 107))
for i in engine.state.output["label_inverted1"]:
np.testing.assert_allclose(
i.astype(np.uint8).astype(np.float32), i.astype(np.float32))
self.assertTupleEqual(i.shape, (1, 100, 101, 107))
# check labels match
reverted = engine.state.output["label_inverted1"][-1].astype(np.int32)
original = LoadImaged(KEYS)(data[-1])["label"]
n_good = np.sum(np.isclose(reverted, original, atol=1e-3))
reverted_name = engine.state.output["label_meta_dict"][
"filename_or_obj"][-1]
original_name = data[-1]["label"]
self.assertEqual(reverted_name, original_name)
print("invert diff", reverted.size - n_good)
# 25300: 2 workers (cpu, non-macos)
# 1812: 0 workers (gpu or macos)
# 1824: torch 1.5.1
self.assertTrue((reverted.size - n_good) in (25300, 1812, 1824),
"diff. in 3 possible values")
# check the case that different items use different interpolation mode to invert transforms
for i in engine.state.output["image_inverted2"]:
# if the interpolation mode is nearest, accumulated diff should be smaller than 1
self.assertLess(
torch.sum(
i.to(torch.float) -
i.to(torch.uint8).to(torch.float)).item(), 1.0)
self.assertTupleEqual(i.shape, (1, 100, 101, 107))
for i in engine.state.output["label_inverted2"]:
# if the interpolation mode is not nearest, accumulated diff should be greater than 10000
self.assertGreater(
torch.sum(
i.to(torch.float) -
i.to(torch.uint8).to(torch.float)).item(), 10000.0)
self.assertTupleEqual(i.shape, (1, 100, 101, 107))
def main():
monai.config.print_config()
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
set_determinism(12345)
device = torch.device("cuda:0")
# load real data
mednist_url = "https://www.dropbox.com/s/5wwskxctvcxiuea/MedNIST.tar.gz?dl=1"
md5_value = "0bc7306e7427e00ad1c5526a6677552d"
extract_dir = "data"
tar_save_path = os.path.join(extract_dir, "MedNIST.tar.gz")
download_and_extract(mednist_url, tar_save_path, extract_dir, md5_value)
hand_dir = os.path.join(extract_dir, "MedNIST", "Hand")
real_data = [{
"hand": os.path.join(hand_dir, filename)
} for filename in os.listdir(hand_dir)]
# define real data transforms
train_transforms = Compose([
LoadPNGD(keys=["hand"]),
AddChannelD(keys=["hand"]),
ScaleIntensityD(keys=["hand"]),
RandRotateD(keys=["hand"], range_x=15, prob=0.5, keep_size=True),
RandFlipD(keys=["hand"], spatial_axis=0, prob=0.5),
RandZoomD(keys=["hand"], min_zoom=0.9, max_zoom=1.1, prob=0.5),
ToTensorD(keys=["hand"]),
])
# create dataset and dataloader
real_dataset = CacheDataset(real_data, train_transforms)
batch_size = 300
real_dataloader = DataLoader(real_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=10)
# define function to process batchdata for input into discriminator
def prepare_batch(batchdata):
"""
Process Dataloader batchdata dict object and return image tensors for D Inferer
"""
return batchdata["hand"]
# define networks
disc_net = Discriminator(in_shape=(1, 64, 64),
channels=(8, 16, 32, 64, 1),
strides=(2, 2, 2, 2, 1),
num_res_units=1,
kernel_size=5).to(device)
latent_size = 64
gen_net = Generator(latent_shape=latent_size,
start_shape=(latent_size, 8, 8),
channels=[32, 16, 8, 1],
strides=[2, 2, 2, 1])
# initialize both networks
disc_net.apply(normal_init)
gen_net.apply(normal_init)
# input images are scaled to [0,1] so enforce the same of generated outputs
gen_net.conv.add_module("activation", torch.nn.Sigmoid())
gen_net = gen_net.to(device)
# create optimizers and loss functions
learning_rate = 2e-4
betas = (0.5, 0.999)
disc_opt = torch.optim.Adam(disc_net.parameters(),
learning_rate,
betas=betas)
gen_opt = torch.optim.Adam(gen_net.parameters(),
learning_rate,
betas=betas)
disc_loss_criterion = torch.nn.BCELoss()
gen_loss_criterion = torch.nn.BCELoss()
real_label = 1
fake_label = 0
def discriminator_loss(gen_images, real_images):
"""
The discriminator loss is calculated by comparing D
prediction for real and generated images.
"""
real = real_images.new_full((real_images.shape[0], 1), real_label)
gen = gen_images.new_full((gen_images.shape[0], 1), fake_label)
realloss = disc_loss_criterion(disc_net(real_images), real)
genloss = disc_loss_criterion(disc_net(gen_images.detach()), gen)
return (genloss + realloss) / 2
def generator_loss(gen_images):
"""
The generator loss is calculated by determining how realistic
the discriminator classifies the generated images.
"""
output = disc_net(gen_images)
cats = output.new_full(output.shape, real_label)
return gen_loss_criterion(output, cats)
# initialize current run dir
run_dir = "model_out"
print("Saving model output to: %s " % run_dir)
# create workflow handlers
handlers = [
StatsHandler(
name="batch_training_loss",
output_transform=lambda x: {
Keys.GLOSS: x[Keys.GLOSS],
Keys.DLOSS: x[Keys.DLOSS]
},
),
CheckpointSaver(
save_dir=run_dir,
save_dict={
"g_net": gen_net,
"d_net": disc_net
},
save_interval=10,
save_final=True,
epoch_level=True,
),
]
# define key metric
key_train_metric = None
# create adversarial trainer
disc_train_steps = 5
num_epochs = 50
trainer = GanTrainer(
device,
num_epochs,
real_dataloader,
gen_net,
gen_opt,
generator_loss,
disc_net,
disc_opt,
discriminator_loss,
d_prepare_batch=prepare_batch,
d_train_steps=disc_train_steps,
latent_shape=latent_size,
key_train_metric=key_train_metric,
train_handlers=handlers,
)
# run GAN training
trainer.run()
# Training completed, save a few random generated images.
print("Saving trained generator sample output.")
test_img_count = 10
test_latents = make_latent(test_img_count, latent_size).to(device)
fakes = gen_net(test_latents)
for i, image in enumerate(fakes):
filename = "gen-fake-final-%d.png" % (i)
save_path = os.path.join(run_dir, filename)
img_array = image[0].cpu().data.numpy()
png_writer.write_png(img_array, save_path, scale=255)
def convert_to_torchscript(
model: nn.Module,
filename_or_obj: Optional[Any] = None,
extra_files: Optional[Dict] = None,
verify: bool = False,
inputs: Optional[Sequence[Any]] = None,
device: Optional[torch.device] = None,
rtol: float = 1e-4,
atol: float = 0.0,
**kwargs,
):
"""
Utility to convert a model into TorchScript model and save to file,
with optional input / output data verification.
Args:
model: source PyTorch model to save.
filename_or_obj: if not None, specify a file-like object (has to implement write and flush)
or a string containing a file path name to save the TorchScript model.
extra_files: map from filename to contents which will be stored as part of the save model file.
works for PyTorch 1.7 or later.
for more details: https://pytorch.org/docs/stable/generated/torch.jit.save.html.
verify: whether to verify the input and output of TorchScript model.
if `filename_or_obj` is not None, load the saved TorchScript model and verify.
inputs: input test data to verify model, should be a sequence of data, every item maps to a argument
of `model()` function.
device: target device to verify the model, if None, use CUDA if available.
rtol: the relative tolerance when comparing the outputs of PyTorch model and TorchScript model.
atol: the absolute tolerance when comparing the outputs of PyTorch model and TorchScript model.
"""
model.eval()
with torch.no_grad():
script_module = torch.jit.script(model)
if filename_or_obj is not None:
if PT_BEFORE_1_7:
torch.jit.save(m=script_module, f=filename_or_obj)
else:
torch.jit.save(m=script_module, f=filename_or_obj, _extra_files=extra_files)
if verify:
if device is None:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if inputs is None:
raise ValueError("missing input data for verification.")
inputs = [i.to(device) if isinstance(i, torch.Tensor) else i for i in inputs]
ts_model = torch.jit.load(filename_or_obj) if filename_or_obj is not None else script_module
ts_model.eval().to(device)
model = model.to(device)
with torch.no_grad():
set_determinism(seed=0)
torch_out = ensure_tuple(model(*inputs))
set_determinism(seed=0)
torchscript_out = ensure_tuple(ts_model(*inputs))
set_determinism(seed=None)
# compare TorchScript and PyTorch results
for r1, r2 in zip(torch_out, torchscript_out):
if isinstance(r1, torch.Tensor) or isinstance(r2, torch.Tensor):
torch.testing.assert_allclose(r1, r2, rtol=rtol, atol=atol)
return script_module
def test_invert(self):
set_determinism(seed=0)
im_fname, seg_fname = [
make_nifti_image(i)
for i in create_test_image_3d(101, 100, 107, noise_max=100)
]
transform = Compose([
LoadImaged(KEYS),
AddChanneld(KEYS),
Orientationd(KEYS, "RPS"),
Spacingd(KEYS,
pixdim=(1.2, 1.01, 0.9),
mode=["bilinear", "nearest"],
dtype=np.float32),
ScaleIntensityd("image", minv=1, maxv=10),
RandFlipd(KEYS, prob=0.5, spatial_axis=[1, 2]),
RandAxisFlipd(KEYS, prob=0.5),
RandRotate90d(KEYS, spatial_axes=(1, 2)),
RandZoomd(KEYS,
prob=0.5,
min_zoom=0.5,
max_zoom=1.1,
keep_size=True),
RandRotated(KEYS,
prob=0.5,
range_x=np.pi,
mode="bilinear",
align_corners=True),
RandAffined(KEYS, prob=0.5, rotate_range=np.pi, mode="nearest"),
ResizeWithPadOrCropd(KEYS, 100),
ToTensord(
"image"
), # test to support both Tensor and Numpy array when inverting
CastToTyped(KEYS, dtype=[torch.uint8, np.uint8]),
])
data = [{"image": im_fname, "label": seg_fname} for _ in range(12)]
# num workers = 0 for mac or gpu transforms
num_workers = 0 if sys.platform == "darwin" or torch.cuda.is_available(
) else 2
dataset = CacheDataset(data, transform=transform, progress=False)
loader = DataLoader(dataset, num_workers=num_workers, batch_size=5)
inverter = Invertd(
keys=["image", "label"],
transform=transform,
loader=loader,
orig_keys="label",
nearest_interp=True,
postfix="inverted",
to_tensor=[True, False],
device="cpu",
num_workers=0
if sys.platform == "darwin" or torch.cuda.is_available() else 2,
)
# execute 1 epoch
for d in loader:
d = inverter(d)
# this unit test only covers basic function, test_handler_transform_inverter covers more
self.assertTupleEqual(d["image"].shape[1:], (1, 100, 100, 100))
self.assertTupleEqual(d["label"].shape[1:], (1, 100, 100, 100))
# check the nearest inerpolation mode
for i in d["image_inverted"]:
torch.testing.assert_allclose(
i.to(torch.uint8).to(torch.float), i.to(torch.float))
self.assertTupleEqual(i.shape, (1, 100, 101, 107))
for i in d["label_inverted"]:
np.testing.assert_allclose(
i.astype(np.uint8).astype(np.float32),
i.astype(np.float32))
self.assertTupleEqual(i.shape, (1, 100, 101, 107))
set_determinism(seed=None)
def setUp(self) -> None:
set_determinism(seed=0)
def test_invert(self):
set_determinism(seed=0)
im_fname, seg_fname = [
make_nifti_image(i)
for i in create_test_image_3d(101, 100, 107, noise_max=100)
]
transform = Compose([
LoadImaged(KEYS),
AddChanneld(KEYS),
Orientationd(KEYS, "RPS"),
Spacingd(KEYS,
pixdim=(1.2, 1.01, 0.9),
mode=["bilinear", "nearest"],
dtype=np.float32),
ScaleIntensityd("image", minv=1, maxv=10),
RandFlipd(KEYS, prob=0.5, spatial_axis=[1, 2]),
RandAxisFlipd(KEYS, prob=0.5),
RandRotate90d(KEYS, spatial_axes=(1, 2)),
RandZoomd(KEYS,
prob=0.5,
min_zoom=0.5,
max_zoom=1.1,
keep_size=True),
RandRotated(KEYS,
prob=0.5,
range_x=np.pi,
mode="bilinear",
align_corners=True),
RandAffined(KEYS, prob=0.5, rotate_range=np.pi, mode="nearest"),
ResizeWithPadOrCropd(KEYS, 100),
ToTensord(KEYS),
CastToTyped(KEYS, dtype=torch.uint8),
])
data = [{"image": im_fname, "label": seg_fname} for _ in range(12)]
# num workers = 0 for mac or gpu transforms
num_workers = 0 if sys.platform == "darwin" or torch.cuda.is_available(
) else 2
dataset = CacheDataset(data, transform=transform, progress=False)
loader = DataLoader(dataset, num_workers=num_workers, batch_size=5)
# set up engine
def _train_func(engine, batch):
self.assertTupleEqual(batch["image"].shape[1:], (1, 100, 100, 100))
engine.state.output = batch
engine.fire_event(IterationEvents.MODEL_COMPLETED)
return engine.state.output
engine = Engine(_train_func)
engine.register_events(*IterationEvents)
# set up testing handler
TransformInverter(
transform=transform,
loader=loader,
output_keys=["image", "label"],
batch_keys="label",
nearest_interp=True,
num_workers=0
if sys.platform == "darwin" or torch.cuda.is_available() else 2,
).attach(engine)
engine.run(loader, max_epochs=1)
set_determinism(seed=None)
self.assertTupleEqual(engine.state.output["image"].shape,
(2, 1, 100, 100, 100))
self.assertTupleEqual(engine.state.output["label"].shape,
(2, 1, 100, 100, 100))
for i in engine.state.output["image_inverted"] + engine.state.output[
"label_inverted"]:
torch.testing.assert_allclose(
i.to(torch.uint8).to(torch.float), i.to(torch.float))
self.assertTupleEqual(i.shape, (1, 100, 101, 107))
# check labels match
reverted = engine.state.output["label_inverted"][-1].detach().cpu(
).numpy()[0].astype(np.int32)
original = LoadImaged(KEYS)(data[-1])["label"]
n_good = np.sum(np.isclose(reverted, original, atol=1e-3))
reverted_name = engine.state.output["label_meta_dict"][
"filename_or_obj"][-1]
original_name = data[-1]["label"]
self.assertEqual(reverted_name, original_name)
print("invert diff", reverted.size - n_good)
self.assertTrue((reverted.size - n_good) in (25300, 1812),
"diff. in two possible values")