def main(hdf_file):
# Use a pad extractor in order to compensate for the valid convolutions of the network. Actual image information is padded
extractor = extr.PadDataExtractor(
(2, 2, 2), extr.DataExtractor(categories=(defs.KEY_IMAGES, )))
# Adapted permutation due to the additional dimension
transform = tfm.Permute(permutation=(3, 0, 1, 2),
entries=(defs.KEY_IMAGES, ))
# Creating patch indexing strategy with patch_shape that equal the network output shape
indexing_strategy = extr.PatchWiseIndexing(patch_shape=(32, 32, 32))
dataset = extr.PymiaDatasource(hdf_file, indexing_strategy, extractor,
transform)
direct_extractor = extr.ComposeExtractor([
extr.ImagePropertiesExtractor(),
extr.DataExtractor(categories=(defs.KEY_LABELS, defs.KEY_IMAGES))
])
assembler = assm.SubjectAssembler(dataset)
# torch specific handling
pytorch_dataset = pymia_torch.PytorchDatasetAdapter(dataset)
loader = torch_data.dataloader.DataLoader(pytorch_dataset,
batch_size=2,
shuffle=False)
# dummy CNN with valid convolutions instead of same convolutions
dummy_network = nn.Sequential(
nn.Conv3d(in_channels=2, out_channels=8, kernel_size=3, padding=0),
nn.Conv3d(in_channels=8, out_channels=1, kernel_size=3, padding=0),
nn.Sigmoid())
torch.set_grad_enabled(False)
nb_batches = len(loader)
# looping over the data in the dataset
for i, batch in enumerate(loader):
x, sample_indices = batch[defs.KEY_IMAGES], batch[
defs.KEY_SAMPLE_INDEX]
prediction = dummy_network(x)
numpy_prediction = prediction.numpy().transpose((0, 2, 3, 4, 1))
is_last = i == nb_batches - 1
assembler.add_batch(numpy_prediction, sample_indices.numpy(), is_last)
for subject_index in assembler.subjects_ready:
subject_prediction = assembler.get_assembled_subject(subject_index)
direct_sample = dataset.direct_extract(direct_extractor,
subject_index)
target, image_properties = direct_sample[
defs.KEY_LABELS], direct_sample[defs.KEY_PROPERTIES]
def main(hdf_file, is_meta):
if not is_meta:
extractor = extr.DataExtractor(categories=(defs.KEY_IMAGES, ))
else:
extractor = extr.FilesystemDataExtractor(
categories=(defs.KEY_IMAGES, ))
transform = tfm.Permute(permutation=(2, 0, 1), entries=(defs.KEY_IMAGES, ))
indexing_strategy = extr.SliceIndexing()
dataset = extr.PymiaDatasource(hdf_file, indexing_strategy, extractor,
transform)
direct_extractor = extr.ComposeExtractor([
extr.ImagePropertiesExtractor(),
extr.DataExtractor(categories=(defs.KEY_LABELS, defs.KEY_IMAGES))
])
assembler = assm.SubjectAssembler(dataset)
# torch specific handling
pytorch_dataset = pymia_torch.PytorchDatasetAdapter(dataset)
loader = torch_data.dataloader.DataLoader(pytorch_dataset,
batch_size=2,
shuffle=False)
dummy_network = nn.Sequential(
nn.Conv2d(in_channels=2, out_channels=8, kernel_size=3, padding=1),
nn.Conv2d(in_channels=8, out_channels=1, kernel_size=3, padding=1),
nn.Sigmoid())
torch.set_grad_enabled(False)
nb_batches = len(loader)
# looping over the data in the dataset
for i, batch in enumerate(loader):
x, sample_indices = batch[defs.KEY_IMAGES], batch[
defs.KEY_SAMPLE_INDEX]
prediction = dummy_network(x)
numpy_prediction = prediction.numpy().transpose((0, 2, 3, 1))
is_last = i == nb_batches - 1
assembler.add_batch(numpy_prediction, sample_indices.numpy(), is_last)
for subject_index in assembler.subjects_ready:
subject_prediction = assembler.get_assembled_subject(subject_index)
direct_sample = dataset.direct_extract(direct_extractor,
subject_index)
target, image_properties = direct_sample[
defs.KEY_LABELS], direct_sample[defs.KEY_PROPERTIES]
def main(hdf_file):
extractor = extr.DataExtractor(categories=(defs.KEY_IMAGES, ))
# no transformation needed because TensorFlow uses channel-last
transform = None
indexing_strategy = extr.SliceIndexing()
dataset = extr.PymiaDatasource(hdf_file, indexing_strategy, extractor,
transform)
direct_extractor = extr.ComposeExtractor([
extr.ImagePropertiesExtractor(),
extr.DataExtractor(categories=(defs.KEY_LABELS, defs.KEY_IMAGES))
])
assembler = assm.SubjectAssembler(dataset)
# TensorFlow specific handling
gen_fn = pymia_tf.get_tf_generator(dataset)
tf_dataset = tf.data.Dataset.from_generator(generator=gen_fn,
output_types={
defs.KEY_IMAGES:
tf.float32,
defs.KEY_SAMPLE_INDEX:
tf.int64
})
tf_dataset = tf_dataset.batch(2)
dummy_network = keras.Sequential([
layers.Conv2D(8, kernel_size=3, padding='same'),
layers.Conv2D(2, kernel_size=3, padding='same', activation='sigmoid')
])
nb_batches = len(dataset) // 2
# looping over the data in the dataset
for i, batch in enumerate(tf_dataset):
x, sample_indices = batch[defs.KEY_IMAGES], batch[
defs.KEY_SAMPLE_INDEX]
prediction = dummy_network(x)
numpy_prediction = prediction.numpy()
is_last = i == nb_batches - 1
assembler.add_batch(numpy_prediction, sample_indices.numpy(), is_last)
for subject_index in assembler.subjects_ready:
subject_prediction = assembler.get_assembled_subject(subject_index)
direct_sample = dataset.direct_extract(direct_extractor,
subject_index)
def main(hdf_file, plot_dir):
os.makedirs(plot_dir, exist_ok=True)
# setup the datasource
extractor = extr.DataExtractor(categories=(defs.KEY_IMAGES, defs.KEY_LABELS))
indexing_strategy = extr.SliceIndexing()
dataset = extr.PymiaDatasource(hdf_file, indexing_strategy, extractor)
seed = 1
np.random.seed(seed)
sample_idx = 55
# set up transformations without augmentation
transforms_augmentation = []
transforms_before_augmentation = [tfm.Permute(permutation=(2, 0, 1)), ] # to have the channel-dimension first
transforms_after_augmentation = [tfm.Squeeze(entries=(defs.KEY_LABELS,)), ] # get rid of the channel-dimension for the labels
train_transforms = tfm.ComposeTransform(transforms_before_augmentation + transforms_augmentation + transforms_after_augmentation)
dataset.set_transform(train_transforms)
sample = dataset[sample_idx]
plot_sample(plot_dir, 'none', sample)
# augmentation with pymia
transforms_augmentation = [augm.RandomRotation90(axes=(-2, -1)), augm.RandomMirror()]
train_transforms = tfm.ComposeTransform(
transforms_before_augmentation + transforms_augmentation + transforms_after_augmentation)
dataset.set_transform(train_transforms)
sample = dataset[sample_idx]
plot_sample(plot_dir, 'pymia', sample)
# augmentation with batchgenerators
transforms_augmentation = [BatchgeneratorsTransform([
bg_tfm.spatial_transforms.MirrorTransform(axes=(0, 1), data_key=defs.KEY_IMAGES, label_key=defs.KEY_LABELS),
bg_tfm.noise_transforms.GaussianBlurTransform(blur_sigma=(0.2, 1.0), data_key=defs.KEY_IMAGES, label_key=defs.KEY_LABELS),
])]
train_transforms = tfm.ComposeTransform(
transforms_before_augmentation + transforms_augmentation + transforms_after_augmentation)
dataset.set_transform(train_transforms)
sample = dataset[sample_idx]
plot_sample(plot_dir, 'batchgenerators', sample)
# augmentation with TorchIO
transforms_augmentation = [TorchIOTransform(
[tio.RandomFlip(axes=('LR'), flip_probability=1.0, keys=(defs.KEY_IMAGES, defs.KEY_LABELS), seed=seed),
tio.RandomAffine(scales=(0.9, 1.2), degrees=(10), isotropic=False, default_pad_value='otsu',
image_interpolation='NEAREST', keys=(defs.KEY_IMAGES, defs.KEY_LABELS), seed=seed),
])]
train_transforms = tfm.ComposeTransform(
transforms_before_augmentation + transforms_augmentation + transforms_after_augmentation)
dataset.set_transform(train_transforms)
sample = dataset[sample_idx]
plot_sample(plot_dir, 'torchio', sample)
def main(hdf_file, log_dir):
# initialize the evaluator with the metrics and the labels to evaluate
metrics = [metric.DiceCoefficient()]
labels = {1: 'WHITEMATTER',
2: 'GREYMATTER',
3: 'HIPPOCAMPUS',
4: 'AMYGDALA',
5: 'THALAMUS'}
evaluator = eval_.SegmentationEvaluator(metrics, labels)
# we want to log the mean and standard deviation of the metrics among all subjects of the dataset
functions = {'MEAN': np.mean, 'STD': np.std}
statistics_aggregator = writer.StatisticsAggregator(functions=functions)
console_writer = writer.ConsoleStatisticsWriter(functions=functions)
# initialize TensorBoard writer
tb = tensorboard.SummaryWriter(os.path.join(log_dir, 'logging-example-torch'))
# setup the training datasource
train_subjects, valid_subjects = ['Subject_1', 'Subject_2', 'Subject_3'], ['Subject_4']
extractor = extr.DataExtractor(categories=(defs.KEY_IMAGES, defs.KEY_LABELS))
indexing_strategy = extr.SliceIndexing()
augmentation_transforms = [augm.RandomElasticDeformation(), augm.RandomMirror()]
transforms = [tfm.Permute(permutation=(2, 0, 1)), tfm.Squeeze(entries=(defs.KEY_LABELS,))]
train_transforms = tfm.ComposeTransform(augmentation_transforms + transforms)
train_dataset = extr.PymiaDatasource(hdf_file, indexing_strategy, extractor, train_transforms,
subject_subset=train_subjects)
# setup the validation datasource
valid_transforms = tfm.ComposeTransform([tfm.Permute(permutation=(2, 0, 1))])
valid_dataset = extr.PymiaDatasource(hdf_file, indexing_strategy, extractor, valid_transforms,
subject_subset=valid_subjects)
direct_extractor = extr.ComposeExtractor(
[extr.SubjectExtractor(),
extr.ImagePropertiesExtractor(),
extr.DataExtractor(categories=(defs.KEY_LABELS,))]
)
assembler = assm.SubjectAssembler(valid_dataset)
# torch specific handling
pytorch_train_dataset = pymia_torch.PytorchDatasetAdapter(train_dataset)
train_loader = torch_data.dataloader.DataLoader(pytorch_train_dataset, batch_size=16, shuffle=True)
pytorch_valid_dataset = pymia_torch.PytorchDatasetAdapter(valid_dataset)
valid_loader = torch_data.dataloader.DataLoader(pytorch_valid_dataset, batch_size=16, shuffle=False)
u_net = unet.UNetModel(ch_in=2, ch_out=6, n_channels=16, n_pooling=3).to(device)
print(u_net)
optimizer = optim.Adam(u_net.parameters(), lr=1e-3)
train_batches = len(train_loader)
# looping over the data in the dataset
epochs = 100
for epoch in range(epochs):
u_net.train()
print(f'Epoch {epoch + 1}/{epochs}')
# training
print('training')
for i, batch in enumerate(train_loader):
x, y = batch[defs.KEY_IMAGES].to(device), batch[defs.KEY_LABELS].to(device).long()
logits = u_net(x)
optimizer.zero_grad()
loss = F.cross_entropy(logits, y)
loss.backward()
optimizer.step()
tb.add_scalar('train/loss', loss.item(), epoch*train_batches + i)
print(f'[{i + 1}/{train_batches}]\tloss: {loss.item()}')
# validation
print('validation')
with torch.no_grad():
u_net.eval()
valid_batches = len(valid_loader)
for i, batch in enumerate(valid_loader):
x, sample_indices = batch[defs.KEY_IMAGES].to(device), batch[defs.KEY_SAMPLE_INDEX]
logits = u_net(x)
prediction = logits.argmax(dim=1, keepdim=True)
numpy_prediction = prediction.cpu().numpy().transpose((0, 2, 3, 1))
is_last = i == valid_batches - 1
assembler.add_batch(numpy_prediction, sample_indices.numpy(), is_last)
for subject_index in assembler.subjects_ready:
subject_prediction = assembler.get_assembled_subject(subject_index)
direct_sample = train_dataset.direct_extract(direct_extractor, subject_index)
target, image_properties = direct_sample[defs.KEY_LABELS], direct_sample[defs.KEY_PROPERTIES]
# evaluate the prediction against the reference
evaluator.evaluate(subject_prediction[..., 0], target[..., 0], direct_sample[defs.KEY_SUBJECT])
# calculate mean and standard deviation of each metric
results = statistics_aggregator.calculate(evaluator.results)
# log to TensorBoard into category train
for result in results:
tb.add_scalar(f'valid/{result.metric}-{result.id_}', result.value, epoch)
console_writer.write(evaluator.results)
# clear results such that the evaluator is ready for the next evaluation
evaluator.clear()
def main(hdf_file: str, log_dir: str):
# initialize the evaluator with the metrics and the labels to evaluate
metrics = [metric.DiceCoefficient()]
labels = {1: 'WHITEMATTER', 2: 'GREYMATTER', 5: 'THALAMUS'}
evaluator = eval_.SegmentationEvaluator(metrics, labels)
# we want to log the mean and standard deviation of the metrics among all subjects of the dataset
functions = {'MEAN': np.mean, 'STD': np.std}
statistics_aggregator = writer.StatisticsAggregator(functions=functions)
# initialize TensorBoard writer
# tb = tensorboard.SummaryWriter(os.path.join(log_dir, 'logging-example-torch'))
tb = tf.summary.create_file_writer(
os.path.join(log_dir, 'logging-example-tensorflow'))
# initialize the data handling
dataset = extr.PymiaDatasource(
hdf_file, extr.SliceIndexing(),
extr.DataExtractor(categories=(defs.KEY_IMAGES, )))
gen_fn = pymia_tf.get_tf_generator(dataset)
tf_dataset = tf.data.Dataset.from_generator(generator=gen_fn,
output_types={
defs.KEY_IMAGES:
tf.float32,
defs.KEY_SAMPLE_INDEX:
tf.int64
})
loader = tf_dataset.batch(100)
assembler = assm.SubjectAssembler(dataset)
direct_extractor = extr.ComposeExtractor([
extr.SubjectExtractor(), # extraction of the subject name
extr.ImagePropertiesExtractor(
), # Extraction of image properties (origin, spacing, etc.) for storage
extr.DataExtractor(
categories=(defs.KEY_LABELS,
)) # Extraction of "labels" entries for evaluation
])
# initialize a dummy network, which returns a random prediction
class DummyNetwork(tf.keras.Model):
def call(self, inputs):
return tf.random.uniform((*inputs.shape[:-1], 1),
0,
6,
dtype=tf.int32)
dummy_network = DummyNetwork()
tf.random.set_seed(0) # set seed for reproducibility
nb_batches = len(dataset) // 2
epochs = 10
for epoch in range(epochs):
print(f'Epoch {epoch + 1}/{epochs}')
for i, batch in enumerate(loader):
# get the data from batch and predict
x, sample_indices = batch[defs.KEY_IMAGES], batch[
defs.KEY_SAMPLE_INDEX]
prediction = dummy_network(x)
# translate the prediction to numpy
numpy_prediction = prediction.numpy()
# add the batch prediction to the assembler
is_last = i == nb_batches - 1
assembler.add_batch(numpy_prediction, sample_indices.numpy(),
is_last)
# process the subjects/images that are fully assembled
for subject_index in assembler.subjects_ready:
subject_prediction = assembler.get_assembled_subject(
subject_index)
# extract the target and image properties via direct extract
direct_sample = dataset.direct_extract(direct_extractor,
subject_index)
reference, image_properties = direct_sample[
defs.KEY_LABELS], direct_sample[defs.KEY_PROPERTIES]
# evaluate the prediction against the reference
evaluator.evaluate(subject_prediction[..., 0], reference[...,
0],
direct_sample[defs.KEY_SUBJECT])
# calculate mean and standard deviation of each metric
results = statistics_aggregator.calculate(evaluator.results)
# log to TensorBoard into category train
for result in results:
with tb.as_default():
tf.summary.scalar(f'train/{result.metric}-{result.id_}',
result.value, epoch)
# clear results such that the evaluator is ready for the next evaluation
evaluator.clear()
def main(hdf_file: str):
extractor = extr.ComposeExtractor([
extr.NamesExtractor(),
extr.DataExtractor(),
extr.SelectiveDataExtractor(),
extr.DataExtractor(('numerical', ), ignore_indexing=True),
extr.DataExtractor(('gender', ), ignore_indexing=True),
extr.DataExtractor(('mask', ), ignore_indexing=False),
extr.SubjectExtractor(),
extr.FilesExtractor(categories=(defs.KEY_IMAGES, defs.KEY_LABELS,
'mask', 'numerical', 'gender')),
extr.IndexingExtractor(),
extr.ImagePropertiesExtractor()
])
dataset = extr.PymiaDatasource(hdf_file, extr.SliceIndexing(), extractor)
for i in range(len(dataset)):
item = dataset[i]
index_expr = item[defs.KEY_INDEX_EXPR] # type: data.IndexExpression
root = item[defs.KEY_FILE_ROOT]
image = None # type: sitk.Image
for i, file in enumerate(
item[defs.KEY_PLACEHOLDER_FILES.format('images')]):
image = sitk.ReadImage(os.path.join(root, file))
np_img = sitk.GetArrayFromImage(image).astype(np.float32)
np_img = (np_img - np_img.mean()) / np_img.std()
np_slice = np_img[index_expr.expression]
if (np_slice != item[defs.KEY_IMAGES][..., i]).any():
raise ValueError('slice not equal')
# for any image
image_properties = conv.ImageProperties(image)
if image_properties != item[defs.KEY_PROPERTIES]:
raise ValueError('image properties not equal')
for file in item[defs.KEY_PLACEHOLDER_FILES.format('labels')]:
image = sitk.ReadImage(os.path.join(root, file))
np_img = sitk.GetArrayFromImage(image)
np_img = np.expand_dims(
np_img, axis=-1
) # due to the convention of having the last dim as number of channels
np_slice = np_img[index_expr.expression]
if (np_slice != item[defs.KEY_LABELS]).any():
raise ValueError('slice not equal')
for file in item[defs.KEY_PLACEHOLDER_FILES.format('mask')]:
image = sitk.ReadImage(os.path.join(root, file))
np_img = sitk.GetArrayFromImage(image)
np_img = np.expand_dims(
np_img, axis=-1
) # due to the convention of having the last dim as number of channels
np_slice = np_img[index_expr.expression]
if (np_slice != item['mask']).any():
raise ValueError('slice not equal')
for file in item[defs.KEY_PLACEHOLDER_FILES.format('numerical')]:
with open(os.path.join(root, file), 'r') as f:
lines = f.readlines()
age = float(lines[0].split(':')[1].strip())
gpa = float(lines[1].split(':')[1].strip())
if age != item['numerical'][0][0] or gpa != item['numerical'][0][1]:
raise ValueError('value not equal')
for file in item[defs.KEY_PLACEHOLDER_FILES.format('gender')]:
with open(os.path.join(root, file), 'r') as f:
gender = f.readlines()[2].split(':')[1].strip()
if gender != str(item['gender'][0]):
raise ValueError('value not equal')
print('All test passed!')
def main(hdf_file: str, log_dir: str):
# initialize the evaluator with the metrics and the labels to evaluate
metrics = [metric.DiceCoefficient()]
labels = {1: 'WHITEMATTER', 2: 'GREYMATTER', 5: 'THALAMUS'}
evaluator = eval_.SegmentationEvaluator(metrics, labels)
# we want to log the mean and standard deviation of the metrics among all subjects of the dataset
functions = {'MEAN': np.mean, 'STD': np.std}
statistics_aggregator = writer.StatisticsAggregator(functions=functions)
# initialize TensorBoard writer
tb = tensorboard.SummaryWriter(
os.path.join(log_dir, 'logging-example-torch'))
# initialize the data handling
transform = tfm.Permute(permutation=(2, 0, 1), entries=(defs.KEY_IMAGES, ))
dataset = extr.PymiaDatasource(
hdf_file, extr.SliceIndexing(),
extr.DataExtractor(categories=(defs.KEY_IMAGES, )), transform)
pytorch_dataset = pymia_torch.PytorchDatasetAdapter(dataset)
loader = torch_data.dataloader.DataLoader(pytorch_dataset,
batch_size=100,
shuffle=False)
assembler = assm.SubjectAssembler(dataset)
direct_extractor = extr.ComposeExtractor([
extr.SubjectExtractor(), # extraction of the subject name
extr.ImagePropertiesExtractor(
), # Extraction of image properties (origin, spacing, etc.) for storage
extr.DataExtractor(
categories=(defs.KEY_LABELS,
)) # Extraction of "labels" entries for evaluation
])
# initialize a dummy network, which returns a random prediction
class DummyNetwork(nn.Module):
def forward(self, x):
return torch.randint(0, 6, (x.size(0), 1, *x.size()[2:]))
dummy_network = DummyNetwork()
torch.manual_seed(0) # set seed for reproducibility
nb_batches = len(loader)
epochs = 10
for epoch in range(epochs):
print(f'Epoch {epoch + 1}/{epochs}')
for i, batch in enumerate(loader):
# get the data from batch and predict
x, sample_indices = batch[defs.KEY_IMAGES], batch[
defs.KEY_SAMPLE_INDEX]
prediction = dummy_network(x)
# translate the prediction to numpy and back to (B)HWC (channel last)
numpy_prediction = prediction.numpy().transpose((0, 2, 3, 1))
# add the batch prediction to the assembler
is_last = i == nb_batches - 1
assembler.add_batch(numpy_prediction, sample_indices.numpy(),
is_last)
# process the subjects/images that are fully assembled
for subject_index in assembler.subjects_ready:
subject_prediction = assembler.get_assembled_subject(
subject_index)
# extract the target and image properties via direct extract
direct_sample = dataset.direct_extract(direct_extractor,
subject_index)
reference, image_properties = direct_sample[
defs.KEY_LABELS], direct_sample[defs.KEY_PROPERTIES]
# evaluate the prediction against the reference
evaluator.evaluate(subject_prediction[..., 0], reference[...,
0],
direct_sample[defs.KEY_SUBJECT])
# calculate mean and standard deviation of each metric
results = statistics_aggregator.calculate(evaluator.results)
# log to TensorBoard into category train
for result in results:
tb.add_scalar(f'train/{result.metric}-{result.id_}', result.value,
epoch)
# clear results such that the evaluator is ready for the next evaluation
evaluator.clear()
def main(hdf_file, log_dir):
# initialize the evaluator with the metrics and the labels to evaluate
metrics = [metric.DiceCoefficient()]
labels = {
1: 'WHITEMATTER',
2: 'GREYMATTER',
3: 'HIPPOCAMPUS',
4: 'AMYGDALA',
5: 'THALAMUS'
}
evaluator = eval_.SegmentationEvaluator(metrics, labels)
# we want to log the mean and standard deviation of the metrics among all subjects of the dataset
functions = {'MEAN': np.mean, 'STD': np.std}
statistics_aggregator = writer.StatisticsAggregator(functions=functions)
console_writer = writer.ConsoleStatisticsWriter(functions=functions)
# initialize TensorBoard writer
summary_writer = tf.summary.create_file_writer(
os.path.join(log_dir, 'logging-example-tensorflow'))
# setup the training datasource
train_subjects, valid_subjects = ['Subject_1', 'Subject_2',
'Subject_3'], ['Subject_4']
extractor = extr.DataExtractor(categories=(defs.KEY_IMAGES,
defs.KEY_LABELS))
indexing_strategy = extr.SliceIndexing()
augmentation_transforms = [
augm.RandomElasticDeformation(),
augm.RandomMirror()
]
transforms = [tfm.Squeeze(entries=(defs.KEY_LABELS, ))]
train_transforms = tfm.ComposeTransform(augmentation_transforms +
transforms)
train_dataset = extr.PymiaDatasource(hdf_file,
indexing_strategy,
extractor,
train_transforms,
subject_subset=train_subjects)
# setup the validation datasource
batch_size = 16
valid_transforms = tfm.ComposeTransform([])
valid_dataset = extr.PymiaDatasource(hdf_file,
indexing_strategy,
extractor,
valid_transforms,
subject_subset=valid_subjects)
direct_extractor = extr.ComposeExtractor([
extr.SubjectExtractor(),
extr.ImagePropertiesExtractor(),
extr.DataExtractor(categories=(defs.KEY_LABELS, ))
])
assembler = assm.SubjectAssembler(valid_dataset)
# tensorflow specific handling
train_gen_fn = pymia_tf.get_tf_generator(train_dataset)
tf_train_dataset = tf.data.Dataset.from_generator(
generator=train_gen_fn,
output_types={
defs.KEY_IMAGES: tf.float32,
defs.KEY_LABELS: tf.int64,
defs.KEY_SAMPLE_INDEX: tf.int64
})
tf_train_dataset = tf_train_dataset.batch(batch_size).shuffle(
len(train_dataset))
valid_gen_fn = pymia_tf.get_tf_generator(valid_dataset)
tf_valid_dataset = tf.data.Dataset.from_generator(
generator=valid_gen_fn,
output_types={
defs.KEY_IMAGES: tf.float32,
defs.KEY_LABELS: tf.int64,
defs.KEY_SAMPLE_INDEX: tf.int64
})
tf_valid_dataset = tf_valid_dataset.batch(batch_size)
u_net = unet.build_model(channels=2,
num_classes=6,
layer_depth=3,
filters_root=16)
optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)
train_loss = tf.keras.metrics.Mean('train_loss', dtype=tf.float32)
train_batches = len(train_dataset) // batch_size
# looping over the data in the dataset
epochs = 100
for epoch in range(epochs):
print(f'Epoch {epoch + 1}/{epochs}')
# training
print('training')
for i, batch in enumerate(tf_train_dataset):
x, y = batch[defs.KEY_IMAGES], batch[defs.KEY_LABELS]
with tf.GradientTape() as tape:
logits = u_net(x, training=True)
loss = tf.keras.losses.sparse_categorical_crossentropy(
y, logits, from_logits=True)
grads = tape.gradient(loss, u_net.trainable_variables)
optimizer.apply_gradients(zip(grads, u_net.trainable_variables))
train_loss(loss)
with summary_writer.as_default():
tf.summary.scalar('train/loss',
train_loss.result(),
step=epoch * train_batches + i)
print(
f'[{i + 1}/{train_batches}]\tloss: {train_loss.result().numpy()}'
)
# validation
print('validation')
valid_batches = len(valid_dataset) // batch_size
for i, batch in enumerate(tf_valid_dataset):
x, sample_indices = batch[defs.KEY_IMAGES], batch[
defs.KEY_SAMPLE_INDEX]
logits = u_net(x)
prediction = tf.expand_dims(tf.math.argmax(logits, -1), -1)
numpy_prediction = prediction.numpy()
is_last = i == valid_batches - 1
assembler.add_batch(numpy_prediction, sample_indices.numpy(),
is_last)
for subject_index in assembler.subjects_ready:
subject_prediction = assembler.get_assembled_subject(
subject_index)
direct_sample = train_dataset.direct_extract(
direct_extractor, subject_index)
target, image_properties = direct_sample[
defs.KEY_LABELS], direct_sample[defs.KEY_PROPERTIES]
# evaluate the prediction against the reference
evaluator.evaluate(subject_prediction[..., 0], target[..., 0],
direct_sample[defs.KEY_SUBJECT])
# calculate mean and standard deviation of each metric
results = statistics_aggregator.calculate(evaluator.results)
# log to TensorBoard into category train
with summary_writer.as_default():
for result in results:
tf.summary.scalar(f'valid/{result.metric}-{result.id_}',
result.value, epoch)
console_writer.write(evaluator.results)
# clear results such that the evaluator is ready for the next evaluation
evaluator.clear()