def generate_and_print_model_summary(config: ModelConfigBase, model: DeviceAwareModule) -> None:
"""
Writes a human readable summary of the present model to logging.info, and logs the number of trainable
parameters to AzureML.
:param config: The configuration for the model.
:param model: The instantiated Pytorch model.
"""
random_state = RandomStateSnapshot.snapshot_random_state()
# There appears to be a bug in apex, where previous use (in training for example) causes problems
# when another model is later built on the CPU (for example, before loading from a checkpoint)
# https://github.com/NVIDIA/apex/issues/694
# Hence, move the model to the GPU before doing model summary.
if config.use_gpu:
model = model.cuda()
if isinstance(config, ScalarModelBase):
# To generate the model summary, read the first item of the dataset. Then use the model's own
# get_model_input function to convert the dataset item to input tensors, and feed them through the model.
train_dataset = config.get_torch_dataset_for_inference(ModelExecutionMode.TRAIN)
train_item_0 = next(iter(train_dataset.as_data_loader(shuffle=False, batch_size=1, num_dataload_workers=0)))
model_inputs = get_scalar_model_inputs_and_labels(config, model, train_item_0).model_inputs
# The model inputs may already be converted to float16, assuming that we would do mixed precision.
# However, the model is not yet converted to float16 when this function is called, hence convert back to float32
summary = ModelSummary(model)
summary.generate_summary(input_tensors=model_inputs, log_summaries_to_files=config.log_summaries_to_files)
elif config.is_segmentation_model:
summary_for_segmentation_models(config, model)
assert model.summarizer
summary = model.summarizer # type: ignore
else:
raise ValueError("Don't know how to generate a summary for this type of model?")
RUN_CONTEXT.log(LoggingColumns.NumTrainableParameters, summary.n_trainable_params)
random_state.restore_random_state()
def test_model_summary_on_classification2() -> None:
image_channels = 2
model = ImageEncoderWithMlp(imaging_feature_type=ImagingFeatureType.Segmentation,
encode_channels_jointly=False,
num_encoder_blocks=3,
initial_feature_channels=2,
num_image_channels=image_channels,
mlp_dropout=0.5)
summarizer = ModelSummary(model)
summarizer.generate_summary(input_sizes=[(image_channels * HDF5_NUM_SEGMENTATION_CLASSES, 6, 32, 32)])
assert summarizer.n_params != 0
assert summarizer.n_trainable_params != 0
def test_unet_summary_generation() -> None:
"""Checks unet summary generation works either in CPU or GPU"""
model = UNet3D(input_image_channels=1,
initial_feature_channels=2,
num_classes=2,
kernel_size=1,
num_downsampling_paths=2)
if machine_has_gpu:
model.cuda()
summary = ModelSummary(model=model).generate_summary(input_sizes=[(1, 4, 4, 4)])
assert summary is not None
def generate_model_summary(self,
crop_size: Optional[TupleInt3] = None,
log_summaries_to_files: bool = False) -> None:
"""
Stores a model summary, containing information about layers, memory consumption and runtime
in the model.summary field.
When called again with the same crop_size, the summary is not created again.
:param crop_size: The crop size for which the summary should be created. If not provided,
the minimum allowed crop size is used.
:param log_summaries_to_files: whether to write the summary to a file
"""
if crop_size is None:
crop_size = self.crop_size_constraints.minimum_size # type: ignore
assert crop_size is not None
input_size = [crop_size]
if self.summary is None or self.summary_crop_size != input_size:
self.summarizer = ModelSummary(self)
self.summary = self.summarizer.generate_summary(
input_sizes=[(self.input_channels, *crop_size)],
log_summaries_to_files=log_summaries_to_files)
self.summary_crop_size = crop_size
def test_unet_model_parallel() -> None:
"""Checks model parallel utilises all the available GPU devices for forward pass"""
if no_gpu_available:
logging.warning("CUDA capable GPU is not available - UNet Model Parallel cannot be tested")
return
model = UNet3D(input_image_channels=1,
initial_feature_channels=2,
num_classes=2,
kernel_size=1,
num_downsampling_paths=2).cuda()
# Partition the network across all available gpu
available_devices = [torch.device('cuda:{}'.format(ii)) for ii in range(torch.cuda.device_count())]
model.generate_model_summary()
model.partition_model(devices=available_devices)
# Verify that all the devices are utilised by the layers of the model
summary = ModelSummary(model=model).generate_summary(input_sizes=[(1, 4, 4, 4)])
layer_devices = set()
for layer_summary in summary.values():
if layer_summary.device:
layer_devices.add(layer_summary.device)
assert layer_devices == set(available_devices)
def test_model_summary_on_classification1() -> None:
model = create_model_with_temperature_scaling(GlaucomaPublic())
ModelSummary(model).generate_summary(input_sizes=[(1, 6, 64, 60)])
def test_image_encoder(test_output_dirs: OutputFolderForTests,
encode_channels_jointly: bool,
use_non_imaging_features: bool,
kernel_size_per_encoding_block: Optional[Union[TupleInt3, List[TupleInt3]]],
stride_size_per_encoding_block: Optional[Union[TupleInt3, List[TupleInt3]]],
reduction_factor: float,
expected_num_reduced_features: int,
aggregation_type: AggregationType) -> None:
"""
Test if the image encoder networks can be trained without errors (including GradCam computation and data
augmentation).
"""
logging_to_stdout()
set_random_seed(0)
dataset_folder = Path(test_output_dirs.make_sub_dir("dataset"))
scan_size = (6, 64, 60)
scan_files: List[str] = []
for s in range(4):
random_scan = np.random.uniform(0, 1, scan_size)
scan_file_name = f"scan{s + 1}{NumpyFile.NUMPY.value}"
np.save(str(dataset_folder / scan_file_name), random_scan)
scan_files.append(scan_file_name)
dataset_contents = """subject,channel,path,label,numerical1,numerical2,categorical1,categorical2
S1,week0,scan1.npy,,1,10,Male,Val1
S1,week1,scan2.npy,True,2,20,Female,Val2
S2,week0,scan3.npy,,3,30,Female,Val3
S2,week1,scan4.npy,False,4,40,Female,Val1
S3,week0,scan1.npy,,5,50,Male,Val2
S3,week1,scan3.npy,True,6,60,Male,Val2
"""
(dataset_folder / "dataset.csv").write_text(dataset_contents)
numerical_columns = ["numerical1", "numerical2"] if use_non_imaging_features else []
categorical_columns = ["categorical1", "categorical2"] if use_non_imaging_features else []
non_image_feature_channels = get_non_image_features_dict(default_channels=["week1", "week0"],
specific_channels={"categorical2": ["week1"]}) \
if use_non_imaging_features else {}
config_for_dataset = ScalarModelBase(
local_dataset=dataset_folder,
image_channels=["week0", "week1"],
image_file_column="path",
label_channels=["week1"],
label_value_column="label",
non_image_feature_channels=non_image_feature_channels,
numerical_columns=numerical_columns,
categorical_columns=categorical_columns,
should_validate=False
)
config_for_dataset.read_dataset_into_dataframe_and_pre_process()
dataset = ScalarDataset(config_for_dataset,
sample_transforms=ScalarItemAugmentation(
RandAugmentSlice(is_transformation_for_segmentation_maps=False)))
assert len(dataset) == 3
config = ImageEncoder(
encode_channels_jointly=encode_channels_jointly,
should_validate=False,
numerical_columns=numerical_columns,
categorical_columns=categorical_columns,
non_image_feature_channels=non_image_feature_channels,
categorical_feature_encoder=config_for_dataset.categorical_feature_encoder,
encoder_dimensionality_reduction_factor=reduction_factor,
aggregation_type=aggregation_type,
scan_size=(6, 64, 60)
)
if kernel_size_per_encoding_block:
config.kernel_size_per_encoding_block = kernel_size_per_encoding_block
if stride_size_per_encoding_block:
config.stride_size_per_encoding_block = stride_size_per_encoding_block
config.set_output_to(test_output_dirs.root_dir)
config.max_batch_grad_cam = 1
model = create_model_with_temperature_scaling(config)
input_size: List[Tuple] = [(len(config.image_channels), *scan_size)]
if use_non_imaging_features:
input_size.append((config.get_total_number_of_non_imaging_features(),))
# Original number output channels (unreduced) is
# num initial channel * (num encoder block - 1) = 4 * (3-1) = 8
if encode_channels_jointly:
# reduced_num_channels + num_non_img_features
assert model.final_num_feature_channels == expected_num_reduced_features + \
config.get_total_number_of_non_imaging_features()
else:
# num_img_channels * reduced_num_channels + num_non_img_features
assert model.final_num_feature_channels == len(config.image_channels) * expected_num_reduced_features + \
config.get_total_number_of_non_imaging_features()
summarizer = ModelSummary(model)
summarizer.generate_summary(input_sizes=input_size)
config.local_dataset = dataset_folder
config.validate()
model_train(config, checkpoint_handler=get_default_checkpoint_handler(model_config=config,
project_root=Path(test_output_dirs.root_dir)))
class BaseSegmentationModel(DeviceAwareModule, ABC):
"""
Base neural network segmentation model.
"""
@initialize_instance_variables
def __init__(self,
name: str,
input_channels: int,
crop_size_constraints: Optional[CropSizeConstraints] = None):
"""
Creates a new instance of the base model class.
:param name: A human readable name of the model.
:param input_channels: The number of image input channels.
:param crop_size_constraints: The size constraints for the training crop size. If not provided,
a minimum crop size of 1 is assumed.
"""
super().__init__()
self.num_dimensions = 3
self.name = name
self.input_channels = input_channels
self.summarizer: Optional[ModelSummary] = None
self.summary: Optional[OrderedDict] = None
self.summary_crop_size: Optional[TupleInt3] = None
if crop_size_constraints is None:
# Allow any size. With this initialization, both multiple_of and minimum_size will be populated.
crop_size_constraints = CropSizeConstraints(multiple_of=1)
self.crop_size_constraints = crop_size_constraints
def get_output_shape(
self, input_shape: Union[TupleInt2, TupleInt3]) -> Tuple[int, ...]:
"""
Computes model's output tensor shape for given input tensor shape.
The argument is expected to be either a 2-tuple or a 3-tuple. A batch dimension (1)
and the number of channels are added as the first dimensions. The result tuple has batch and channel dimension
stripped off.
:param input_shape: A tuple (2D or 3D) representing incoming tensor shape.
"""
# Create a sample tensor for inference
batch_size = 1
if len(input_shape) not in [2, 3]:
raise ValueError(
"Input shape has to be in 2D or 3D, found {}".format(
len(input_shape)))
input_tensors = \
[torch.zeros(batch_size, self.input_channels, *input_shape, dtype=torch.float)]
# Perform a forward pass then restore the state of the module
output_shape = forward_preserve_state(module=self,
inputs=input_tensors).size()
return tuple(output_shape[2:])
def partition_model(self,
devices: Optional[List[torch.device]] = None) -> None:
"""A method to partition a neural network model across multiple devices.
If no list of devices is given, use all available GPU devices."""
pass
def validate_crop_size(self,
crop_size: TupleInt3,
message_prefix: Optional[str] = None) -> None:
"""
Checks if the given crop size is a valid crop size for the present model.
If it is not valid, throw a ValueError.
:param crop_size: The crop size that should be checked.
:param message_prefix: A string prefix for the error message if the crop size is found to be invalid.
"""
if self.crop_size_constraints is not None:
self.crop_size_constraints.validate(crop_size, message_prefix)
def generate_model_summary(self,
crop_size: Optional[TupleInt3] = None,
log_summaries_to_files: bool = False) -> None:
"""
Stores a model summary, containing information about layers, memory consumption and runtime
in the model.summary field.
When called again with the same crop_size, the summary is not created again.
:param crop_size: The crop size for which the summary should be created. If not provided,
the minimum allowed crop size is used.
:param log_summaries_to_files: whether to write the summary to a file
"""
if crop_size is None:
crop_size = self.crop_size_constraints.minimum_size # type: ignore
assert crop_size is not None
input_size = [crop_size]
if self.summary is None or self.summary_crop_size != input_size:
self.summarizer = ModelSummary(self)
self.summary = self.summarizer.generate_summary(
input_sizes=[(self.input_channels, *crop_size)],
log_summaries_to_files=log_summaries_to_files)
self.summary_crop_size = crop_size
@abc.abstractmethod
def forward(self, input: Any) -> Any: # type: ignore
raise NotImplementedError("forward must be implemented by subclasses")
def get_all_child_layers(self) -> List[torch.nn.Module]:
raise NotImplementedError(
"get_all_child_layers must be implemented by subclasses")
