Add Covid configs

CHANGELOG.md

@@ -69,6 +69,7 @@ with only minimum code changes required. See [the MD documentation](docs/bring_y
   Detection Challenge datasets. See
   [SSL doc](https://github.com/microsoft/InnerEye-DeepLearning/blob/main/docs/self_supervised_models.md) for more
   details.
+- ([#456](https://github.com/microsoft/InnerEye-DeepLearning/pull/456)) Adding configs to train Covid detection models
 
 ### Changed
 

InnerEye/ML/SSL/datamodules_and_datasets/cxr_datasets.py

@@ -9,6 +9,7 @@
 import numpy as np
 import pandas as pd
 from PIL import Image
+from torch.utils.data import Subset
 from torchvision.datasets import VisionDataset
 
 from InnerEye.Common.type_annotations import PathOrString
@@ -175,3 +176,39 @@ def _prepare_dataset(self) -> None:
         self.dataset_dataframe.Path = self.dataset_dataframe.Path.apply(lambda x: x[strip_n:])
         self.indices = np.arange(len(self.dataset_dataframe))
         self.filenames = [self.root / p for p in self.dataset_dataframe.Path.values]
+
+
+class CovidDataset(InnerEyeCXRDatasetWithReturnIndex):
+    """
+    Dataset class to load CovidDataset dataset as datamodule for monitoring SSL training quality directly on
+    CovidDataset data.
+    We use CVX03 against CVX12 as proxy task.
+    """
+
+    def _prepare_dataset(self):
+        self.dataset_dataframe = pd.read_csv(self.root / "dataset.csv")
+        mapping = {"0": 0, "3": 0, "1": 1, "2": 1}
+        self.dataset_dataframe["final_label"] = self.dataset_dataframe.final_label.astype(int).astype(str).apply(
+            lambda x: mapping[x])
+        self.indices = np.arange(len(self.dataset_dataframe))
+        self.subject_ids = self.dataset_dataframe.subject.values
+        self.filenames = [self.root / file for file in self.dataset_dataframe.filepath.values]
+        self.targets = self.dataset_dataframe.final_label.values.astype(np.int64).reshape(-1)
+
+    @property
+    def num_classes(self) -> int:
+        return 2
+
+    def _split_dataset(self, val_split: float, seed: int):
+        """
+        Implements val - train split.
+        :param val_split: proportion to use for validation
+        :param seed: random seed for splitting
+        :return: dataset_train, dataset_val
+        """
+        shuffled_subject_ids = np.random.RandomState(seed).permutation(np.unique(self.subject_ids))
+        n_val = int(len(shuffled_subject_ids) * val_split)
+        val_subjects, train_subjects = shuffled_subject_ids[:n_val], shuffled_subject_ids[n_val:]
+        train_ids, val_ids = np.where(np.isin(self.subject_ids, train_subjects))[0], \
+                             np.where(np.isin(self.subject_ids, val_subjects))[0]
+        return Subset(self, train_ids), Subset(self, val_ids)

InnerEye/ML/SSL/lightning_containers/ssl_container.py

@@ -13,7 +13,7 @@
 from yacs.config import CfgNode
 
 from InnerEye.ML.SSL.datamodules_and_datasets.cifar_datasets import InnerEyeCIFAR10, InnerEyeCIFAR100
-from InnerEye.ML.SSL.datamodules_and_datasets.cxr_datasets import CheXpert, NIHCXR, RSNAKaggleCXR
+from InnerEye.ML.SSL.datamodules_and_datasets.cxr_datasets import CheXpert, CovidDataset, NIHCXR, RSNAKaggleCXR
 from InnerEye.ML.SSL.datamodules_and_datasets.datamodules import CombinedDataModule, InnerEyeVisionDataModule
 from InnerEye.ML.SSL.datamodules_and_datasets.transforms_utils import InnerEyeCIFARLinearHeadTransform, \
     InnerEyeCIFARTrainTransform, \
@@ -42,11 +42,12 @@ class EncoderName(Enum):
 
 
 class SSLDatasetName(Enum):
-    RSNAKaggleCXR = "RSNAKaggleCXR"
-    NIHCXR = "NIHCXR"
     CIFAR10 = "CIFAR10"
     CIFAR100 = "CIFAR100"
+    RSNAKaggleCXR = "RSNAKaggleCXR"
+    NIHCXR = "NIHCXR"
     CheXpert = "CheXpert"
+    Covid = "CovidDataset"
 
 
 InnerEyeDataModuleTypes = Union[InnerEyeVisionDataModule, CombinedDataModule]
@@ -62,11 +63,12 @@ class SSLContainer(LightningContainer):
     Note that this container is also used as the base class for SSLImageClassifier (finetuning container) as they share
     setup and datamodule methods.
     """
-    _SSLDataClassMappings = {SSLDatasetName.RSNAKaggleCXR.value: RSNAKaggleCXR,
-                             SSLDatasetName.NIHCXR.value: NIHCXR,
-                             SSLDatasetName.CIFAR10.value: InnerEyeCIFAR10,
+    _SSLDataClassMappings = {SSLDatasetName.CIFAR10.value: InnerEyeCIFAR10,
                              SSLDatasetName.CIFAR100.value: InnerEyeCIFAR100,
-                             SSLDatasetName.CheXpert.value: CheXpert}
+                             SSLDatasetName.RSNAKaggleCXR.value: RSNAKaggleCXR,
+                             SSLDatasetName.NIHCXR.value: NIHCXR,
+                             SSLDatasetName.CheXpert.value: CheXpert,
+                             SSLDatasetName.Covid.value: CovidDataset}
 
     ssl_augmentation_config = param.ClassSelector(class_=Path, allow_None=True,
                                                   doc="The path to the yaml config defining the parameters of the "
@@ -87,11 +89,13 @@ class SSLContainer(LightningContainer):
                                        "Used for debugging and tests.")
     linear_head_augmentation_config = param.ClassSelector(class_=Path,
                                                           doc="The path to the yaml config for the linear head "
-                                                             "augmentations")
+                                                              "augmentations")
     linear_head_dataset_name = param.ClassSelector(class_=SSLDatasetName,
                                                    doc="Name of the dataset to use for the linear head training")
     linear_head_batch_size = param.Integer(default=256, doc="Batch size for linear head tuning")
-    learning_rate_linear_head_during_ssl_training = param.Number(default=1e-4, doc="Learning rate for linear head training during SSL training.")
+    learning_rate_linear_head_during_ssl_training = param.Number(default=1e-4,
+                                                                 doc="Learning rate for linear head training during "
+                                                                     "SSL training.")
 
     def setup(self) -> None:
         from InnerEye.ML.SSL.lightning_containers.ssl_image_classifier import SSLClassifierContainer
@@ -173,7 +177,8 @@ def get_data_module(self) -> InnerEyeDataModuleTypes:
             return self.data_module
         encoder_data_module = self._create_ssl_data_modules(is_ssl_encoder_module=True)
         linear_data_module = self._create_ssl_data_modules(is_ssl_encoder_module=False)
-        return CombinedDataModule(encoder_data_module, linear_data_module, self.use_balanced_binary_loss_for_linear_head)
+        return CombinedDataModule(encoder_data_module, linear_data_module,
+                                  self.use_balanced_binary_loss_for_linear_head)
 
     def _create_ssl_data_modules(self, is_ssl_encoder_module: bool) -> InnerEyeVisionDataModule:
         """
@@ -220,7 +225,10 @@ def _get_transforms(self, augmentation_config: Optional[CfgNode],
         applied on. If False, return only one transformation.
         :return: training transformation pipeline and validation transformation pipeline.
         """
-        if dataset_name in [SSLDatasetName.RSNAKaggleCXR.value, SSLDatasetName.NIHCXR.value, SSLDatasetName.CheXpert.value]:
+        if dataset_name in [SSLDatasetName.RSNAKaggleCXR.value,
+                            SSLDatasetName.NIHCXR.value,
+                            SSLDatasetName.CheXpert.value,
+                            SSLDatasetName.Covid.value]:
             assert augmentation_config is not None
             train_transforms, val_transforms = get_cxr_ssl_transforms(augmentation_config,
                                                                       return_two_views_per_sample=is_ssl_encoder_module,

InnerEye/ML/configs/classification/CovidHierarchicalModel.py

@@ -0,0 +1,233 @@
+import codecs
+import logging
+import pickle
+from pathlib import Path
+
+from typing import Any, Callable
+
+import PIL
+import pandas as pd
+import param
+import torch
+from PIL import Image
+from pytorch_lightning import LightningModule
+from torchvision.transforms import Compose
+
+from InnerEye.Common.common_util import ModelProcessing, get_best_epoch_results_path
+from InnerEye.ML.SSL.datamodules_and_datasets.transforms_utils import create_chest_xray_transform
+from InnerEye.ML.SSL.lightning_containers.ssl_container import EncoderName
+
+from InnerEye.ML.SSL.lightning_modules.ssl_classifier_module import SSLClassifier
+from InnerEye.ML.SSL.utils import create_ssl_encoder, create_ssl_image_classifier, load_ssl_augmentation_config
+from InnerEye.ML.common import ModelExecutionMode
+
+from InnerEye.ML.configs.ssl.CXR_SSL_configs import path_linear_head_augmentation_cxr
+from InnerEye.ML.deep_learning_config import LRSchedulerType, MultiprocessingStartMethod, \
+    OptimizerType
+
+from InnerEye.ML.model_config_base import ModelTransformsPerExecutionMode
+from InnerEye.ML.model_testing import MODEL_OUTPUT_CSV
+
+from InnerEye.ML.models.architectures.classification.image_encoder_with_mlp import ImagingFeatureType
+from InnerEye.ML.reports.notebook_report import generate_notebook, get_ipynb_report_name, str_or_empty
+
+from InnerEye.ML.scalar_config import ScalarLoss, ScalarModelBase
+from InnerEye.ML.utils.augmentation import ScalarItemAugmentation
+from InnerEye.ML.utils.run_recovery import RunRecovery
+from InnerEye.ML.utils.split_dataset import DatasetSplits
+
+from InnerEye.ML.configs.ssl.CovidContainers import COVID_DATASET_ID
+from InnerEye.Common import fixed_paths as fixed_paths_innereye
+
+
+class CovidHierarchicalModel(ScalarModelBase):
+    """
+    Model to train a CovidDataset model from scratch or finetune from SSL-pretrained model.
+
+    For AML you need to provide the run_id of your SSL training job as a command line argument
+    --pretraining_run_recovery_id=id_of_your_ssl_model, this will download the checkpoints of the run to your
+    machine and load the corresponding pretrained model.
+
+    To recover from a particular checkpoint from your SSL run e.g. "recovery_epoch=499.ckpt" please use hte
+    --name_of_checkpoint argument.
+    """
+    use_pretrained_model = param.Boolean(default=False, doc="TBD")
+    freeze_encoder = param.Boolean(default=False, doc="Whether to freeze the pretrained encoder or not.")
+    name_of_checkpoint = param.String(default=None, doc="Filename of checkpoint to use for recovery")
+
+    def __init__(self, covid_dataset_id=COVID_DATASET_ID, **kwargs: Any):
+        learning_rate = 1e-5 if self.use_pretrained_model else 1e-4
+        super().__init__(target_names=['CVX03vs12', 'CVX0vs3', 'CVX1vs2'],
+                         loss_type=ScalarLoss.CustomClassification,
+                         class_names=['CVX0', 'CVX1', 'CVX2', 'CVX3'],
+                         max_num_gpus=1,
+                         azure_dataset_id=covid_dataset_id,
+                         subject_column="series",
+                         image_file_column="filepath",
+                         label_value_column="final_label",
+                         non_image_feature_channels=[],
+                         numerical_columns=[],
+                         use_mixed_precision=False,
+                         num_dataload_workers=12,
+                         multiprocessing_start_method=MultiprocessingStartMethod.fork,
+                         train_batch_size=64,
+                         optimizer_type=OptimizerType.Adam,
+                         num_epochs=50,
+                         l_rate_scheduler=LRSchedulerType.Step,
+                         l_rate_step_gamma=1.0,
+                         l_rate=learning_rate,
+                         l_rate_multi_step_milestones=None,
+                         **kwargs)
+        self.num_classes = 3
+        if not self.use_pretrained_model and self.freeze_encoder:
+            raise ValueError("You specified training from scratch but you asked to freeze the encoder.")
+
+    def should_generate_multilabel_report(self) -> bool:
+        return False
+
+    def get_model_train_test_dataset_splits(self, dataset_df: pd.DataFrame) -> DatasetSplits:
+        test_df = pd.read_csv(self.local_dataset / "test_series_ids.csv")
+
+        in_test_set = dataset_df.series.isin(test_df.series)
+        train_ids = dataset_df.series[~in_test_set].values
+        test_ids = dataset_df.series[in_test_set].values
+
+        num_val_samples = 400
+        val_ids = train_ids[:num_val_samples]
+        train_ids = train_ids[num_val_samples:]
+
+        return DatasetSplits.from_subject_ids(dataset_df, train_ids=train_ids, val_ids=val_ids, test_ids=test_ids,
+                                              subject_column="series", group_column="subject")
+
+    # noinspection PyTypeChecker
+    def get_image_sample_transforms(self) -> ModelTransformsPerExecutionMode:
+        config = load_ssl_augmentation_config(path_linear_head_augmentation_cxr)
+        train_transforms = ScalarItemAugmentation(
+            Compose(
+                [DicomPreparation(), create_chest_xray_transform(config, apply_augmentations=True)]))
+        val_transforms = ScalarItemAugmentation(
+            Compose(
+                [DicomPreparation(), create_chest_xray_transform(config, apply_augmentations=False)]))
+
+        return ModelTransformsPerExecutionMode(train=train_transforms,
+                                               val=val_transforms,
+                                               test=val_transforms)
+
+    def create_model(self) -> LightningModule:
+        """
+        This method must create the actual Lightning model that will be trained.
+        """
+        if self.use_pretrained_model:
+            path_to_checkpoint = self._get_ssl_checkpoint_path()
+
+            model = create_ssl_image_classifier(
+                num_classes=self.num_classes,
+                pl_checkpoint_path=str(path_to_checkpoint),
+                freeze_encoder=self.freeze_encoder)
+
+        else:
+            encoder = create_ssl_encoder(encoder_name=EncoderName.densenet121.value)
+            model = SSLClassifier(num_classes=self.num_classes,
+                                  encoder=encoder,
+                                  freeze_encoder=self.freeze_encoder,
+                                  class_weights=None)
+        # Next args are just here because we are using this model within an InnerEyeContainer
+        model.imaging_feature_type = ImagingFeatureType.Image
+        model.num_non_image_features = 0
+        model.encode_channels_jointly = True
+        return model
+
+    def _get_ssl_checkpoint_path(self):
+        # Get the SSL weights from the AML run provided via "pretraining_run_recovery_id" command line argument.
+        # Accessible via extra_downloaded_run_id field of the config.
+        assert self.extra_downloaded_run_id is not None
+        assert isinstance(self.extra_downloaded_run_id, RunRecovery)
+        ssl_path = self.checkpoint_folder / "ssl_checkpoint.ckpt"
+        logging.info(f"name_of_checkpoint: {self.name_of_checkpoint}")
+        if not ssl_path.exists():  # for test (when it is already present) we don't need to redo this.
+            if self.name_of_checkpoint is not None:
+                path_to_checkpoint = self.extra_downloaded_run_id.checkpoints_roots[0] / self.name_of_checkpoint
+            else:
+                try:
+                    path_to_checkpoint = self.extra_downloaded_run_id.get_best_checkpoint_paths()[0]
+                    assert path_to_checkpoint.exists()
+                except AssertionError:
+                    path_to_checkpoint = self.extra_downloaded_run_id.get_recovery_checkpoint_paths()[0]
+            assert path_to_checkpoint.exists()
+            path_to_checkpoint.rename(ssl_path)
+        return ssl_path
+
+    def pre_process_dataset_dataframe(self) -> None:
+        pass
+
+    @staticmethod
+    def get_posthoc_label_transform() -> Callable:
+        import torch
+
+        def multiclass_to_hierarchical_labels(classes: torch.Tensor) -> torch.Tensor:
+            classes = classes.clone()
+            cvx03vs12 = classes[..., 1] + classes[..., 2]
+            cvx0vs3 = classes[..., 3]
+            cvx1vs2 = classes[..., 2]
+            cvx0vs3[cvx03vs12 == 1] = float('nan')  # CVX0vs3 only gets gradient for CVX03
+            cvx1vs2[cvx03vs12 == 0] = float('nan')  # CVX1vs2 only gets gradient for CVX12
+            return torch.stack([cvx03vs12, cvx0vs3, cvx1vs2], -1)
+
+        return multiclass_to_hierarchical_labels
+
+    @staticmethod
+    def get_loss_function() -> Callable:
+        import torch
+        import torch.nn.functional as F
+
+        def nan_bce_with_logits(output: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
+            """Compute BCE with logits, ignoring NaN values"""
+            valid = labels.isfinite()
+            losses = F.binary_cross_entropy_with_logits(output[valid], labels[valid], reduction='none')
+            return losses.sum() / labels.shape[0]
+
+        return nan_bce_with_logits
+
+    def generate_custom_report(self, report_dir: Path, model_proc: ModelProcessing) -> Path:
+        """
+        Generate a custom report for the CovidDataset Hierarchical model. At the moment, this report will read the
+        file model_output.csv generated for the training, validation or test sets and compute a 4 class accuracy
+        and confusion matrix based on this.
+        :param report_dir: Directory report is to be written to
+        :param model_proc: Whether this is a single or ensemble model (model_output.csv will be located in different
+        paths for single vs ensemble runs.)
+        """
+
+        def get_output_csv_path(mode: ModelExecutionMode) -> Path:
+            p = get_best_epoch_results_path(mode=mode, model_proc=model_proc)
+            return self.outputs_folder / p / MODEL_OUTPUT_CSV
+
+        train_metrics = get_output_csv_path(ModelExecutionMode.TRAIN)
+        val_metrics = get_output_csv_path(ModelExecutionMode.VAL)
+        test_metrics = get_output_csv_path(ModelExecutionMode.TEST)
+
+        notebook_params = \
+            {
+                'innereye_path': str(fixed_paths_innereye.repository_root_directory()),
+                'train_metrics_csv': str_or_empty(train_metrics),
+                'val_metrics_csv': str_or_empty(val_metrics),
+                'test_metrics_csv': str_or_empty(test_metrics),
+                "config": codecs.encode(pickle.dumps(self), "base64").decode(),
+                "is_crossval_report": False
+            }
+        template = Path(__file__).absolute().parent.parent / "utils" / "HierarchicalCovidModelReport.ipynb"
+        return generate_notebook(template,
+                                 notebook_params=notebook_params,
+                                 result_notebook=report_dir / get_ipynb_report_name(
+                                     f"{self.model_category.value}_hierarchical"))
+
+
+class DicomPreparation:
+    def __call__(self, item: torch.Tensor) -> PIL.Image:
+        # Item will be of dimension [C, Z, X, Y]
+        images = item.numpy()
+        assert images.shape[0] == 1 and images.shape[1] == 1
+        images = images.reshape(images.shape[2:])
+        normalized_image = (images - images.min()) * 255. / (images.max() - images.min())
+        image = Image.fromarray(normalized_image).convert("L")
+        return image