synapse_net.training.supervised_training

  1import os
  2from glob import glob
  3from typing import Optional, Tuple
  4
  5import torch
  6import torch_em
  7from sklearn.model_selection import train_test_split
  8from torch_em.model import AnisotropicUNet, UNet2d
  9
 10
 11def get_3d_model(
 12    out_channels: int,
 13    in_channels: int = 1,
 14    scale_factors: Tuple[Tuple[int, int, int]] = [[1, 2, 2], [2, 2, 2], [2, 2, 2], [2, 2, 2]],
 15    initial_features: int = 32,
 16    final_activation: str = "Sigmoid",
 17) -> torch.nn.Module:
 18    """Get the U-Net model for 3D segmentation tasks.
 19
 20    Args:
 21        out_channels: The number of output channels of the network.
 22        scale_factors: The downscaling factors for each level of the U-Net encoder.
 23        initial_features: The number of features in the first level of the U-Net.
 24            The number of features increases by a factor of two in each level.
 25        final_activation: The activation applied to the last output layer.
 26
 27    Returns:
 28        The U-Net.
 29    """
 30    model = AnisotropicUNet(
 31        scale_factors=scale_factors,
 32        in_channels=in_channels,
 33        out_channels=out_channels,
 34        initial_features=initial_features,
 35        gain=2,
 36        final_activation=final_activation,
 37    )
 38    return model
 39
 40
 41def get_2d_model(
 42    out_channels: int,
 43    in_channels: int = 1,
 44    initial_features: int = 32,
 45    final_activation: str = "Sigmoid",
 46) -> torch.nn.Module:
 47    """Get the U-Net model for 2D segmentation tasks.
 48
 49    Args:
 50        out_channels: The number of output channels of the network.
 51        initial_features: The number of features in the first level of the U-Net.
 52            The number of features increases by a factor of two in each level.
 53        final_activation: The activation applied to the last output layer.
 54
 55    Returns:
 56        The U-Net.
 57    """
 58    model = UNet2d(
 59        in_channels=in_channels,
 60        out_channels=out_channels,
 61        initial_features=initial_features,
 62        gain=2,
 63        depth=4,
 64        final_activation=final_activation,
 65    )
 66    return model
 67
 68
 69def _adjust_patch_shape(data_shape, patch_shape):
 70    # If data is 2D and patch_shape is 3D, drop the extra dimension in patch_shape
 71    if data_shape == 2 and len(patch_shape) == 3:
 72        return patch_shape[1:]  # Remove the leading dimension in patch_shape
 73    return patch_shape  # Return the original patch_shape for 3D data
 74
 75
 76def _determine_ndim(patch_shape):
 77    # Check for 2D or 3D training
 78    try:
 79        z, y, x = patch_shape
 80    except ValueError:
 81        y, x = patch_shape
 82        z = 1
 83    is_2d = z == 1
 84    ndim = 2 if is_2d else 3
 85    return is_2d, ndim
 86
 87
 88def get_supervised_loader(
 89    data_paths: Tuple[str],
 90    raw_key: str,
 91    label_key: str,
 92    patch_shape: Tuple[int, int, int],
 93    batch_size: int,
 94    n_samples: Optional[int],
 95    add_boundary_transform: bool = True,
 96    label_dtype=torch.float32,
 97    rois: Optional[Tuple[Tuple[slice]]] = None,
 98    sampler: Optional[callable] = None,
 99    ignore_label: Optional[int] = None,
100    label_transform: Optional[callable] = None,
101    label_paths: Optional[Tuple[str]] = None,
102    **loader_kwargs,
103) -> torch.utils.data.DataLoader:
104    """Get a dataloader for supervised segmentation training.
105
106    Args:
107        data_paths: The filepaths to the hdf5 files containing the training data.
108        raw_key: The key that holds the raw data inside of the hdf5.
109        label_key: The key that holds the labels inside of the hdf5.
110        patch_shape: The patch shape used for a training example.
111            In order to run 2d training pass a patch shape with a singleton in the z-axis,
112            e.g. 'patch_shape = [1, 512, 512]'.
113        batch_size: The batch size for training.
114        n_samples: The number of samples per epoch. By default this will be estimated
115            based on the patch_shape and size of the volumes used for training.
116        add_boundary_transform: Whether to add a boundary channel to the training data.
117        label_dtype: The datatype of the labels returned by the dataloader.
118        rois: Optional region of interests for training.
119        sampler: Optional sampler for selecting blocks for training.
120            By default a minimum instance sampler will be used.
121        ignore_label: Ignore label in the ground-truth. The areas marked by this label will be
122            ignored in the loss computation. By default this option is not used.
123        label_transform: Label transform that is applied to the segmentation to compute the targets.
124            If no label transform is passed (the default) a boundary transform is used.
125        label_paths: Optional paths containing the labels / annotations for training.
126            If not given, the labels are expected to be contained in the `data_paths`.
127        loader_kwargs: Additional keyword arguments for the dataloader.
128
129    Returns:
130        The PyTorch dataloader.
131    """
132    _, ndim = _determine_ndim(patch_shape)
133    if label_transform is not None:  # A specific label transform was passed, do nothing.
134        pass
135    elif add_boundary_transform:
136        if ignore_label is None:
137            label_transform = torch_em.transform.BoundaryTransform(add_binary_target=True)
138        else:
139            label_transform = torch_em.transform.label.BoundaryTransformWithIgnoreLabel(
140                add_binary_target=True, ignore_label=ignore_label
141            )
142
143    else:
144        if ignore_label is not None:
145            raise NotImplementedError
146        label_transform = torch_em.transform.label.connected_components
147
148    if ndim == 2:
149        adjusted_patch_shape = _adjust_patch_shape(ndim, patch_shape)
150        transform = torch_em.transform.Compose(
151            torch_em.transform.PadIfNecessary(adjusted_patch_shape), torch_em.transform.get_augmentations(2)
152        )
153    else:
154        transform = torch_em.transform.Compose(
155            torch_em.transform.PadIfNecessary(patch_shape), torch_em.transform.get_augmentations(3)
156        )
157
158    num_workers = loader_kwargs.pop("num_workers", 4 * batch_size)
159    shuffle = loader_kwargs.pop("shuffle", True)
160
161    if sampler is None:
162        sampler = torch_em.data.sampler.MinInstanceSampler(min_num_instances=4)
163
164    if label_paths is None:
165        label_paths = data_paths
166    elif len(label_paths) != len(data_paths):
167        raise ValueError(f"Data paths and label paths don't match: {len(data_paths)} != {len(label_paths)}")
168
169    loader = torch_em.default_segmentation_loader(
170        data_paths, raw_key,
171        label_paths, label_key, sampler=sampler,
172        batch_size=batch_size, patch_shape=patch_shape, ndim=ndim,
173        is_seg_dataset=True, label_transform=label_transform, transform=transform,
174        num_workers=num_workers, shuffle=shuffle, n_samples=n_samples,
175        label_dtype=label_dtype, rois=rois, **loader_kwargs,
176    )
177    return loader
178
179
180def supervised_training(
181    name: str,
182    train_paths: Tuple[str],
183    val_paths: Tuple[str],
184    label_key: str,
185    patch_shape: Tuple[int, int, int],
186    save_root: Optional[str] = None,
187    raw_key: str = "raw",
188    batch_size: int = 1,
189    lr: float = 1e-4,
190    n_iterations: int = int(1e5),
191    train_label_paths: Optional[Tuple[str]] = None,
192    val_label_paths: Optional[Tuple[str]] = None,
193    train_rois: Optional[Tuple[Tuple[slice]]] = None,
194    val_rois: Optional[Tuple[Tuple[slice]]] = None,
195    sampler: Optional[callable] = None,
196    n_samples_train: Optional[int] = None,
197    n_samples_val: Optional[int] = None,
198    check: bool = False,
199    ignore_label: Optional[int] = None,
200    label_transform: Optional[callable] = None,
201    in_channels: int = 1,
202    out_channels: int = 2,
203    mask_channel: bool = False,
204    checkpoint_path: Optional[str] = None,
205    **loader_kwargs,
206):
207    """Run supervised segmentation training.
208
209    This function trains a UNet for predicting outputs for segmentation.
210    Expects instance labels and converts them to boundary targets.
211    This behaviour can be changed by passing custom arguments for `label_transform`
212    and/or `out_channels`.
213
214    Args:
215        name: The name for the checkpoint to be trained.
216        train_paths: Filepaths to the hdf5 files for the training data.
217        val_paths: Filepaths to the df5 files for the validation data.
218        label_key: The key that holds the labels inside of the hdf5.
219        patch_shape: The patch shape used for a training example.
220            In order to run 2d training pass a patch shape with a singleton in the z-axis,
221            e.g. 'patch_shape = [1, 512, 512]'.
222        save_root: Folder where the checkpoint will be saved.
223        raw_key: The key that holds the raw data inside of the hdf5.
224        batch_size: The batch size for training.
225        lr: The initial learning rate.
226        n_iterations: The number of iterations to train for.
227        train_label_paths: Optional paths containing the label data for training.
228            If not given, the labels are expected to be part of `train_paths`.
229        val_label_paths: Optional paths containing the label data for validation.
230            If not given, the labels are expected to be part of `val_paths`.
231        train_rois: Optional region of interests for training.
232        val_rois: Optional region of interests for validation.
233        sampler: Optional sampler for selecting blocks for training.
234            By default a minimum instance sampler will be used.
235        n_samples_train: The number of train samples per epoch. By default this will be estimated
236            based on the patch_shape and size of the volumes used for training.
237        n_samples_val: The number of val samples per epoch. By default this will be estimated
238            based on the patch_shape and size of the volumes used for validation.
239        check: Whether to check the training and validation loaders instead of running training.
240        ignore_label: Ignore label in the ground-truth. The areas marked by this label will be
241            ignored in the loss computation. By default this option is not used.
242        label_transform: Label transform that is applied to the segmentation to compute the targets.
243            If no label transform is passed (the default) a boundary transform is used.
244        out_channels: The number of output channels of the UNet.
245        mask_channel: Whether the last channels in the labels should be used for masking the loss.
246            This can be used to implement more complex masking operations and is not compatible with `ignore_label`.
247        checkpoint_path: Path to the directory where 'best.pt' resides; continue training this model.
248        loader_kwargs: Additional keyword arguments for the dataloader.
249    """
250    train_loader = get_supervised_loader(train_paths, raw_key, label_key, patch_shape, batch_size,
251                                         n_samples=n_samples_train, rois=train_rois, sampler=sampler,
252                                         ignore_label=ignore_label, label_transform=label_transform,
253                                         label_paths=train_label_paths, **loader_kwargs)
254    val_loader = get_supervised_loader(val_paths, raw_key, label_key, patch_shape, batch_size,
255                                       n_samples=n_samples_val, rois=val_rois, sampler=sampler,
256                                       ignore_label=ignore_label, label_transform=label_transform,
257                                       label_paths=val_label_paths, **loader_kwargs)
258
259    if check:
260        from torch_em.util.debug import check_loader
261        check_loader(train_loader, n_samples=4)
262        check_loader(val_loader, n_samples=4)
263        return
264
265    is_2d, _ = _determine_ndim(patch_shape)
266    if is_2d:
267        model = get_2d_model(out_channels=out_channels, in_channels=in_channels)
268    else:
269        model = get_3d_model(out_channels=out_channels, in_channels=in_channels)
270    
271    if checkpoint_path:
272        model = torch_em.util.load_model(checkpoint=checkpoint_path)
273
274    loss, metric = None, None
275    # No ignore label -> we can use default loss.
276    if ignore_label is None and not mask_channel:
277        pass
278    # If we have an ignore label the loss and metric have to be modified
279    # so that the ignore mask is not used in the gradient calculation.
280    elif ignore_label is not None:
281        loss = torch_em.loss.LossWrapper(
282            loss=torch_em.loss.DiceLoss(),
283            transform=torch_em.loss.wrapper.MaskIgnoreLabel(
284                ignore_label=ignore_label, masking_method="multiply",
285            )
286        )
287        metric = loss
288    elif mask_channel:
289        loss = torch_em.loss.LossWrapper(
290            loss=torch_em.loss.DiceLoss(),
291            transform=torch_em.loss.wrapper.ApplyAndRemoveMask(
292                masking_method="crop" if out_channels == 1 else "multiply")
293        )
294        metric = loss
295    else:
296        raise ValueError
297
298    trainer = torch_em.default_segmentation_trainer(
299        name=name,
300        model=model,
301        train_loader=train_loader,
302        val_loader=val_loader,
303        learning_rate=lr,
304        mixed_precision=True,
305        log_image_interval=100,
306        compile_model=False,
307        save_root=save_root,
308        loss=loss,
309        metric=metric,
310    )
311    trainer.fit(n_iterations)
312
313
314def _derive_key_from_files(files, key):
315    # Get all file extensions (general wild-cards may pick up files with multiple extensions).
316    extensions = list(set([os.path.splitext(ff)[1] for ff in files]))
317
318    # If we have more than 1 file extension we just use the key that was passed,
319    # as it is unclear how to derive a consistent key.
320    if len(extensions) > 1:
321        return files, key
322
323    ext = extensions[0]
324    extension_to_key = {".tif": None, ".mrc": "data", ".rec": "data"}
325
326    # Derive the key from the extension if the key is None.
327    if key is None and ext in extension_to_key:
328        key = extension_to_key[ext]
329    # If the key is None and can't be derived raise an error.
330    elif key is None and ext not in extension_to_key:
331        raise ValueError(
332            f"You have not passed a key for the data in {ext} format, for which the key cannot be derived."
333        )
334    # If the key was passed and doesn't match the extension raise an error.
335    elif key is not None and ext in extension_to_key and key != extension_to_key[ext]:
336        raise ValueError(
337            f"The expected key {extension_to_key[ext]} for format {ext} did not match the passed key {key}."
338        )
339    return files, key
340
341
342def _parse_input_folder(folder, pattern, key):
343    files = sorted(glob(os.path.join(folder, "**", pattern), recursive=True))
344    return _derive_key_from_files(files, key)
345
346
347def _parse_input_files(args):
348    train_image_paths, raw_key = _parse_input_folder(args.train_folder, args.image_file_pattern, args.raw_key)
349    train_label_paths, label_key = _parse_input_folder(args.label_folder, args.label_file_pattern, args.label_key)
350    if len(train_image_paths) != len(train_label_paths):
351        raise ValueError(
352            f"The image and label paths parsed from {args.train_folder} and {args.label_folder} don't match."
353            f"The image folder contains {len(train_image_paths)}, the label folder contains {len(train_label_paths)}."
354        )
355
356    if args.val_folder is None:
357        if args.val_label_folder is not None:
358            raise ValueError("You have passed a val_label_folder, but not a val_folder.")
359        train_image_paths, val_image_paths, train_label_paths, val_label_paths = train_test_split(
360            train_image_paths, train_label_paths, test_size=args.val_fraction, random_state=42
361        )
362    else:
363        if args.val_label_folder is None:
364            raise ValueError("You have passed a val_folder, but not a val_label_folder.")
365        val_image_paths = _parse_input_folder(args.val_image_folder, args.image_file_pattern, raw_key)
366        val_label_paths = _parse_input_folder(args.val_label_folder, args.label_file_pattern, label_key)
367
368    return train_image_paths, train_label_paths, val_image_paths, val_label_paths, raw_key, label_key
369
370
371# TODO enable initialization with a pre-trained model.
372def main():
373    """@private
374    """
375    import argparse
376
377    parser = argparse.ArgumentParser(
378        description="Train a model for foreground and boundary segmentation via supervised learning.\n\n"
379        "You can use this function to train a model for vesicle segmentation, or another segmentation task, like this:\n"  # noqa
380        "synapse_net.run_supervised_training -n my_model -i /path/to/images -l /path/to/labels --patch_shape 32 192 192\n"  # noqa
381        "The trained model will be saved in the folder 'checkpoints/my_model' (or whichever name you pass to the '-n' argument)."  # noqa
382        "You can then use this model for segmentation with the SynapseNet GUI or CLI. "
383        "Check out the information below for details on the arguments of this function.",
384        formatter_class=argparse.RawTextHelpFormatter
385    )
386    parser.add_argument("-n", "--name", required=True, help="The name of the model to be trained.")
387    parser.add_argument("-p", "--patch_shape", nargs=3, type=int, help="The patch shape for training.")
388
389    # Folders with training data, containing raw/image data and labels.
390    parser.add_argument("-i", "--train_folder", required=True, help="The input folder with the training image data.")
391    parser.add_argument("--image_file_pattern", default="*",
392                        help="The pattern for selecting image files. For example, '*.mrc' to select all mrc files.")
393    parser.add_argument("--raw_key",
394                        help="The internal path for the raw data. If not given, will be determined based on the file extension.")  # noqa
395    parser.add_argument("-l", "--label_folder", required=True, help="The input folder with the training labels.")
396    parser.add_argument("--label_file_pattern", default="*",
397                        help="The pattern for selecting label files. For example, '*.tif' to select all tif files.")
398    parser.add_argument("--label_key",
399                        help="The internal path for the label data. If not given, will be determined based on the file extension.")  # noqa
400
401    # Optional folders with validation data. If not given the training data is split into train/val.
402    parser.add_argument("--val_folder",
403                        help="The input folder with the validation data. If not given the training data will be split for validation")  # noqa
404    parser.add_argument("--val_label_folder",
405                        help="The input folder with the validation labels. If not given the training data will be split for validation.")  # noqa
406
407    # More optional argument:
408    parser.add_argument("--batch_size", type=int, default=1, help="The batch size for training.")
409    parser.add_argument("--n_samples_train", type=int, help="The number of samples per epoch for training. If not given will be derived from the data size.")  # noqa
410    parser.add_argument("--n_samples_val", type=int, help="The number of samples per epoch for validation. If not given will be derived from the data size.")  # noqa
411    parser.add_argument("--val_fraction", type=float, default=0.15, help="The fraction of the data to use for validation. This has no effect if 'val_folder' and 'val_label_folder' were passed.")  # noqa
412    parser.add_argument("--check", action="store_true", help="Visualize samples from the data loaders to ensure correct data instead of running training.")  # noqa
413    args = parser.parse_args()
414
415    train_image_paths, train_label_paths, val_image_paths, val_label_paths, raw_key, label_key =\
416        _parse_input_files(args)
417
418    supervised_training(
419        name=args.name, train_paths=train_image_paths, val_paths=val_image_paths,
420        train_label_paths=train_label_paths, val_label_paths=val_label_paths,
421        raw_key=raw_key, label_key=label_key, patch_shape=args.patch_shape, batch_size=args.batch_size,
422        n_samples_train=args.n_samples_train, n_samples_val=args.n_samples_val,
423        check=args.check,
424    )