[NF] Auto-encoders - streamlines - FINTA

dwi_ml/data/hdf5/hdf5_creation.py

@@ -634,7 +634,7 @@ def _create_streamline_groups(self, ref, subj_input_dir, subj_id,
                     raise ValueError(
                         "The data_per_streamline key '{}' was not found in "
                         "the sft. Check your tractogram file.".format(dps_key))
-                
+
                 logging.debug("    Include dps \"{}\" in the HDF5.".format(dps_key))
                 streamlines_group.create_dataset('dps_' + dps_key,
                                                  data=sft.data_per_streamline[dps_key])
@@ -669,8 +669,6 @@ def _process_one_streamline_group(
             Reference used to load and send the streamlines in voxel space and
             to create final merged SFT. If the file is a .trk, 'same' is used
             instead.
-        remove_invalid : bool
-            If True, invalid streamlines will be removed
 
         Returns
         -------

dwi_ml/models/projects/ae_models.py

@@ -0,0 +1,189 @@
+# -*- coding: utf-8 -*-
+import logging
+from typing import List
+
+import torch
+from torch.nn import functional as F
+
+from dwi_ml.models.main_models import MainModelAbstract
+
+
+class ModelAE(MainModelAbstract):
+    """
+    Recurrent tracking model.
+
+    Composed of an embedding for the imaging data's input + for the previous
+    direction's input, an RNN model to process the sequences, and a direction
+    getter model to convert the RNN outputs to the right structure, e.g.
+    deterministic (3D vectors) or probabilistic (based on probability
+    distribution parameters).
+    """
+    def __init__(self, kernel_size, latent_space_dims,
+                 experiment_name: str,
+                 # Target preprocessing params for the batch loader + tracker
+                 step_size: float = None,
+                 compress_lines: float = False,
+                 # Other
+                 log_level=logging.root.level):
+        super().__init__(experiment_name, step_size, compress_lines, log_level)
+
+        self.kernel_size = kernel_size
+        self.latent_space_dims = latent_space_dims
+
+        self.pad = torch.nn.ReflectionPad1d(1)
+
+        def pre_pad(m):
+            return torch.nn.Sequential(self.pad, m)
+
+        self.fc1 = torch.nn.Linear(8192,
+                                   self.latent_space_dims)  # 8192 = 1024*8
+        self.fc2 = torch.nn.Linear(self.latent_space_dims, 8192)
+
+        """
+        Encode convolutions
+        """
+        self.encod_conv1 = pre_pad(
+            torch.nn.Conv1d(3, 32, self.kernel_size, stride=2, padding=0)
+        )
+        self.encod_conv2 = pre_pad(
+            torch.nn.Conv1d(32, 64, self.kernel_size, stride=2, padding=0)
+        )
+        self.encod_conv3 = pre_pad(
+            torch.nn.Conv1d(64, 128, self.kernel_size, stride=2, padding=0)
+        )
+        self.encod_conv4 = pre_pad(
+            torch.nn.Conv1d(128, 256, self.kernel_size, stride=2, padding=0)
+        )
+        self.encod_conv5 = pre_pad(
+            torch.nn.Conv1d(256, 512, self.kernel_size, stride=2, padding=0)
+        )
+        self.encod_conv6 = pre_pad(
+            torch.nn.Conv1d(512, 1024, self.kernel_size, stride=1, padding=0)
+        )
+
+        """
+        Decode convolutions
+        """
+        self.decod_conv1 = pre_pad(
+            torch.nn.Conv1d(1024, 512, self.kernel_size, stride=1, padding=0)
+        )
+        self.upsampl1 = torch.nn.Upsample(
+            scale_factor=2, mode="linear", align_corners=False
+        )
+        self.decod_conv2 = pre_pad(
+            torch.nn.Conv1d(512, 256, self.kernel_size, stride=1, padding=0)
+        )
+        self.upsampl2 = torch.nn.Upsample(
+            scale_factor=2, mode="linear", align_corners=False
+        )
+        self.decod_conv3 = pre_pad(
+            torch.nn.Conv1d(256, 128, self.kernel_size, stride=1, padding=0)
+        )
+        self.upsampl3 = torch.nn.Upsample(
+            scale_factor=2, mode="linear", align_corners=False
+        )
+        self.decod_conv4 = pre_pad(
+            torch.nn.Conv1d(128, 64, self.kernel_size, stride=1, padding=0)
+        )
+        self.upsampl4 = torch.nn.Upsample(
+            scale_factor=2, mode="linear", align_corners=False
+        )
+        self.decod_conv5 = pre_pad(
+            torch.nn.Conv1d(64, 32, self.kernel_size, stride=1, padding=0)
+        )
+        self.upsampl5 = torch.nn.Upsample(
+            scale_factor=2, mode="linear", align_corners=False
+        )
+        self.decod_conv6 = pre_pad(
+            torch.nn.Conv1d(32, 3, self.kernel_size, stride=1, padding=0)
+        )
+
+    @property
+    def params_for_checkpoint(self):
+        """All parameters necessary to create again the same model. Will be
+        used in the trainer, when saving the checkpoint state. Params here
+        will be used to re-create the model when starting an experiment from
+        checkpoint. You should be able to re-create an instance of your
+        model with those params."""
+        # p = super().params_for_checkpoint()
+        p = {'kernel_size': self.kernel_size,
+             'latent_space_dims': self.latent_space_dims,
+             'experiment_name': self.experiment_name}
+        return p
+
+    @classmethod
+    def _load_params(cls, model_dir):
+        p = super()._load_params(model_dir)
+        p['kernel_size'] = 3
+        p['latent_space_dims'] = 32
+        return p
+
+    def forward(self,
+                input_streamlines: List[torch.tensor],
+                ):
+        """Run the model on a batch of sequences.
+
+        Parameters
+        ----------
+        input_streamlines: List[torch.tensor],
+            Batch of streamlines. Only used if previous directions are added to
+            the model. Used to compute directions; its last point will not be
+            used.
+
+        Returns
+        -------
+        model_outputs : List[Tensor]
+            Output data, ready to be passed to either `compute_loss()` or
+            `get_tracking_directions()`.
+        """
+
+        x = self.decode(self.encode(input_streamlines))
+        return x
+
+    def encode(self, x):
+        # x: list of tensors
+        x = torch.stack(x)
+        x = torch.swapaxes(x, 1, 2)
+
+        h1 = F.relu(self.encod_conv1(x))
+        h2 = F.relu(self.encod_conv2(h1))
+        h3 = F.relu(self.encod_conv3(h2))
+        h4 = F.relu(self.encod_conv4(h3))
+        h5 = F.relu(self.encod_conv5(h4))
+        h6 = self.encod_conv6(h5)
+
+        self.encoder_out_size = (h6.shape[1], h6.shape[2])
+
+        # Flatten
+        h7 = h6.view(-1, self.encoder_out_size[0] * self.encoder_out_size[1])
+
+        fc1 = self.fc1(h7)
+
+        return fc1
+
+    def decode(self, z):
+        fc = self.fc2(z)
+        fc_reshape = fc.view(
+            -1, self.encoder_out_size[0], self.encoder_out_size[1]
+        )
+        h1 = F.relu(self.decod_conv1(fc_reshape))
+        h2 = self.upsampl1(h1)
+        h3 = F.relu(self.decod_conv2(h2))
+        h4 = self.upsampl2(h3)
+        h5 = F.relu(self.decod_conv3(h4))
+        h6 = self.upsampl3(h5)
+        h7 = F.relu(self.decod_conv4(h6))
+        h8 = self.upsampl4(h7)
+        h9 = F.relu(self.decod_conv5(h8))
+        h10 = self.upsampl5(h9)
+        h11 = self.decod_conv6(h10)
+
+        return h11
+
+    def compute_loss(self, model_outputs, targets, average_results=True):
+
+        targets = torch.stack(targets)
+        targets = torch.swapaxes(targets, 1, 2)
+        reconstruction_loss = torch.nn.MSELoss(reduction="sum")
+        mse = reconstruction_loss(model_outputs, targets)
+        return mse, 1

dwi_ml/training/batch_loaders.py

@@ -58,7 +58,7 @@
 logger = logging.getLogger('batch_loader_logger')
 
 
-class DWIMLAbstractBatchLoader:
+class DWIMLStreamlinesBatchLoader:
     def __init__(self, dataset: MultiSubjectDataset, model: MainModelAbstract,
                  streamline_group_name: str, rng: int,
                  split_ratio: float = 0.,
@@ -197,7 +197,11 @@ def _data_augmentation_sft(self, sft):
                 self.context_subset.compress == self.model.compress_lines:
             logger.debug("Compression rate is the same as when creating "
                          "the hdf5 dataset. Not compressing again.")
-        else:
+        elif self.model.step_size is not None and \
+                self.model.compress_lines is not None:
+            logger.debug("Resample streamlines using: \n" +
+                         "- step_size: {}\n".format(self.model.step_size) +
+                         "- compress_lines: {}".format(self.model.compress_lines))
             sft = resample_or_compress(sft, self.model.step_size,
                                        self.model.nb_points,
                                        self.model.compress_lines)
@@ -314,6 +318,7 @@ def load_batch_streamlines(
             sft.to_vox()
             sft.to_corner()
             batch_streamlines.extend(sft.streamlines)
+
         batch_streamlines = [torch.as_tensor(s) for s in batch_streamlines]
 
         return batch_streamlines, final_s_ids_per_subj
@@ -351,7 +356,7 @@ def load_batch_connectivity_matrices(
                 connectivity_nb_blocs, connectivity_labels)
 
 
-class DWIMLBatchLoaderOneInput(DWIMLAbstractBatchLoader):
+class DWIMLBatchLoaderOneInput(DWIMLStreamlinesBatchLoader):
     """
     Loads:
         input = one volume group

dwi_ml/training/trainers.py

@@ -17,7 +17,7 @@
 from dwi_ml.models.main_models import (MainModelAbstract,
                                        ModelWithDirectionGetter)
 from dwi_ml.training.batch_loaders import (
-    DWIMLAbstractBatchLoader, DWIMLBatchLoaderOneInput)
+    DWIMLStreamlinesBatchLoader, DWIMLBatchLoaderOneInput)
 from dwi_ml.training.batch_samplers import DWIMLBatchIDSampler
 from dwi_ml.training.utils.gradient_norm import compute_gradient_norm
 from dwi_ml.training.utils.monitoring import (
@@ -53,7 +53,7 @@ class DWIMLAbstractTrainer:
     def __init__(self,
                  model: MainModelAbstract, experiments_path: str,
                  experiment_name: str, batch_sampler: DWIMLBatchIDSampler,
-                 batch_loader: DWIMLAbstractBatchLoader,
+                 batch_loader: DWIMLStreamlinesBatchLoader,
                  learning_rates: Union[List, float] = None,
                  weight_decay: float = 0.01,
                  optimizer: str = 'Adam', max_epochs: int = 10,
@@ -78,7 +78,7 @@ def __init__(self,
         batch_sampler: DWIMLBatchIDSampler
             Instantiated class used for sampling batches.
             Data in batch_sampler.dataset must be already loaded.
-        batch_loader: DWIMLAbstractBatchLoader
+        batch_loader: DWIMLStreamlinesBatchLoader
             Instantiated class with a load_batch method able to load data
             associated to sampled batch ids. Data in batch_sampler.dataset must
             be already loaded.
@@ -461,7 +461,7 @@ def _prepare_checkpoint_info(self) -> dict:
     def init_from_checkpoint(
             cls, model: MainModelAbstract, experiments_path, experiment_name,
             batch_sampler: DWIMLBatchIDSampler,
-            batch_loader: DWIMLAbstractBatchLoader,
+            batch_loader: DWIMLStreamlinesBatchLoader,
             checkpoint_state: dict, new_patience, new_max_epochs, log_level):
         """
         Loads checkpoint information (parameters and states) to instantiate
@@ -1013,7 +1013,51 @@ def run_one_batch(self, data):
             Any other data returned when computing loss. Not used in the
             trainer, but could be useful anywhere else.
         """
-        raise NotImplementedError
+        # Data interpolation has not been done yet. GPU computations are done
+        # here in the main thread.
+        targets, ids_per_subj = data
+
+        # Dataloader always works on CPU. Sending to right device.
+        # (model is already moved).
+        targets = [s.to(self.device, non_blocking=True, dtype=torch.float)
+                   for s in targets]
+
+        # Getting the inputs points from the volumes.
+        # Uses the model's method, with the batch_loader's data.
+        # Possibly skipping the last point if not useful.
+        streamlines_f = targets
+        if isinstance(self.model, ModelWithDirectionGetter) and \
+                not self.model.direction_getter.add_eos:
+            # No EOS = We don't use the last coord because it does not have an
+            # associated target direction.
+            streamlines_f = [s[:-1, :] for s in streamlines_f]
+
+        # Possibly add noise to inputs here.
+        logger.debug('*** Computing forward propagation')
+
+        # Now possibly add noise to streamlines (training / valid)
+        streamlines_f = self.batch_loader.add_noise_streamlines_forward(
+            streamlines_f, self.device)
+
+        # Possibly computing directions twice (during forward and loss)
+        # but ok, shouldn't be too heavy. Easier to deal with multiple
+        # projects' requirements by sending whole streamlines rather
+        # than only directions.
+        model_outputs = self.model(streamlines_f)
+        del streamlines_f
+
+        logger.debug('*** Computing loss')
+        targets = self.batch_loader.add_noise_streamlines_loss(
+            targets, self.device)
+
+        results = self.model.compute_loss(model_outputs, targets,
+                                          average_results=True)
+
+        if self.use_gpu:
+            log_gpu_memory_usage(logger)
+
+        # The mean tensor is a single value. Converting to float using item().
+        return results
 
     def fix_parameters(self):
         """