Skip to content

Training Setup

Defines the TrainingSetup class, which is an interface for training setups.

Functions that raise NotImplementedError are meant to be overridden. To create a custom training setup, define a new class in the bfm.training.setups folder that inherits from TrainingSetup and implements the required methods.

TrainingSetup

Interface for training setups.

Must implement the following methods
  • initialize_model: Initializes the model components and sets self.model_components.
  • calculate_pretrain_loss: Calculates the pretraining loss for a batch of data.
  • generate_frozen_features: Generates frozen features for a batch of data.
Source code in bfm/training/training_setup.py 
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
class TrainingSetup:
    '''
    Interface for training setups.

    Must implement the following methods:
        - initialize_model: Initializes the model components and sets self.model_components.
        - calculate_pretrain_loss: Calculates the pretraining loss for a batch of data.
        - generate_frozen_features: Generates frozen features for a batch of data.
    '''
    def __init__(self, all_subjects, config, verbose=True):
        self.config = config
        self.all_subjects = all_subjects
        self.verbose = verbose

        self.model_components = {}

    def initialize_model(self):
        """
            This function initializes the model.

            It must set the self.model_components dictionary to a dictionary of the model components, like
            {"model": model, "electrode_embeddings": electrode_embeddings}, where model and electrode_embeddings are PyTorch modules (those classes must inherit from bfm.model.base)
        """
        raise NotImplementedError("This function is not (yet) implemented for this training setup.")

    def calculate_pretrain_loss(self, batch, output_accuracy=True):
        """
        Calculate the pretraining loss for a batch of data.

        Args:
            batch (dict): Dictionary containing:
                data (torch.Tensor): Shape (batch_size, n_electrodes, n_timesamples)
                electrode_index (torch.Tensor): Shape (batch_size, n_electrodes)
                metadata (dict): Contains subject identifier, trial id, sampling rate, etc.
            output_accuracy (bool): Whether to output accuracy metrics

        Returns:
            dict: Dictionary of losses where keys are loss names and values are loss values.
                 The final loss is the mean of all losses. Accuracies are exempt and used only for logging.
        """
        raise NotImplementedError("This function is not (yet) implemented for this training setup.")

    def generate_frozen_features(self, batch):
        """
        Generate frozen features (meaning, the weights of the model are frozen) for a batch of data. This function is used for the model evaluation on the benchmarks.

        Args:
            batch (dict): Dictionary containing:
                data (torch.Tensor): Shape (batch_size, n_electrodes, n_timesamples)
                electrode_labels (list): List of length 1 (since it's the same across the batch), each element is a list of electrode labels
                metadata (dict): Contains subject identifier, trial id, sampling rate, etc.

        Returns:
            features (torch.Tensor): Shape (batch_size, n_electrodes or n_electrodes+1, n_timebins, *) where * can be arbitrary
                                   if n_electrodes+1, then the first dimension is the cls token
        """
        raise NotImplementedError("This function is not (yet) implemented for this training setup.")

    def _preprocess_laplacian_rereference(self, batch):
        laplacian_rereference_batch(batch, remove_non_laplacian=False, inplace=True)
        return batch
    def _preprocess_normalize_voltage(self, batch):
        batch['data'] = batch['data'] - torch.mean(batch['data'], dim=[0, 2], keepdim=True)
        batch['data'] = batch['data'] / (torch.std(batch['data'], dim=[0, 2], keepdim=True) + 1)
        return batch

    def _preprocess_subset_electrodes(self, batch, output_selected_idx=False, keys=['data'], allowed_remove_idx=None, force_remove_idx=None):
        # Find minimum number of electrodes in batch
        batch_size = batch[keys[0]].shape[0]
        n_electrodes = batch[keys[0]].shape[1]

        if allowed_remove_idx is None:
            allowed_remove_idx = torch.arange(n_electrodes, device=batch[keys[0]].device, dtype=torch.long)
        if force_remove_idx is None:
            force_remove_idx = torch.tensor([], device=batch[keys[0]].device, dtype=torch.long)
        elif not isinstance(force_remove_idx, torch.Tensor):
            force_remove_idx = torch.tensor(force_remove_idx, device=batch[keys[0]].device, dtype=torch.long)

        subset_n_electrodes = min(n_electrodes, self.config['training']['max_n_electrodes']) if self.config['training']['max_n_electrodes']>0 else n_electrodes
        remove_n_electrodes = max(0, n_electrodes - subset_n_electrodes - len(force_remove_idx))

        selected_remove_idx = torch.cat([allowed_remove_idx[torch.randperm(len(allowed_remove_idx))[:remove_n_electrodes]], force_remove_idx])
        selected_idx = torch.arange(n_electrodes, device=batch[keys[0]].device, dtype=torch.long)
        selected_idx = selected_idx[~torch.isin(selected_idx, selected_remove_idx)]

        if 'electrode_labels' in batch:
            batch['electrode_labels'] = [[batch['electrode_labels'][0][i] for i in selected_idx]] * batch_size
        for key in keys:
            batch[key] = batch[key][:, selected_idx]

        if output_selected_idx:
            return batch, selected_idx
        else:
            return batch

    def get_preprocess_functions(self, pretraining=False):
        """
        Get the preprocess functions for the training setup.
        Default: only subset electrodes to the maximum number during pretraining (during eval, pass all electrodes).

        This must be an array of functions, each takes just batch as input and returns the (modified) batch. Modifying it in place is fine (but still return the batch).
        """
        return [self._preprocess_subset_electrodes] if pretraining else []

    def generate_state_dicts(self):
        """
            This function generates the state dicts for the model components. It is used for saving and retrieving the model.
        """
        return {"state_dicts": 
            {
                model_component_name: model_component.state_dict() 
                for model_component_name, model_component in self.model_components.items()
            }
        }

    def save_model(self, epoch, eval_results={}, save_in_dir=RUN_DIR, training_statistics_store=None):
        path = f"{save_in_dir}/{self.config['cluster']['dir_name']}"
        model_path = f"{path}/model_epoch_{epoch}.pth"
        os.makedirs(path, exist_ok=True)
        torch.save({
            'eval_results': eval_results,
            'epoch': epoch,
            'state_dicts': self.generate_state_dicts()['state_dicts'],
            'config': convert_dtypes(self.config),
        }, model_path)
        if training_statistics_store is not None:
            with open(f"{path}/training_statistics.json", 'w') as f:
                json.dump(training_statistics_store, f)

    def load_model(self, epoch, load_from_dir=RUN_DIR):
        model_path = f"{load_from_dir}/{self.config['cluster']['dir_name']}/model_epoch_{epoch}.pth"
        state_dicts = torch.load(model_path, map_location=self.config['device'])
        for model_component_name, model_component in self.model_components.items():
            if "state_dicts" in state_dicts:
                model_component.load_state_dict(state_dicts["state_dicts"][model_component_name])
            else:
                model_component.load_state_dict(state_dicts[model_component_name+"_state_dict"]) # XXX: this is only here for backwards compatibility, can remove soon
            model_component.to(self.config['device'], dtype=self.config['model']['dtype'])

    def model_parameters(self, verbose=False):
        """
            This function returns all of the parameters in the model, and stores the number of parameters in the config.
            It must output a list of all parameters in the model.
        """
        if 'n_params' not in self.config['model']: self.config['model']['n_params'] = {}
        all_params = []
        for model_component_name, model_component in self.model_components.items():
            all_params.extend(list(model_component.parameters()))
            if verbose: log(f"{model_component_name} parameters: {sum(p.numel() for p in model_component.parameters()):,}", priority=0)
            self.config['model']['n_params'][model_component_name] = sum(p.numel() for p in model_component.parameters())
        total_n_params = sum(p.numel() for p in all_params)
        log(f"Total model parameters: {total_n_params:,}", priority=0)
        self.config['model']['n_params']['total'] = total_n_params
        return all_params

    def train_mode(self):
        for model_component in self.model_components.values():
            model_component.train()

    def eval_mode(self):
        for model_component in self.model_components.values():
            model_component.eval()

    def calculate_pretrain_test_loss(self):
        """
        Calculate the pretraining test loss. This function uses the calculate_pretrain_loss function.

        Returns:
            dict: Dictionary of losses where keys are loss names and values are loss values.
                 The final loss is the mean of all losses. Accuracies are exempt and used only for logging.
        """
        losses = {}
        n_batches = 0
        for batch in self.test_dataloader:
            loss = self.calculate_pretrain_loss(batch)

            for key, value in loss.items():
                if key not in losses: losses[key] = 0
                losses[key] += value
            n_batches += 1

        # Handle case where test dataloader is empty
        if n_batches == 0:
            return {}

        return {k: v / n_batches for k, v in losses.items()}

    def load_dataloaders(self):
        """
            This function loads the dataloaders for the training and test sets.

            It must set the self.train_dataloader and self.test_dataloader attributes to the dataloaders (they are used in the pretraining code in pretrain.py)
        """
        config = self.config

        # Step 1: Load datasets
        datasets = []
        for subject_identifier, trial_id in config['training']['train_subject_trials']:
            if self.verbose: log(f"loading dataset for {subject_identifier}_{trial_id}...", indent=1, priority=1)
            datasets.append(
                SubjectTrialDataset(
                    self.all_subjects[subject_identifier], 
                    trial_id, 
                    int(config['model']['context_length'] * self.all_subjects[subject_identifier].get_sampling_rate(trial_id)), 
                    dtype=config['training']['data_dtype'], 
                    output_metadata=True,
                    output_electrode_labels=True
                )
            )
            if self.verbose: log(f"finished loading dataset for {subject_identifier}_{trial_id}", indent=1, priority=1)

        # Step 2: Split into train and test
        train_datasets = []
        test_datasets = []
        for dataset in datasets:
            train_size = int(len(dataset) * (1 - config['training']['p_test']))
            train_dataset, test_dataset = torch.utils.data.random_split(dataset, [train_size, len(dataset) - train_size])
            train_datasets.append(train_dataset)
            test_datasets.append(test_dataset)
        train_dataset = ConcatDataset(train_datasets)
        test_dataset = ConcatDataset(test_datasets)

        # Step 3: Create dataloaders with custom sampler
        num_workers_dataloader_test = max(int(config['cluster']['num_workers_dataloaders'] * 0.15), 1)
        num_workers_dataloader_train = config['cluster']['num_workers_dataloaders'] - num_workers_dataloader_test
        train_dataloader = DataLoader(
            train_dataset,
            batch_sampler=SubjectBatchSampler(
                # length of each dataset = number of windows in the subject
                [len(ds) for ds in train_datasets],
                batch_size=config['training']['batch_size'],
                shuffle=True
            ),
            num_workers=num_workers_dataloader_train,
            pin_memory=True,  # Pin memory for faster GPU transfer
            persistent_workers=True,  # Keep worker processes alive between iterations
            prefetch_factor=config['cluster']['prefetch_factor'],
            collate_fn=PreprocessCollator(preprocess_functions=self.get_preprocess_functions(pretraining=True))
        )
        test_dataloader = DataLoader(
            test_dataset,
            batch_sampler=SubjectBatchSampler(
                # length of each dataset = number of windows in the subject
                [len(ds) for ds in test_datasets],
                batch_size=config['training']['batch_size'],
                shuffle=False
            ),
            num_workers=num_workers_dataloader_test,
            pin_memory=True,
            persistent_workers=True,
            prefetch_factor=config['cluster']['prefetch_factor'],
            collate_fn=PreprocessCollator(preprocess_functions=self.get_preprocess_functions(pretraining=True))
        )

        self.train_dataloader = train_dataloader
        self.test_dataloader = test_dataloader

calculate_pretrain_loss(batch, output_accuracy=True)

Calculate the pretraining loss for a batch of data.

Parameters:

Name Type Description Default
batch dict

Dictionary containing: data (torch.Tensor): Shape (batch_size, n_electrodes, n_timesamples) electrode_index (torch.Tensor): Shape (batch_size, n_electrodes) metadata (dict): Contains subject identifier, trial id, sampling rate, etc.

 required
output_accuracy bool

Whether to output accuracy metrics

 True

Returns:

Name Type Description
dict

Dictionary of losses where keys are loss names and values are loss values. The final loss is the mean of all losses. Accuracies are exempt and used only for logging.
Source code in bfm/training/training_setup.py 
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
def calculate_pretrain_loss(self, batch, output_accuracy=True):
    """
    Calculate the pretraining loss for a batch of data.

    Args:
        batch (dict): Dictionary containing:
            data (torch.Tensor): Shape (batch_size, n_electrodes, n_timesamples)
            electrode_index (torch.Tensor): Shape (batch_size, n_electrodes)
            metadata (dict): Contains subject identifier, trial id, sampling rate, etc.
        output_accuracy (bool): Whether to output accuracy metrics

    Returns:
        dict: Dictionary of losses where keys are loss names and values are loss values.
             The final loss is the mean of all losses. Accuracies are exempt and used only for logging.
    """
    raise NotImplementedError("This function is not (yet) implemented for this training setup.")

calculate_pretrain_test_loss()

Calculate the pretraining test loss. This function uses the calculate_pretrain_loss function.

Returns:

Name Type Description
dict

Dictionary of losses where keys are loss names and values are loss values. The final loss is the mean of all losses. Accuracies are exempt and used only for logging.
Source code in bfm/training/training_setup.py 
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
def calculate_pretrain_test_loss(self):
    """
    Calculate the pretraining test loss. This function uses the calculate_pretrain_loss function.

    Returns:
        dict: Dictionary of losses where keys are loss names and values are loss values.
             The final loss is the mean of all losses. Accuracies are exempt and used only for logging.
    """
    losses = {}
    n_batches = 0
    for batch in self.test_dataloader:
        loss = self.calculate_pretrain_loss(batch)

        for key, value in loss.items():
            if key not in losses: losses[key] = 0
            losses[key] += value
        n_batches += 1

    # Handle case where test dataloader is empty
    if n_batches == 0:
        return {}

    return {k: v / n_batches for k, v in losses.items()}

generate_frozen_features(batch)

Generate frozen features (meaning, the weights of the model are frozen) for a batch of data. This function is used for the model evaluation on the benchmarks.

Parameters:

Name Type Description Default
batch dict

Dictionary containing: data (torch.Tensor): Shape (batch_size, n_electrodes, n_timesamples) electrode_labels (list): List of length 1 (since it's the same across the batch), each element is a list of electrode labels metadata (dict): Contains subject identifier, trial id, sampling rate, etc.

 required

Returns:

Name Type Description
features Tensor

Shape (batch_size, n_electrodes or n_electrodes+1, n_timebins, *) where * can be arbitrary if n_electrodes+1, then the first dimension is the cls token
Source code in bfm/training/training_setup.py 
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
def generate_frozen_features(self, batch):
    """
    Generate frozen features (meaning, the weights of the model are frozen) for a batch of data. This function is used for the model evaluation on the benchmarks.

    Args:
        batch (dict): Dictionary containing:
            data (torch.Tensor): Shape (batch_size, n_electrodes, n_timesamples)
            electrode_labels (list): List of length 1 (since it's the same across the batch), each element is a list of electrode labels
            metadata (dict): Contains subject identifier, trial id, sampling rate, etc.

    Returns:
        features (torch.Tensor): Shape (batch_size, n_electrodes or n_electrodes+1, n_timebins, *) where * can be arbitrary
                               if n_electrodes+1, then the first dimension is the cls token
    """
    raise NotImplementedError("This function is not (yet) implemented for this training setup.")

generate_state_dicts()

This function generates the state dicts for the model components. It is used for saving and retrieving the model.

Source code in bfm/training/training_setup.py 
131
132
133
134
135
136
137
138
139
140
def generate_state_dicts(self):
    """
        This function generates the state dicts for the model components. It is used for saving and retrieving the model.
    """
    return {"state_dicts": 
        {
            model_component_name: model_component.state_dict() 
            for model_component_name, model_component in self.model_components.items()
        }
    }

get_preprocess_functions(pretraining=False)

Get the preprocess functions for the training setup. Default: only subset electrodes to the maximum number during pretraining (during eval, pass all electrodes).

This must be an array of functions, each takes just batch as input and returns the (modified) batch. Modifying it in place is fine (but still return the batch).

Source code in bfm/training/training_setup.py 
122
123
124
125
126
127
128
129
def get_preprocess_functions(self, pretraining=False):
    """
    Get the preprocess functions for the training setup.
    Default: only subset electrodes to the maximum number during pretraining (during eval, pass all electrodes).

    This must be an array of functions, each takes just batch as input and returns the (modified) batch. Modifying it in place is fine (but still return the batch).
    """
    return [self._preprocess_subset_electrodes] if pretraining else []

initialize_model()

This function initializes the model.

It must set the self.model_components dictionary to a dictionary of the model components, like {"model": model, "electrode_embeddings": electrode_embeddings}, where model and electrode_embeddings are PyTorch modules (those classes must inherit from bfm.model.base)

Source code in bfm/training/training_setup.py 
43
44
45
46
47
48
49
50
def initialize_model(self):
    """
        This function initializes the model.

        It must set the self.model_components dictionary to a dictionary of the model components, like
        {"model": model, "electrode_embeddings": electrode_embeddings}, where model and electrode_embeddings are PyTorch modules (those classes must inherit from bfm.model.base)
    """
    raise NotImplementedError("This function is not (yet) implemented for this training setup.")

load_dataloaders()

This function loads the dataloaders for the training and test sets.

It must set the self.train_dataloader and self.test_dataloader attributes to the dataloaders (they are used in the pretraining code in pretrain.py)

Source code in bfm/training/training_setup.py 
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
def load_dataloaders(self):
    """
        This function loads the dataloaders for the training and test sets.

        It must set the self.train_dataloader and self.test_dataloader attributes to the dataloaders (they are used in the pretraining code in pretrain.py)
    """
    config = self.config

    # Step 1: Load datasets
    datasets = []
    for subject_identifier, trial_id in config['training']['train_subject_trials']:
        if self.verbose: log(f"loading dataset for {subject_identifier}_{trial_id}...", indent=1, priority=1)
        datasets.append(
            SubjectTrialDataset(
                self.all_subjects[subject_identifier], 
                trial_id, 
                int(config['model']['context_length'] * self.all_subjects[subject_identifier].get_sampling_rate(trial_id)), 
                dtype=config['training']['data_dtype'], 
                output_metadata=True,
                output_electrode_labels=True
            )
        )
        if self.verbose: log(f"finished loading dataset for {subject_identifier}_{trial_id}", indent=1, priority=1)

    # Step 2: Split into train and test
    train_datasets = []
    test_datasets = []
    for dataset in datasets:
        train_size = int(len(dataset) * (1 - config['training']['p_test']))
        train_dataset, test_dataset = torch.utils.data.random_split(dataset, [train_size, len(dataset) - train_size])
        train_datasets.append(train_dataset)
        test_datasets.append(test_dataset)
    train_dataset = ConcatDataset(train_datasets)
    test_dataset = ConcatDataset(test_datasets)

    # Step 3: Create dataloaders with custom sampler
    num_workers_dataloader_test = max(int(config['cluster']['num_workers_dataloaders'] * 0.15), 1)
    num_workers_dataloader_train = config['cluster']['num_workers_dataloaders'] - num_workers_dataloader_test
    train_dataloader = DataLoader(
        train_dataset,
        batch_sampler=SubjectBatchSampler(
            # length of each dataset = number of windows in the subject
            [len(ds) for ds in train_datasets],
            batch_size=config['training']['batch_size'],
            shuffle=True
        ),
        num_workers=num_workers_dataloader_train,
        pin_memory=True,  # Pin memory for faster GPU transfer
        persistent_workers=True,  # Keep worker processes alive between iterations
        prefetch_factor=config['cluster']['prefetch_factor'],
        collate_fn=PreprocessCollator(preprocess_functions=self.get_preprocess_functions(pretraining=True))
    )
    test_dataloader = DataLoader(
        test_dataset,
        batch_sampler=SubjectBatchSampler(
            # length of each dataset = number of windows in the subject
            [len(ds) for ds in test_datasets],
            batch_size=config['training']['batch_size'],
            shuffle=False
        ),
        num_workers=num_workers_dataloader_test,
        pin_memory=True,
        persistent_workers=True,
        prefetch_factor=config['cluster']['prefetch_factor'],
        collate_fn=PreprocessCollator(preprocess_functions=self.get_preprocess_functions(pretraining=True))
    )

    self.train_dataloader = train_dataloader
    self.test_dataloader = test_dataloader

model_parameters(verbose=False)

This function returns all of the parameters in the model, and stores the number of parameters in the config. It must output a list of all parameters in the model.

Source code in bfm/training/training_setup.py 
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
def model_parameters(self, verbose=False):
    """
        This function returns all of the parameters in the model, and stores the number of parameters in the config.
        It must output a list of all parameters in the model.
    """
    if 'n_params' not in self.config['model']: self.config['model']['n_params'] = {}
    all_params = []
    for model_component_name, model_component in self.model_components.items():
        all_params.extend(list(model_component.parameters()))
        if verbose: log(f"{model_component_name} parameters: {sum(p.numel() for p in model_component.parameters()):,}", priority=0)
        self.config['model']['n_params'][model_component_name] = sum(p.numel() for p in model_component.parameters())
    total_n_params = sum(p.numel() for p in all_params)
    log(f"Total model parameters: {total_n_params:,}", priority=0)
    self.config['model']['n_params']['total'] = total_n_params
    return all_params