Skip to content

dualagg

ContrastiveLoss

Bases: Module

Contrastive loss function.

Source code in nebula/core/models/cifar10/dualagg.py
25
26
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
class ContrastiveLoss(torch.nn.Module):
    """
    Contrastive loss function.
    """

    def __init__(self, mu=0.5):
        super().__init__()
        self.mu = mu
        self.cross_entropy_loss = torch.nn.CrossEntropyLoss()

    def forward(self, local_out, global_out, historical_out, labels):
        """
        Calculates the contrastive loss between the local output, global output, and historical output.

        Args:
            local_out (torch.Tensor): The local output tensor of shape (batch_size, embedding_size).
            global_out (torch.Tensor): The global output tensor of shape (batch_size, embedding_size).
            historical_out (torch.Tensor): The historical output tensor of shape (batch_size, embedding_size).
            labels (torch.Tensor): The ground truth labels tensor of shape (batch_size,).

        Returns:
            torch.Tensor: The contrastive loss value.

        Raises:
            ValueError: If the input tensors have different shapes.

        Notes:
            - The contrastive loss is calculated as the difference between the mean cosine similarity of the local output
                with the historical output and the mean cosine similarity of the local output with the global output,
                multiplied by a scaling factor mu.
            - The cosine similarity values represent the similarity between the corresponding vectors in the input tensors.
            Higher values indicate greater similarity, while lower values indicate less similarity.
        """
        if local_out.shape != global_out.shape or local_out.shape != historical_out.shape:
            raise ValueError("Input tensors must have the same shape.")

        # Cross-entropy loss
        ce_loss = self.cross_entropy_loss(local_out, labels)
        # if round > 1:
        # Positive cosine similarity
        pos_cos_sim = F.cosine_similarity(local_out, historical_out, dim=1).mean()

        # Negative cosine similarity
        neg_cos_sim = -F.cosine_similarity(local_out, global_out, dim=1).mean()

        # Combined loss
        contrastive_loss = ce_loss + self.mu * 0.5 * (pos_cos_sim + neg_cos_sim)

        logging_training.debug(
            f"Contrastive loss (mu={self.mu}) with 0.5 of factor: ce_loss: {ce_loss}, pos_cos_sim_local_historical: {pos_cos_sim}, neg_cos_sim_local_global: {neg_cos_sim}, loss: {contrastive_loss}"
        )
        return contrastive_loss

forward(local_out, global_out, historical_out, labels)

Calculates the contrastive loss between the local output, global output, and historical output.

Parameters:

Name Type Description Default
local_out Tensor

The local output tensor of shape (batch_size, embedding_size).

required
global_out Tensor

The global output tensor of shape (batch_size, embedding_size).

required
historical_out Tensor

The historical output tensor of shape (batch_size, embedding_size).

required
labels Tensor

The ground truth labels tensor of shape (batch_size,).

required

Returns:

Type Description

torch.Tensor: The contrastive loss value.

Raises:

Type Description
ValueError

If the input tensors have different shapes.

Notes
  • The contrastive loss is calculated as the difference between the mean cosine similarity of the local output with the historical output and the mean cosine similarity of the local output with the global output, multiplied by a scaling factor mu.
  • The cosine similarity values represent the similarity between the corresponding vectors in the input tensors. Higher values indicate greater similarity, while lower values indicate less similarity.
Source code in nebula/core/models/cifar10/dualagg.py
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
def forward(self, local_out, global_out, historical_out, labels):
    """
    Calculates the contrastive loss between the local output, global output, and historical output.

    Args:
        local_out (torch.Tensor): The local output tensor of shape (batch_size, embedding_size).
        global_out (torch.Tensor): The global output tensor of shape (batch_size, embedding_size).
        historical_out (torch.Tensor): The historical output tensor of shape (batch_size, embedding_size).
        labels (torch.Tensor): The ground truth labels tensor of shape (batch_size,).

    Returns:
        torch.Tensor: The contrastive loss value.

    Raises:
        ValueError: If the input tensors have different shapes.

    Notes:
        - The contrastive loss is calculated as the difference between the mean cosine similarity of the local output
            with the historical output and the mean cosine similarity of the local output with the global output,
            multiplied by a scaling factor mu.
        - The cosine similarity values represent the similarity between the corresponding vectors in the input tensors.
        Higher values indicate greater similarity, while lower values indicate less similarity.
    """
    if local_out.shape != global_out.shape or local_out.shape != historical_out.shape:
        raise ValueError("Input tensors must have the same shape.")

    # Cross-entropy loss
    ce_loss = self.cross_entropy_loss(local_out, labels)
    # if round > 1:
    # Positive cosine similarity
    pos_cos_sim = F.cosine_similarity(local_out, historical_out, dim=1).mean()

    # Negative cosine similarity
    neg_cos_sim = -F.cosine_similarity(local_out, global_out, dim=1).mean()

    # Combined loss
    contrastive_loss = ce_loss + self.mu * 0.5 * (pos_cos_sim + neg_cos_sim)

    logging_training.debug(
        f"Contrastive loss (mu={self.mu}) with 0.5 of factor: ce_loss: {ce_loss}, pos_cos_sim_local_historical: {pos_cos_sim}, neg_cos_sim_local_global: {neg_cos_sim}, loss: {contrastive_loss}"
    )
    return contrastive_loss

DualAggModel

Bases: LightningModule

Source code in nebula/core/models/cifar10/dualagg.py
 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
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
class DualAggModel(pl.LightningModule):
    def process_metrics(self, phase, y_pred, y, loss=None, mode="local"):
        """
        Calculate and log metrics for the given phase.
        Args:
            phase (str): One of 'Train', 'Validation', or 'Test'
            y_pred (torch.Tensor): Model predictions
            y (torch.Tensor): Ground truth labels
            loss (torch.Tensor, optional): Loss value
        """

        y_pred_classes = torch.argmax(y_pred, dim=1)
        if phase == "Train":
            if mode == "local":
                output = self.local_train_metrics(y_pred_classes, y)
            elif mode == "historical":
                output = self.historical_train_metrics(y_pred_classes, y)
            elif mode == "global":
                output = self.global_train_metrics(y_pred_classes, y)
        elif phase == "Validation":
            if mode == "local":
                output = self.local_val_metrics(y_pred_classes, y)
            elif mode == "historical":
                output = self.historical_val_metrics(y_pred_classes, y)
            elif mode == "global":
                output = self.global_val_metrics(y_pred_classes, y)
        elif phase == "Test":
            if mode == "local":
                output = self.local_test_metrics(y_pred_classes, y)
            elif mode == "historical":
                output = self.historical_test_metrics(y_pred_classes, y)
            elif mode == "global":
                output = self.global_test_metrics(y_pred_classes, y)
        else:
            raise NotImplementedError
        # print(f"y_pred shape: {y_pred.shape}, y_pred_classes shape: {y_pred_classes.shape}, y shape: {y.shape}")  # Debug print
        output = {
            f"{mode}/{phase}/{key.replace('Multiclass', '').split('/')[-1]}": value for key, value in output.items()
        }
        self.log_dict(output, prog_bar=True, logger=True)

        if self.local_cm is not None and self.historical_cm is not None and self.global_cm is not None:
            if mode == "local":
                self.local_cm.update(y_pred_classes, y)
            elif mode == "historical":
                self.historical_cm.update(y_pred_classes, y)
            elif mode == "global":
                self.global_cm.update(y_pred_classes, y)

    def log_metrics_by_epoch(self, phase, print_cm=False, plot_cm=False, mode="local"):
        """
        Log all metrics at the end of an epoch for the given phase.
        Args:
            phase (str): One of 'Train', 'Validation', or 'Test'
            :param phase:
            :param plot_cm:
        """
        if mode == "local":
            print(f"Epoch end: {mode} {phase}, epoch number: {self.local_epoch_global_number[phase]}")
        elif mode == "historical":
            print(f"Epoch end: {mode} {phase}, epoch number: {self.historical_epoch_global_number[phase]}")
        elif mode == "global":
            print(f"Epoch end: {mode} {phase}, epoch number: {self.global_epoch_global_number[phase]}")

        if phase == "Train":
            if mode == "local":
                output = self.local_train_metrics.compute()
                self.local_train_metrics.reset()
            elif mode == "historical":
                output = self.historical_train_metrics.compute()
                self.historical_train_metrics.reset()
            elif mode == "global":
                output = self.global_train_metrics.compute()
                self.global_train_metrics.reset()
        elif phase == "Validation":
            if mode == "local":
                output = self.local_val_metrics.compute()
                self.local_val_metrics.reset()
            elif mode == "historical":
                output = self.historical_val_metrics.compute()
                self.historical_val_metrics.reset()
            elif mode == "global":
                output = self.global_val_metrics.compute()
                self.global_val_metrics.reset()
        elif phase == "Test":
            if mode == "local":
                output = self.local_test_metrics.compute()
                self.local_test_metrics.reset()
            elif mode == "historical":
                output = self.historical_test_metrics.compute()
                self.historical_test_metrics.reset()
            elif mode == "global":
                output = self.global_test_metrics.compute()
                self.global_test_metrics.reset()
        else:
            raise NotImplementedError

        output = {
            f"{mode}/{phase}Epoch/{key.replace('Multiclass', '').split('/')[-1]}": value
            for key, value in output.items()
        }

        self.log_dict(output, prog_bar=True, logger=True)

        if self.local_cm is not None and self.historical_cm is not None and self.global_cm is not None:
            if mode == "local":
                cm = self.local_cm.compute().cpu()
            elif mode == "historical":
                cm = self.historical_cm.compute().cpu()
            elif mode == "global":
                cm = self.global_cm.compute().cpu()
            print(f"{mode}/{phase}Epoch/CM\n", cm) if print_cm else None
            if plot_cm:
                plt.figure(figsize=(10, 7))
                ax = sns.heatmap(cm.numpy(), annot=True, fmt="d", cmap="Blues")
                ax.set_xlabel("Predicted labels")
                ax.set_ylabel("True labels")
                ax.set_title("Confusion Matrix")
                ax.set_xticks(range(self.num_classes))
                ax.set_yticks(range(self.num_classes))
                ax.xaxis.set_ticklabels([i for i in range(self.num_classes)])
                ax.yaxis.set_ticklabels([i for i in range(self.num_classes)])
                if mode == "local":
                    self.logger.experiment.add_figure(
                        f"{mode}/{phase}Epoch/CM",
                        ax.get_figure(),
                        global_step=self.local_epoch_global_number[phase],
                    )
                elif mode == "historical":
                    self.logger.experiment.add_figure(
                        f"{mode}/{phase}Epoch/CM",
                        ax.get_figure(),
                        global_step=self.historical_epoch_global_number[phase],
                    )
                elif mode == "global":
                    self.logger.experiment.add_figure(
                        f"{mode}/{phase}Epoch/CM",
                        ax.get_figure(),
                        global_step=self.global_epoch_global_number[phase],
                    )
                plt.close()

        if mode == "local":
            self.local_epoch_global_number[phase] += 1
        elif mode == "historical":
            self.historical_epoch_global_number[phase] += 1
        elif mode == "global":
            self.global_epoch_global_number[phase] += 1

    def __init__(
        self,
        input_channels=3,
        num_classes=10,
        learning_rate=1e-3,
        mu=0.5,
        metrics=None,
        confusion_matrix=None,
        seed=None,
    ):
        super().__init__()

        self.input_channels = input_channels
        self.num_classes = num_classes
        self.learning_rate = learning_rate
        self.mu = mu

        if metrics is None:
            metrics = MetricCollection([
                MulticlassAccuracy(num_classes=num_classes),
                MulticlassPrecision(num_classes=num_classes),
                MulticlassRecall(num_classes=num_classes),
                MulticlassF1Score(num_classes=num_classes),
            ])

        # Define metrics
        self.local_train_metrics = metrics.clone(prefix="Local/Train/")
        self.local_val_metrics = metrics.clone(prefix="Local/Validation/")
        self.local_test_metrics = metrics.clone(prefix="Local/Test/")

        self.historical_train_metrics = metrics.clone(prefix="Historical/Train/")
        self.historical_val_metrics = metrics.clone(prefix="Historical/Validation/")
        self.historical_test_metrics = metrics.clone(prefix="Historical/Test/")

        self.global_train_metrics = metrics.clone(prefix="Global/Train/")
        self.global_val_metrics = metrics.clone(prefix="Global/Validation/")
        self.global_test_metrics = metrics.clone(prefix="Global/Test/")

        if confusion_matrix is None:
            self.local_cm = MulticlassConfusionMatrix(num_classes=num_classes)
            self.historical_cm = MulticlassConfusionMatrix(num_classes=num_classes)
            self.global_cm = MulticlassConfusionMatrix(num_classes=num_classes)

        # Set seed for reproducibility initialization
        if seed is not None:
            torch.manual_seed(seed)
            torch.cuda.manual_seed_all(seed)

        self.local_epoch_global_number = {"Train": 0, "Validation": 0, "Test": 0}
        self.historical_epoch_global_number = {"Train": 0, "Validation": 0, "Test": 0}
        self.global_epoch_global_number = {"Train": 0, "Validation": 0, "Test": 0}

        self.config = {"beta1": 0.851436, "beta2": 0.999689, "amsgrad": True}

        self.example_input_array = torch.rand(1, 3, 32, 32)
        self.learning_rate = learning_rate
        self.criterion = ContrastiveLoss(mu=self.mu)

        # Define layers of the model
        self.model = torch.nn.Sequential(
            torch.nn.Conv2d(input_channels, 16, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Conv2d(16, 32, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Conv2d(32, 64, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Flatten(),
            torch.nn.Linear(64 * 4 * 4, 512),
            torch.nn.ReLU(),
            torch.nn.Linear(512, num_classes),
        )

        # Siamese models of the model above
        self.historical_model = torch.nn.Sequential(
            torch.nn.Conv2d(input_channels, 16, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Conv2d(16, 32, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Conv2d(32, 64, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Flatten(),
            torch.nn.Linear(64 * 4 * 4, 512),
            torch.nn.ReLU(),
            torch.nn.Linear(512, num_classes),
        )
        self.global_model = torch.nn.Sequential(
            torch.nn.Conv2d(input_channels, 16, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Conv2d(16, 32, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Conv2d(32, 64, 3, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
            torch.nn.Flatten(),
            torch.nn.Linear(64 * 4 * 4, 512),
            torch.nn.ReLU(),
            torch.nn.Linear(512, num_classes),
        )
        # self.historical_model = copy.deepcopy(self.model)
        # self.global_model = copy.deepcopy(self.model)

    def forward(self, x, mode="local"):
        """Forward pass of the model."""
        if mode == "local":
            return self.model(x)
        elif mode == "global":
            return self.global_model(x)
        elif mode == "historical":
            return self.historical_model(x)
        else:
            raise NotImplementedError

    def configure_optimizers(self):
        """ """
        optimizer = torch.optim.Adam(
            self.parameters(),
            lr=self.learning_rate,
            betas=(self.config["beta1"], self.config["beta2"]),
            amsgrad=self.config["amsgrad"],
        )
        return optimizer

    def step(self, batch, batch_idx, phase):
        images, labels = batch
        images = images.to(self.device)
        labels = labels.to(self.device)
        local_out = self.forward(images, mode="local")
        with torch.no_grad():
            historical_out = self.forward(images, mode="historical")
            global_out = self.forward(images, mode="global")

        loss = self.criterion(local_out, global_out, historical_out, labels)

        # Get metrics for each batch and log them
        self.log(f"{phase}/ConstrastiveLoss", loss, prog_bar=True, logger=True)  # Constrastive loss
        self.process_metrics(phase, local_out, labels, loss, mode="local")
        self.process_metrics(phase, historical_out, labels, loss, mode="historical")
        self.process_metrics(phase, global_out, labels, loss, mode="global")

        return loss

    def training_step(self, batch, batch_id):
        """
        Training step for the model.
        Args:
            batch:
            batch_id:

        Returns:
        """
        return self.step(batch, batch_id, "Train")

    def on_train_epoch_end(self):
        self.log_metrics_by_epoch("Train", print_cm=True, plot_cm=True, mode="local")
        self.log_metrics_by_epoch("Train", print_cm=True, plot_cm=True, mode="historical")
        self.log_metrics_by_epoch("Train", print_cm=True, plot_cm=True, mode="global")

    def validation_step(self, batch, batch_idx):
        """
        Validation step for the model.
        Args:
            batch:
            batch_idx:

        Returns:
        """
        return self.step(batch, batch_idx, "Validation")

    def on_validation_epoch_end(self):
        self.log_metrics_by_epoch("Validation", print_cm=True, plot_cm=False, mode="local")
        self.log_metrics_by_epoch("Validation", print_cm=True, plot_cm=False, mode="historical")
        self.log_metrics_by_epoch("Validation", print_cm=True, plot_cm=False, mode="global")

    def test_step(self, batch, batch_idx):
        """
        Test step for the model.
        Args:
            batch:
            batch_idx:

        Returns:
        """
        return self.step(batch, batch_idx, "Test")

    def on_test_epoch_end(self):
        self.log_metrics_by_epoch("Test", print_cm=True, plot_cm=True, mode="local")
        self.log_metrics_by_epoch("Test", print_cm=True, plot_cm=True, mode="historical")
        self.log_metrics_by_epoch("Test", print_cm=True, plot_cm=True, mode="global")

    def save_historical_model(self):
        """
        Save the current local model as the historical model.
        """
        logging_training.info("Copying local model to historical model.")
        self.historical_model.load_state_dict(self.model.state_dict())

    def global_load_state_dict(self, state_dict):
        """
        Load the given state dictionary into the global model.
        Args:
            state_dict (dict): The state dictionary to load into the global model.
        """
        logging_training.info("Loading state dict into global model.")
        adapted_state_dict = self.adapt_state_dict_for_model(state_dict, "model")
        self.global_model.load_state_dict(adapted_state_dict)

    def historical_load_state_dict(self, state_dict):
        """
        Load the given state dictionary into the historical model.
        Args:
            state_dict (dict): The state dictionary to load into the historical model.
        """
        logging_training.info("Loading state dict into historical model.")
        adapted_state_dict = self.adapt_state_dict_for_model(state_dict, "model")
        self.historical_model.load_state_dict(adapted_state_dict)

    def adapt_state_dict_for_model(self, state_dict, model_prefix):
        """
        Adapt the keys in the provided state_dict to match the structure expected by the model.
        """
        new_state_dict = {}
        prefix = f"{model_prefix}."
        for key, value in state_dict.items():
            if key.startswith(prefix):
                # Remove the specific prefix from each key
                new_key = key[len(prefix) :]
                new_state_dict[new_key] = value
        return new_state_dict

    def get_global_model_parameters(self):
        """
        Get the parameters of the global model.
        """
        return self.global_model.state_dict()

    def print_summary(self):
        """
        Print a summary of local, historical and global models to check if they are the same.
        """
        logging_training.info("Local model summary:")
        logging_training.info(self.model)
        logging_training.info("Historical model summary:")
        logging_training.info(self.historical_model)
        logging_training.info("Global model summary:")
        logging_training.info(self.global_model)

adapt_state_dict_for_model(state_dict, model_prefix)

Adapt the keys in the provided state_dict to match the structure expected by the model.

Source code in nebula/core/models/cifar10/dualagg.py
455
456
457
458
459
460
461
462
463
464
465
466
def adapt_state_dict_for_model(self, state_dict, model_prefix):
    """
    Adapt the keys in the provided state_dict to match the structure expected by the model.
    """
    new_state_dict = {}
    prefix = f"{model_prefix}."
    for key, value in state_dict.items():
        if key.startswith(prefix):
            # Remove the specific prefix from each key
            new_key = key[len(prefix) :]
            new_state_dict[new_key] = value
    return new_state_dict

configure_optimizers()

Source code in nebula/core/models/cifar10/dualagg.py
351
352
353
354
355
356
357
358
359
def configure_optimizers(self):
    """ """
    optimizer = torch.optim.Adam(
        self.parameters(),
        lr=self.learning_rate,
        betas=(self.config["beta1"], self.config["beta2"]),
        amsgrad=self.config["amsgrad"],
    )
    return optimizer

forward(x, mode='local')

Forward pass of the model.

Source code in nebula/core/models/cifar10/dualagg.py
340
341
342
343
344
345
346
347
348
349
def forward(self, x, mode="local"):
    """Forward pass of the model."""
    if mode == "local":
        return self.model(x)
    elif mode == "global":
        return self.global_model(x)
    elif mode == "historical":
        return self.historical_model(x)
    else:
        raise NotImplementedError

get_global_model_parameters()

Get the parameters of the global model.

Source code in nebula/core/models/cifar10/dualagg.py
468
469
470
471
472
def get_global_model_parameters(self):
    """
    Get the parameters of the global model.
    """
    return self.global_model.state_dict()

global_load_state_dict(state_dict)

Load the given state dictionary into the global model. Args: state_dict (dict): The state dictionary to load into the global model.

Source code in nebula/core/models/cifar10/dualagg.py
435
436
437
438
439
440
441
442
443
def global_load_state_dict(self, state_dict):
    """
    Load the given state dictionary into the global model.
    Args:
        state_dict (dict): The state dictionary to load into the global model.
    """
    logging_training.info("Loading state dict into global model.")
    adapted_state_dict = self.adapt_state_dict_for_model(state_dict, "model")
    self.global_model.load_state_dict(adapted_state_dict)

historical_load_state_dict(state_dict)

Load the given state dictionary into the historical model. Args: state_dict (dict): The state dictionary to load into the historical model.

Source code in nebula/core/models/cifar10/dualagg.py
445
446
447
448
449
450
451
452
453
def historical_load_state_dict(self, state_dict):
    """
    Load the given state dictionary into the historical model.
    Args:
        state_dict (dict): The state dictionary to load into the historical model.
    """
    logging_training.info("Loading state dict into historical model.")
    adapted_state_dict = self.adapt_state_dict_for_model(state_dict, "model")
    self.historical_model.load_state_dict(adapted_state_dict)

log_metrics_by_epoch(phase, print_cm=False, plot_cm=False, mode='local')

Log all metrics at the end of an epoch for the given phase. Args: phase (str): One of 'Train', 'Validation', or 'Test' :param phase: :param plot_cm:

Source code in nebula/core/models/cifar10/dualagg.py
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
def log_metrics_by_epoch(self, phase, print_cm=False, plot_cm=False, mode="local"):
    """
    Log all metrics at the end of an epoch for the given phase.
    Args:
        phase (str): One of 'Train', 'Validation', or 'Test'
        :param phase:
        :param plot_cm:
    """
    if mode == "local":
        print(f"Epoch end: {mode} {phase}, epoch number: {self.local_epoch_global_number[phase]}")
    elif mode == "historical":
        print(f"Epoch end: {mode} {phase}, epoch number: {self.historical_epoch_global_number[phase]}")
    elif mode == "global":
        print(f"Epoch end: {mode} {phase}, epoch number: {self.global_epoch_global_number[phase]}")

    if phase == "Train":
        if mode == "local":
            output = self.local_train_metrics.compute()
            self.local_train_metrics.reset()
        elif mode == "historical":
            output = self.historical_train_metrics.compute()
            self.historical_train_metrics.reset()
        elif mode == "global":
            output = self.global_train_metrics.compute()
            self.global_train_metrics.reset()
    elif phase == "Validation":
        if mode == "local":
            output = self.local_val_metrics.compute()
            self.local_val_metrics.reset()
        elif mode == "historical":
            output = self.historical_val_metrics.compute()
            self.historical_val_metrics.reset()
        elif mode == "global":
            output = self.global_val_metrics.compute()
            self.global_val_metrics.reset()
    elif phase == "Test":
        if mode == "local":
            output = self.local_test_metrics.compute()
            self.local_test_metrics.reset()
        elif mode == "historical":
            output = self.historical_test_metrics.compute()
            self.historical_test_metrics.reset()
        elif mode == "global":
            output = self.global_test_metrics.compute()
            self.global_test_metrics.reset()
    else:
        raise NotImplementedError

    output = {
        f"{mode}/{phase}Epoch/{key.replace('Multiclass', '').split('/')[-1]}": value
        for key, value in output.items()
    }

    self.log_dict(output, prog_bar=True, logger=True)

    if self.local_cm is not None and self.historical_cm is not None and self.global_cm is not None:
        if mode == "local":
            cm = self.local_cm.compute().cpu()
        elif mode == "historical":
            cm = self.historical_cm.compute().cpu()
        elif mode == "global":
            cm = self.global_cm.compute().cpu()
        print(f"{mode}/{phase}Epoch/CM\n", cm) if print_cm else None
        if plot_cm:
            plt.figure(figsize=(10, 7))
            ax = sns.heatmap(cm.numpy(), annot=True, fmt="d", cmap="Blues")
            ax.set_xlabel("Predicted labels")
            ax.set_ylabel("True labels")
            ax.set_title("Confusion Matrix")
            ax.set_xticks(range(self.num_classes))
            ax.set_yticks(range(self.num_classes))
            ax.xaxis.set_ticklabels([i for i in range(self.num_classes)])
            ax.yaxis.set_ticklabels([i for i in range(self.num_classes)])
            if mode == "local":
                self.logger.experiment.add_figure(
                    f"{mode}/{phase}Epoch/CM",
                    ax.get_figure(),
                    global_step=self.local_epoch_global_number[phase],
                )
            elif mode == "historical":
                self.logger.experiment.add_figure(
                    f"{mode}/{phase}Epoch/CM",
                    ax.get_figure(),
                    global_step=self.historical_epoch_global_number[phase],
                )
            elif mode == "global":
                self.logger.experiment.add_figure(
                    f"{mode}/{phase}Epoch/CM",
                    ax.get_figure(),
                    global_step=self.global_epoch_global_number[phase],
                )
            plt.close()

    if mode == "local":
        self.local_epoch_global_number[phase] += 1
    elif mode == "historical":
        self.historical_epoch_global_number[phase] += 1
    elif mode == "global":
        self.global_epoch_global_number[phase] += 1

print_summary()

Print a summary of local, historical and global models to check if they are the same.

Source code in nebula/core/models/cifar10/dualagg.py
474
475
476
477
478
479
480
481
482
483
def print_summary(self):
    """
    Print a summary of local, historical and global models to check if they are the same.
    """
    logging_training.info("Local model summary:")
    logging_training.info(self.model)
    logging_training.info("Historical model summary:")
    logging_training.info(self.historical_model)
    logging_training.info("Global model summary:")
    logging_training.info(self.global_model)

process_metrics(phase, y_pred, y, loss=None, mode='local')

Calculate and log metrics for the given phase. Args: phase (str): One of 'Train', 'Validation', or 'Test' y_pred (torch.Tensor): Model predictions y (torch.Tensor): Ground truth labels loss (torch.Tensor, optional): Loss value

Source code in nebula/core/models/cifar10/dualagg.py
 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
def process_metrics(self, phase, y_pred, y, loss=None, mode="local"):
    """
    Calculate and log metrics for the given phase.
    Args:
        phase (str): One of 'Train', 'Validation', or 'Test'
        y_pred (torch.Tensor): Model predictions
        y (torch.Tensor): Ground truth labels
        loss (torch.Tensor, optional): Loss value
    """

    y_pred_classes = torch.argmax(y_pred, dim=1)
    if phase == "Train":
        if mode == "local":
            output = self.local_train_metrics(y_pred_classes, y)
        elif mode == "historical":
            output = self.historical_train_metrics(y_pred_classes, y)
        elif mode == "global":
            output = self.global_train_metrics(y_pred_classes, y)
    elif phase == "Validation":
        if mode == "local":
            output = self.local_val_metrics(y_pred_classes, y)
        elif mode == "historical":
            output = self.historical_val_metrics(y_pred_classes, y)
        elif mode == "global":
            output = self.global_val_metrics(y_pred_classes, y)
    elif phase == "Test":
        if mode == "local":
            output = self.local_test_metrics(y_pred_classes, y)
        elif mode == "historical":
            output = self.historical_test_metrics(y_pred_classes, y)
        elif mode == "global":
            output = self.global_test_metrics(y_pred_classes, y)
    else:
        raise NotImplementedError
    # print(f"y_pred shape: {y_pred.shape}, y_pred_classes shape: {y_pred_classes.shape}, y shape: {y.shape}")  # Debug print
    output = {
        f"{mode}/{phase}/{key.replace('Multiclass', '').split('/')[-1]}": value for key, value in output.items()
    }
    self.log_dict(output, prog_bar=True, logger=True)

    if self.local_cm is not None and self.historical_cm is not None and self.global_cm is not None:
        if mode == "local":
            self.local_cm.update(y_pred_classes, y)
        elif mode == "historical":
            self.historical_cm.update(y_pred_classes, y)
        elif mode == "global":
            self.global_cm.update(y_pred_classes, y)

save_historical_model()

Save the current local model as the historical model.

Source code in nebula/core/models/cifar10/dualagg.py
428
429
430
431
432
433
def save_historical_model(self):
    """
    Save the current local model as the historical model.
    """
    logging_training.info("Copying local model to historical model.")
    self.historical_model.load_state_dict(self.model.state_dict())

test_step(batch, batch_idx)

Test step for the model. Args: batch: batch_idx:

Returns:

Source code in nebula/core/models/cifar10/dualagg.py
412
413
414
415
416
417
418
419
420
421
def test_step(self, batch, batch_idx):
    """
    Test step for the model.
    Args:
        batch:
        batch_idx:

    Returns:
    """
    return self.step(batch, batch_idx, "Test")

training_step(batch, batch_id)

Training step for the model. Args: batch: batch_id:

Returns:

Source code in nebula/core/models/cifar10/dualagg.py
380
381
382
383
384
385
386
387
388
389
def training_step(self, batch, batch_id):
    """
    Training step for the model.
    Args:
        batch:
        batch_id:

    Returns:
    """
    return self.step(batch, batch_id, "Train")

validation_step(batch, batch_idx)

Validation step for the model. Args: batch: batch_idx:

Returns:

Source code in nebula/core/models/cifar10/dualagg.py
396
397
398
399
400
401
402
403
404
405
def validation_step(self, batch, batch_idx):
    """
    Validation step for the model.
    Args:
        batch:
        batch_idx:

    Returns:
    """
    return self.step(batch, batch_idx, "Validation")