model_utils.py 20 KB
Newer Older
1
# @author lucasmiranda42
2
3
4
5
6
7
8
9
# encoding: utf-8
# module deepof

"""

Functions and general utilities for the deepof tensorflow models. See documentation for details

"""
10

11
from itertools import combinations
lucas_miranda's avatar
lucas_miranda committed
12
from typing import Any, Tuple
lucas_miranda's avatar
lucas_miranda committed
13
14
15
16
17

import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
import tensorflow_probability as tfp
18
from tensorflow.keras import backend as K
19
20
21
from tensorflow.keras.constraints import Constraint
from tensorflow.keras.layers import Layer

22
tfd = tfp.distributions
23
tfpl = tfp.layers
24

lucas_miranda's avatar
lucas_miranda committed
25

26
# Helper functions and classes
27
28
29
30
@tf.function
def compute_shannon_entropy(tensor):
    """Computes Shannon entropy for a given tensor"""
    tensor = tf.cast(tensor, tf.dtypes.int32)
31
32
33
    bins = (
        tf.math.bincount(tensor, dtype=tf.dtypes.float32)
        / tf.cast(tf.shape(tensor), tf.dtypes.float32)[0]
34
35
36
37
38
39
40
41
42
43
44
    )
    return -tf.reduce_sum(bins * tf.math.log(bins + 1e-5))


@tf.function
def get_k_nearest_neighbors(tensor, k, index):
    """Retrieve indices of the k nearest neighbors in tensor to the vector with the specified index"""
    query = tensor[index]
    distances = tf.norm(tensor - query, axis=1)
    max_distance = tf.sort(distances)[k]
    neighbourhood_mask = distances < max_distance
45
46
47
48
49
50
    return tf.squeeze(tf.where(neighbourhood_mask))


@tf.function
def get_neighbourhood_entropy(tensor, clusters, k, index):
    neighborhood = get_k_nearest_neighbors(tensor, k, index)
51
    cluster_z = tf.gather(clusters, neighborhood)
52
53
    neigh_entropy = compute_shannon_entropy(cluster_z)
    return neigh_entropy
54
55


56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
class exponential_learning_rate(tf.keras.callbacks.Callback):
    """Simple class that allows to grow learning rate exponentially during training"""

    def __init__(self, factor):
        super().__init__()
        self.factor = factor
        self.rates = []
        self.losses = []

    # noinspection PyMethodOverriding
    def on_batch_end(self, batch, logs):
        """This callback acts after processing each batch"""

        self.rates.append(K.get_value(self.model.optimizer.lr))
        self.losses.append(logs["loss"])
        K.set_value(self.model.optimizer.lr, self.model.optimizer.lr * self.factor)


def find_learning_rate(
75
    model, X, y, epochs=1, batch_size=32, min_rate=10 ** -5, max_rate=10
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
):
    """Trains the provided model for an epoch with an exponentially increasing learning rate"""

    init_weights = model.get_weights()
    iterations = len(X) // batch_size * epochs
    factor = K.exp(K.log(max_rate / min_rate) / iterations)
    init_lr = K.get_value(model.optimizer.lr)
    K.set_value(model.optimizer.lr, min_rate)
    exp_lr = exponential_learning_rate(factor)
    model.fit(X, y, epochs=epochs, batch_size=batch_size, callbacks=[exp_lr])
    K.set_value(model.optimizer.lr, init_lr)
    model.set_weights(init_weights)
    return exp_lr.rates, exp_lr.losses


def plot_lr_vs_loss(rates, losses):  # pragma: no cover
    """Plots learing rate versus the loss function of the model"""

    plt.plot(rates, losses)
    plt.gca().set_xscale("log")
    plt.hlines(min(losses), min(rates), max(rates))
    plt.axis([min(rates), max(rates), min(losses), (losses[0] + min(losses)) / 2])
    plt.xlabel("Learning rate")
    plt.ylabel("Loss")


lucas_miranda's avatar
lucas_miranda committed
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
def compute_kernel(x: tf.Tensor, y: tf.Tensor) -> tf.Tensor:
    """

    Computes the MMD between the two specified vectors using a gaussian kernel.

        Parameters:
            - x (tf.Tensor): left tensor
            - y (tf.Tensor): right tensor

        Returns
            - kernel (tf.Tensor): returns the result of applying the kernel, for
            each training instance

    """

117
118
119
120
121
122
123
124
125
    x_size = tf.shape(x)[0]
    y_size = tf.shape(y)[0]
    dim = tf.shape(x)[1]
    tiled_x = tf.tile(
        tf.reshape(x, tf.stack([x_size, 1, dim])), tf.stack([1, y_size, 1])
    )
    tiled_y = tf.tile(
        tf.reshape(y, tf.stack([1, y_size, dim])), tf.stack([x_size, 1, 1])
    )
lucas_miranda's avatar
lucas_miranda committed
126
    kernel = tf.exp(
127
        -tf.reduce_mean(tf.square(tiled_x - tiled_y), axis=2) / tf.cast(dim, tf.float32)
128
    )
lucas_miranda's avatar
lucas_miranda committed
129
    return kernel
130
131


132
@tf.function
133
def compute_mmd(tensors: Tuple[Any]) -> tf.Tensor:
lucas_miranda's avatar
lucas_miranda committed
134
135
    """

136
    Computes the MMD between the two specified vectors using a gaussian kernel.
lucas_miranda's avatar
lucas_miranda committed
137

138
139
        Parameters:
            - tensors (tuple): tuple containing two tf.Tensor objects
lucas_miranda's avatar
lucas_miranda committed
140

141
142
143
        Returns
            - mmd (tf.Tensor): returns the maximum mean discrepancy for each
            training instance
lucas_miranda's avatar
lucas_miranda committed
144

145
    """
146
147
148
149

    x = tensors[0]
    y = tensors[1]

150
151
152
    x_kernel = compute_kernel(x, x)
    y_kernel = compute_kernel(y, y)
    xy_kernel = compute_kernel(x, y)
lucas_miranda's avatar
lucas_miranda committed
153
    mmd = (
154
155
156
        tf.reduce_mean(x_kernel)
        + tf.reduce_mean(y_kernel)
        - 2 * tf.reduce_mean(xy_kernel)
157
    )
lucas_miranda's avatar
lucas_miranda committed
158
    return mmd
159
160


161
# Custom auxiliary classes
lucas_miranda's avatar
lucas_miranda committed
162
163
164
165
166
167
168
169
class one_cycle_scheduler(tf.keras.callbacks.Callback):
    """

    One cycle learning rate scheduler.
    Based on https://arxiv.org/pdf/1506.01186.pdf

    """

170
    def __init__(
171
172
173
174
175
176
177
        self,
        iterations: int,
        max_rate: float,
        start_rate: float = None,
        last_iterations: int = None,
        last_rate: float = None,
        log_dir: str = ".",
178
    ):
lucas_miranda's avatar
lucas_miranda committed
179
        super().__init__()
180
181
182
183
184
185
186
        self.iterations = iterations
        self.max_rate = max_rate
        self.start_rate = start_rate or max_rate / 10
        self.last_iterations = last_iterations or iterations // 10 + 1
        self.half_iteration = (iterations - self.last_iterations) // 2
        self.last_rate = last_rate or self.start_rate / 1000
        self.iteration = 0
187
        self.history = {}
lucas_miranda's avatar
lucas_miranda committed
188
        self.log_dir = log_dir
189

lucas_miranda's avatar
lucas_miranda committed
190
    def _interpolate(self, iter1: int, iter2: int, rate1: float, rate2: float) -> float:
191
192
        return (rate2 - rate1) * (self.iteration - iter1) / (iter2 - iter1) + rate1

lucas_miranda's avatar
lucas_miranda committed
193
194
195
    # noinspection PyMethodOverriding,PyTypeChecker
    def on_batch_begin(self, batch: int, logs):
        """ Defines computations to perform for each batch """
196
197
198

        self.history.setdefault("lr", []).append(K.get_value(self.model.optimizer.lr))

199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
        if self.iteration < self.half_iteration:
            rate = self._interpolate(
                0, self.half_iteration, self.start_rate, self.max_rate
            )
        elif self.iteration < 2 * self.half_iteration:
            rate = self._interpolate(
                self.half_iteration,
                2 * self.half_iteration,
                self.max_rate,
                self.start_rate,
            )
        else:
            rate = self._interpolate(
                2 * self.half_iteration,
                self.iterations,
                self.start_rate,
                self.last_rate,
            )
            rate = max(rate, self.last_rate)
        self.iteration += 1
        K.set_value(self.model.optimizer.lr, rate)
220

lucas_miranda's avatar
lucas_miranda committed
221
222
223
224
225
226
227
    def on_epoch_end(self, epoch, logs=None):
        """Logs the learning rate to tensorboard"""

        writer = tf.summary.create_file_writer(self.log_dir)

        with writer.as_default():
            tf.summary.scalar(
lucas_miranda's avatar
lucas_miranda committed
228
229
230
                "learning_rate",
                data=self.model.optimizer.lr,
                step=epoch,
lucas_miranda's avatar
lucas_miranda committed
231
            )
232
233


234
class neighbor_latent_entropy(tf.keras.callbacks.Callback):
235
236
    """

237
238
    Latent space entropy callback. Computes the entropy of cluster assignment across k nearest neighbors of a subset
    of samples in the latent space.
239
240
241

    """

242
    def __init__(
243
244
245
246
247
248
        self,
        encoding_dim: int,
        validation_data: np.ndarray = None,
        k: int = 100,
        samples: int = 10000,
        log_dir: str = ".",
249
    ):
250
        super().__init__()
251
        self.enc = encoding_dim
lucas_miranda's avatar
lucas_miranda committed
252
        self.validation_data = validation_data
253
        self.k = k
lucas_miranda's avatar
lucas_miranda committed
254
        self.samples = samples
lucas_miranda's avatar
lucas_miranda committed
255
        self.log_dir = log_dir
256
257

    # noinspection PyMethodOverriding,PyTypeChecker
lucas_miranda's avatar
lucas_miranda committed
258
    def on_epoch_end(self, epoch, logs=None):
259
260
        """ Passes samples through the encoder and computes cluster purity on the latent embedding """

261
        if self.validation_data is not None:
lucas_miranda's avatar
lucas_miranda committed
262
263

            # Get encoer and grouper from full model
264
265
266
267
            latent_distribution = [
                layer
                for layer in self.model.layers
                if layer.name == "latent_distribution"
lucas_miranda's avatar
lucas_miranda committed
268
269
270
271
272
273
274
275
            ][0]
            cluster_assignment = [
                layer
                for layer in self.model.layers
                if layer.name == "cluster_assignment"
            ][0]

            encoder = tf.keras.models.Model(
276
                self.model.layers[0].input, latent_distribution.output
lucas_miranda's avatar
lucas_miranda committed
277
            )
lucas_miranda's avatar
lucas_miranda committed
278
279
            grouper = tf.keras.models.Model(
                self.model.layers[0].input, cluster_assignment.output
lucas_miranda's avatar
lucas_miranda committed
280
281
            )

lucas_miranda's avatar
lucas_miranda committed
282
283
284
285
            # Use encoder and grouper to predict on validation data
            encoding = encoder.predict(self.validation_data)
            groups = grouper.predict(self.validation_data)
            hard_groups = groups.argmax(axis=1)
286
            max_groups = groups.max(axis=1)
lucas_miranda's avatar
lucas_miranda committed
287

288
            # Iterate over samples and compute purity across neighbourhood
289
            self.samples = np.min([self.samples, encoding.shape[0]])
lucas_miranda's avatar
lucas_miranda committed
290
291
292
            random_idxs = np.random.choice(
                range(encoding.shape[0]), self.samples, replace=False
            )
293

294
            # Add result to pre allocated array
295
296
297
298
299
300
            purity_vector = np.zeros(self.samples)

            for i, sample in enumerate(random_idxs):
                purity_vector[i] = get_neighbourhood_entropy(
                    encodings, hard_groups, self.k, sample
                )
lucas_miranda's avatar
lucas_miranda committed
301

lucas_miranda's avatar
lucas_miranda committed
302
303
            writer = tf.summary.create_file_writer(self.log_dir)
            with writer.as_default():
304
305
                tf.summary.scalar(
                    "number_of_populated_clusters",
306
                    data=len(set(hard_groups[max_groups >= 0.25])),
307
308
                    step=epoch,
                )
lucas_miranda's avatar
lucas_miranda committed
309
                tf.summary.scalar(
310
                    "average_neighborhood_cluster_entropy",
311
                    data=np.average(purity_vector, weights=max_groups[random_idxs]),
312
313
314
315
316
317
318
                    step=epoch,
                )
                tf.summary.scalar(
                    "average_confidence_in_selected_cluster",
                    data=np.average(max_groups),
                    step=epoch,
                )
319
320


lucas_miranda's avatar
lucas_miranda committed
321
322
323
class uncorrelated_features_constraint(Constraint):
    """

324
    tf.keras.constraints.Constraint subclass that forces a layer to have uncorrelated features.
lucas_miranda's avatar
lucas_miranda committed
325
326
327
328
    Useful, among others, for auto encoder bottleneck layers

    """

329
330
331
332
    def __init__(self, encoding_dim, weightage=1.0):
        self.encoding_dim = encoding_dim
        self.weightage = weightage

333
    def get_config(self):  # pragma: no cover
334
        """Updates Constraint metadata"""
335
336

        config = super().get_config().copy()
337
        config.update({"encoding_dim": self.encoding_dim, "weightage": self.weightage})
338
339
340
        return config

    def get_covariance(self, x):
341
342
        """Computes the covariance of the elements of the passed layer"""

343
344
345
        x_centered_list = []

        for i in range(self.encoding_dim):
346
            x_centered_list.append(x[:, i] - K.mean(x[:, i]))
347
348

        x_centered = tf.stack(x_centered_list)
349
        covariance = K.dot(x_centered, K.transpose(x_centered)) / tf.cast(
350
351
352
353
354
355
            x_centered.get_shape()[0], tf.float32
        )

        return covariance

    # Constraint penalty
356
    # noinspection PyUnusedLocal
357
    def uncorrelated_feature(self, x):
358
359
        """Adds a penalty on feature correlation, forcing more independent sets of weights"""

360
        if self.encoding_dim <= 1:  # pragma: no cover
361
362
            return 0.0
        else:
363
364
            output = K.sum(
                K.square(
365
                    self.covariance
366
                    - tf.math.multiply(self.covariance, tf.eye(self.encoding_dim))
367
368
369
370
371
372
373
374
375
                )
            )
            return output

    def __call__(self, x):
        self.covariance = self.get_covariance(x)
        return self.weightage * self.uncorrelated_feature(x)


376
377
# Custom Layers
class MCDropout(tf.keras.layers.Dropout):
378
379
380
    """Equivalent to tf.keras.layers.Dropout, but with training mode enabled at prediction time.
    Useful for Montecarlo predictions"""

381
    def call(self, inputs, **kwargs):
382
        """Overrides the call method of the subclassed function"""
383
384
385
386
        return super().call(inputs, training=True)


class DenseTranspose(Layer):
387
388
389
390
    """Mirrors a tf.keras.layers.Dense instance with transposed weights.
    Useful for decoder layers in autoencoders, to force structure and
    decrease the effective number of parameters to train"""

391
392
393
394
395
396
    def __init__(self, dense, output_dim, activation=None, **kwargs):
        self.dense = dense
        self.output_dim = output_dim
        self.activation = tf.keras.activations.get(activation)
        super().__init__(**kwargs)

397
    def get_config(self):  # pragma: no cover
398
399
        """Updates Constraint metadata"""

400
401
402
403
404
405
406
407
408
409
        config = super().get_config().copy()
        config.update(
            {
                "dense": self.dense,
                "output_dim": self.output_dim,
                "activation": self.activation,
            }
        )
        return config

410
    # noinspection PyAttributeOutsideInit
411
    def build(self, batch_input_shape):
412
413
        """Updates Layer's build method"""

414
        self.biases = self.add_weight(
lucas_miranda's avatar
lucas_miranda committed
415
            name="bias",
lucas_miranda's avatar
lucas_miranda committed
416
            shape=self.dense.get_input_at(-1).get_shape().as_list()[1:],
lucas_miranda's avatar
lucas_miranda committed
417
            initializer="zeros",
418
419
420
421
        )
        super().build(batch_input_shape)

    def call(self, inputs, **kwargs):
422
423
        """Updates Layer's call method"""

424
425
426
        z = tf.matmul(inputs, self.dense.weights[0], transpose_b=True)
        return self.activation(z + self.biases)

427
    def compute_output_shape(self, input_shape):  # pragma: no cover
428
429
        """Outputs the transposed shape"""

430
431
432
        return input_shape[0], self.output_dim


433
class KLDivergenceLayer(tfpl.KLDivergenceAddLoss):
434
    """
435
436
    Identity transform layer that adds KL Divergence
    to the final model loss.
437
438
    """

439
    def __init__(self, iters, warm_up_iters, annealing_mode, *args, **kwargs):
440
        super(KLDivergenceLayer, self).__init__(*args, **kwargs)
441
442
443
        self.is_placeholder = True
        self._iters = iters
        self._warm_up_iters = warm_up_iters
444
        self._annealing_mode = annealing_mode
445

446
    def get_config(self):  # pragma: no cover
447
448
        """Updates Constraint metadata"""

449
        config = super().get_config().copy()
450
        config.update({"is_placeholder": self.is_placeholder})
451
452
        config.update({"_iters": self._iters})
        config.update({"_warm_up_iters": self._warm_up_iters})
453
        config.update({"_annealing_mode": self._annealing_mode})
454
455
456
        return config

    def call(self, distribution_a):
457
458
        """Updates Layer's call method"""

459
460
        # Define and update KL weight for warmup
        if self._warm_up_iters > 0:
461
            if self._annealing_mode in ["linear", "sigmoid"]:
462
463
464
                kl_weight = tf.cast(
                    K.min([self._iters / self._warm_up_iters, 1.0]), tf.float32
                )
465
                if self._annealing_mode == "sigmoid":
466
467
468
                    kl_weight = tf.math.sigmoid(
                        (2 * kl_weight - 1) / (kl_weight - kl_weight ** 2)
                    )
469
            else:
470
471
472
                raise NotImplementedError(
                    "annealing_mode must be one of 'linear' and 'sigmoid'"
                )
473
474
475
476
477
        else:
            kl_weight = tf.cast(1.0, tf.float32)

        kl_batch = kl_weight * self._regularizer(distribution_a)

478
479
        self.add_loss(kl_batch, inputs=[distribution_a])
        self.add_metric(
480
481
482
            kl_batch,
            aggregation="mean",
            name="kl_divergence",
483
        )
484
        # noinspection PyProtectedMember
485
        self.add_metric(kl_weight, aggregation="mean", name="kl_rate")
486
487
488
489

        return distribution_a


490
class MMDiscrepancyLayer(Layer):
491
    """
492
    Identity transform layer that adds MM Discrepancy
493
494
495
    to the final model loss.
    """

496
    def __init__(
497
        self, batch_size, prior, iters, warm_up_iters, annealing_mode, *args, **kwargs
498
    ):
499
        super(MMDiscrepancyLayer, self).__init__(*args, **kwargs)
500
        self.is_placeholder = True
501
        self.batch_size = batch_size
502
        self.prior = prior
503
504
        self._iters = iters
        self._warm_up_iters = warm_up_iters
505
        self._annealing_mode = annealing_mode
506

507
    def get_config(self):  # pragma: no cover
508
509
        """Updates Constraint metadata"""

510
        config = super().get_config().copy()
511
        config.update({"batch_size": self.batch_size})
512
513
        config.update({"_iters": self._iters})
        config.update({"_warmup_iters": self._warm_up_iters})
514
        config.update({"prior": self.prior})
515
        config.update({"_annealing_mode": self._annealing_mode})
516
517
        return config

518
    def call(self, z, **kwargs):
519
520
        """Updates Layer's call method"""

521
        true_samples = self.prior.sample(self.batch_size)
522

523
524
        # Define and update MMD weight for warmup
        if self._warm_up_iters > 0:
525
526
527
528
529
            if self._annealing_mode in ["linear", "sigmoid"]:
                mmd_weight = tf.cast(
                    K.min([self._iters / self._warm_up_iters, 1.0]), tf.float32
                )
                if self._annealing_mode == "sigmoid":
530
531
532
                    mmd_weight = tf.math.sigmoid(
                        (2 * mmd_weight - 1) / (mmd_weight - mmd_weight ** 2)
                    )
533
534
535
536
            else:
                raise NotImplementedError(
                    "annealing_mode must be one of 'linear' and 'sigmoid'"
                )
537
538
539
        else:
            mmd_weight = tf.cast(1.0, tf.float32)

540
        mmd_batch = mmd_weight * compute_mmd((true_samples, z))
541

542
        self.add_loss(K.mean(mmd_batch), inputs=z)
543
        self.add_metric(mmd_batch, aggregation="mean", name="mmd")
544
        self.add_metric(mmd_weight, aggregation="mean", name="mmd_rate")
545
546

        return z
547
548


549
class ClusterOverlap(Layer):
550
551
    """
    Identity layer that measures the overlap between the components of the latent Gaussian Mixture
552
553
    using the the entropy of the nearest neighbourhood. If self.loss_weight > 0, it adds a regularization
    penalty to the loss function
554
555
    """

556
557
558
559
560
561
562
563
564
565
566
567
    def __init__(
        self,
        encoding_dim: int,
        k: int = 100,
        loss_weight: float = False,
        samples: int = 512,
        *args,
        **kwargs
    ):
        self.enc = encoding_dim
        self.k = k
        self.loss_weight = loss_weight
568
        self.min_confidence = 0.25
569
        self.samples = samples
570
        super(ClusterOverlap, self).__init__(*args, **kwargs)
571

572
    def get_config(self):  # pragma: no cover
lucas_miranda's avatar
lucas_miranda committed
573
574
        """Updates Constraint metadata"""

575
        config = super().get_config().copy()
576
577
578
        config.update({"enc": self.enc})
        config.update({"k": self.k})
        config.update({"loss_weight": self.loss_weight})
579
        config.update({"min_confidence": self.min_confidence})
580
581
582
        config.update({"samples": self.samples})
        return config

lucas_miranda's avatar
lucas_miranda committed
583
    @tf.function
584
    def call(self, encodings, categorical, **kwargs):
lucas_miranda's avatar
lucas_miranda committed
585
        """Updates Layer's call method"""
586

587
588
        hard_groups = tf.math.argmax(categorical, axis=1)
        max_groups = tf.reduce_max(categorical, axis=1)
589

590
591
592
593
        # Iterate over samples and compute purity across neighbourhood
        self.samples = tf.reduce_min([self.samples, encodings.shape[0]])
        random_idxs = range(encoding.shape[0])
        random_idxs = tf.random.categorical(
594
            tf.expand_dims(random_idxs / tf.reduce_sum(random_idxs), 0), self.samples
595
        )
596

597
        purity_vector = tf.map_fn(get_neighbourhood_entropy, random_idxs)
598

599
600
601
602
        for i, sample in enumerate(random_idxs):
            purity_vector[i] = get_neighbourhood_entropy(
                encodings, hard_groups, self.k, sample
            )
603
604
605
606
607
608
609
610
611
612
613
614
615

        neighbourhood_entropy = purity_vector * max_groups[random_idxs]

        self.add_metric(
            len(set(hard_groups[max_groups >= self.min_confidence])),
            aggregation="mean",
            name="number_of_populated_clusters",
        )

        self.add_metric(
            max_groups,
            aggregation="mean",
            name="average_confidence_in_selected_cluster",
616
        )
617

618
        self.add_metric(
619
            neighbourhood_entropy, aggregation="mean", name="neighbourhood_entropy"
620
        )
621

622
        if self.loss_weight:
623
            self.add_loss(neighbourhood_entropy, inputs=[target, categorical])
624

625
        return encodings