model_utils.py 15.7 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 functools import partial
lucas_miranda's avatar
lucas_miranda committed
12
from typing import Any, Tuple
lucas_miranda's avatar
lucas_miranda committed
13
14
15
16

import matplotlib.pyplot as plt
import tensorflow as tf
import tensorflow_probability as tfp
17
from tensorflow.keras import backend as K
18
19
from tensorflow.keras.layers import Layer

20
tfd = tfp.distributions
21
tfpl = tfp.layers
22

lucas_miranda's avatar
lucas_miranda committed
23

24
# Helper functions and classes
25
26
27
28
@tf.function
def compute_shannon_entropy(tensor):
    """Computes Shannon entropy for a given tensor"""
    tensor = tf.cast(tensor, tf.dtypes.int32)
29
30
31
    bins = (
        tf.math.bincount(tensor, dtype=tf.dtypes.float32)
        / tf.cast(tf.shape(tensor), tf.dtypes.float32)[0]
32
33
34
35
36
37
38
    )
    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"""
39
    query = tf.gather(tensor, index, batch_dims=0)
40
41
42
    distances = tf.norm(tensor - query, axis=1)
    max_distance = tf.sort(distances)[k]
    neighbourhood_mask = distances < max_distance
43
44
45
46
    return tf.squeeze(tf.where(neighbourhood_mask))


@tf.function
47
def get_neighbourhood_entropy(index, tensor, clusters, k):
48
    neighborhood = get_k_nearest_neighbors(tensor, k, index)
49
    cluster_z = tf.gather(clusters, neighborhood, batch_dims=0)
50
51
    neigh_entropy = compute_shannon_entropy(cluster_z)
    return neigh_entropy
52
53


54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
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(
73
    model, X, y, epochs=1, batch_size=32, min_rate=10 ** -5, max_rate=10
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
):
    """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
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
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

    """

115
116
117
118
119
120
121
122
123
    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
124
    kernel = tf.exp(
125
        -tf.reduce_mean(tf.square(tiled_x - tiled_y), axis=2) / tf.cast(dim, tf.float32)
126
    )
lucas_miranda's avatar
lucas_miranda committed
127
    return kernel
128
129


130
@tf.function
131
def compute_mmd(tensors: Tuple[Any]) -> tf.Tensor:
lucas_miranda's avatar
lucas_miranda committed
132
133
    """

134
    Computes the MMD between the two specified vectors using a gaussian kernel.
lucas_miranda's avatar
lucas_miranda committed
135

136
137
        Parameters:
            - tensors (tuple): tuple containing two tf.Tensor objects
lucas_miranda's avatar
lucas_miranda committed
138

139
140
141
        Returns
            - mmd (tf.Tensor): returns the maximum mean discrepancy for each
            training instance
lucas_miranda's avatar
lucas_miranda committed
142

143
    """
144
145
146
147

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

148
149
150
    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
151
    mmd = (
152
153
154
        tf.reduce_mean(x_kernel)
        + tf.reduce_mean(y_kernel)
        - 2 * tf.reduce_mean(xy_kernel)
155
    )
lucas_miranda's avatar
lucas_miranda committed
156
    return mmd
157
158


159
# Custom auxiliary classes
lucas_miranda's avatar
lucas_miranda committed
160
161
162
163
164
165
166
167
class one_cycle_scheduler(tf.keras.callbacks.Callback):
    """

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

    """

168
    def __init__(
169
170
171
172
173
174
175
        self,
        iterations: int,
        max_rate: float,
        start_rate: float = None,
        last_iterations: int = None,
        last_rate: float = None,
        log_dir: str = ".",
176
    ):
lucas_miranda's avatar
lucas_miranda committed
177
        super().__init__()
178
179
180
181
182
183
184
        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
185
        self.history = {}
lucas_miranda's avatar
lucas_miranda committed
186
        self.log_dir = log_dir
187

lucas_miranda's avatar
lucas_miranda committed
188
    def _interpolate(self, iter1: int, iter2: int, rate1: float, rate2: float) -> float:
189
190
        return (rate2 - rate1) * (self.iteration - iter1) / (iter2 - iter1) + rate1

lucas_miranda's avatar
lucas_miranda committed
191
192
193
    # noinspection PyMethodOverriding,PyTypeChecker
    def on_batch_begin(self, batch: int, logs):
        """ Defines computations to perform for each batch """
194
195
196

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

197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
        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)
218

lucas_miranda's avatar
lucas_miranda committed
219
220
221
222
223
224
225
    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
226
227
228
                "learning_rate",
                data=self.model.optimizer.lr,
                step=epoch,
lucas_miranda's avatar
lucas_miranda committed
229
            )
230
231


232
233
# Custom Layers
class MCDropout(tf.keras.layers.Dropout):
234
235
236
    """Equivalent to tf.keras.layers.Dropout, but with training mode enabled at prediction time.
    Useful for Montecarlo predictions"""

237
    def call(self, inputs, **kwargs):
238
        """Overrides the call method of the subclassed function"""
239
240
241
242
        return super().call(inputs, training=True)


class DenseTranspose(Layer):
243
244
245
246
    """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"""

247
248
249
250
251
252
    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)

253
    def get_config(self):  # pragma: no cover
254
255
        """Updates Constraint metadata"""

256
257
258
259
260
261
262
263
264
265
        config = super().get_config().copy()
        config.update(
            {
                "dense": self.dense,
                "output_dim": self.output_dim,
                "activation": self.activation,
            }
        )
        return config

266
    # noinspection PyAttributeOutsideInit
267
    def build(self, batch_input_shape):
268
269
        """Updates Layer's build method"""

270
        self.biases = self.add_weight(
lucas_miranda's avatar
lucas_miranda committed
271
            name="bias",
lucas_miranda's avatar
lucas_miranda committed
272
            shape=self.dense.get_input_at(-1).get_shape().as_list()[1:],
lucas_miranda's avatar
lucas_miranda committed
273
            initializer="zeros",
274
275
276
277
        )
        super().build(batch_input_shape)

    def call(self, inputs, **kwargs):
278
279
        """Updates Layer's call method"""

280
281
282
        z = tf.matmul(inputs, self.dense.weights[0], transpose_b=True)
        return self.activation(z + self.biases)

283
    def compute_output_shape(self, input_shape):  # pragma: no cover
284
285
        """Outputs the transposed shape"""

286
287
288
        return input_shape[0], self.output_dim


289
class KLDivergenceLayer(tfpl.KLDivergenceAddLoss):
290
    """
291
292
    Identity transform layer that adds KL Divergence
    to the final model loss.
293
294
    """

295
    def __init__(self, iters, warm_up_iters, annealing_mode, *args, **kwargs):
296
        super(KLDivergenceLayer, self).__init__(*args, **kwargs)
297
298
299
        self.is_placeholder = True
        self._iters = iters
        self._warm_up_iters = warm_up_iters
300
        self._annealing_mode = annealing_mode
301

302
    def get_config(self):  # pragma: no cover
303
304
        """Updates Constraint metadata"""

305
        config = super().get_config().copy()
306
        config.update({"is_placeholder": self.is_placeholder})
307
308
        config.update({"_iters": self._iters})
        config.update({"_warm_up_iters": self._warm_up_iters})
309
        config.update({"_annealing_mode": self._annealing_mode})
310
311
312
        return config

    def call(self, distribution_a):
313
314
        """Updates Layer's call method"""

315
316
        # Define and update KL weight for warmup
        if self._warm_up_iters > 0:
317
            if self._annealing_mode in ["linear", "sigmoid"]:
318
319
320
                kl_weight = tf.cast(
                    K.min([self._iters / self._warm_up_iters, 1.0]), tf.float32
                )
321
                if self._annealing_mode == "sigmoid":
322
323
324
                    kl_weight = tf.math.sigmoid(
                        (2 * kl_weight - 1) / (kl_weight - kl_weight ** 2)
                    )
325
            else:
326
327
328
                raise NotImplementedError(
                    "annealing_mode must be one of 'linear' and 'sigmoid'"
                )
329
330
331
332
333
        else:
            kl_weight = tf.cast(1.0, tf.float32)

        kl_batch = kl_weight * self._regularizer(distribution_a)

334
335
        self.add_loss(kl_batch, inputs=[distribution_a])
        self.add_metric(
336
337
338
            kl_batch,
            aggregation="mean",
            name="kl_divergence",
339
        )
340
        # noinspection PyProtectedMember
341
        self.add_metric(kl_weight, aggregation="mean", name="kl_rate")
342
343
344
345

        return distribution_a


346
class MMDiscrepancyLayer(Layer):
347
    """
348
    Identity transform layer that adds MM Discrepancy
349
350
351
    to the final model loss.
    """

352
    def __init__(
353
        self, batch_size, prior, iters, warm_up_iters, annealing_mode, *args, **kwargs
354
    ):
355
        super(MMDiscrepancyLayer, self).__init__(*args, **kwargs)
356
        self.is_placeholder = True
357
        self.batch_size = batch_size
358
        self.prior = prior
359
360
        self._iters = iters
        self._warm_up_iters = warm_up_iters
361
        self._annealing_mode = annealing_mode
362

363
    def get_config(self):  # pragma: no cover
364
365
        """Updates Constraint metadata"""

366
        config = super().get_config().copy()
367
        config.update({"batch_size": self.batch_size})
368
369
        config.update({"_iters": self._iters})
        config.update({"_warmup_iters": self._warm_up_iters})
370
        config.update({"prior": self.prior})
371
        config.update({"_annealing_mode": self._annealing_mode})
372
373
        return config

374
    def call(self, z, **kwargs):
375
376
        """Updates Layer's call method"""

377
        true_samples = self.prior.sample(self.batch_size)
378

379
380
        # Define and update MMD weight for warmup
        if self._warm_up_iters > 0:
381
382
383
384
385
            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":
386
387
388
                    mmd_weight = tf.math.sigmoid(
                        (2 * mmd_weight - 1) / (mmd_weight - mmd_weight ** 2)
                    )
389
390
391
392
            else:
                raise NotImplementedError(
                    "annealing_mode must be one of 'linear' and 'sigmoid'"
                )
393
394
395
        else:
            mmd_weight = tf.cast(1.0, tf.float32)

396
        mmd_batch = mmd_weight * compute_mmd((true_samples, z))
397

398
        self.add_loss(K.mean(mmd_batch), inputs=z)
399
        self.add_metric(mmd_batch, aggregation="mean", name="mmd")
400
        self.add_metric(mmd_weight, aggregation="mean", name="mmd_rate")
401
402

        return z
403
404


405
class ClusterOverlap(Layer):
406
407
    """
    Identity layer that measures the overlap between the components of the latent Gaussian Mixture
408
409
    using the the entropy of the nearest neighbourhood. If self.loss_weight > 0, it adds a regularization
    penalty to the loss function
410
411
    """

412
413
    def __init__(
        self,
414
        batch_size: int,
415
        encoding_dim: int,
416
        k: int = 25,
417
        loss_weight: float = 0.0,
418
419
420
        *args,
        **kwargs
    ):
421
        self.batch_size = batch_size
422
423
424
        self.enc = encoding_dim
        self.k = k
        self.loss_weight = loss_weight
425
426
        self.min_confidence = 0.25
        super(ClusterOverlap, self).__init__(*args, **kwargs)
427

428
    def get_config(self):  # pragma: no cover
lucas_miranda's avatar
lucas_miranda committed
429
430
        """Updates Constraint metadata"""

431
        config = super().get_config().copy()
432
        config.update({"batch_size": self.batch_size})
433
434
435
        config.update({"enc": self.enc})
        config.update({"k": self.k})
        config.update({"loss_weight": self.loss_weight})
436
        config.update({"min_confidence": self.min_confidence})
437
438
439
        config.update({"samples": self.samples})
        return config

lucas_miranda's avatar
lucas_miranda committed
440
    def call(self, inputs, **kwargs):
lucas_miranda's avatar
lucas_miranda committed
441
        """Updates Layer's call method"""
442

443
444
        encodings, categorical = inputs[0], inputs[1]

lucas_miranda's avatar
lucas_miranda committed
445
446
447
448
449
450
451
452
453
        hard_groups = tf.math.argmax(categorical, axis=1)
        max_groups = tf.reduce_max(categorical, axis=1)

        get_local_neighbourhood_entropy = partial(
            get_neighbourhood_entropy,
            tensor=encodings,
            clusters=hard_groups,
            k=self.k,
        )
454

lucas_miranda's avatar
lucas_miranda committed
455
456
457
458
459
        purity_vector = tf.map_fn(
            get_local_neighbourhood_entropy,
            tf.constant(list(range(self.batch_size))),
            dtype=tf.dtypes.float32,
        )
460

lucas_miranda's avatar
lucas_miranda committed
461
462
463
464
465
466
467
468
469
470
471
        ### CANDIDATE FOR REMOVAL. EXPLORE HOW USEFUL THIS REALLY IS ###
        neighbourhood_entropy = purity_vector * max_groups

        number_of_clusters = tf.cast(
            tf.shape(
                tf.unique(
                    tf.reshape(
                        tf.gather(
                            tf.cast(hard_groups, tf.dtypes.float32),
                            tf.where(max_groups >= self.min_confidence),
                            batch_dims=0,
472
                        ),
lucas_miranda's avatar
lucas_miranda committed
473
474
                        [-1],
                    ),
475
                )[0],
lucas_miranda's avatar
lucas_miranda committed
476
477
478
            )[0],
            tf.dtypes.float32,
        )
479

lucas_miranda's avatar
lucas_miranda committed
480
481
482
483
        self.add_metric(
            number_of_clusters,
            name="number_of_populated_clusters",
        )
484

lucas_miranda's avatar
lucas_miranda committed
485
486
487
488
489
        self.add_metric(
            max_groups,
            aggregation="mean",
            name="average_confidence_in_selected_cluster",
        )
490

lucas_miranda's avatar
lucas_miranda committed
491
492
493
        self.add_metric(
            neighbourhood_entropy, aggregation="mean", name="neighbourhood_entropy"
        )
494

lucas_miranda's avatar
lucas_miranda committed
495
496
497
498
499
        if self.loss_weight:
            # minimize local entropy
            self.add_loss(self.loss_weight * tf.reduce_mean(neighbourhood_entropy))
            # maximize number of clusters
            # self.add_loss(-self.loss_weight * tf.reduce_mean(number_of_clusters))
500

501
        return encodings