train_utils.py 18.1 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
# @author lucasmiranda42
# encoding: utf-8
# module deepof

"""

Simple utility functions used in deepof example scripts. These are not part of the main package

"""

lucas_miranda's avatar
lucas_miranda committed
11
12
import json
import os
13
from datetime import date, datetime
lucas_miranda's avatar
lucas_miranda committed
14
15
16
17
from typing import Tuple, Union, Any, List

import numpy as np
import tensorflow as tf
18
from kerastuner import BayesianOptimization, Hyperband
19
from kerastuner_tensorboard_logger import TensorBoardLogger
lucas_miranda's avatar
lucas_miranda committed
20
from sklearn.metrics import roc_auc_score
21
from tensorboard.plugins.hparams import api as hp
lucas_miranda's avatar
lucas_miranda committed
22

23
24
25
import deepof.hypermodels
import deepof.model_utils

26
27
28
29
30
# Ignore warning with no downstream effect
tf.get_logger().setLevel("ERROR")
tf.autograph.set_verbosity(0)


31
class CustomStopper(tf.keras.callbacks.EarlyStopping):
32
    """ Custom early stopping callback. Prevents the model from stopping before warmup is over """
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49

    def __init__(self, start_epoch, *args, **kwargs):
        super(CustomStopper, self).__init__(*args, **kwargs)
        self.start_epoch = start_epoch

    def get_config(self):  # pragma: no cover
        """Updates callback metadata"""

        config = super().get_config().copy()
        config.update({"start_epoch": self.start_epoch})
        return config

    def on_epoch_end(self, epoch, logs=None):
        if epoch > self.start_epoch:
            super().on_epoch_end(epoch, logs)


50
51
52
53
54
def load_treatments(train_path):
    """Loads a dictionary containing the treatments per individual,
    to be loaded as metadata in the coordinates class"""
    try:
        with open(
55
56
57
58
59
            os.path.join(
                train_path,
                [i for i in os.listdir(train_path) if i.endswith(".json")][0],
            ),
            "r",
60
        ) as handle:
61
            treatment_dict = json.load(handle)
62
63
64
65
66
67
68
    except IndexError:
        treatment_dict = None

    return treatment_dict


def get_callbacks(
69
70
71
72
73
74
75
76
77
78
79
80
81
82
    X_train: np.array,
    batch_size: int,
    variational: bool,
    phenotype_class: float,
    predictor: float,
    loss: str,
    X_val: np.array = None,
    cp: bool = False,
    reg_cat_clusters: bool = False,
    reg_cluster_variance: bool = False,
    entropy_samples: int = 15000,
    entropy_knn: int = 100,
    logparam: dict = None,
    outpath: str = ".",
83
) -> List[Union[Any]]:
84
    """Generates callbacks for model training, including:
85
86
87
88
    - run_ID: run name, with coarse parameter details;
    - tensorboard_callback: for real-time visualization;
    - cp_callback: for checkpoint saving,
    - onecycle: for learning rate scheduling"""
89

90
91
92
93
94
95
96
97
    latreg = "none"
    if reg_cat_clusters and not reg_cluster_variance:
        latreg = "categorical"
    elif reg_cluster_variance and not reg_cat_clusters:
        latreg = "variance"
    elif reg_cat_clusters and reg_cluster_variance:
        latreg = "categorical+variance"

98
    run_ID = "{}{}{}{}{}{}{}_{}".format(
99
        ("GMVAE" if variational else "AE"),
lucas_miranda's avatar
lucas_miranda committed
100
        ("_Pred={}".format(predictor) if predictor > 0 and variational else ""),
101
        ("_Pheno={}".format(phenotype_class) if phenotype_class > 0 else ""),
102
        ("_loss={}".format(loss) if variational else ""),
103
104
        ("_encoding={}".format(logparam["encoding"]) if logparam is not None else ""),
        ("_k={}".format(logparam["k"]) if logparam is not None else ""),
105
        ("_latreg={}".format(latreg)),
106
        ("entknn={}".format(entropy_knn)),
107
        (datetime.now().strftime("%Y%m%d-%H%M%S")),
108
109
    )

110
    log_dir = os.path.abspath(os.path.join(outpath, "fit", run_ID))
111
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
112
113
114
        log_dir=log_dir,
        histogram_freq=1,
        profile_batch=2,
115
116
    )

117
    entropy = deepof.model_utils.neighbor_latent_entropy(
118
        encoding_dim=logparam["encoding"],
119
        k=entropy_knn,
120
        samples=entropy_samples,
lucas_miranda's avatar
lucas_miranda committed
121
        validation_data=X_val,
122
        log_dir=os.path.join(outpath, "metrics", run_ID),
123
        variational=variational,
lucas_miranda's avatar
lucas_miranda committed
124
125
    )

126
    onecycle = deepof.model_utils.one_cycle_scheduler(
127
128
        X_train.shape[0] // batch_size * 250,
        max_rate=0.005,
129
        log_dir=os.path.join(outpath, "metrics", run_ID),
130
131
    )

132
    callbacks = [run_ID, tensorboard_callback, entropy, onecycle]
133
134
135

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
136
            os.path.join(outpath, "checkpoints", run_ID + "/cp-{epoch:04d}.ckpt"),
137
138
139
140
141
142
143
144
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
145
146


lucas_miranda's avatar
lucas_miranda committed
147
def log_hyperparameters(phenotype_class: float, rec: str):
lucas_miranda's avatar
lucas_miranda committed
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
    """Blueprint for hyperparameter and metric logging in tensorboard during hyperparameter tuning"""

    logparams = [
        hp.HParam(
            "encoding",
            hp.Discrete([2, 4, 6, 8, 12, 16]),
            display_name="encoding",
            description="encoding size dimensionality",
        ),
        hp.HParam(
            "k",
            hp.IntInterval(min_value=1, max_value=25),
            display_name="k",
            description="cluster_number",
        ),
        hp.HParam(
            "loss",
            hp.Discrete(["ELBO", "MMD", "ELBO+MMD"]),
            display_name="loss function",
            description="loss function",
        ),
    ]

    metrics = [
        hp.Metric("val_{}mae".format(rec), display_name="val_{}mae".format(rec)),
        hp.Metric("val_{}mse".format(rec), display_name="val_{}mse".format(rec)),
    ]
    if phenotype_class:
        logparams.append(
            hp.HParam(
                "pheno_weight",
                hp.RealInterval(min_value=0.0, max_value=1000.0),
                display_name="pheno weight",
                description="weight applied to phenotypic classifier from the latent space",
            )
        )
        metrics += [
            hp.Metric(
                "phenotype_prediction_accuracy",
                display_name="phenotype_prediction_accuracy",
            ),
            hp.Metric(
                "phenotype_prediction_auc",
                display_name="phenotype_prediction_auc",
            ),
        ]

    return logparams, metrics


# noinspection PyUnboundLocalVariable
lucas_miranda's avatar
lucas_miranda committed
199
def tensorboard_metric_logging(
200
201
202
203
204
205
206
207
    run_dir: str,
    hpms: Any,
    ae: Any,
    X_val: np.ndarray,
    y_val: np.ndarray,
    phenotype_class: float,
    predictor: float,
    rec: str,
lucas_miranda's avatar
lucas_miranda committed
208
):
lucas_miranda's avatar
lucas_miranda committed
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
    """Autoencoder metric logging in tensorboard"""

    output = ae.predict(X_val)
    if phenotype_class or predictor:
        reconstruction = output[0]
        prediction = output[1]
        pheno = output[-1]
    else:
        reconstruction = output

    with tf.summary.create_file_writer(run_dir).as_default():
        hp.hparams(hpms)  # record the values used in this trial
        val_mae = tf.reduce_mean(
            tf.keras.metrics.mean_absolute_error(X_val, reconstruction)
        )
        val_mse = tf.reduce_mean(
            tf.keras.metrics.mean_squared_error(X_val, reconstruction)
        )
        tf.summary.scalar("val_{}mae".format(rec), val_mae, step=1)
        tf.summary.scalar("val_{}mse".format(rec), val_mse, step=1)

        if predictor:
            pred_mae = tf.reduce_mean(
                tf.keras.metrics.mean_absolute_error(X_val, prediction)
            )
            pred_mse = tf.reduce_mean(
                tf.keras.metrics.mean_squared_error(X_val, prediction)
            )
            tf.summary.scalar("val_prediction_mae".format(rec), pred_mae, step=1)
            tf.summary.scalar("val_prediction_mse".format(rec), pred_mse, step=1)

        if phenotype_class:
            pheno_acc = tf.keras.metrics.binary_accuracy(y_val, tf.squeeze(pheno))
            pheno_auc = roc_auc_score(y_val, pheno)

            tf.summary.scalar("phenotype_prediction_accuracy", pheno_acc, step=1)
            tf.summary.scalar("phenotype_prediction_auc", pheno_auc, step=1)


248
def autoencoder_fitting(
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
    preprocessed_object: Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray],
    batch_size: int,
    encoding_size: int,
    epochs: int,
    hparams: dict,
    kl_warmup: int,
    log_history: bool,
    log_hparams: bool,
    loss: str,
    mmd_warmup: int,
    montecarlo_kl: int,
    n_components: int,
    output_path: str,
    phenotype_class: float,
    predictor: float,
    pretrained: str,
    save_checkpoints: bool,
    save_weights: bool,
    variational: bool,
    reg_cat_clusters: bool,
    reg_cluster_variance: bool,
    entropy_samples: int,
    entropy_knn: int,
272
):
273
274
    """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""

275
    # Load data
276
277
278
279
280
    X_train, y_train, X_val, y_val = preprocessed_object

    # To avoid stability issues
    tf.keras.backend.clear_session()

281
    # Defines what to log on tensorboard (useful for trying out different models)
282
283
    logparam = {
        "encoding": encoding_size,
284
        "k": n_components,
285
286
287
288
289
        "loss": loss,
    }
    if phenotype_class:
        logparam["pheno_weight"] = phenotype_class

290
    # Load callbacks
291
    run_ID, *cbacks = get_callbacks(
292
        X_train=X_train,
lucas_miranda's avatar
lucas_miranda committed
293
        X_val=(X_val if X_val.shape != (0,) else None),
294
295
296
        batch_size=batch_size,
        cp=save_checkpoints,
        variational=variational,
297
        phenotype_class=phenotype_class,
298
299
        predictor=predictor,
        loss=loss,
300
        entropy_samples=entropy_samples,
301
        entropy_knn=entropy_knn,
302
        reg_cat_clusters=reg_cat_clusters,
303
        reg_cluster_variance=reg_cluster_variance,
304
305
306
        logparam=logparam,
        outpath=output_path,
    )
307
308
    if not log_history:
        cbacks = cbacks[1:]
309

310
    # Logs hyperparameters to tensorboard
lucas_miranda's avatar
lucas_miranda committed
311
    rec = "reconstruction_" if phenotype_class else ""
312
    if log_hparams:
lucas_miranda's avatar
lucas_miranda committed
313
        logparams, metrics = log_hyperparameters(phenotype_class, rec)
314
315

        with tf.summary.create_file_writer(
316
            os.path.join(output_path, "hparams", run_ID)
317
318
319
320
321
        ).as_default():
            hp.hparams_config(
                hparams=logparams,
                metrics=metrics,
            )
322

323
    # Build models
324
325
326
327
328
329
330
    if not variational:
        encoder, decoder, ae = deepof.models.SEQ_2_SEQ_AE(
            ({} if hparams is None else hparams)
        ).build(X_train.shape)
        return_list = (encoder, decoder, ae)

    else:
331
332
333
334
335
336
337
338
        (
            encoder,
            generator,
            grouper,
            ae,
            prior,
            posterior,
        ) = deepof.models.SEQ_2_SEQ_GMVAE(
339
340
341
342
343
344
345
346
347
348
349
350
351
            architecture_hparams=({} if hparams is None else hparams),
            batch_size=batch_size,
            compile_model=True,
            encoding=encoding_size,
            kl_warmup_epochs=kl_warmup,
            loss=loss,
            mmd_warmup_epochs=mmd_warmup,
            montecarlo_kl=montecarlo_kl,
            neuron_control=False,
            number_of_components=n_components,
            overlap_loss=False,
            phenotype_prediction=phenotype_class,
            predictor=predictor,
352
353
            reg_cat_clusters=reg_cat_clusters,
            reg_cluster_variance=reg_cluster_variance,
354
355
356
        ).build(
            X_train.shape
        )
357
358
359
        return_list = (encoder, generator, grouper, ae)

    if pretrained:
360
        # If pretrained models are specified, load weights and return
361
362
363
364
365
366
367
368
369
        ae.load_weights(pretrained)
        return return_list

    else:
        if not variational:

            ae.fit(
                x=X_train,
                y=X_train,
370
                epochs=epochs,
371
372
373
                batch_size=batch_size,
                verbose=1,
                validation_data=(X_val, X_val),
374
                callbacks=cbacks
375
376
377
378
379
380
381
382
                + [
                    CustomStopper(
                        monitor="val_loss",
                        patience=5,
                        restore_best_weights=True,
                        start_epoch=max(kl_warmup, mmd_warmup),
                    ),
                ],
383
384
            )

385
386
387
            if save_weights:
                ae.save_weights("{}_final_weights.h5".format(run_ID))

388
389
        else:

390
            callbacks_ = cbacks + [
391
392
393
394
                CustomStopper(
                    monitor="val_loss",
                    patience=5,
                    restore_best_weights=True,
395
                    start_epoch=max(kl_warmup, mmd_warmup),
396
397
398
                ),
            ]

399
400
401
402
403
404
405
            Xs, ys = [X_train], [X_train]
            Xvals, yvals = [X_val], [X_val]

            if predictor > 0.0:
                Xs, ys = X_train[:-1], [X_train[:-1], X_train[1:]]
                Xvals, yvals = X_val[:-1], [X_val[:-1], X_val[1:]]

406
            if phenotype_class > 0.0:
407
408
409
                ys += [y_train]
                yvals += [y_val]

410
            ae.fit(
411
412
                x=Xs,
                y=ys,
413
                epochs=epochs,
414
415
416
417
418
419
420
421
422
                batch_size=batch_size,
                verbose=1,
                validation_data=(
                    Xvals,
                    yvals,
                ),
                callbacks=callbacks_,
            )

423
            if not os.path.exists(os.path.join(output_path, "trained_weights")):
424
425
                os.makedirs("trained_weights")

426
            if save_weights:
427
428
                ae.save_weights(
                    os.path.join(
429
430
431
                        "{}".format(output_path),
                        "trained_weights",
                        "{}_final_weights.h5".format(run_ID),
432
433
                    )
                )
434

435
436
437
438
439
            if log_hparams:
                # Logparams to tensorboard
                tensorboard_metric_logging(
                    os.path.join(output_path, "hparams", run_ID),
                    logparam,
lucas_miranda's avatar
lucas_miranda committed
440
                    ae,
lucas_miranda's avatar
lucas_miranda committed
441
442
                    Xvals,
                    yvals[-1],
lucas_miranda's avatar
lucas_miranda committed
443
444
445
                    phenotype_class,
                    predictor,
                    rec,
446
                )
447

448
449
450
    return return_list


451
def tune_search(
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
    data: List[np.array],
    encoding_size: int,
    hypertun_trials: int,
    hpt_type: str,
    hypermodel: str,
    k: int,
    kl_warmup_epochs: int,
    loss: str,
    mmd_warmup_epochs: int,
    overlap_loss: float,
    phenotype_class: float,
    predictor: float,
    project_name: str,
    callbacks: List,
    n_epochs: int = 30,
    n_replicas: int = 1,
    outpath: str = ".",
469
) -> Union[bool, Tuple[Any, Any]]:
470
471
    """Define the search space using keras-tuner and bayesian optimization

472
473
474
475
476
477
478
479
480
481
482
    Parameters:
        - train (np.array): dataset to train the model on
        - test (np.array): dataset to validate the model on
        - hypertun_trials (int): number of Bayesian optimization iterations to run
        - hpt_type (str): specify one of Bayesian Optimization (bayopt) and Hyperband (hyperband)
        - hypermodel (str): hypermodel to load. Must be one of S2SAE (plain autoencoder)
        or S2SGMVAE (Gaussian Mixture Variational autoencoder).
        - k (int) number of components of the Gaussian Mixture
        - loss (str): one of [ELBO, MMD, ELBO+MMD]
        - overlap_loss (float): assigns as weight to an extra loss term which
        penalizes overlap between GM components
483
        - phenotype_class (float): adds an extra regularizing neural network to the model,
484
485
486
487
488
489
490
491
492
493
494
495
        which tries to predict the phenotype of the animal from which the sequence comes
        - predictor (float): adds an extra regularizing neural network to the model,
        which tries to predict the next frame from the current one
        - project_name (str): ID of the current run
        - callbacks (list): list of callbacks for the training loop
        - n_epochs (int): optional. Number of epochs to train each run for
        - n_replicas (int): optional. Number of replicas per parameter set. Higher values
         will yield more robust results, but will affect performance severely

    Returns:
        - best_hparams (dict): dictionary with the best retrieved hyperparameters
        - best_run (tf.keras.Model): trained instance of the best model found
496
497
498

    """

499
500
    X_train, y_train, X_val, y_val = data

501
502
503
504
    assert hpt_type in ["bayopt", "hyperband"], (
        "Invalid hyperparameter tuning framework. " "Select one of bayopt and hyperband"
    )

lucas_miranda's avatar
lucas_miranda committed
505
    if hypermodel == "S2SAE":  # pragma: no cover
506
        assert (
507
            predictor == 0.0 and phenotype_class == 0.0
508
        ), "Prediction branches are only available for variational models. See documentation for more details"
509
        batch_size = 1
510
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_AE(input_shape=X_train.shape)
511
512

    elif hypermodel == "S2SGMVAE":
513
        batch_size = 64
514
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_GMVAE(
515
            input_shape=X_train.shape,
516
            encoding=encoding_size,
517
            kl_warmup_epochs=kl_warmup_epochs,
518
            loss=loss,
519
            mmd_warmup_epochs=mmd_warmup_epochs,
520
            number_of_components=k,
521
            overlap_loss=overlap_loss,
522
            phenotype_predictor=phenotype_class,
523
            predictor=predictor,
524
        )
lucas_miranda's avatar
lucas_miranda committed
525

526
527
528
    else:
        return False

529
530
531
    hpt_params = {
        "hypermodel": hypermodel,
        "executions_per_trial": n_replicas,
532
533
534
        "logger": TensorBoardLogger(
            metrics=["val_mae"], logdir=os.path.join(outpath, "logged_hparams")
        ),
535
536
537
538
539
540
541
        "objective": "val_mae",
        "project_name": project_name,
        "tune_new_entries": True,
    }

    if hpt_type == "hyperband":
        tuner = Hyperband(
542
543
544
            directory=os.path.join(
                outpath, "HyperBandx_{}_{}".format(loss, str(date.today()))
            ),
545
546
            max_epochs=35,
            hyperband_iterations=hypertun_trials,
547
            factor=3,
548
549
550
551
            **hpt_params
        )
    else:
        tuner = BayesianOptimization(
552
553
554
            directory=os.path.join(
                outpath, "BayOpt_{}_{}".format(loss, str(date.today()))
            ),
555
556
557
            max_trials=hypertun_trials,
            **hpt_params
        )
558
559
560

    print(tuner.search_space_summary())

561
562
563
564
565
566
567
    Xs, ys = [X_train], [X_train]
    Xvals, yvals = [X_val], [X_val]

    if predictor > 0.0:
        Xs, ys = X_train[:-1], [X_train[:-1], X_train[1:]]
        Xvals, yvals = X_val[:-1], [X_val[:-1], X_val[1:]]

568
    if phenotype_class > 0.0:
569
570
571
        ys += [y_train]
        yvals += [y_val]

572
    tuner.search(
573
574
        Xs,
        ys,
575
        epochs=n_epochs,
576
        validation_data=(Xvals, yvals),
577
        verbose=1,
578
        batch_size=batch_size,
lucas_miranda's avatar
lucas_miranda committed
579
        callbacks=callbacks,
580
581
582
583
584
    )

    best_hparams = tuner.get_best_hyperparameters(num_trials=1)[0]
    best_run = tuner.hypermodel.build(best_hparams)

lucas_miranda's avatar
lucas_miranda committed
585
586
    print(tuner.results_summary())

587
    return best_hparams, best_run