train_utils.py 20.9 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
    X_train: np.array,
    batch_size: int,
    variational: bool,
72
73
74
    phenotype_prediction: float,
    next_sequence_prediction: float,
    rule_based_prediction: float,
75
76
77
78
79
80
81
82
83
    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 = ".",
84
) -> List[Union[Any]]:
85
    """Generates callbacks for model training, including:
86
87
88
89
    - run_ID: run name, with coarse parameter details;
    - tensorboard_callback: for real-time visualization;
    - cp_callback: for checkpoint saving,
    - onecycle: for learning rate scheduling"""
90

91
92
93
94
95
96
97
98
    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"

99
    run_ID = "{}{}{}{}{}{}{}_{}".format(
100
        ("GMVAE" if variational else "AE"),
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
        (
            "_NextSeqPred={}".format(next_sequence_prediction)
            if next_sequence_prediction > 0 and variational
            else ""
        ),
        (
            "_PhenoPred={}".format(phenotype_prediction)
            if phenotype_prediction > 0
            else ""
        ),
        (
            "_RuleBasedPred={}".format(rule_based_prediction)
            if rule_based_prediction > 0
            else ""
        ),
116
        ("_loss={}".format(loss) if variational else ""),
117
118
        ("_encoding={}".format(logparam["encoding"]) if logparam is not None else ""),
        ("_k={}".format(logparam["k"]) if logparam is not None else ""),
119
        ("_latreg={}".format(latreg)),
120
        ("entknn={}".format(entropy_knn)),
121
        (datetime.now().strftime("%Y%m%d-%H%M%S")),
122
123
    )

124
    log_dir = os.path.abspath(os.path.join(outpath, "fit", run_ID))
125
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
126
127
128
        log_dir=log_dir,
        histogram_freq=1,
        profile_batch=2,
129
130
    )

131
    entropy = deepof.model_utils.neighbor_latent_entropy(
132
        encoding_dim=logparam["encoding"],
133
        k=entropy_knn,
134
        samples=entropy_samples,
lucas_miranda's avatar
lucas_miranda committed
135
        validation_data=X_val,
136
        log_dir=os.path.join(outpath, "metrics", run_ID),
137
        variational=variational,
lucas_miranda's avatar
lucas_miranda committed
138
139
    )

140
    onecycle = deepof.model_utils.one_cycle_scheduler(
141
142
        X_train.shape[0] // batch_size * 250,
        max_rate=0.005,
143
        log_dir=os.path.join(outpath, "metrics", run_ID),
144
145
    )

146
    callbacks = [run_ID, tensorboard_callback, entropy, onecycle]
147
148
149

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
150
            os.path.join(outpath, "checkpoints", run_ID + "/cp-{epoch:04d}.ckpt"),
151
152
153
154
155
156
157
158
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
159
160


lucas_miranda's avatar
lucas_miranda committed
161
def log_hyperparameters(phenotype_class: float, rec: str):
lucas_miranda's avatar
lucas_miranda committed
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
    """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
213
def tensorboard_metric_logging(
214
215
216
217
218
    run_dir: str,
    hpms: Any,
    ae: Any,
    X_val: np.ndarray,
    y_val: np.ndarray,
219
220
221
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
222
    rec: str,
lucas_miranda's avatar
lucas_miranda committed
223
):
lucas_miranda's avatar
lucas_miranda committed
224
225
226
    """Autoencoder metric logging in tensorboard"""

    output = ae.predict(X_val)
227
228
229
230
231
232

    if any([next_sequence_prediction, phenotype_prediction, rule_based_prediction]):
        reconstruction_pred, reconstruction_true = output[0], y_val[0]
        nextseq_pred, nextseq_true = output[1], y_val[1]
        pheno_pred, pheno_true = output[2], y_val[2]
        rules_pred, rules_true = output[3], y_val[3]
lucas_miranda's avatar
lucas_miranda committed
233
234
235
236
237
238
    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(
239
240
241
            tf.keras.metrics.mean_absolute_error(
                reconstruction_pred, reconstruction_true
            )
lucas_miranda's avatar
lucas_miranda committed
242
243
        )
        val_mse = tf.reduce_mean(
244
245
246
            tf.keras.metrics.mean_squared_error(
                reconstruction_pred, reconstruction_true
            )
lucas_miranda's avatar
lucas_miranda committed
247
248
249
250
        )
        tf.summary.scalar("val_{}mae".format(rec), val_mae, step=1)
        tf.summary.scalar("val_{}mse".format(rec), val_mse, step=1)

251
        if next_sequence_prediction:
lucas_miranda's avatar
lucas_miranda committed
252
            pred_mae = tf.reduce_mean(
253
                tf.keras.metrics.mean_absolute_error(nextseq_pred, nextseq_true)
lucas_miranda's avatar
lucas_miranda committed
254
255
            )
            pred_mse = tf.reduce_mean(
256
257
258
259
260
261
262
                tf.keras.metrics.mean_squared_error(nextseq_pred, nextseq_true)
            )
            tf.summary.scalar(
                "val_next_sequence_prediction_mae".format(rec), pred_mae, step=1
            )
            tf.summary.scalar(
                "val_next_sequence_prediction_mse".format(rec), pred_mse, step=1
lucas_miranda's avatar
lucas_miranda committed
263
264
            )

265
266
267
268
269
270
271
        if phenotype_prediction:
            pheno_acc = tf.keras.metrics.binary_accuracy(
                pheno_true, tf.squeeze(pheno_pred)
            )
            pheno_auc = tf.keras.metrics.AUC()
            pheno_auc.update_state(pheno_true, pheno_pred)
            pheno_auc = pheno_auc.result().numpy()
lucas_miranda's avatar
lucas_miranda committed
272
273
274
275

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

276
277
278
279
280
281
282
283
284
285
        if rule_based_prediction:
            rules_mae = tf.reduce_mean(
                tf.keras.metrics.mean_absolute_error(rules_pred, rules_true)
            )
            rules_mse = tf.reduce_mean(
                tf.keras.metrics.mean_squared_error(rules_pred, rules_true)
            )
            tf.summary.scalar("val_prediction_mae".format(rec), rules_mae, step=1)
            tf.summary.scalar("val_prediction_mse".format(rec), rules_mse, step=1)

lucas_miranda's avatar
lucas_miranda committed
286

287
def autoencoder_fitting(
288
289
290
291
292
293
294
295
296
297
298
299
300
    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,
301
302
303
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
304
305
306
307
308
309
310
311
    pretrained: str,
    save_checkpoints: bool,
    save_weights: bool,
    variational: bool,
    reg_cat_clusters: bool,
    reg_cluster_variance: bool,
    entropy_samples: int,
    entropy_knn: int,
312
):
313
314
    """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""

315
    # Load data
316
317
318
319
320
    X_train, y_train, X_val, y_val = preprocessed_object

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

321
    # Defines what to log on tensorboard (useful for trying out different models)
322
323
    logparam = {
        "encoding": encoding_size,
324
        "k": n_components,
325
326
        "loss": loss,
    }
327
328
    if phenotype_prediction:
        logparam["pheno_weight"] = phenotype_prediction
329

330
    # Load callbacks
331
    run_ID, *cbacks = get_callbacks(
332
        X_train=X_train,
lucas_miranda's avatar
lucas_miranda committed
333
        X_val=(X_val if X_val.shape != (0,) else None),
334
335
336
        batch_size=batch_size,
        cp=save_checkpoints,
        variational=variational,
337
338
339
        next_sequence_prediction=next_sequence_prediction,
        phenotype_prediction=phenotype_prediction,
        rule_based_prediction=rule_based_prediction,
340
        loss=loss,
341
        entropy_samples=entropy_samples,
342
        entropy_knn=entropy_knn,
343
        reg_cat_clusters=reg_cat_clusters,
344
        reg_cluster_variance=reg_cluster_variance,
345
346
347
        logparam=logparam,
        outpath=output_path,
    )
348
349
    if not log_history:
        cbacks = cbacks[1:]
350

351
    # Logs hyperparameters to tensorboard
352
    rec = "reconstruction_" if phenotype_prediction else ""
353
    if log_hparams:
354
        logparams, metrics = log_hyperparameters(phenotype_prediction, rec)
355
356

        with tf.summary.create_file_writer(
357
            os.path.join(output_path, "hparams", run_ID)
358
359
360
361
362
        ).as_default():
            hp.hparams_config(
                hparams=logparams,
                metrics=metrics,
            )
363

364
    # Build models
365
366
367
368
369
370
371
    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:
372
373
374
375
376
377
378
379
        (
            encoder,
            generator,
            grouper,
            ae,
            prior,
            posterior,
        ) = deepof.models.SEQ_2_SEQ_GMVAE(
380
381
382
383
384
385
386
387
388
389
390
            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,
391
392
393
394
395
396
            next_sequence_prediction=next_sequence_prediction,
            phenotype_prediction=phenotype_prediction,
            rule_based_prediction=rule_based_prediction,
            rule_based_features=(
                y_train.shape[1] if not phenotype_prediction else y_train.shape[1] - 1
            ),
397
398
            reg_cat_clusters=reg_cat_clusters,
            reg_cluster_variance=reg_cluster_variance,
399
400
401
        ).build(
            X_train.shape
        )
402
403
404
        return_list = (encoder, generator, grouper, ae)

    if pretrained:
405
        # If pretrained models are specified, load weights and return
406
407
408
409
410
411
412
413
414
        ae.load_weights(pretrained)
        return return_list

    else:
        if not variational:

            ae.fit(
                x=X_train,
                y=X_train,
415
                epochs=epochs,
416
417
418
                batch_size=batch_size,
                verbose=1,
                validation_data=(X_val, X_val),
419
                callbacks=cbacks
420
421
422
423
424
425
426
427
                + [
                    CustomStopper(
                        monitor="val_loss",
                        patience=5,
                        restore_best_weights=True,
                        start_epoch=max(kl_warmup, mmd_warmup),
                    ),
                ],
428
429
            )

430
431
432
            if save_weights:
                ae.save_weights("{}_final_weights.h5".format(run_ID))

433
434
        else:

435
            callbacks_ = cbacks + [
436
437
438
439
                CustomStopper(
                    monitor="val_loss",
                    patience=5,
                    restore_best_weights=True,
440
                    start_epoch=max(kl_warmup, mmd_warmup),
441
442
443
                ),
            ]

444
445
446
            Xs, ys = [X_train], [X_train]
            Xvals, yvals = [X_val], [X_val]

447
            if next_sequence_prediction > 0.0:
448
449
450
                Xs, ys = X_train[:-1], [X_train[:-1], X_train[1:]]
                Xvals, yvals = X_val[:-1], [X_val[:-1], X_val[1:]]

451
452
453
454
455
456
457
458
459
            if phenotype_prediction > 0.0:
                ys += [y_train[:, 0]]
                yvals += [y_val[:, 0]]

                # Remove the used column (phenotype) from both y arrays
                y_train = y_train[:, 1:]
                y_val = y_val[:, 1:]

            if rule_based_prediction > 0.0:
460
461
462
                ys += [y_train]
                yvals += [y_val]

463
            ae.fit(
464
465
                x=Xs,
                y=ys,
466
                epochs=epochs,
467
468
469
470
471
472
473
474
475
                batch_size=batch_size,
                verbose=1,
                validation_data=(
                    Xvals,
                    yvals,
                ),
                callbacks=callbacks_,
            )

476
            if not os.path.exists(os.path.join(output_path, "trained_weights")):
477
478
                os.makedirs("trained_weights")

479
            if save_weights:
480
481
                ae.save_weights(
                    os.path.join(
482
483
484
                        "{}".format(output_path),
                        "trained_weights",
                        "{}_final_weights.h5".format(run_ID),
485
486
                    )
                )
487

488
489
490
            if log_hparams:
                # Logparams to tensorboard
                tensorboard_metric_logging(
491
492
493
494
495
496
497
498
499
                    run_dir=os.path.join(output_path, "hparams", run_ID),
                    hpms=logparam,
                    ae=ae,
                    X_val=Xvals,
                    y_val=yvals,
                    next_sequence_prediction=next_sequence_prediction,
                    phenotype_prediction=phenotype_prediction,
                    rule_based_prediction=rule_based_prediction,
                    rec=rec,
500
                )
501

502
503
504
    return return_list


505
def tune_search(
506
507
508
509
510
511
512
513
514
515
    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,
516
517
518
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
519
520
521
522
523
    project_name: str,
    callbacks: List,
    n_epochs: int = 30,
    n_replicas: int = 1,
    outpath: str = ".",
524
) -> Union[bool, Tuple[Any, Any]]:
525
526
    """Define the search space using keras-tuner and bayesian optimization

527
528
529
530
531
532
533
534
535
536
537
    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
538
        - phenotype_class (float): adds an extra regularizing neural network to the model,
539
540
541
542
543
544
545
546
547
548
549
550
        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
551
552
553

    """

554
555
    X_train, y_train, X_val, y_val = data

556
557
558
559
    assert hpt_type in ["bayopt", "hyperband"], (
        "Invalid hyperparameter tuning framework. " "Select one of bayopt and hyperband"
    )

lucas_miranda's avatar
lucas_miranda committed
560
    if hypermodel == "S2SAE":  # pragma: no cover
561
        assert (
562
            next_sequence_prediction == 0.0 and phenotype_prediction == 0.0
563
        ), "Prediction branches are only available for variational models. See documentation for more details"
564
        batch_size = 1
565
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_AE(input_shape=X_train.shape)
566
567

    elif hypermodel == "S2SGMVAE":
568
        batch_size = 64
569
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_GMVAE(
570
            input_shape=X_train.shape,
571
            encoding=encoding_size,
572
            kl_warmup_epochs=kl_warmup_epochs,
573
            loss=loss,
574
            mmd_warmup_epochs=mmd_warmup_epochs,
575
            number_of_components=k,
576
            overlap_loss=overlap_loss,
577
578
579
            next_sequence_prediction=next_sequence_prediction,
            phenotype_prediction=phenotype_prediction,
            rule_based_prediction=rule_based_prediction,
580
581
582
            rule_based_features=(
                y_train.shape[1] if not phenotype_prediction else y_train.shape[1] - 1
            ),
583
        )
lucas_miranda's avatar
lucas_miranda committed
584

585
586
587
    else:
        return False

588
589
590
    hpt_params = {
        "hypermodel": hypermodel,
        "executions_per_trial": n_replicas,
591
592
593
        "logger": TensorBoardLogger(
            metrics=["val_mae"], logdir=os.path.join(outpath, "logged_hparams")
        ),
594
595
596
597
598
599
600
        "objective": "val_mae",
        "project_name": project_name,
        "tune_new_entries": True,
    }

    if hpt_type == "hyperband":
        tuner = Hyperband(
601
602
603
            directory=os.path.join(
                outpath, "HyperBandx_{}_{}".format(loss, str(date.today()))
            ),
604
605
            max_epochs=35,
            hyperband_iterations=hypertun_trials,
606
            factor=3,
607
608
609
610
            **hpt_params
        )
    else:
        tuner = BayesianOptimization(
611
612
613
            directory=os.path.join(
                outpath, "BayOpt_{}_{}".format(loss, str(date.today()))
            ),
614
615
616
            max_trials=hypertun_trials,
            **hpt_params
        )
617
618
619

    print(tuner.search_space_summary())

620
621
622
    Xs, ys = [X_train], [X_train]
    Xvals, yvals = [X_val], [X_val]

623
    if next_sequence_prediction > 0.0:
624
625
626
        Xs, ys = X_train[:-1], [X_train[:-1], X_train[1:]]
        Xvals, yvals = X_val[:-1], [X_val[:-1], X_val[1:]]

627
628
629
630
631
632
633
634
635
    if phenotype_prediction > 0.0:
        ys += [y_train[:, 0]]
        yvals += [y_val[:, 0]]

        # Remove the used column (phenotype) from both y arrays
        y_train = y_train[:, 1:]
        y_val = y_val[:, 1:]

    if rule_based_prediction > 0.0:
636
637
638
        ys += [y_train]
        yvals += [y_val]

639
    tuner.search(
640
641
        Xs,
        ys,
642
        epochs=n_epochs,
643
        validation_data=(Xvals, yvals),
644
        verbose=1,
645
        batch_size=batch_size,
lucas_miranda's avatar
lucas_miranda committed
646
        callbacks=callbacks,
647
648
649
650
651
    )

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

lucas_miranda's avatar
lucas_miranda committed
652
653
    print(tuner.results_summary())

654
    return best_hparams, best_run