train_utils.py 20.8 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
    run: int = False,
85
) -> List[Union[Any]]:
86
    """Generates callbacks for model training, including:
87
88
89
90
    - run_ID: run name, with coarse parameter details;
    - tensorboard_callback: for real-time visualization;
    - cp_callback: for checkpoint saving,
    - onecycle: for learning rate scheduling"""
91

92
93
94
95
96
97
98
99
    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"

100
    run_ID = "{}{}{}{}{}{}{}_{}".format(
101
        ("GMVAE" if variational else "AE"),
102
103
104
        ("_NextSeqPred={}".format(next_sequence_prediction) if variational else ""),
        ("_PhenoPred={}".format(phenotype_prediction) if variational else ""),
        ("_RuleBasedPred={}".format(rule_based_prediction) if variational else ""),
105
        ("_loss={}".format(loss) if variational else ""),
106
107
        ("_encoding={}".format(logparam["encoding"]) if logparam is not None else ""),
        ("_k={}".format(logparam["k"]) if logparam is not None else ""),
108
        ("_latreg={}".format(latreg)),
109
110
        ("_entknn={}".format(entropy_knn)),
        ("_run={}".format(run) if run else ""),
111
        ("_{}".format(datetime.now().strftime("%Y%m%d-%H%M%S")) if not run else ""),
112
113
    )

114
    log_dir = os.path.abspath(os.path.join(outpath, "fit", run_ID))
115
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
116
117
118
        log_dir=log_dir,
        histogram_freq=1,
        profile_batch=2,
119
120
    )

121
    entropy = deepof.model_utils.neighbor_latent_entropy(
122
        encoding_dim=logparam["encoding"],
123
        k=entropy_knn,
124
        samples=entropy_samples,
lucas_miranda's avatar
lucas_miranda committed
125
        validation_data=X_val,
126
        log_dir=os.path.join(outpath, "metrics", run_ID),
127
        variational=variational,
lucas_miranda's avatar
lucas_miranda committed
128
129
    )

130
    onecycle = deepof.model_utils.one_cycle_scheduler(
131
132
        X_train.shape[0] // batch_size * 250,
        max_rate=0.005,
133
        log_dir=os.path.join(outpath, "metrics", run_ID),
134
135
    )

136
    callbacks = [run_ID, tensorboard_callback, entropy, onecycle]
137
138
139

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
140
            os.path.join(outpath, "checkpoints", run_ID + "/cp-{epoch:04d}.ckpt"),
141
142
143
144
145
146
147
148
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
149
150


lucas_miranda's avatar
lucas_miranda committed
151
def log_hyperparameters(phenotype_class: float, rec: str):
lucas_miranda's avatar
lucas_miranda committed
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
    """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
203
def tensorboard_metric_logging(
204
205
206
207
208
    run_dir: str,
    hpms: Any,
    ae: Any,
    X_val: np.ndarray,
    y_val: np.ndarray,
209
210
211
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
212
    rec: str,
lucas_miranda's avatar
lucas_miranda committed
213
):
lucas_miranda's avatar
lucas_miranda committed
214
215
    """Autoencoder metric logging in tensorboard"""

216
217
    outputs = ae.predict(X_val)
    idx_generator = (idx for idx in range(len(outputs)))
lucas_miranda's avatar
lucas_miranda committed
218
219
220

    with tf.summary.create_file_writer(run_dir).as_default():
        hp.hparams(hpms)  # record the values used in this trial
221
222
        idx = next(idx_generator)

lucas_miranda's avatar
lucas_miranda committed
223
        val_mae = tf.reduce_mean(
224
            tf.keras.metrics.mean_absolute_error(y_val[idx], outputs[idx])
lucas_miranda's avatar
lucas_miranda committed
225
226
        )
        val_mse = tf.reduce_mean(
227
            tf.keras.metrics.mean_squared_error(y_val[idx], outputs[idx])
lucas_miranda's avatar
lucas_miranda committed
228
229
230
231
        )
        tf.summary.scalar("val_{}mae".format(rec), val_mae, step=1)
        tf.summary.scalar("val_{}mse".format(rec), val_mse, step=1)

232
        if next_sequence_prediction:
233
            idx = next(idx_generator)
lucas_miranda's avatar
lucas_miranda committed
234
            pred_mae = tf.reduce_mean(
235
                tf.keras.metrics.mean_absolute_error(y_val[idx], outputs[idx])
lucas_miranda's avatar
lucas_miranda committed
236
237
            )
            pred_mse = tf.reduce_mean(
238
                tf.keras.metrics.mean_squared_error(y_val[idx], outputs[idx])
239
240
241
242
243
244
            )
            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
245
246
            )

247
        if phenotype_prediction:
248
            idx = next(idx_generator)
249
250
251
            pheno_acc = tf.keras.metrics.binary_accuracy(
                y_val[idx], tf.squeeze(outputs[idx])
            )
252
            pheno_auc = tf.keras.metrics.AUC()
253
            pheno_auc.update_state(y_val[idx], outputs[idx])
254
            pheno_auc = pheno_auc.result().numpy()
lucas_miranda's avatar
lucas_miranda committed
255
256
257
258

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

259
        if rule_based_prediction:
260
            idx = next(idx_generator)
261
            rules_mae = tf.reduce_mean(
262
                tf.keras.metrics.mean_absolute_error(y_val[idx], outputs[idx])
263
264
            )
            rules_mse = tf.reduce_mean(
265
                tf.keras.metrics.mean_squared_error(y_val[idx], outputs[idx])
266
267
268
269
            )
            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
270

271
def autoencoder_fitting(
272
273
274
275
276
277
278
279
280
281
282
283
284
    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,
285
286
287
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
288
289
290
291
292
293
294
295
    pretrained: str,
    save_checkpoints: bool,
    save_weights: bool,
    variational: bool,
    reg_cat_clusters: bool,
    reg_cluster_variance: bool,
    entropy_samples: int,
    entropy_knn: int,
296
    run: int = run,
297
):
298
299
    """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""

300
    # Load data
301
302
303
304
305
    X_train, y_train, X_val, y_val = preprocessed_object

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

306
    # Defines what to log on tensorboard (useful for trying out different models)
307
308
    logparam = {
        "encoding": encoding_size,
309
        "k": n_components,
310
311
        "loss": loss,
    }
312
313
    if phenotype_prediction:
        logparam["pheno_weight"] = phenotype_prediction
314

315
    # Load callbacks
316
    run_ID, *cbacks = get_callbacks(
317
        X_train=X_train,
lucas_miranda's avatar
lucas_miranda committed
318
        X_val=(X_val if X_val.shape != (0,) else None),
319
320
321
        batch_size=batch_size,
        cp=save_checkpoints,
        variational=variational,
322
323
324
        next_sequence_prediction=next_sequence_prediction,
        phenotype_prediction=phenotype_prediction,
        rule_based_prediction=rule_based_prediction,
325
        loss=loss,
326
        entropy_samples=entropy_samples,
327
        entropy_knn=entropy_knn,
328
        reg_cat_clusters=reg_cat_clusters,
329
        reg_cluster_variance=reg_cluster_variance,
330
331
        logparam=logparam,
        outpath=output_path,
332
        run=run,
333
    )
334
335
    if not log_history:
        cbacks = cbacks[1:]
336

337
    # Logs hyperparameters to tensorboard
338
    rec = "reconstruction_" if phenotype_prediction else ""
339
    if log_hparams:
340
        logparams, metrics = log_hyperparameters(phenotype_prediction, rec)
341
342

        with tf.summary.create_file_writer(
343
            os.path.join(output_path, "hparams", run_ID)
344
345
346
347
348
        ).as_default():
            hp.hparams_config(
                hparams=logparams,
                metrics=metrics,
            )
349

350
351
352
353
354
355
356
357
    # Gets the number of rule-based features
    try:
        rule_based_features = (
            y_train.shape[1] if not phenotype_prediction else y_train.shape[1] - 1
        )
    except IndexError:
        rule_based_features = 0

358
    # Build models
359
360
361
362
363
364
365
    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:
366
367
368
369
370
371
372
373
        (
            encoder,
            generator,
            grouper,
            ae,
            prior,
            posterior,
        ) = deepof.models.SEQ_2_SEQ_GMVAE(
374
375
376
377
378
379
380
381
382
383
384
            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,
385
386
387
            next_sequence_prediction=next_sequence_prediction,
            phenotype_prediction=phenotype_prediction,
            rule_based_prediction=rule_based_prediction,
388
            rule_based_features=rule_based_features,
389
390
            reg_cat_clusters=reg_cat_clusters,
            reg_cluster_variance=reg_cluster_variance,
391
392
393
        ).build(
            X_train.shape
        )
394
395
396
        return_list = (encoder, generator, grouper, ae)

    if pretrained:
397
        # If pretrained models are specified, load weights and return
398
399
400
401
402
403
404
405
406
        ae.load_weights(pretrained)
        return return_list

    else:
        if not variational:

            ae.fit(
                x=X_train,
                y=X_train,
407
                epochs=epochs,
408
409
410
                batch_size=batch_size,
                verbose=1,
                validation_data=(X_val, X_val),
411
                callbacks=cbacks
412
413
414
415
416
417
418
419
                + [
                    CustomStopper(
                        monitor="val_loss",
                        patience=5,
                        restore_best_weights=True,
                        start_epoch=max(kl_warmup, mmd_warmup),
                    ),
                ],
420
421
            )

422
423
424
            if save_weights:
                ae.save_weights("{}_final_weights.h5".format(run_ID))

425
426
        else:

427
            callbacks_ = cbacks + [
428
429
430
431
                CustomStopper(
                    monitor="val_loss",
                    patience=5,
                    restore_best_weights=True,
432
                    start_epoch=max(kl_warmup, mmd_warmup),
433
434
435
                ),
            ]

436
437
438
            Xs, ys = [X_train], [X_train]
            Xvals, yvals = [X_val], [X_val]

439
            if next_sequence_prediction > 0.0:
440
441
442
                Xs, ys = X_train[:-1], [X_train[:-1], X_train[1:]]
                Xvals, yvals = X_val[:-1], [X_val[:-1], X_val[1:]]

443
            if phenotype_prediction > 0.0:
444
445
                ys += [y_train[-Xs.shape[0] :, 0]]
                yvals += [y_val[-Xs.shape[0] :, 0]]
446
447
448
449
450
451

                # 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:
452
453
                ys += [y_train[-Xs.shape[0] :]]
                yvals += [y_val[-Xs.shape[0] :]]
454

455
            ae.fit(
456
457
                x=Xs,
                y=ys,
458
                epochs=epochs,
459
460
461
462
463
464
465
466
467
                batch_size=batch_size,
                verbose=1,
                validation_data=(
                    Xvals,
                    yvals,
                ),
                callbacks=callbacks_,
            )

468
            if not os.path.exists(os.path.join(output_path, "trained_weights")):
469
470
                os.makedirs("trained_weights")

471
            if save_weights:
472
473
                ae.save_weights(
                    os.path.join(
474
475
476
                        "{}".format(output_path),
                        "trained_weights",
                        "{}_final_weights.h5".format(run_ID),
477
478
                    )
                )
479

480
481
482
            if log_hparams:
                # Logparams to tensorboard
                tensorboard_metric_logging(
483
484
485
486
487
488
489
490
491
                    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,
492
                )
493

494
495
496
    return return_list


497
def tune_search(
498
499
500
501
502
503
504
505
506
507
    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,
508
509
510
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
511
512
513
514
515
    project_name: str,
    callbacks: List,
    n_epochs: int = 30,
    n_replicas: int = 1,
    outpath: str = ".",
516
) -> Union[bool, Tuple[Any, Any]]:
517
518
    """Define the search space using keras-tuner and bayesian optimization

519
520
521
522
523
524
525
526
527
528
529
    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
530
        - phenotype_class (float): adds an extra regularizing neural network to the model,
531
532
533
534
535
536
537
538
539
540
541
542
        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
543
544
545

    """

546
547
    X_train, y_train, X_val, y_val = data

548
549
550
551
    assert hpt_type in ["bayopt", "hyperband"], (
        "Invalid hyperparameter tuning framework. " "Select one of bayopt and hyperband"
    )

lucas_miranda's avatar
lucas_miranda committed
552
    if hypermodel == "S2SAE":  # pragma: no cover
553
        assert (
554
            next_sequence_prediction == 0.0 and phenotype_prediction == 0.0
555
        ), "Prediction branches are only available for variational models. See documentation for more details"
556
        batch_size = 1
557
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_AE(input_shape=X_train.shape)
558
559

    elif hypermodel == "S2SGMVAE":
560
        batch_size = 64
561
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_GMVAE(
562
            input_shape=X_train.shape,
563
            encoding=encoding_size,
564
            kl_warmup_epochs=kl_warmup_epochs,
565
            loss=loss,
566
            mmd_warmup_epochs=mmd_warmup_epochs,
567
            number_of_components=k,
568
            overlap_loss=overlap_loss,
569
570
571
            next_sequence_prediction=next_sequence_prediction,
            phenotype_prediction=phenotype_prediction,
            rule_based_prediction=rule_based_prediction,
572
573
574
            rule_based_features=(
                y_train.shape[1] if not phenotype_prediction else y_train.shape[1] - 1
            ),
575
        )
lucas_miranda's avatar
lucas_miranda committed
576

577
578
579
    else:
        return False

580
581
582
    hpt_params = {
        "hypermodel": hypermodel,
        "executions_per_trial": n_replicas,
583
584
585
        "logger": TensorBoardLogger(
            metrics=["val_mae"], logdir=os.path.join(outpath, "logged_hparams")
        ),
586
587
588
589
590
591
592
        "objective": "val_mae",
        "project_name": project_name,
        "tune_new_entries": True,
    }

    if hpt_type == "hyperband":
        tuner = Hyperband(
593
594
595
            directory=os.path.join(
                outpath, "HyperBandx_{}_{}".format(loss, str(date.today()))
            ),
596
597
            max_epochs=35,
            hyperband_iterations=hypertun_trials,
598
            factor=3,
599
600
601
602
            **hpt_params
        )
    else:
        tuner = BayesianOptimization(
603
604
605
            directory=os.path.join(
                outpath, "BayOpt_{}_{}".format(loss, str(date.today()))
            ),
606
607
608
            max_trials=hypertun_trials,
            **hpt_params
        )
609
610
611

    print(tuner.search_space_summary())

612
613
614
    Xs, ys = [X_train], [X_train]
    Xvals, yvals = [X_val], [X_val]

615
    if next_sequence_prediction > 0.0:
616
617
618
        Xs, ys = X_train[:-1], [X_train[:-1], X_train[1:]]
        Xvals, yvals = X_val[:-1], [X_val[:-1], X_val[1:]]

619
620
621
622
623
624
625
626
627
    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:
628
629
630
        ys += [y_train]
        yvals += [y_val]

631
    tuner.search(
632
633
        Xs,
        ys,
634
        epochs=n_epochs,
635
        validation_data=(Xvals, yvals),
636
        verbose=1,
637
        batch_size=batch_size,
lucas_miranda's avatar
lucas_miranda committed
638
        callbacks=callbacks,
639
640
641
642
643
    )

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

lucas_miranda's avatar
lucas_miranda committed
644
645
    print(tuner.results_summary())

646
    return best_hparams, best_run