train_utils.py 18.7 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

"""

11
from datetime import date, datetime
12
from kerastuner import BayesianOptimization, Hyperband
13
from kerastuner import HyperParameters
14
from kerastuner_tensorboard_logger import TensorBoardLogger
lucas_miranda's avatar
lucas_miranda committed
15
from sklearn.metrics import roc_auc_score
16
from tensorboard.plugins.hparams import api as hp
17
from typing import Tuple, Union, Any, List
18
19
20
21
22
23
24
import deepof.hypermodels
import deepof.model_utils
import numpy as np
import os
import pickle
import tensorflow as tf

25
26
27
28
29
# Ignore warning with no downstream effect
tf.get_logger().setLevel("ERROR")
tf.autograph.set_verbosity(0)


30
class CustomStopper(tf.keras.callbacks.EarlyStopping):
31
    """ Custom early stopping callback. Prevents the model from stopping before warmup is over """
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48

    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)


49
50
51
52
53
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(
54
55
56
57
58
            os.path.join(
                train_path,
                [i for i in os.listdir(train_path) if i.endswith(".pkl")][0],
            ),
            "rb",
59
60
61
62
63
64
65
66
67
        ) as handle:
            treatment_dict = pickle.load(handle)
    except IndexError:
        treatment_dict = None

    return treatment_dict


def get_callbacks(
68
69
70
71
72
73
74
75
76
77
78
79
80
81
    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_radius: float = None,
    logparam: dict = None,
    outpath: str = ".",
82
) -> List[Union[Any]]:
83
    """Generates callbacks for model training, including:
84
85
86
87
    - run_ID: run name, with coarse parameter details;
    - tensorboard_callback: for real-time visualization;
    - cp_callback: for checkpoint saving,
    - onecycle: for learning rate scheduling"""
88

89
90
91
92
93
94
95
96
    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"

97
    run_ID = "{}{}{}{}{}{}{}_{}".format(
98
        ("GMVAE" if variational else "AE"),
lucas_miranda's avatar
lucas_miranda committed
99
        ("_Pred={}".format(predictor) if predictor > 0 and variational else ""),
100
        ("_Pheno={}".format(phenotype_class) if phenotype_class > 0 else ""),
101
        ("_loss={}".format(loss) if variational else ""),
102
103
        ("_encoding={}".format(logparam["encoding"]) if logparam is not None else ""),
        ("_k={}".format(logparam["k"]) if logparam is not None else ""),
104
        ("_latreg={}".format(latreg)),
105
        (datetime.now().strftime("%Y%m%d-%H%M%S")),
106
107
    )

108
    log_dir = os.path.abspath(os.path.join(outpath, "fit", run_ID))
109
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
110
111
112
        log_dir=log_dir,
        histogram_freq=1,
        profile_batch=2,
113
114
    )

115
    entropy = deepof.model_utils.neighbor_cluster_purity(
116
117
118
119
        r=(
            entropy_radius
            if entropy_radius is not None
            else 0.15 * logparam["encoding"]
120
            - 0.18  # equation derived empirically to keep neighbor number constant.
121
122
            # See examples/set_default_entropy_radius.ipynb for details
        ),
123
        samples=entropy_samples,
lucas_miranda's avatar
lucas_miranda committed
124
        validation_data=X_val,
125
        log_dir=os.path.join(outpath, "metrics", run_ID),
126
        variational=variational,
lucas_miranda's avatar
lucas_miranda committed
127
128
    )

129
    onecycle = deepof.model_utils.one_cycle_scheduler(
130
131
        X_train.shape[0] // batch_size * 250,
        max_rate=0.005,
132
        log_dir=os.path.join(outpath, "metrics", run_ID),
133
134
    )

135
    callbacks = [run_ID, tensorboard_callback, entropy, onecycle]
136
137
138

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
139
            os.path.join(outpath, "checkpoints", run_ID + "/cp-{epoch:04d}.ckpt"),
140
141
142
143
144
145
146
147
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
148
149


lucas_miranda's avatar
lucas_miranda committed
150
def log_hyperparameters(phenotype_class: float, rec: str):
lucas_miranda's avatar
lucas_miranda committed
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
199
200
201
    """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
202
def tensorboard_metric_logging(
203
204
205
206
207
208
209
210
    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
211
):
lucas_miranda's avatar
lucas_miranda committed
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
248
249
250
    """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)


251
def autoencoder_fitting(
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
    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_radius: float,
    entropy_samples: int,
275
):
276
277
    """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""

278
    # Load data
279
280
281
282
283
    X_train, y_train, X_val, y_val = preprocessed_object

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

284
    # Defines what to log on tensorboard (useful for trying out different models)
285
286
    logparam = {
        "encoding": encoding_size,
287
        "k": n_components,
288
289
290
291
292
        "loss": loss,
    }
    if phenotype_class:
        logparam["pheno_weight"] = phenotype_class

293
    # Load callbacks
294
    run_ID, *cbacks = get_callbacks(
295
        X_train=X_train,
lucas_miranda's avatar
lucas_miranda committed
296
        X_val=(X_val if X_val.shape != (0,) else None),
297
298
299
        batch_size=batch_size,
        cp=save_checkpoints,
        variational=variational,
300
        phenotype_class=phenotype_class,
301
302
        predictor=predictor,
        loss=loss,
303
304
        entropy_radius=entropy_radius,
        entropy_samples=entropy_samples,
305
        reg_cat_clusters=reg_cat_clusters,
306
        reg_cluster_variance=reg_cluster_variance,
307
308
309
        logparam=logparam,
        outpath=output_path,
    )
310
311
    if not log_history:
        cbacks = cbacks[1:]
312

313
    # Logs hyperparameters to tensorboard
lucas_miranda's avatar
lucas_miranda committed
314
    rec = "reconstruction_" if phenotype_class else ""
315
    if log_hparams:
lucas_miranda's avatar
lucas_miranda committed
316
        logparams, metrics = log_hyperparameters(phenotype_class, rec)
317
318

        with tf.summary.create_file_writer(
319
            os.path.join(output_path, "hparams", run_ID)
320
321
322
323
324
        ).as_default():
            hp.hparams_config(
                hparams=logparams,
                metrics=metrics,
            )
325

326
    # Build models
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
    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:
        (
            encoder,
            generator,
            grouper,
            ae,
            kl_warmup_callback,
            mmd_warmup_callback,
        ) = deepof.models.SEQ_2_SEQ_GMVAE(
            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,
355
356
            reg_cat_clusters=reg_cat_clusters,
            reg_cluster_variance=reg_cluster_variance,
357
358
359
360
361
362
        ).build(
            X_train.shape
        )
        return_list = (encoder, generator, grouper, ae)

    if pretrained:
363
        # If pretrained models are specified, load weights and return
364
365
366
367
368
369
370
371
372
        ae.load_weights(pretrained)
        return return_list

    else:
        if not variational:

            ae.fit(
                x=X_train,
                y=X_train,
373
                epochs=epochs,
374
375
376
                batch_size=batch_size,
                verbose=1,
                validation_data=(X_val, X_val),
377
                callbacks=cbacks
378
379
380
381
382
383
384
385
                + [
                    CustomStopper(
                        monitor="val_loss",
                        patience=5,
                        restore_best_weights=True,
                        start_epoch=max(kl_warmup, mmd_warmup),
                    ),
                ],
386
387
            )

388
389
390
            if save_weights:
                ae.save_weights("{}_final_weights.h5".format(run_ID))

391
392
        else:

393
            callbacks_ = cbacks + [
394
395
396
397
                CustomStopper(
                    monitor="val_loss",
                    patience=5,
                    restore_best_weights=True,
398
                    start_epoch=max(kl_warmup, mmd_warmup),
399
400
401
                ),
            ]

402
            if "ELBO" in loss and kl_warmup > 0:
403
404
                # noinspection PyUnboundLocalVariable
                callbacks_.append(kl_warmup_callback)
405
            if "MMD" in loss and mmd_warmup > 0:
406
407
408
                # noinspection PyUnboundLocalVariable
                callbacks_.append(mmd_warmup_callback)

409
410
411
412
413
414
415
            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:]]

416
            if phenotype_class > 0.0:
417
418
419
                ys += [y_train]
                yvals += [y_val]

420
            ae.fit(
421
422
                x=Xs,
                y=ys,
423
                epochs=epochs,
424
425
426
427
428
429
430
431
432
                batch_size=batch_size,
                verbose=1,
                validation_data=(
                    Xvals,
                    yvals,
                ),
                callbacks=callbacks_,
            )

433
            if not os.path.exists(os.path.join(output_path, "trained_weights")):
434
435
                os.makedirs("trained_weights")

436
            if save_weights:
437
438
                ae.save_weights(
                    os.path.join(
439
440
441
                        "{}".format(output_path),
                        "trained_weights",
                        "{}_final_weights.h5".format(run_ID),
442
443
                    )
                )
444

445
446
447
448
449
            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
450
                    ae,
lucas_miranda's avatar
lucas_miranda committed
451
452
                    Xvals,
                    yvals[-1],
lucas_miranda's avatar
lucas_miranda committed
453
454
455
                    phenotype_class,
                    predictor,
                    rec,
456
                )
457

458
459
460
    return return_list


461
def tune_search(
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
    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 = ".",
479
) -> Union[bool, Tuple[Any, Any]]:
480
481
    """Define the search space using keras-tuner and bayesian optimization

482
483
484
485
486
487
488
489
490
491
492
    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
493
        - phenotype_class (float): adds an extra regularizing neural network to the model,
494
495
496
497
498
499
500
501
502
503
504
505
        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
506
507
508

    """

509
510
    X_train, y_train, X_val, y_val = data

511
512
513
514
    assert hpt_type in ["bayopt", "hyperband"], (
        "Invalid hyperparameter tuning framework. " "Select one of bayopt and hyperband"
    )

lucas_miranda's avatar
lucas_miranda committed
515
    if hypermodel == "S2SAE":  # pragma: no cover
516
        assert (
517
            predictor == 0.0 and phenotype_class == 0.0
518
        ), "Prediction branches are only available for variational models. See documentation for more details"
519
        batch_size = 1
520
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_AE(input_shape=X_train.shape)
521
522

    elif hypermodel == "S2SGMVAE":
523
        batch_size = 64
524
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_GMVAE(
525
            input_shape=X_train.shape,
526
            encoding=encoding_size,
527
            kl_warmup_epochs=kl_warmup_epochs,
528
            loss=loss,
529
            mmd_warmup_epochs=mmd_warmup_epochs,
530
            number_of_components=k,
531
            overlap_loss=overlap_loss,
532
            phenotype_predictor=phenotype_class,
533
            predictor=predictor,
534
        )
lucas_miranda's avatar
lucas_miranda committed
535

536
537
538
    else:
        return False

539
540
541
    hpt_params = {
        "hypermodel": hypermodel,
        "executions_per_trial": n_replicas,
542
543
544
        "logger": TensorBoardLogger(
            metrics=["val_mae"], logdir=os.path.join(outpath, "logged_hparams")
        ),
545
546
547
548
549
550
551
        "objective": "val_mae",
        "project_name": project_name,
        "tune_new_entries": True,
    }

    if hpt_type == "hyperband":
        tuner = Hyperband(
552
553
554
            directory=os.path.join(
                outpath, "HyperBandx_{}_{}".format(loss, str(date.today()))
            ),
555
556
            max_epochs=35,
            hyperband_iterations=hypertun_trials,
557
            factor=3,
558
559
560
561
            **hpt_params
        )
    else:
        tuner = BayesianOptimization(
562
563
564
            directory=os.path.join(
                outpath, "BayOpt_{}_{}".format(loss, str(date.today()))
            ),
565
566
567
            max_trials=hypertun_trials,
            **hpt_params
        )
568
569
570

    print(tuner.search_space_summary())

571
572
573
574
575
576
577
    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:]]

578
    if phenotype_class > 0.0:
579
580
581
        ys += [y_train]
        yvals += [y_val]

582
    tuner.search(
583
584
        Xs,
        ys,
585
        epochs=n_epochs,
586
        validation_data=(Xvals, yvals),
587
        verbose=1,
588
        batch_size=batch_size,
lucas_miranda's avatar
lucas_miranda committed
589
        callbacks=callbacks,
590
591
592
593
594
    )

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

lucas_miranda's avatar
lucas_miranda committed
595
596
    print(tuner.results_summary())

597
    return best_hparams, best_run