train_utils.py 20.1 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
    X_train: np.array,
    batch_size: int,
71
72
73
    phenotype_prediction: float,
    next_sequence_prediction: float,
    rule_based_prediction: float,
74
    loss: str,
75
76
    loss_warmup: int = 0,
    warmup_mode: str = "none",
77
    X_val: np.array = None,
78
    input_type: str = False,
79
80
81
82
83
84
85
    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 = ".",
86
    run: int = False,
87
) -> List[Union[Any]]:
88
    """Generates callbacks for model training, including:
89
90
    - run_ID: run name, with coarse parameter details;
    - tensorboard_callback: for real-time visualization;
91
92
93
94
    - cp_callback: for checkpoint saving;
    - onecycle: for learning rate scheduling;
    - entropy: neighborhood entropy in the latent space;
    """
95

96
97
98
99
100
101
102
103
    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"

104
    run_ID = "{}{}{}{}{}{}{}{}{}{}{}{}{}{}{}".format(
105
        ("deepof_GMVAE"),
106
        ("_input_type={}".format(input_type) if input_type else "coords"),
107
        ("_window_size={}".format(X_train.shape[1])),
108
109
110
111
        ("_NextSeqPred={}".format(next_sequence_prediction)),
        ("_PhenoPred={}".format(phenotype_prediction)),
        ("_RuleBasedPred={}".format(rule_based_prediction)),
        ("_loss={}".format(loss)),
112
113
        ("_loss_warmup={}".format(loss_warmup)),
        ("_warmup_mode={}".format(warmup_mode)),
114
115
        ("_encoding={}".format(logparam["encoding"]) if logparam is not None else ""),
        ("_k={}".format(logparam["k"]) if logparam is not None else ""),
116
        ("_latreg={}".format(latreg)),
117
118
        ("_entknn={}".format(entropy_knn)),
        ("_run={}".format(run) if run else ""),
119
        ("_{}".format(datetime.now().strftime("%Y%m%d-%H%M%S")) if not run else ""),
120
121
    )

122
    log_dir = os.path.abspath(os.path.join(outpath, "fit", run_ID))
123
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
124
125
126
        log_dir=log_dir,
        histogram_freq=1,
        profile_batch=2,
127
128
    )

129
    entropy = deepof.model_utils.neighbor_latent_entropy(
130
        encoding_dim=logparam["encoding"],
131
        k=entropy_knn,
132
        samples=entropy_samples,
lucas_miranda's avatar
lucas_miranda committed
133
        validation_data=X_val,
134
        log_dir=os.path.join(outpath, "metrics", run_ID),
lucas_miranda's avatar
lucas_miranda committed
135
136
    )

137
    onecycle = deepof.model_utils.one_cycle_scheduler(
138
139
        X_train.shape[0] // batch_size * 250,
        max_rate=0.005,
140
        log_dir=os.path.join(outpath, "metrics", run_ID),
141
142
    )

143
    callbacks = [run_ID, tensorboard_callback, entropy, onecycle]
144
145
146

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
147
            os.path.join(outpath, "checkpoints", run_ID + "/cp-{epoch:04d}.ckpt"),
148
149
150
151
152
153
154
155
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
156
157


lucas_miranda's avatar
lucas_miranda committed
158
def log_hyperparameters(phenotype_class: float, rec: str):
lucas_miranda's avatar
lucas_miranda committed
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
203
204
205
206
207
208
209
    """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
210
def tensorboard_metric_logging(
211
212
213
214
215
    run_dir: str,
    hpms: Any,
    ae: Any,
    X_val: np.ndarray,
    y_val: np.ndarray,
216
217
218
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
219
    rec: str,
lucas_miranda's avatar
lucas_miranda committed
220
):
lucas_miranda's avatar
lucas_miranda committed
221
222
    """Autoencoder metric logging in tensorboard"""

223
224
    outputs = ae.predict(X_val)
    idx_generator = (idx for idx in range(len(outputs)))
lucas_miranda's avatar
lucas_miranda committed
225
226
227

    with tf.summary.create_file_writer(run_dir).as_default():
        hp.hparams(hpms)  # record the values used in this trial
228
229
        idx = next(idx_generator)

lucas_miranda's avatar
lucas_miranda committed
230
        val_mae = tf.reduce_mean(
231
            tf.keras.metrics.mean_absolute_error(y_val[idx], outputs[idx])
lucas_miranda's avatar
lucas_miranda committed
232
233
        )
        val_mse = tf.reduce_mean(
234
            tf.keras.metrics.mean_squared_error(y_val[idx], outputs[idx])
lucas_miranda's avatar
lucas_miranda committed
235
236
237
238
        )
        tf.summary.scalar("val_{}mae".format(rec), val_mae, step=1)
        tf.summary.scalar("val_{}mse".format(rec), val_mse, step=1)

239
        if next_sequence_prediction:
240
            idx = next(idx_generator)
lucas_miranda's avatar
lucas_miranda committed
241
            pred_mae = tf.reduce_mean(
242
                tf.keras.metrics.mean_absolute_error(y_val[idx], outputs[idx])
lucas_miranda's avatar
lucas_miranda committed
243
244
            )
            pred_mse = tf.reduce_mean(
245
                tf.keras.metrics.mean_squared_error(y_val[idx], outputs[idx])
246
247
248
249
250
251
            )
            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
252
253
            )

254
        if phenotype_prediction:
255
            idx = next(idx_generator)
256
257
258
            pheno_acc = tf.keras.metrics.binary_accuracy(
                y_val[idx], tf.squeeze(outputs[idx])
            )
259
            pheno_auc = tf.keras.metrics.AUC()
260
            pheno_auc.update_state(y_val[idx], outputs[idx])
261
            pheno_auc = pheno_auc.result().numpy()
lucas_miranda's avatar
lucas_miranda committed
262
263
264
265

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

266
        if rule_based_prediction:
267
            idx = next(idx_generator)
268
            rules_mae = tf.reduce_mean(
269
                tf.keras.metrics.mean_absolute_error(y_val[idx], outputs[idx])
270
271
            )
            rules_mse = tf.reduce_mean(
272
                tf.keras.metrics.mean_squared_error(y_val[idx], outputs[idx])
273
274
275
276
            )
            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
277

278
def autoencoder_fitting(
279
280
281
282
283
    preprocessed_object: Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray],
    batch_size: int,
    encoding_size: int,
    epochs: int,
    hparams: dict,
284
    kl_annealing_mode: str,
285
286
287
288
    kl_warmup: int,
    log_history: bool,
    log_hparams: bool,
    loss: str,
289
    mmd_annealing_mode: str,
290
291
292
293
    mmd_warmup: int,
    montecarlo_kl: int,
    n_components: int,
    output_path: str,
294
    overlap_loss: float,
295
296
297
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
298
299
300
301
302
303
304
    pretrained: str,
    save_checkpoints: bool,
    save_weights: bool,
    reg_cat_clusters: bool,
    reg_cluster_variance: bool,
    entropy_samples: int,
    entropy_knn: int,
305
    input_type: str,
306
    run: int = 0,
307
    strategy: tf.distribute.Strategy = tf.distribute.MirroredStrategy(),
308
):
309
310
    """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""

311
    # Load data
312
313
314
315
316
    X_train, y_train, X_val, y_val = preprocessed_object

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

317
318
319
320
321
322
323
324
325
326
    # Set options for tf.data.Datasets
    options = tf.data.Options()
    options.experimental_distribute.auto_shard_policy = (
        tf.data.experimental.AutoShardPolicy.DATA
    )

    # Generate validation dataset for callback usage
    X_val_dataset = (
        tf.data.Dataset.from_tensor_slices(X_val)
        .with_options(options)
327
        .batch(batch_size * strategy.num_replicas_in_sync)
328
329
    )

330
    # Defines what to log on tensorboard (useful for trying out different models)
331
332
    logparam = {
        "encoding": encoding_size,
333
        "k": n_components,
334
335
        "loss": loss,
    }
336
337
    if phenotype_prediction:
        logparam["pheno_weight"] = phenotype_prediction
338

339
    # Load callbacks
340
    run_ID, *cbacks = get_callbacks(
341
342
        X_train=X_train,
        batch_size=batch_size,
343
        phenotype_prediction=phenotype_prediction,
344
        next_sequence_prediction=next_sequence_prediction,
345
        rule_based_prediction=rule_based_prediction,
346
        loss=loss,
347
348
        loss_warmup=kl_warmup,
        warmup_mode=kl_annealing_mode,
349
        input_type=input_type,
350
        X_val=(X_val_dataset if X_val.shape != (0,) else None),
351
        cp=save_checkpoints,
352
        reg_cat_clusters=reg_cat_clusters,
353
        reg_cluster_variance=reg_cluster_variance,
354
355
        entropy_samples=entropy_samples,
        entropy_knn=entropy_knn,
356
357
        logparam=logparam,
        outpath=output_path,
358
        run=run,
359
    )
360
361
    if not log_history:
        cbacks = cbacks[1:]
362

363
    # Logs hyperparameters to tensorboard
364
    rec = "reconstruction_" if phenotype_prediction else ""
365
    if log_hparams:
366
        logparams, metrics = log_hyperparameters(phenotype_prediction, rec)
367
368

        with tf.summary.create_file_writer(
369
            os.path.join(output_path, "hparams", run_ID)
370
371
372
373
374
        ).as_default():
            hp.hparams_config(
                hparams=logparams,
                metrics=metrics,
            )
375

376
377
378
379
380
381
382
383
    # 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

384
385
386
387
388
389
390
391
392
393
394
395
396
397
    # Build model
    with strategy.scope():
        (encoder, generator, grouper, ae, prior, posterior,) = deepof.models.GMVAE(
            architecture_hparams=({} if hparams is None else hparams),
            batch_size=batch_size * strategy.num_replicas_in_sync,
            compile_model=True,
            encoding=encoding_size,
            kl_annealing_mode=kl_annealing_mode,
            kl_warmup_epochs=kl_warmup,
            loss=loss,
            mmd_annealing_mode=mmd_annealing_mode,
            mmd_warmup_epochs=mmd_warmup,
            montecarlo_kl=montecarlo_kl,
            number_of_components=n_components,
398
            overlap_loss=overlap_loss,
399
400
401
402
403
404
            next_sequence_prediction=next_sequence_prediction,
            phenotype_prediction=phenotype_prediction,
            rule_based_prediction=rule_based_prediction,
            rule_based_features=rule_based_features,
            reg_cat_clusters=reg_cat_clusters,
            reg_cluster_variance=reg_cluster_variance,
405
        ).build(X_train.shape)
406
        return_list = (encoder, generator, grouper, ae)
407
408

    if pretrained:
409
        # If pretrained models are specified, load weights and return
410
411
412
        ae.load_weights(pretrained)
        return return_list

413
414
415
416
417
418
419
420
    callbacks_ = cbacks + [
        CustomStopper(
            monitor="val_loss",
            patience=15,
            restore_best_weights=True,
            start_epoch=max(kl_warmup, mmd_warmup),
        ),
    ]
421

422
423
424
425
426
427
428
429
430
431
432
433
434
435
    Xs, ys = X_train, [X_train]
    Xvals, yvals = X_val, [X_val]

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

    if phenotype_prediction > 0.0:
        ys += [y_train[-Xs.shape[0] :, 0]]
        yvals += [y_val[-Xvals.shape[0] :, 0]]

        # Remove the used column (phenotype) from both y arrays
        y_train = y_train[:, 1:]
        y_val = y_val[:, 1:]
436

437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
    if rule_based_prediction > 0.0:
        ys += [y_train[-Xs.shape[0] :]]
        yvals += [y_val[-Xvals.shape[0] :]]

    # Convert data to tf.data.Dataset objects
    train_dataset = (
        tf.data.Dataset.from_tensor_slices((Xs, tuple(ys)))
        .batch(batch_size * strategy.num_replicas_in_sync)
        .shuffle(buffer_size=X_train.shape[0])
        .with_options(options)
    )
    val_dataset = (
        tf.data.Dataset.from_tensor_slices((Xvals, tuple(yvals)))
        .batch(batch_size * strategy.num_replicas_in_sync)
        .with_options(options)
    )

    ae.fit(
        x=train_dataset,
        epochs=epochs,
        verbose=1,
        validation_data=val_dataset,
        callbacks=callbacks_,
    )

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

    if save_weights:
        ae.save_weights(
            os.path.join(
                "{}".format(output_path),
                "trained_weights",
                "{}_final_weights.h5".format(run_ID),
471
            )
472
        )
473

474
475
476
477
478
479
480
481
482
483
484
485
486
    if log_hparams:
        # Logparams to tensorboard
        tensorboard_metric_logging(
            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,
        )
487

488
489
490
    return return_list


491
def tune_search(
492
493
494
495
496
497
498
499
500
    data: List[np.array],
    encoding_size: int,
    hypertun_trials: int,
    hpt_type: str,
    k: int,
    kl_warmup_epochs: int,
    loss: str,
    mmd_warmup_epochs: int,
    overlap_loss: float,
501
502
503
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
504
505
506
507
508
    project_name: str,
    callbacks: List,
    n_epochs: int = 30,
    n_replicas: int = 1,
    outpath: str = ".",
509
) -> Union[bool, Tuple[Any, Any]]:
510
511
    """Define the search space using keras-tuner and bayesian optimization

512
513
514
515
516
517
518
519
520
    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)
        - 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
521
        - phenotype_class (float): adds an extra regularizing neural network to the model,
522
523
524
525
526
527
528
529
530
531
532
533
        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
534
535
536

    """

537
538
    X_train, y_train, X_val, y_val = data

539
540
541
542
    assert hpt_type in ["bayopt", "hyperband"], (
        "Invalid hyperparameter tuning framework. " "Select one of bayopt and hyperband"
    )

543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
    batch_size = 64
    hypermodel = deepof.hypermodels.GMVAE(
        input_shape=X_train.shape,
        encoding=encoding_size,
        kl_warmup_epochs=kl_warmup_epochs,
        loss=loss,
        mmd_warmup_epochs=mmd_warmup_epochs,
        number_of_components=k,
        overlap_loss=overlap_loss,
        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
        ),
    )
559

560
561
562
    hpt_params = {
        "hypermodel": hypermodel,
        "executions_per_trial": n_replicas,
563
564
565
        "logger": TensorBoardLogger(
            metrics=["val_mae"], logdir=os.path.join(outpath, "logged_hparams")
        ),
566
567
568
569
570
571
572
        "objective": "val_mae",
        "project_name": project_name,
        "tune_new_entries": True,
    }

    if hpt_type == "hyperband":
        tuner = Hyperband(
573
574
575
            directory=os.path.join(
                outpath, "HyperBandx_{}_{}".format(loss, str(date.today()))
            ),
576
            max_epochs=30,
577
            hyperband_iterations=hypertun_trials,
578
            factor=3,
579
580
581
582
            **hpt_params
        )
    else:
        tuner = BayesianOptimization(
583
584
585
            directory=os.path.join(
                outpath, "BayOpt_{}_{}".format(loss, str(date.today()))
            ),
586
587
588
            max_trials=hypertun_trials,
            **hpt_params
        )
589
590
591

    print(tuner.search_space_summary())

592
593
594
    Xs, ys = [X_train], [X_train]
    Xvals, yvals = [X_val], [X_val]

595
    if next_sequence_prediction > 0.0:
596
597
598
        Xs, ys = X_train[:-1], [X_train[:-1], X_train[1:]]
        Xvals, yvals = X_val[:-1], [X_val[:-1], X_val[1:]]

599
    if phenotype_prediction > 0.0:
600
601
        ys += [y_train[-Xs.shape[0] :, 0]]
        yvals += [y_val[-Xvals.shape[0] :, 0]]
602
603
604
605
606
607

        # 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:
608
609
        ys += [y_train[-Xs.shape[0] :]]
        yvals += [y_val[-Xvals.shape[0] :]]
610

611
    tuner.search(
612
613
        Xs,
        ys,
614
        epochs=n_epochs,
615
        validation_data=(Xvals, yvals),
616
        verbose=1,
617
        batch_size=batch_size,
lucas_miranda's avatar
lucas_miranda committed
618
        callbacks=callbacks,
619
620
621
622
623
    )

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

lucas_miranda's avatar
lucas_miranda committed
624
625
    print(tuner.results_summary())

626
    return best_hparams, best_run