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, Objective
19
from kerastuner_tensorboard_logger import TensorBoardLogger
20
from tensorboard.plugins.hparams import api as hp
lucas_miranda's avatar
lucas_miranda committed
21

22
23
24
import deepof.hypermodels
import deepof.model_utils

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(".json")][0],
            ),
            "r",
59
        ) as handle:
60
            treatment_dict = json.load(handle)
61
62
63
64
65
66
67
    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
82
83
84
    X_train: np.array,
    batch_size: int,
    phenotype_prediction: float,
    next_sequence_prediction: float,
    rule_based_prediction: float,
    overlap_loss: float,
    loss: str,
    loss_warmup: int = 0,
    warmup_mode: str = "none",
    input_type: str = False,
    cp: bool = False,
    reg_cat_clusters: bool = False,
    reg_cluster_variance: bool = False,
    entropy_knn: int = 100,
    logparam: dict = None,
    outpath: str = ".",
    run: int = False,
85
) -> List[Union[Any]]:
86
    """Generates callbacks for model training, including:
87
88
    - run_ID: run name, with coarse parameter details;
    - tensorboard_callback: for real-time visualization;
89
90
91
92
    - cp_callback: for checkpoint saving;
    - onecycle: for learning rate scheduling;
    - entropy: neighborhood entropy in the latent space;
    """
93

94
95
96
97
98
99
100
101
    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"

102
103
    run_ID = "{}{}{}{}{}{}{}{}{}{}{}{}{}{}{}{}".format(
        "deepof_GMVAE",
104
        ("_input_type={}".format(input_type) if input_type else "coords"),
105
        ("_window_size={}".format(X_train.shape[1])),
106
107
108
        ("_NSPred={}".format(next_sequence_prediction)),
        ("_PPred={}".format(phenotype_prediction)),
        ("_RBPred={}".format(rule_based_prediction)),
109
        ("_loss={}".format(loss)),
110
        ("_overlap_loss={}".format(overlap_loss)),
111
112
        ("_loss_warmup={}".format(loss_warmup)),
        ("_warmup_mode={}".format(warmup_mode)),
113
114
        ("_encoding={}".format(logparam["encoding"]) if logparam is not None else ""),
        ("_k={}".format(logparam["k"]) if logparam is not None else ""),
115
        ("_latreg={}".format(latreg)),
116
117
        ("_entknn={}".format(entropy_knn)),
        ("_run={}".format(run) if run else ""),
118
        ("_{}".format(datetime.now().strftime("%Y%m%d-%H%M%S")) if not run else ""),
119
120
    )

121
    log_dir = os.path.abspath(os.path.join(outpath, "fit", run_ID))
122
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
123
124
125
        log_dir=log_dir,
        histogram_freq=1,
        profile_batch=2,
126
127
128
    )

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

134
    callbacks = [run_ID, tensorboard_callback, onecycle]
135
136
137

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
138
            os.path.join(outpath, "checkpoints", run_ID + "/cp-{epoch:04d}.ckpt"),
139
140
141
142
143
144
145
146
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
147
148


lucas_miranda's avatar
lucas_miranda committed
149
def log_hyperparameters(phenotype_class: float, rec: str):
lucas_miranda's avatar
lucas_miranda committed
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
    """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
201
def tensorboard_metric_logging(
202
203
204
205
206
207
208
209
210
    run_dir: str,
    hpms: Any,
    ae: Any,
    X_val: np.ndarray,
    y_val: np.ndarray,
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
    rec: str,
lucas_miranda's avatar
lucas_miranda committed
211
):
lucas_miranda's avatar
lucas_miranda committed
212
213
    """Autoencoder metric logging in tensorboard"""

214
215
    outputs = ae.predict(X_val)
    idx_generator = (idx for idx in range(len(outputs)))
lucas_miranda's avatar
lucas_miranda committed
216
217
218

    with tf.summary.create_file_writer(run_dir).as_default():
        hp.hparams(hpms)  # record the values used in this trial
219
220
        idx = next(idx_generator)

lucas_miranda's avatar
lucas_miranda committed
221
        val_mae = tf.reduce_mean(
222
            tf.keras.metrics.mean_absolute_error(y_val[idx], outputs[idx])
lucas_miranda's avatar
lucas_miranda committed
223
224
        )
        val_mse = tf.reduce_mean(
225
            tf.keras.metrics.mean_squared_error(y_val[idx], outputs[idx])
lucas_miranda's avatar
lucas_miranda committed
226
227
228
229
        )
        tf.summary.scalar("val_{}mae".format(rec), val_mae, step=1)
        tf.summary.scalar("val_{}mse".format(rec), val_mse, step=1)

230
        if next_sequence_prediction:
231
            idx = next(idx_generator)
lucas_miranda's avatar
lucas_miranda committed
232
            pred_mae = tf.reduce_mean(
233
                tf.keras.metrics.mean_absolute_error(y_val[idx], outputs[idx])
lucas_miranda's avatar
lucas_miranda committed
234
235
            )
            pred_mse = tf.reduce_mean(
236
                tf.keras.metrics.mean_squared_error(y_val[idx], outputs[idx])
237
238
239
240
241
242
            )
            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
243
244
            )

245
        if phenotype_prediction:
246
            idx = next(idx_generator)
247
248
249
            pheno_acc = tf.keras.metrics.binary_accuracy(
                y_val[idx], tf.squeeze(outputs[idx])
            )
250
            pheno_auc = tf.keras.metrics.AUC()
251
            pheno_auc.update_state(y_val[idx], outputs[idx])
252
            pheno_auc = pheno_auc.result().numpy()
lucas_miranda's avatar
lucas_miranda committed
253
254
255
256

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

257
        if rule_based_prediction:
258
            idx = next(idx_generator)
259
            rules_mae = tf.reduce_mean(
260
                tf.keras.metrics.mean_absolute_error(y_val[idx], outputs[idx])
261
262
            )
            rules_mse = tf.reduce_mean(
263
                tf.keras.metrics.mean_squared_error(y_val[idx], outputs[idx])
264
265
266
267
            )
            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
268

269
def autoencoder_fitting(
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
    preprocessed_object: Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray],
    batch_size: int,
    encoding_size: int,
    epochs: int,
    hparams: dict,
    kl_annealing_mode: str,
    kl_warmup: int,
    log_history: bool,
    log_hparams: bool,
    loss: str,
    mmd_annealing_mode: str,
    mmd_warmup: int,
    montecarlo_kl: int,
    n_components: int,
    output_path: str,
    overlap_loss: float,
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
    pretrained: str,
    save_checkpoints: bool,
    save_weights: bool,
    reg_cat_clusters: bool,
    reg_cluster_variance: bool,
    entropy_knn: int,
    input_type: str,
    run: int = 0,
    strategy: tf.distribute.Strategy = tf.distribute.MirroredStrategy(),
298
):
299
300
    """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""

301
    # Load data
302
303
304
305
306
    X_train, y_train, X_val, y_val = preprocessed_object

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

307
308
309
310
311
312
    # Set options for tf.data.Datasets
    options = tf.data.Options()
    options.experimental_distribute.auto_shard_policy = (
        tf.data.experimental.AutoShardPolicy.DATA
    )

313
    # Defines what to log on tensorboard (useful for trying out different models)
314
315
    logparam = {
        "encoding": encoding_size,
316
        "k": n_components,
317
318
        "loss": loss,
    }
319
320
    if phenotype_prediction:
        logparam["pheno_weight"] = phenotype_prediction
321

322
    # Load callbacks
323
    run_ID, *cbacks = get_callbacks(
324
325
        X_train=X_train,
        batch_size=batch_size,
326
        phenotype_prediction=phenotype_prediction,
327
        next_sequence_prediction=next_sequence_prediction,
328
        rule_based_prediction=rule_based_prediction,
329
        loss=loss,
330
        loss_warmup=kl_warmup,
331
        overlap_loss=overlap_loss,
332
        warmup_mode=kl_annealing_mode,
333
334
        input_type=input_type,
        cp=save_checkpoints,
335
        reg_cat_clusters=reg_cat_clusters,
336
        reg_cluster_variance=reg_cluster_variance,
337
        entropy_knn=entropy_knn,
338
339
        logparam=logparam,
        outpath=output_path,
340
        run=run,
341
    )
342
343
    if not log_history:
        cbacks = cbacks[1:]
344

345
    # Logs hyperparameters to tensorboard
346
    rec = "reconstruction_" if phenotype_prediction else ""
347
    if log_hparams:
348
        logparams, metrics = log_hyperparameters(phenotype_prediction, rec)
349
350

        with tf.summary.create_file_writer(
351
            os.path.join(output_path, "hparams", run_ID)
352
353
354
355
356
        ).as_default():
            hp.hparams_config(
                hparams=logparams,
                metrics=metrics,
            )
357

358
359
360
361
362
363
364
365
    # 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

366
367
368
369
370
371
372
373
374
375
376
377
378
379
    # 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,
380
            overlap_loss=overlap_loss,
381
382
383
384
385
386
            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,
387
        ).build(X_train.shape)
388
        return_list = (encoder, generator, grouper, ae)
389
390

    if pretrained:
391
        # If pretrained models are specified, load weights and return
392
393
394
        ae.load_weights(pretrained)
        return return_list

395
396
397
398
399
400
401
402
    callbacks_ = cbacks + [
        CustomStopper(
            monitor="val_loss",
            patience=15,
            restore_best_weights=True,
            start_epoch=max(kl_warmup, mmd_warmup),
        ),
    ]
403

404
405
406
407
408
409
410
411
    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:
412
413
        ys += [y_train[-Xs.shape[0] :, 0]]
        yvals += [y_val[-Xvals.shape[0] :, 0]]
414
415
416
417

        # Remove the used column (phenotype) from both y arrays
        y_train = y_train[:, 1:]
        y_val = y_val[:, 1:]
418

419
    if rule_based_prediction > 0.0:
420
421
        ys += [y_train[-Xs.shape[0] :]]
        yvals += [y_val[-Xvals.shape[0] :]]
422
423
424
425

    # Convert data to tf.data.Dataset objects
    train_dataset = (
        tf.data.Dataset.from_tensor_slices((Xs, tuple(ys)))
426
427
428
        .batch(batch_size * strategy.num_replicas_in_sync, drop_remainder=True)
        .shuffle(buffer_size=X_train.shape[0])
        .with_options(options)
429
430
431
    )
    val_dataset = (
        tf.data.Dataset.from_tensor_slices((Xvals, tuple(yvals)))
432
433
        .batch(batch_size * strategy.num_replicas_in_sync, drop_remainder=True)
        .with_options(options)
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
    )

    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),
453
            )
454
        )
455

456
457
458
459
460
461
462
463
464
465
466
467
468
    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,
        )
469

470
471
472
    return return_list


473
def tune_search(
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
    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,
    next_sequence_prediction: float,
    phenotype_prediction: float,
    rule_based_prediction: float,
    project_name: str,
    callbacks: List,
    n_epochs: int = 30,
    n_replicas: int = 1,
    outpath: str = ".",
491
) -> Union[bool, Tuple[Any, Any]]:
492
493
    """Define the search space using keras-tuner and bayesian optimization

494
495
496
497
498
499
500
501
502
    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
503
        - phenotype_class (float): adds an extra regularizing neural network to the model,
504
505
506
507
508
509
510
511
512
513
514
515
        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
516
517
518

    """

519
520
    X_train, y_train, X_val, y_val = data

521
522
523
524
    assert hpt_type in ["bayopt", "hyperband"], (
        "Invalid hyperparameter tuning framework. " "Select one of bayopt and hyperband"
    )

525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
    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
        ),
    )
541

542
543
544
545
    tuner_objective = (
        "val_mae" if not next_sequence_prediction else "val_reconstruction_mae"
    )

546
547
548
    hpt_params = {
        "hypermodel": hypermodel,
        "executions_per_trial": n_replicas,
549
        "logger": TensorBoardLogger(
550
            metrics=[tuner_objective], logdir=os.path.join(outpath, "logged_hparams")
551
        ),
552
        "objective": Objective(tuner_objective, direction="min"),
553
554
555
556
557
558
        "project_name": project_name,
        "tune_new_entries": True,
    }

    if hpt_type == "hyperband":
        tuner = Hyperband(
559
560
561
            directory=os.path.join(
                outpath, "HyperBandx_{}_{}".format(loss, str(date.today()))
            ),
562
            max_epochs=30,
563
            hyperband_iterations=hypertun_trials,
564
            factor=3,
565
566
567
568
            **hpt_params
        )
    else:
        tuner = BayesianOptimization(
569
570
571
            directory=os.path.join(
                outpath, "BayOpt_{}_{}".format(loss, str(date.today()))
            ),
572
573
574
            max_trials=hypertun_trials,
            **hpt_params
        )
575
576
577

    print(tuner.search_space_summary())

578
579
    Xs, ys = X_train, [X_train]
    Xvals, yvals = X_val, [X_val]
580

581
    if next_sequence_prediction > 0.0:
582
583
584
        Xs, ys = X_train[:-1], [X_train[:-1], X_train[1:]]
        Xvals, yvals = X_val[:-1], [X_val[:-1], X_val[1:]]

585
    if phenotype_prediction > 0.0:
586
587
        ys += [y_train[-Xs.shape[0] :, 0]]
        yvals += [y_val[-Xvals.shape[0] :, 0]]
588
589
590
591
592
593

        # 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:
594
595
        ys += [y_train[-Xs.shape[0] :]]
        yvals += [y_val[-Xvals.shape[0] :]]
596

597
598
599
600
601
602
    # Convert data to tf.data.Dataset objects
    train_dataset = (
        tf.data.Dataset.from_tensor_slices((Xs, tuple(ys)))
        .batch(batch_size, drop_remainder=True)
        .shuffle(buffer_size=X_train.shape[0])
    )
603
604
    val_dataset = tf.data.Dataset.from_tensor_slices((Xvals, tuple(yvals))).batch(
        batch_size, drop_remainder=True
605
606
    )

607
    # Convert data to tf.data.Dataset objects
608
    tuner.search(
609
        train_dataset,
610
        epochs=n_epochs,
611
        validation_data=val_dataset,
612
        verbose=1,
613
        batch_size=batch_size,
lucas_miranda's avatar
lucas_miranda committed
614
        callbacks=callbacks,
615
616
617
618
619
    )

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

lucas_miranda's avatar
lucas_miranda committed
620
621
    print(tuner.results_summary())

622
    return best_hparams, best_run