train_utils.py 18.4 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(
lucas_miranda's avatar
lucas_miranda committed
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(
lucas_miranda's avatar
lucas_miranda committed
69
70
71
72
73
74
75
76
77
78
79
80
81
82
        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_knn: int = 100,
        logparam: dict = None,
        outpath: str = ".",
83
) -> List[Union[Any]]:
84
    """Generates callbacks for model training, including:
85
86
87
88
    - run_ID: run name, with coarse parameter details;
    - tensorboard_callback: for real-time visualization;
    - cp_callback: for checkpoint saving,
    - onecycle: for learning rate scheduling"""
89

90
91
92
93
94
95
96
97
    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"

98
    run_ID = "{}{}{}{}{}{}{}_{}".format(
99
        ("GMVAE" if variational else "AE"),
lucas_miranda's avatar
lucas_miranda committed
100
        ("_Pred={}".format(predictor) if predictor > 0 and variational else ""),
101
        ("_Pheno={}".format(phenotype_class) if phenotype_class > 0 else ""),
102
        ("_loss={}".format(loss) if variational else ""),
103
104
        ("_encoding={}".format(logparam["encoding"]) if logparam is not None else ""),
        ("_k={}".format(logparam["k"]) if logparam is not None else ""),
105
        ("_latreg={}".format(latreg)),
106
        ("entknn={}".format(entropy_knn)),
107
        (datetime.now().strftime("%Y%m%d-%H%M%S")),
108
109
    )

110
    log_dir = os.path.abspath(os.path.join(outpath, "fit", run_ID))
111
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
112
113
114
        log_dir=log_dir,
        histogram_freq=1,
        profile_batch=2,
115
116
    )

117
    entropy = deepof.model_utils.neighbor_latent_entropy(
118
        encoding_dim=logparam["encoding"],
119
        k=entropy_knn,
120
        samples=entropy_samples,
lucas_miranda's avatar
lucas_miranda committed
121
        validation_data=X_val,
122
        log_dir=os.path.join(outpath, "metrics", run_ID),
123
        variational=variational,
lucas_miranda's avatar
lucas_miranda committed
124
125
    )

126
    onecycle = deepof.model_utils.one_cycle_scheduler(
127
128
        X_train.shape[0] // batch_size * 250,
        max_rate=0.005,
129
        log_dir=os.path.join(outpath, "metrics", run_ID),
130
131
    )

132
    callbacks = [run_ID, tensorboard_callback, entropy, onecycle]
133
134
135

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
136
            os.path.join(outpath, "checkpoints", run_ID + "/cp-{epoch:04d}.ckpt"),
137
138
139
140
141
142
143
144
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
145
146


lucas_miranda's avatar
lucas_miranda committed
147
def log_hyperparameters(phenotype_class: float, rec: str):
lucas_miranda's avatar
lucas_miranda committed
148
149
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
    """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
199
def tensorboard_metric_logging(
lucas_miranda's avatar
lucas_miranda committed
200
201
202
203
204
205
206
207
        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
208
):
lucas_miranda's avatar
lucas_miranda committed
209
210
211
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
    """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)


248
def autoencoder_fitting(
lucas_miranda's avatar
lucas_miranda committed
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
        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_samples: int,
        entropy_knn: int,
272
):
273
274
    """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""

275
    # Load data
276
277
278
279
280
    X_train, y_train, X_val, y_val = preprocessed_object

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

281
    # Defines what to log on tensorboard (useful for trying out different models)
282
283
    logparam = {
        "encoding": encoding_size,
284
        "k": n_components,
285
286
287
288
289
        "loss": loss,
    }
    if phenotype_class:
        logparam["pheno_weight"] = phenotype_class

290
    # Load callbacks
291
    run_ID, *cbacks = get_callbacks(
292
        X_train=X_train,
lucas_miranda's avatar
lucas_miranda committed
293
        X_val=(X_val if X_val.shape != (0,) else None),
294
295
296
        batch_size=batch_size,
        cp=save_checkpoints,
        variational=variational,
297
        phenotype_class=phenotype_class,
298
299
        predictor=predictor,
        loss=loss,
300
        entropy_samples=entropy_samples,
301
        entropy_knn=entropy_knn,
302
        reg_cat_clusters=reg_cat_clusters,
303
        reg_cluster_variance=reg_cluster_variance,
304
305
306
        logparam=logparam,
        outpath=output_path,
    )
307
308
    if not log_history:
        cbacks = cbacks[1:]
309

310
    # Logs hyperparameters to tensorboard
lucas_miranda's avatar
lucas_miranda committed
311
    rec = "reconstruction_" if phenotype_class else ""
312
    if log_hparams:
lucas_miranda's avatar
lucas_miranda committed
313
        logparams, metrics = log_hyperparameters(phenotype_class, rec)
314
315

        with tf.summary.create_file_writer(
lucas_miranda's avatar
lucas_miranda committed
316
                os.path.join(output_path, "hparams", run_ID)
317
318
319
320
321
        ).as_default():
            hp.hparams_config(
                hparams=logparams,
                metrics=metrics,
            )
322

323
    # Build models
324
325
326
327
328
329
330
    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:
331
        (encoder, generator, grouper, ae,) = deepof.models.SEQ_2_SEQ_GMVAE(
332
333
334
335
336
337
338
339
340
341
342
343
344
            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,
345
346
            reg_cat_clusters=reg_cat_clusters,
            reg_cluster_variance=reg_cluster_variance,
347
        ).build(X_train.shape)
348
349
350
        return_list = (encoder, generator, grouper, ae)

    if pretrained:
351
        # If pretrained models are specified, load weights and return
352
353
354
355
356
357
358
359
360
        ae.load_weights(pretrained)
        return return_list

    else:
        if not variational:

            ae.fit(
                x=X_train,
                y=X_train,
361
                epochs=epochs,
362
363
364
                batch_size=batch_size,
                verbose=1,
                validation_data=(X_val, X_val),
365
                callbacks=cbacks
lucas_miranda's avatar
lucas_miranda committed
366
367
368
369
370
371
372
373
                          + [
                              CustomStopper(
                                  monitor="val_loss",
                                  patience=5,
                                  restore_best_weights=True,
                                  start_epoch=max(kl_warmup, mmd_warmup),
                              ),
                          ],
374
375
            )

376
377
378
            if save_weights:
                ae.save_weights("{}_final_weights.h5".format(run_ID))

379
380
        else:

381
            callbacks_ = cbacks + [
382
383
384
385
                CustomStopper(
                    monitor="val_loss",
                    patience=5,
                    restore_best_weights=True,
386
                    start_epoch=max(kl_warmup, mmd_warmup),
387
388
389
                ),
            ]

390
391
392
393
394
395
396
            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:]]

397
            if phenotype_class > 0.0:
398
399
400
                ys += [y_train]
                yvals += [y_val]

401
            ae.fit(
402
403
                x=Xs,
                y=ys,
404
                epochs=epochs,
405
406
407
408
409
410
411
412
413
                batch_size=batch_size,
                verbose=1,
                validation_data=(
                    Xvals,
                    yvals,
                ),
                callbacks=callbacks_,
            )

414
            if not os.path.exists(os.path.join(output_path, "trained_weights")):
415
416
                os.makedirs("trained_weights")

417
            if save_weights:
418
419
                ae.save_weights(
                    os.path.join(
420
421
422
                        "{}".format(output_path),
                        "trained_weights",
                        "{}_final_weights.h5".format(run_ID),
423
424
                    )
                )
425

426
427
428
429
430
            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
431
                    ae,
lucas_miranda's avatar
lucas_miranda committed
432
433
                    Xvals,
                    yvals[-1],
lucas_miranda's avatar
lucas_miranda committed
434
435
436
                    phenotype_class,
                    predictor,
                    rec,
437
                )
438

439
440
441
    return return_list


442
def tune_search(
lucas_miranda's avatar
lucas_miranda committed
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
        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 = ".",
460
) -> Union[bool, Tuple[Any, Any]]:
461
462
    """Define the search space using keras-tuner and bayesian optimization

463
464
465
466
467
468
469
470
471
472
473
    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
474
        - phenotype_class (float): adds an extra regularizing neural network to the model,
475
476
477
478
479
480
481
482
483
484
485
486
        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
487
488
489

    """

490
491
    X_train, y_train, X_val, y_val = data

492
493
494
495
    assert hpt_type in ["bayopt", "hyperband"], (
        "Invalid hyperparameter tuning framework. " "Select one of bayopt and hyperband"
    )

lucas_miranda's avatar
lucas_miranda committed
496
    if hypermodel == "S2SAE":  # pragma: no cover
497
        assert (
lucas_miranda's avatar
lucas_miranda committed
498
                predictor == 0.0 and phenotype_class == 0.0
499
        ), "Prediction branches are only available for variational models. See documentation for more details"
500
        batch_size = 1
501
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_AE(input_shape=X_train.shape)
502
503

    elif hypermodel == "S2SGMVAE":
504
        batch_size = 64
505
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_GMVAE(
506
            input_shape=X_train.shape,
507
            encoding=encoding_size,
508
            kl_warmup_epochs=kl_warmup_epochs,
509
            loss=loss,
510
            mmd_warmup_epochs=mmd_warmup_epochs,
511
            number_of_components=k,
512
            overlap_loss=overlap_loss,
513
            phenotype_predictor=phenotype_class,
514
            predictor=predictor,
515
        )
lucas_miranda's avatar
lucas_miranda committed
516

517
518
519
    else:
        return False

520
521
522
    hpt_params = {
        "hypermodel": hypermodel,
        "executions_per_trial": n_replicas,
523
524
525
        "logger": TensorBoardLogger(
            metrics=["val_mae"], logdir=os.path.join(outpath, "logged_hparams")
        ),
526
527
528
529
530
531
532
        "objective": "val_mae",
        "project_name": project_name,
        "tune_new_entries": True,
    }

    if hpt_type == "hyperband":
        tuner = Hyperband(
533
534
535
            directory=os.path.join(
                outpath, "HyperBandx_{}_{}".format(loss, str(date.today()))
            ),
536
537
            max_epochs=35,
            hyperband_iterations=hypertun_trials,
538
            factor=3,
539
540
541
542
            **hpt_params
        )
    else:
        tuner = BayesianOptimization(
543
544
545
            directory=os.path.join(
                outpath, "BayOpt_{}_{}".format(loss, str(date.today()))
            ),
546
547
548
            max_trials=hypertun_trials,
            **hpt_params
        )
549
550
551

    print(tuner.search_space_summary())

552
553
554
555
556
557
558
    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:]]

559
    if phenotype_class > 0.0:
560
561
562
        ys += [y_train]
        yvals += [y_val]

563
    tuner.search(
564
565
        Xs,
        ys,
566
        epochs=n_epochs,
567
        validation_data=(Xvals, yvals),
568
        verbose=1,
569
        batch_size=batch_size,
lucas_miranda's avatar
lucas_miranda committed
570
        callbacks=callbacks,
571
572
573
574
575
    )

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

lucas_miranda's avatar
lucas_miranda committed
576
577
    print(tuner.results_summary())

578
    return best_hparams, best_run