train_utils.py 18.2 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
def load_hparams(hparams):
50
51
52
53
54
55
56
57
    """Loads hyperparameters from a custom dictionary pickled on disc.
    Thought to be used with the output of hyperparameter_tuning.py"""

    if hparams is not None:
        with open(hparams, "rb") as handle:
            hparams = pickle.load(handle)
    else:
        hparams = {
58
59
60
61
62
            "bidirectional_merge": "ave",
            "clipvalue": 1.0,
            "dense_activation": "relu",
            "dense_layers_per_branch": 1,
            "dropout_rate": 1e-3,
63
            "learning_rate": 1e-3,
64
65
66
            "units_conv": 160,
            "units_dense2": 120,
            "units_lstm": 300,
67
68
69
70
71
72
73
74
75
76
77
        }
    return hparams


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(
            os.path.join(
                train_path,
78
                [i for i in os.listdir(train_path) if i.endswith(".pkl")][0],
79
80
81
82
83
84
85
86
87
88
89
            ),
            "rb",
        ) as handle:
            treatment_dict = pickle.load(handle)
    except IndexError:
        treatment_dict = None

    return treatment_dict


def get_callbacks(
lucas_miranda's avatar
lucas_miranda committed
90
91
92
93
    X_train: np.array,
    batch_size: int,
    cp: bool,
    variational: bool,
94
    phenotype_class: float,
lucas_miranda's avatar
lucas_miranda committed
95
96
    predictor: float,
    loss: str,
97
98
    reg_cat_clusters: bool,
    reg_cluster_variance: bool,
99
    logparam: dict = None,
100
    outpath: str = ".",
101
) -> List[Union[Any]]:
102
    """Generates callbacks for model training, including:
103
104
105
106
    - run_ID: run name, with coarse parameter details;
    - tensorboard_callback: for real-time visualization;
    - cp_callback: for checkpoint saving,
    - onecycle: for learning rate scheduling"""
107

108
109
110
111
112
113
114
115
    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"

116
    run_ID = "{}{}{}{}{}{}{}_{}".format(
117
        ("GMVAE" if variational else "AE"),
lucas_miranda's avatar
lucas_miranda committed
118
        ("_Pred={}".format(predictor) if predictor > 0 and variational else ""),
119
        ("_Pheno={}".format(phenotype_class) if phenotype_class > 0 else ""),
120
        ("_loss={}".format(loss) if variational else ""),
121
122
        ("_encoding={}".format(logparam["encoding"]) if logparam is not None else ""),
        ("_k={}".format(logparam["k"]) if logparam is not None else ""),
123
        ("_latreg={}".format(latreg)),
124
        (datetime.now().strftime("%Y%m%d-%H%M%S")),
125
126
    )

127
    log_dir = os.path.abspath(os.path.join(outpath, "fit", run_ID))
128
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
129
130
131
        log_dir=log_dir,
        histogram_freq=1,
        profile_batch=2,
132
133
134
    )

    onecycle = deepof.model_utils.one_cycle_scheduler(
135
136
        X_train.shape[0] // batch_size * 250,
        max_rate=0.005,
137
138
    )

139
140
141
142
    callbacks = [run_ID, tensorboard_callback, onecycle]

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
143
            os.path.join(outpath, "checkpoints", run_ID + "/cp-{epoch:04d}.ckpt"),
144
145
146
147
148
149
150
151
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
152
153


lucas_miranda's avatar
lucas_miranda committed
154
def log_hyperparameters(phenotype_class: float, rec: str):
lucas_miranda's avatar
lucas_miranda committed
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
    """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
206
207
208
209
210
211
212
213
214
215
def tensorboard_metric_logging(
    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
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
251
252
253
254
    """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)


255
def autoencoder_fitting(
256
    preprocessed_object: Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray],
257
258
    batch_size: int,
    encoding_size: int,
259
    epochs: int,
260
261
262
263
264
    hparams: dict,
    kl_warmup: int,
    log_history: bool,
    log_hparams: bool,
    loss: str,
265
266
267
268
269
    mmd_warmup: int,
    montecarlo_kl: int,
    n_components: int,
    output_path: str,
    phenotype_class: float,
270
271
272
    predictor: float,
    pretrained: str,
    save_checkpoints: bool,
273
    save_weights: bool,
274
    variational: bool,
275
276
    reg_cat_clusters: bool,
    reg_cluster_variance: bool,
277
):
278
279
    """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""

280
    # Load data
281
282
283
284
285
    X_train, y_train, X_val, y_val = preprocessed_object

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

286
    # Defines what to log on tensorboard (useful for trying out different models)
287
288
    logparam = {
        "encoding": encoding_size,
289
        "k": n_components,
290
291
292
293
294
        "loss": loss,
    }
    if phenotype_class:
        logparam["pheno_weight"] = phenotype_class

295
    # Load callbacks
296
    run_ID, *cbacks = get_callbacks(
297
298
299
300
        X_train=X_train,
        batch_size=batch_size,
        cp=save_checkpoints,
        variational=variational,
301
        phenotype_class=phenotype_class,
302
303
        predictor=predictor,
        loss=loss,
304
        reg_cat_clusters=reg_cat_clusters,
305
        reg_cluster_variance=reg_cluster_variance,
306
307
308
        logparam=logparam,
        outpath=output_path,
    )
309
310
    if not log_history:
        cbacks = cbacks[1:]
311

312
    # Logs hyperparameters to tensorboard
lucas_miranda's avatar
lucas_miranda committed
313
    rec = "reconstruction_" if phenotype_class else ""
314
    if log_hparams:
lucas_miranda's avatar
lucas_miranda committed
315
        logparams, metrics = log_hyperparameters(phenotype_class, rec)
316
317
318
319
320
321
322
323

        with tf.summary.create_file_writer(
            os.path.join(output_path, "hparams", run_ID)
        ).as_default():
            hp.hparams_config(
                hparams=logparams,
                metrics=metrics,
            )
324

325
    # Build models
326
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
    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,
354
355
            reg_cat_clusters=reg_cat_clusters,
            reg_cluster_variance=reg_cluster_variance,
356
357
358
359
360
361
        ).build(
            X_train.shape
        )
        return_list = (encoder, generator, grouper, ae)

    if pretrained:
362
        # If pretrained models are specified, load weights and return
363
364
365
366
367
368
369
370
371
        ae.load_weights(pretrained)
        return return_list

    else:
        if not variational:

            ae.fit(
                x=X_train,
                y=X_train,
372
                epochs=epochs,
373
374
375
                batch_size=batch_size,
                verbose=1,
                validation_data=(X_val, X_val),
376
377
                callbacks=cbacks
                + [
378
379
380
381
                    CustomStopper(
                        monitor="val_loss",
                        patience=5,
                        restore_best_weights=True,
382
                        start_epoch=max(kl_warmup, mmd_warmup),
383
384
385
386
                    ),
                ],
            )

387
388
389
            if save_weights:
                ae.save_weights("{}_final_weights.h5".format(run_ID))

390
391
        else:

392
            callbacks_ = cbacks + [
393
394
395
396
                CustomStopper(
                    monitor="val_loss",
                    patience=5,
                    restore_best_weights=True,
397
                    start_epoch=max(kl_warmup, mmd_warmup),
398
399
400
                ),
            ]

401
            if "ELBO" in loss and kl_warmup > 0:
402
403
                # noinspection PyUnboundLocalVariable
                callbacks_.append(kl_warmup_callback)
404
            if "MMD" in loss and mmd_warmup > 0:
405
406
407
                # noinspection PyUnboundLocalVariable
                callbacks_.append(mmd_warmup_callback)

408
409
410
411
412
413
414
            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:]]

415
            if phenotype_class > 0.0:
416
417
418
                ys += [y_train]
                yvals += [y_val]

419
            ae.fit(
420
421
                x=Xs,
                y=ys,
422
                epochs=epochs,
423
424
425
426
427
428
429
430
431
                batch_size=batch_size,
                verbose=1,
                validation_data=(
                    Xvals,
                    yvals,
                ),
                callbacks=callbacks_,
            )

432
            if save_weights:
433
434
                ae.save_weights(
                    os.path.join(
435
436
437
                        "{}".format(output_path),
                        "trained_weights",
                        "{}_final_weights.h5".format(run_ID),
438
439
                    )
                )
440

441
442
443
444
445
            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
446
                    ae,
lucas_miranda's avatar
lucas_miranda committed
447
448
                    Xvals,
                    yvals[-1],
lucas_miranda's avatar
lucas_miranda committed
449
450
451
                    phenotype_class,
                    predictor,
                    rec,
452
                )
453

454
455
456
    return return_list


457
def tune_search(
458
    data: List[np.array],
459
    encoding_size: int,
460
461
    hypertun_trials: int,
    hpt_type: str,
462
463
    hypermodel: str,
    k: int,
464
    kl_warmup_epochs: int,
465
    loss: str,
466
    mmd_warmup_epochs: int,
467
    overlap_loss: float,
468
    phenotype_class: float,
469
470
    predictor: float,
    project_name: str,
471
    callbacks: List,
472
    n_epochs: int = 30,
473
    n_replicas: int = 1,
474
) -> Union[bool, Tuple[Any, Any]]:
475
476
    """Define the search space using keras-tuner and bayesian optimization

477
478
479
480
481
482
483
484
485
486
487
    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
488
        - phenotype_class (float): adds an extra regularizing neural network to the model,
489
490
491
492
493
494
495
496
497
498
499
500
        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
501
502
503

    """

504
505
    X_train, y_train, X_val, y_val = data

506
507
508
509
    assert hpt_type in ["bayopt", "hyperband"], (
        "Invalid hyperparameter tuning framework. " "Select one of bayopt and hyperband"
    )

lucas_miranda's avatar
lucas_miranda committed
510
    if hypermodel == "S2SAE":  # pragma: no cover
511
        assert (
512
            predictor == 0.0 and phenotype_class == 0.0
513
        ), "Prediction branches are only available for variational models. See documentation for more details"
514
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_AE(input_shape=X_train.shape)
515
516
517

    elif hypermodel == "S2SGMVAE":
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_GMVAE(
518
            input_shape=X_train.shape,
519
            encoding=encoding_size,
520
            kl_warmup_epochs=kl_warmup_epochs,
521
            loss=loss,
522
            mmd_warmup_epochs=mmd_warmup_epochs,
523
            number_of_components=k,
524
            overlap_loss=overlap_loss,
525
            phenotype_predictor=phenotype_class,
526
            predictor=predictor,
527
        )
lucas_miranda's avatar
lucas_miranda committed
528

529
530
531
    else:
        return False

532
533
534
535
536
537
538
539
540
541
542
543
544
    hpt_params = {
        "hypermodel": hypermodel,
        "executions_per_trial": n_replicas,
        "logger": TensorBoardLogger(metrics=["val_mae"], logdir="./logs/hparams"),
        "objective": "val_mae",
        "project_name": project_name,
        "seed": 42,
        "tune_new_entries": True,
    }

    if hpt_type == "hyperband":
        tuner = Hyperband(
            directory="HyperBandx_{}_{}".format(loss, str(date.today())),
545
546
            max_epochs=35,
            hyperband_iterations=hypertun_trials,
lucas_miranda's avatar
lucas_miranda committed
547
            factor=2,
548
549
550
551
552
553
554
555
            **hpt_params
        )
    else:
        tuner = BayesianOptimization(
            directory="BayOpt_{}_{}".format(loss, str(date.today())),
            max_trials=hypertun_trials,
            **hpt_params
        )
556
557
558

    print(tuner.search_space_summary())

559
560
561
562
563
564
565
    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:]]

566
    if phenotype_class > 0.0:
567
568
569
        ys += [y_train]
        yvals += [y_val]

570
    tuner.search(
571
572
        Xs,
        ys,
573
        epochs=n_epochs,
574
        validation_data=(Xvals, yvals),
575
576
        verbose=1,
        batch_size=256,
lucas_miranda's avatar
lucas_miranda committed
577
        callbacks=callbacks,
578
579
580
581
582
    )

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

lucas_miranda's avatar
lucas_miranda committed
583
584
    print(tuner.results_summary())

585
    return best_hparams, best_run