train_utils.py 17.7 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
15
from tensorboard.plugins.hparams import api as hp
16
from typing import Tuple, Union, Any, List
17
18
19
20
21
22
23
import deepof.hypermodels
import deepof.model_utils
import numpy as np
import os
import pickle
import tensorflow as tf

24
25
26
27
28
# Ignore warning with no downstream effect
tf.get_logger().setLevel("ERROR")
tf.autograph.set_verbosity(0)


29
class CustomStopper(tf.keras.callbacks.EarlyStopping):
30
    """ Custom early stopping callback. Prevents the model from stopping before warmup is over """
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47

    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)


48
def load_hparams(hparams):
49
50
51
52
53
54
55
56
    """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 = {
57
58
59
60
61
            "bidirectional_merge": "ave",
            "clipvalue": 1.0,
            "dense_activation": "relu",
            "dense_layers_per_branch": 1,
            "dropout_rate": 1e-3,
62
            "learning_rate": 1e-3,
63
64
65
            "units_conv": 160,
            "units_dense2": 120,
            "units_lstm": 300,
66
67
68
69
70
71
72
73
74
75
76
        }
    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,
77
                [i for i in os.listdir(train_path) if i.endswith(".pkl")][0],
78
79
80
81
82
83
84
85
86
87
88
            ),
            "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
89
90
91
92
    X_train: np.array,
    batch_size: int,
    cp: bool,
    variational: bool,
93
    phenotype_class: float,
lucas_miranda's avatar
lucas_miranda committed
94
95
    predictor: float,
    loss: str,
96
    logparam: dict = None,
97
    outpath: str = ".",
98
) -> List[Union[Any]]:
99
    """Generates callbacks for model training, including:
100
101
102
103
    - run_ID: run name, with coarse parameter details;
    - tensorboard_callback: for real-time visualization;
    - cp_callback: for checkpoint saving,
    - onecycle: for learning rate scheduling"""
104

105
    run_ID = "{}{}{}{}{}{}_{}".format(
106
        ("GMVAE" if variational else "AE"),
107
108
        ("Pred={}".format(predictor) if predictor > 0 and variational else ""),
        ("_Pheno={}".format(phenotype_class) if phenotype_class > 0 else ""),
109
        ("_loss={}".format(loss) if variational else ""),
110
111
        ("_encoding={}".format(logparam["encoding"]) if logparam is not None else ""),
        ("_k={}".format(logparam["k"]) if logparam is not None else ""),
112
        (datetime.now().strftime("%Y%m%d-%H%M%S")),
113
114
    )

115
    log_dir = os.path.abspath(os.path.join(outpath, "fit", run_ID))
116
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
117
118
119
        log_dir=log_dir,
        histogram_freq=1,
        profile_batch=2,
120
121
122
    )

    onecycle = deepof.model_utils.one_cycle_scheduler(
123
124
        X_train.shape[0] // batch_size * 250,
        max_rate=0.005,
125
126
    )

127
128
129
130
    callbacks = [run_ID, tensorboard_callback, onecycle]

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
131
            os.path.join(outpath, "checkpoints", run_ID + "/cp-{epoch:04d}.ckpt"),
132
133
134
135
136
137
138
139
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
140
141


142
def autoencoder_fitting(
143
    preprocessed_object: Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray],
144
145
    batch_size: int,
    encoding_size: int,
146
    epochs: int,
147
148
149
150
151
152
153
154
155
156
157
158
159
    hparams: dict,
    kl_warmup: int,
    log_history: bool,
    log_hparams: bool,
    loss: str,
    mmd_warmup,
    montecarlo_kl,
    n_components,
    output_path,
    phenotype_class,
    predictor: float,
    pretrained: str,
    save_checkpoints: bool,
160
    save_weights: bool,
161
    variational: bool,
162
):
163
164
    """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""

165
    # Load data
166
167
168
169
170
    X_train, y_train, X_val, y_val = preprocessed_object

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

171
    # Defines what to log on tensorboard (useful for trying out different models)
172

173
174
    logparam = {
        "encoding": encoding_size,
175
        "k": n_components,
176
177
178
179
180
        "loss": loss,
    }
    if phenotype_class:
        logparam["pheno_weight"] = phenotype_class

181
    # Load callbacks
182
    run_ID, *cbacks = get_callbacks(
183
184
185
186
        X_train=X_train,
        batch_size=batch_size,
        cp=save_checkpoints,
        variational=variational,
187
        phenotype_class=phenotype_class,
188
189
190
191
192
        predictor=predictor,
        loss=loss,
        logparam=logparam,
        outpath=output_path,
    )
193
194
    if not log_history:
        cbacks = cbacks[1:]
195

196
    # Logs hyperparameters to tensorboard
197
198
    if log_hparams:
        logparams = [
199
            hp.HParam(
200
201
202
203
                "encoding",
                hp.Discrete([2, 4, 6, 8, 12, 16]),
                display_name="encoding",
                description="encoding size dimensionality",
204
            ),
205
206
207
208
209
210
211
212
213
214
215
            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",
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
        rec = "reconstruction_" if phenotype_class else ""
        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",
                ),
            ]

        with tf.summary.create_file_writer(
            os.path.join(output_path, "hparams", run_ID)
        ).as_default():
            hp.hparams_config(
                hparams=logparams,
                metrics=metrics,
            )
251

252
    # Build models
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
    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,
        ).build(
            X_train.shape
        )
        return_list = (encoder, generator, grouper, ae)

    if pretrained:
287
        # If pretrained models are specified, load weights and return
288
289
290
291
292
293
294
295
296
        ae.load_weights(pretrained)
        return return_list

    else:
        if not variational:

            ae.fit(
                x=X_train,
                y=X_train,
297
                epochs=epochs,
298
299
300
                batch_size=batch_size,
                verbose=1,
                validation_data=(X_val, X_val),
301
302
                callbacks=cbacks
                + [
303
304
305
306
                    CustomStopper(
                        monitor="val_loss",
                        patience=5,
                        restore_best_weights=True,
307
                        start_epoch=max(kl_warmup, mmd_warmup),
308
309
310
311
                    ),
                ],
            )

312
313
314
            if save_weights:
                ae.save_weights("{}_final_weights.h5".format(run_ID))

315
316
        else:

317
            callbacks_ = cbacks + [
318
319
320
321
                CustomStopper(
                    monitor="val_loss",
                    patience=5,
                    restore_best_weights=True,
322
                    start_epoch=max(kl_warmup, mmd_warmup),
323
324
325
                ),
            ]

326
            if "ELBO" in loss and kl_warmup > 0:
327
328
                # noinspection PyUnboundLocalVariable
                callbacks_.append(kl_warmup_callback)
329
            if "MMD" in loss and mmd_warmup > 0:
330
331
332
                # noinspection PyUnboundLocalVariable
                callbacks_.append(mmd_warmup_callback)

333
334
335
336
337
338
339
            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:]]

340
            if phenotype_class > 0.0:
341
342
343
                ys += [y_train]
                yvals += [y_val]

344
            ae.fit(
345
346
                x=Xs,
                y=ys,
347
                epochs=epochs,
348
349
350
351
352
353
354
355
356
                batch_size=batch_size,
                verbose=1,
                validation_data=(
                    Xvals,
                    yvals,
                ),
                callbacks=callbacks_,
            )

357
358
359
            if save_weights:
                ae.save_weights("{}_final_weights.h5".format(run_ID))

360
361
362
            if log_hparams:
                # noinspection PyUnboundLocalVariable
                def tensorboard_metric_logging(run_dir: str, hpms: Any):
363
                    output = ae.predict(X_val)
364
365
366
367
368
369
370
371
372
373
374
                    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)
375
                        )
376
377
                        val_mse = tf.reduce_mean(
                            tf.keras.metrics.mean_squared_error(X_val, reconstruction)
378
                        )
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
                        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
                            )

                # Logparams to tensorboard
                tensorboard_metric_logging(
                    os.path.join(output_path, "hparams", run_ID),
                    logparam,
                )
414

415
416
417
    return return_list


418
def tune_search(
419
    data: List[np.array],
420
    encoding_size: int,
421
422
    hypertun_trials: int,
    hpt_type: str,
423
424
    hypermodel: str,
    k: int,
425
    kl_warmup_epochs: int,
426
    loss: str,
427
    mmd_warmup_epochs: int,
428
    overlap_loss: float,
429
    phenotype_class: float,
430
431
    predictor: float,
    project_name: str,
432
    callbacks: List,
433
    n_epochs: int = 30,
434
    n_replicas: int = 1,
435
) -> Union[bool, Tuple[Any, Any]]:
436
437
    """Define the search space using keras-tuner and bayesian optimization

438
439
440
441
442
443
444
445
446
447
448
    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
449
        - phenotype_class (float): adds an extra regularizing neural network to the model,
450
451
452
453
454
455
456
457
458
459
460
461
        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
462
463
464

    """

465
466
    X_train, y_train, X_val, y_val = data

467
468
469
470
    assert hpt_type in ["bayopt", "hyperband"], (
        "Invalid hyperparameter tuning framework. " "Select one of bayopt and hyperband"
    )

lucas_miranda's avatar
lucas_miranda committed
471
    if hypermodel == "S2SAE":  # pragma: no cover
472
        assert (
473
            predictor == 0.0 and phenotype_class == 0.0
474
        ), "Prediction branches are only available for variational models. See documentation for more details"
475
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_AE(input_shape=X_train.shape)
476
477
478

    elif hypermodel == "S2SGMVAE":
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_GMVAE(
479
            input_shape=X_train.shape,
480
            encoding=encoding_size,
481
            kl_warmup_epochs=kl_warmup_epochs,
482
            loss=loss,
483
            mmd_warmup_epochs=mmd_warmup_epochs,
484
            number_of_components=k,
485
            overlap_loss=overlap_loss,
486
            phenotype_predictor=phenotype_class,
487
            predictor=predictor,
488
        )
lucas_miranda's avatar
lucas_miranda committed
489

490
491
492
    else:
        return False

493
494
495
496
497
498
499
500
501
502
503
504
505
506
    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())),
            max_epochs=hypertun_trials,
507
            hyperband_iterations=3,
lucas_miranda's avatar
lucas_miranda committed
508
            factor=2,
509
510
511
512
513
514
515
516
            **hpt_params
        )
    else:
        tuner = BayesianOptimization(
            directory="BayOpt_{}_{}".format(loss, str(date.today())),
            max_trials=hypertun_trials,
            **hpt_params
        )
517
518
519

    print(tuner.search_space_summary())

520
521
522
523
524
525
526
    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:]]

527
    if phenotype_class > 0.0:
528
529
530
        ys += [y_train]
        yvals += [y_val]

531
    tuner.search(
532
533
        Xs,
        ys,
534
        epochs=n_epochs,
535
        validation_data=(Xvals, yvals),
536
537
        verbose=1,
        batch_size=256,
lucas_miranda's avatar
lucas_miranda committed
538
        callbacks=callbacks,
539
540
541
542
543
    )

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

lucas_miranda's avatar
lucas_miranda committed
544
545
    print(tuner.results_summary())

546
    return best_hparams, best_run