train_utils.py 7.05 KB
Newer Older
1
2
3
4
5
6
7
8
9
# @author lucasmiranda42
# encoding: utf-8
# module deepof

"""

Simple utility functions used in deepof example scripts. These are not part of the main package

"""
10
from datetime import date, datetime
11

12
from kerastuner import BayesianOptimization, Hyperband
13
from kerastuner import HyperParameters
14
from kerastuner_tensorboard_logger import TensorBoardLogger
15
from typing import Tuple, Union, Any, List
16
17
18
19
20
21
22
import deepof.hypermodels
import deepof.model_utils
import numpy as np
import os
import pickle
import tensorflow as tf

23
24
hp = HyperParameters()

25

26
def load_hparams(hparams):
27
28
29
30
31
32
33
34
35
36
37
38
    """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 = {
            "units_conv": 256,
            "units_lstm": 256,
            "units_dense2": 64,
            "dropout_rate": 0.25,
39
            "encoding": 16,
40
41
42
43
44
45
46
47
48
49
50
51
            "learning_rate": 1e-3,
        }
    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,
52
                [i for i in os.listdir(train_path) if i.endswith(".pkl")][0],
53
54
55
56
57
58
59
60
61
62
63
            ),
            "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
64
65
66
67
68
69
    X_train: np.array,
    batch_size: int,
    cp: bool,
    variational: bool,
    predictor: float,
    loss: str,
70
) -> List[Union[Any]]:
71
72
73
74
75
76
    """Generates callbacks for model training, including:
        - run_ID: run name, with coarse parameter details;
        - tensorboard_callback: for real-time visualization;
        - cp_callback: for checkpoint saving,
        - onecycle: for learning rate scheduling"""

77
    run_ID = "{}{}{}_{}".format(
78
79
80
81
82
83
84
        ("GMVAE" if variational else "AE"),
        ("P" if predictor > 0 and variational else ""),
        ("_loss={}".format(loss) if variational else ""),
        datetime.now().strftime("%Y%m%d-%H%M%S"),
    )

    log_dir = os.path.abspath("logs/fit/{}".format(run_ID))
85
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
86
87
88
89
90
91
92
        log_dir=log_dir, histogram_freq=1, profile_batch=2,
    )

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

93
94
95
96
97
98
99
100
101
102
103
104
105
    callbacks = [run_ID, tensorboard_callback, onecycle]

    if cp:
        cp_callback = tf.keras.callbacks.ModelCheckpoint(
            "./logs/checkpoints/" + run_ID + "/cp-{epoch:04d}.ckpt",
            verbose=1,
            save_best_only=False,
            save_weights_only=True,
            save_freq="epoch",
        )
        callbacks.append(cp_callback)

    return callbacks
106
107
108


def tune_search(
109
    data: List[np.array],
110
111
112
    bayopt_trials: int,
    hypermodel: str,
    k: int,
113
    kl_warmup_epochs: int,
114
    loss: str,
115
    mmd_warmup_epochs: int,
116
    overlap_loss: float,
117
    pheno_class: float,
118
119
    predictor: float,
    project_name: str,
120
    callbacks: List,
121
    n_epochs: int = 30,
122
    n_replicas: int = 1,
123
) -> Union[bool, Tuple[Any, Any]]:
124
125
126
127
128
129
    """Define the search space using keras-tuner and bayesian optimization

        Parameters:
            - train (np.array): dataset to train the model on
            - test (np.array): dataset to validate the model on
            - bayopt_trials (int): number of Bayesian optimization iterations to run
130
131
            - hypermodel (str): hypermodel to load. Must be one of S2SAE (plain autoencoder)
            or S2SGMVAE (Gaussian Mixture Variational autoencoder).
132
133
134
135
            - 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
136
137
            - pheno_class (float): adds an extra regularizing neural network to the model,
            which tries to predict the phenotype of the animal from which the sequence comes
138
139
140
141
            - 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
142
143
144
            - 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
145
146
147
148
149
150
151

        Returns:
            - best_hparams (dict): dictionary with the best retrieved hyperparameters
            - best_run (tf.keras.Model): trained instance of the best model found

    """

152
153
    X_train, y_train, X_val, y_val = data

lucas_miranda's avatar
lucas_miranda committed
154
    if hypermodel == "S2SAE":  # pragma: no cover
155
156
157
        assert (
            predictor == 0.0 and pheno_class == 0.0
        ), "Prediction branches are only available for variational models. See documentation for more details"
158
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_AE(input_shape=X_train.shape)
159
160
161

    elif hypermodel == "S2SGMVAE":
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_GMVAE(
162
            input_shape=X_train.shape,
163
            kl_warmup_epochs=kl_warmup_epochs,
164
            loss=loss,
165
            mmd_warmup_epochs=mmd_warmup_epochs,
166
            number_of_components=k,
167
            overlap_loss=overlap_loss,
168
            phenotype_predictor=pheno_class,
169
            predictor=predictor,
170
        )
lucas_miranda's avatar
lucas_miranda committed
171

172
173
174
    else:
        return False

175
    tuner = Hyperband(
176
        hypermodel,
177
        directory="HyperBandx_{}_{}".format(loss, str(date.today())),
178
        executions_per_trial=n_replicas,
179
        logger=TensorBoardLogger(metrics=["val_mae"], logdir="./logs/hparams"),
180
        max_epochs=bayopt_trials,
181
182
        objective="val_mae",
        project_name=project_name,
183
        seed=42,
184
        tune_new_entries=True,
185
186
187
188
    )

    print(tuner.search_space_summary())

189
190
191
192
193
194
195
196
197
198
199
    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:]]

    if pheno_class > 0.0:
        ys += [y_train]
        yvals += [y_val]

200
    tuner.search(
201
202
        Xs,
        ys,
203
        epochs=n_epochs,
204
        validation_data=(Xvals, yvals),
205
206
        verbose=1,
        batch_size=256,
lucas_miranda's avatar
lucas_miranda committed
207
        callbacks=callbacks,
208
209
210
211
212
    )

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

lucas_miranda's avatar
lucas_miranda committed
213
214
    print(tuner.results_summary())

215
216
217
218
219
220
221
222
    return best_hparams, best_run


# TODO:
#    - load_treatments should be part of the main data module. If available in the main directory,
#    a table (preferrable in csv) should be loaded as metadata of the coordinates automatically.
#    This becomes particularly important por the supervised models that include phenotype classification
#    alongside the encoding.