train_utils.py 6.51 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
# @author lucasmiranda42
# encoding: utf-8
# module deepof

"""

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

"""
from datetime import datetime

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


23
def load_hparams(hparams):
24
25
26
27
28
29
30
31
32
33
34
35
    """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,
36
            "encoding": 16,
37
38
39
40
41
42
43
44
45
46
47
48
            "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,
49
                [i for i in os.listdir(train_path) if i.endswith(".pkl")][0],
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
            ),
            "rb",
        ) as handle:
            treatment_dict = pickle.load(handle)
    except IndexError:
        treatment_dict = None

    return treatment_dict


def get_callbacks(
    X_train: np.array,
    batch_size: int,
    variational: bool,
    predictor: float,
    k: int,
    loss: str,
    kl_wu: int,
    mmd_wu: int,
) -> Tuple:
    """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"""

    run_ID = "{}{}{}{}{}{}_{}".format(
        ("GMVAE" if variational else "AE"),
        ("P" if predictor > 0 and variational else ""),
        ("_components={}".format(k) if variational else ""),
        ("_loss={}".format(loss) if variational else ""),
        ("_kl_warmup={}".format(kl_wu) if variational else ""),
        ("_mmd_warmup={}".format(mmd_wu) if variational else ""),
        datetime.now().strftime("%Y%m%d-%H%M%S"),
    )

    log_dir = os.path.abspath("logs/fit/{}".format(run_ID))
    tensorboard_callback = keras.callbacks.TensorBoard(
        log_dir=log_dir, histogram_freq=1, profile_batch=2,
    )

91
92
93
94
95
96
    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",
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
    )

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

    return run_ID, tensorboard_callback, cp_callback, onecycle


def tune_search(
    train: np.array,
    test: np.array,
    bayopt_trials: int,
    hypermodel: str,
    k: int,
    kl_wu: int,
    loss: str,
    mmd_wu: int,
    overlap_loss: float,
    predictor: float,
    project_name: str,
118
    callbacks: List,
119
    n_epochs: int = 40,
120
    n_replicas: int = 1,
121
) -> Union[bool, Tuple[Any, Any]]:
122
123
124
125
126
127
    """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
128
129
            - hypermodel (str): hypermodel to load. Must be one of S2SAE (plain autoencoder)
            or S2SGMVAE (Gaussian Mixture Variational autoencoder).
130
131
132
133
134
135
136
137
138
139
            - k (int) number of components of the Gaussian Mixture
            - kl_wu (int): number of epochs for KL divergence warm up
            - loss (str): one of [ELBO, MMD, ELBO+MMD]
            - mmd_wu (int): number of epochs for MMD warm up
            - overlap_loss (float): assigns as weight to an extra loss term which
            penalizes overlap between GM components
            - 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
140
141
142
            - 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
143
144
145
146
147
148
149
150

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

    """

    tensorboard_callback, cp_callback, onecycle = callbacks
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170

    if hypermodel == "S2SAE":
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_AE(input_shape=train.shape)

    elif hypermodel == "S2SGMVAE":
        hypermodel = deepof.hypermodels.SEQ_2_SEQ_GMVAE(
            input_shape=train.shape,
            loss=loss,
            number_of_components=k,
            kl_warmup_epochs=kl_wu,
            mmd_warmup_epochs=mmd_wu,
            predictor=predictor,
            overlap_loss=overlap_loss,
        )
    else:
        return False

    tuner = BayesianOptimization(
        hypermodel,
        max_trials=bayopt_trials,
171
        executions_per_trial=n_replicas,
172
173
174
175
176
177
178
179
180
181
182
        objective="val_mae",
        seed=42,
        directory="BayesianOptx",
        project_name=project_name,
    )

    print(tuner.search_space_summary())

    tuner.search(
        train,
        train,
183
        epochs=n_epochs,
184
185
186
187
188
        validation_data=(test, test),
        verbose=1,
        batch_size=256,
        callbacks=[
            tensorboard_callback,
189
190
191
192
193
            tf.keras.callbacks.EarlyStopping(
                "val_mae", patience=10, restore_best_weights=True
            ),
            cp_callback,
            onecycle,
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
        ],
    )

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

    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.