Mercurial > repos > bgruening > create_tool_recommendation_model

comparison optimise_hyperparameters.py @ 5:9ec705bd11cb draft default tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
planemo upload for repository https://github.com/bgruening/galaxytools/tree/recommendation_training/tools/tool_recommendation_model commit 24bab7a797f53fe4bcc668b18ee0326625486164
author bgruening
date Sun, 16 Oct 2022 11:51:32 +0000
parents f0da532be419
children
comparison
equal deleted inserted replaced
4:f0da532be419 5:9ec705bd11cb
1 """
2 Find the optimal combination of hyperparameters
3 """
4
5 import numpy as np
6 import utils
7 from hyperopt import fmin, hp, STATUS_OK, tpe, Trials
8 from tensorflow.keras.callbacks import EarlyStopping
9 from tensorflow.keras.layers import Dense, Dropout, Embedding, GRU, SpatialDropout1D
10 from tensorflow.keras.models import Sequential
11 from tensorflow.keras.optimizers import RMSprop
12
13
14 class HyperparameterOptimisation:
15 def __init__(self):
16 """ Init method. """
17
18 def train_model(
19 self,
20 config,
21 reverse_dictionary,
22 train_data,
23 train_labels,
24 test_data,
25 test_labels,
26 tool_tr_samples,
27 class_weights,
28 ):
29 """
30 Train a model and report accuracy
31 """
32 # convert items to integer
33 l_batch_size = list(map(int, config["batch_size"].split(",")))
34 l_embedding_size = list(map(int, config["embedding_size"].split(",")))
35 l_units = list(map(int, config["units"].split(",")))
36
37 # convert items to float
38 l_learning_rate = list(map(float, config["learning_rate"].split(",")))
39 l_dropout = list(map(float, config["dropout"].split(",")))
40 l_spatial_dropout = list(map(float, config["spatial_dropout"].split(",")))
41 l_recurrent_dropout = list(map(float, config["recurrent_dropout"].split(",")))
42
43 optimize_n_epochs = int(config["optimize_n_epochs"])
44
45 # get dimensions
46 dimensions = len(reverse_dictionary) + 1
47 best_model_params = dict()
48 early_stopping = EarlyStopping(
49 monitor="val_loss",
50 mode="min",
51 verbose=1,
52 min_delta=1e-1,
53 restore_best_weights=True,
54 )
55
56 # specify the search space for finding the best combination of parameters using Bayesian optimisation
57 params = {
58 "embedding_size": hp.quniform(
59 "embedding_size", l_embedding_size[0], l_embedding_size[1], 1
60 ),
61 "units": hp.quniform("units", l_units[0], l_units[1], 1),
62 "batch_size": hp.quniform(
63 "batch_size", l_batch_size[0], l_batch_size[1], 1
64 ),
65 "learning_rate": hp.loguniform(
66 "learning_rate", np.log(l_learning_rate[0]), np.log(l_learning_rate[1])
67 ),
68 "dropout": hp.uniform("dropout", l_dropout[0], l_dropout[1]),
69 "spatial_dropout": hp.uniform(
70 "spatial_dropout", l_spatial_dropout[0], l_spatial_dropout[1]
71 ),
72 "recurrent_dropout": hp.uniform(
73 "recurrent_dropout", l_recurrent_dropout[0], l_recurrent_dropout[1]
74 ),
75 }
76
77 def create_model(params):
78 model = Sequential()
79 model.add(
80 Embedding(dimensions, int(params["embedding_size"]), mask_zero=True)
81 )
82 model.add(SpatialDropout1D(params["spatial_dropout"]))
83 model.add(
84 GRU(
85 int(params["units"]),
86 dropout=params["dropout"],
87 recurrent_dropout=params["recurrent_dropout"],
88 return_sequences=True,
89 activation="elu",
90 )
91 )
92 model.add(Dropout(params["dropout"]))
93 model.add(
94 GRU(
95 int(params["units"]),
96 dropout=params["dropout"],
97 recurrent_dropout=params["recurrent_dropout"],
98 return_sequences=False,
99 activation="elu",
100 )
101 )
102 model.add(Dropout(params["dropout"]))
103 model.add(Dense(2 * dimensions, activation="sigmoid"))
104 optimizer_rms = RMSprop(lr=params["learning_rate"])
105 batch_size = int(params["batch_size"])
106 model.compile(
107 loss=utils.weighted_loss(class_weights), optimizer=optimizer_rms
108 )
109 print(model.summary())
110 model_fit = model.fit(
111 utils.balanced_sample_generator(
112 train_data,
113 train_labels,
114 batch_size,
115 tool_tr_samples,
116 reverse_dictionary,
117 ),
118 steps_per_epoch=len(train_data) // batch_size,
119 epochs=optimize_n_epochs,
120 callbacks=[early_stopping],
121 validation_data=(test_data, test_labels),
122 verbose=2,
123 shuffle=True,
124 )
125 return {
126 "loss": model_fit.history["val_loss"][-1],
127 "status": STATUS_OK,
128 "model": model,
129 }
130
131 # minimize the objective function using the set of parameters above
132 trials = Trials()
133 learned_params = fmin(
134 create_model,
135 params,
136 trials=trials,
137 algo=tpe.suggest,
138 max_evals=int(config["max_evals"]),
139 )
140 best_model = trials.results[np.argmin([r["loss"] for r in trials.results])][
141 "model"
142 ]
143 # set the best params with respective values
144 for item in learned_params:
145 item_val = learned_params[item]
146 best_model_params[item] = item_val
147 return best_model_params, best_model