Coverage for biobb_ml/neural_networks/recurrent_neural_network.py: 88%
139 statements
« prev ^ index » next coverage.py v7.5.1, created at 2024-05-07 09:39 +0000
1#!/usr/bin/env python3
3"""Module containing the RecurrentNeuralNetwork class and the command line interface."""
4import argparse
5import h5py
6import json
7import numpy as np
8import pandas as pd
9from biobb_common.generic.biobb_object import BiobbObject
10from tensorflow.python.keras.saving import hdf5_format
11from tensorflow.keras import Sequential
12from tensorflow.keras.layers import Dense
13from tensorflow.keras.layers import LSTM
14from tensorflow.keras.callbacks import EarlyStopping
15from biobb_common.configuration import settings
16from biobb_common.tools import file_utils as fu
17from biobb_common.tools.file_utils import launchlogger
18from biobb_ml.neural_networks.common import check_input_path, check_output_path, getHeader, getTargetValue, split_sequence, plotResultsReg
21class RecurrentNeuralNetwork(BiobbObject):
22 """
23 | biobb_ml RecurrentNeuralNetwork
24 | Wrapper of the TensorFlow Keras LSTM method using Recurrent Neural Networks.
25 | Trains and tests a given dataset and save the complete model for a Recurrent Neural Network. Visit the `LSTM documentation page <https://www.tensorflow.org/api_docs/python/tf/keras/layers/LSTM>`_ in the TensorFlow Keras official website for further information.
27 Args:
28 input_dataset_path (str): Path to the input dataset. File type: input. `Sample file <https://github.com/bioexcel/biobb_ml/raw/master/biobb_ml/test/data/neural_networks/dataset_recurrent.csv>`_. Accepted formats: csv (edam:format_3752).
29 output_model_path (str): Path to the output model file. File type: output. `Sample file <https://github.com/bioexcel/biobb_ml/raw/master/biobb_ml/test/reference/neural_networks/ref_output_model_recurrent.h5>`_. Accepted formats: h5 (edam:format_3590).
30 output_test_table_path (str) (Optional): Path to the test table file. File type: output. `Sample file <https://github.com/bioexcel/biobb_ml/raw/master/biobb_ml/test/reference/neural_networks/ref_output_test_recurrent.csv>`_. Accepted formats: csv (edam:format_3752).
31 output_plot_path (str) (Optional): Loss, accuracy and MSE plots. File type: output. `Sample file <https://github.com/bioexcel/biobb_ml/raw/master/biobb_ml/test/reference/neural_networks/ref_output_plot_recurrent.png>`_. Accepted formats: png (edam:format_3603).
32 properties (dic - Python dictionary object containing the tool parameters, not input/output files):
33 * **target** (*dict*) - ({}) Dependent variable you want to predict from your dataset. You can specify either a column name or a column index. Formats: { "column": "column3" } or { "index": 21 }. In case of mulitple formats, the first one will be picked.
34 * **validation_size** (*float*) - (0.2) [0~1|0.05] Represents the proportion of the dataset to include in the validation split. It should be between 0.0 and 1.0.
35 * **window_size** (*int*) - (5) [0~100|1] Number of steps for each window of training model.
36 * **test_size** (*int*) - (5) [0~100000|1] Represents the number of samples of the dataset to include in the test split.
37 * **hidden_layers** (*list*) - (None) List of dictionaries with hidden layers values. Format: [ { 'size': 50, 'activation': 'relu' } ].
38 * **optimizer** (*string*) - ("Adam") Name of optimizer instance. Values: Adadelta (Adadelta optimization is a stochastic gradient descent method that is based on adaptive learning rate per dimension to address two drawbacks: the continual decay of learning rates throughout training and the need for a manually selected global learning rate), Adagrad (Adagrad is an optimizer with parameter-specific learning rates; which are adapted relative to how frequently a parameter gets updated during training. The more updates a parameter receives; the smaller the updates), Adam (Adam optimization is a stochastic gradient descent method that is based on adaptive estimation of first-order and second-order moments), Adamax (It is a variant of Adam based on the infinity norm. Default parameters follow those provided in the paper. Adamax is sometimes superior to adam; specially in models with embeddings), Ftrl (Optimizer that implements the FTRL algorithm), Nadam (Much like Adam is essentially RMSprop with momentum; Nadam is Adam with Nesterov momentum), RMSprop (Optimizer that implements the RMSprop algorithm), SGD (Gradient descent -with momentum- optimizer).
39 * **learning_rate** (*float*) - (0.02) [0~100|0.01] Determines the step size at each iteration while moving toward a minimum of a loss function
40 * **batch_size** (*int*) - (100) [0~1000|1] Number of samples per gradient update.
41 * **max_epochs** (*int*) - (100) [0~1000|1] Number of epochs to train the model. As the early stopping is enabled, this is a maximum.
42 * **normalize_cm** (*bool*) - (False) Whether or not to normalize the confusion matrix.
43 * **remove_tmp** (*bool*) - (True) [WF property] Remove temporal files.
44 * **restart** (*bool*) - (False) [WF property] Do not execute if output files exist.
46 Examples:
47 This is a use example of how to use the building block from Python::
49 from biobb_ml.neural_networks.recurrent_neural_network import recurrent_neural_network
50 prop = {
51 'target': {
52 'column': 'target'
53 },
54 'window_size': 5,
55 'validation_size': 0.2,
56 'test_size': 0.2,
57 'hidden_layers': [
58 {
59 'size': 10,
60 'activation': 'relu'
61 },
62 {
63 'size': 8,
64 'activation': 'relu'
65 }
66 ],
67 'optimizer': 'Adam',
68 'learning_rate': 0.01,
69 'batch_size': 32,
70 'max_epochs': 150
71 }
72 recurrent_neural_network(input_dataset_path='/path/to/myDataset.csv',
73 output_model_path='/path/to/newModel.h5',
74 output_test_table_path='/path/to/newTable.csv',
75 output_plot_path='/path/to/newPlot.png',
76 properties=prop)
78 Info:
79 * wrapped_software:
80 * name: TensorFlow Keras LSTM
81 * version: >2.1.0
82 * license: MIT
83 * ontology:
84 * name: EDAM
85 * schema: http://edamontology.org/EDAM.owl
87 """
89 def __init__(self, input_dataset_path, output_model_path,
90 output_test_table_path=None, output_plot_path=None, properties=None, **kwargs) -> None:
91 properties = properties or {}
93 # Call parent class constructor
94 super().__init__(properties)
95 self.locals_var_dict = locals().copy()
97 # Input/Output files
98 self.io_dict = {
99 "in": {"input_dataset_path": input_dataset_path},
100 "out": {"output_model_path": output_model_path, "output_test_table_path": output_test_table_path, "output_plot_path": output_plot_path}
101 }
103 # Properties specific for BB
104 self.target = properties.get('target', '')
105 self.validation_size = properties.get('validation_size', 0.1)
106 self.window_size = properties.get('window_size', 5)
107 self.test_size = properties.get('test_size', 5)
108 self.hidden_layers = properties.get('hidden_layers', [])
109 self.optimizer = properties.get('optimizer', 'Adam')
110 self.learning_rate = properties.get('learning_rate', 0.02)
111 self.batch_size = properties.get('batch_size', 100)
112 self.max_epochs = properties.get('max_epochs', 100)
113 self.normalize_cm = properties.get('normalize_cm', False)
114 self.properties = properties
116 # Check the properties
117 self.check_properties(properties)
118 self.check_arguments()
120 def check_data_params(self, out_log, err_log):
121 """ Checks all the input/output paths and parameters """
122 self.io_dict["in"]["input_dataset_path"] = check_input_path(self.io_dict["in"]["input_dataset_path"], "input_dataset_path", False, out_log, self.__class__.__name__)
123 self.io_dict["out"]["output_model_path"] = check_output_path(self.io_dict["out"]["output_model_path"], "output_model_path", False, out_log, self.__class__.__name__)
124 self.io_dict["out"]["output_test_table_path"] = check_output_path(self.io_dict["out"]["output_test_table_path"], "output_test_table_path", True, out_log, self.__class__.__name__)
125 self.io_dict["out"]["output_plot_path"] = check_output_path(self.io_dict["out"]["output_plot_path"], "output_plot_path", True, out_log, self.__class__.__name__)
127 def build_model(self, input_shape):
128 """ Builds Neural network according to hidden_layers property """
130 # create model
131 model = Sequential([])
133 # if no hidden_layers provided, create manually a hidden layer with default values
134 if not self.hidden_layers:
135 self.hidden_layers = [{'size': 50, 'activation': 'relu'}]
137 # generate hidden_layers
138 for i, layer in enumerate(self.hidden_layers):
139 if i == 0:
140 model.add(LSTM(layer['size'], activation=layer['activation'], kernel_initializer='he_normal', input_shape=input_shape)) # 1st hidden layer
141 else:
142 model.add(Dense(layer['size'], activation=layer['activation'], kernel_initializer='he_normal'))
144 model.add(Dense(1)) # output layer
146 return model
148 @launchlogger
149 def launch(self) -> int:
150 """Execute the :class:`RecurrentNeuralNetwork <neural_networks.recurrent_neural_network.RecurrentNeuralNetwork>` neural_networks.recurrent_neural_network.RecurrentNeuralNetwork object."""
152 # check input/output paths and parameters
153 self.check_data_params(self.out_log, self.err_log)
155 # Setup Biobb
156 if self.check_restart():
157 return 0
158 self.stage_files()
160 # load dataset
161 fu.log('Getting dataset from %s' % self.io_dict["in"]["input_dataset_path"], self.out_log, self.global_log)
162 if 'column' in self.target:
163 labels = getHeader(self.io_dict["in"]["input_dataset_path"])
164 skiprows = 1
165 else:
166 labels = None
167 skiprows = None
168 data = pd.read_csv(self.io_dict["in"]["input_dataset_path"], header=None, sep="\\s+|;|:|,|\t", engine="python", skiprows=skiprows, names=labels)
170 # get target column
171 target = data[getTargetValue(self.target)].to_numpy()
173 # split into samples
174 X, y = split_sequence(target, self.window_size)
175 # reshape into [samples, timesteps, features]
176 X = X.reshape((X.shape[0], X.shape[1], 1))
178 # train / test split
179 fu.log('Creating train and test sets', self.out_log, self.global_log)
180 X_train, X_test, y_train, y_test = X[:-self.test_size], X[-self.test_size:], y[:-self.test_size], y[-self.test_size:]
182 # build model
183 fu.log('Building model', self.out_log, self.global_log)
184 model = self.build_model((X_train.shape[1], 1))
186 # model summary
187 stringlist = []
188 model.summary(print_fn=lambda x: stringlist.append(x))
189 model_summary = "\n".join(stringlist)
190 fu.log('Model summary:\n\n%s\n' % model_summary, self.out_log, self.global_log)
192 # get optimizer
193 mod = __import__('tensorflow.keras.optimizers', fromlist=[self.optimizer])
194 opt_class = getattr(mod, self.optimizer)
195 opt = opt_class(lr=self.learning_rate)
196 # compile model
197 model.compile(optimizer=opt, loss='mse', metrics=['mse', 'mae'])
199 # fitting
200 fu.log('Training model', self.out_log, self.global_log)
201 # set an early stopping mechanism
202 # set patience=2, to be a bit tolerant against random validation loss increases
203 early_stopping = EarlyStopping(patience=2)
204 # fit the model
205 mf = model.fit(X_train,
206 y_train,
207 batch_size=self.batch_size,
208 epochs=self.max_epochs,
209 callbacks=[early_stopping],
210 validation_split=self.validation_size,
211 verbose=1)
213 train_metrics = pd.DataFrame()
214 train_metrics['metric'] = ['Train loss', ' Train MAE', 'Train MSE', 'Validation loss', 'Validation MAE', 'Validation MSE']
215 train_metrics['coefficient'] = [mf.history['loss'][-1], mf.history['mae'][-1], mf.history['mse'][-1], mf.history['val_loss'][-1], mf.history['val_mae'][-1], mf.history['val_mse'][-1]]
217 fu.log('Training metrics\n\nTRAINING METRICS TABLE\n\n%s\n' % train_metrics, self.out_log, self.global_log)
219 # testing
220 fu.log('Testing model', self.out_log, self.global_log)
221 test_loss, test_mse, test_mae = model.evaluate(X_test, y_test)
223 # predict data from X_test
224 test_predictions = model.predict(X_test)
225 test_predictions = np.around(test_predictions, decimals=2)
226 tpr = np.squeeze(np.asarray(test_predictions))
228 test_metrics = pd.DataFrame()
229 test_metrics['metric'] = ['Test loss', 'Test MAE', 'Test MSE']
230 test_metrics['coefficient'] = [test_loss, test_mae, test_mse]
232 fu.log('Testing metrics\n\nTESTING METRICS TABLE\n\n%s\n' % test_metrics, self.out_log, self.global_log)
234 test_table = pd.DataFrame()
235 test_table['prediction'] = tpr
236 test_table['target'] = y_test
237 test_table['residual'] = test_table['target'] - test_table['prediction']
238 test_table['difference %'] = np.absolute(test_table['residual']/test_table['target']*100)
239 pd.set_option('display.float_format', lambda x: '%.2f' % x)
240 # sort by difference in %
241 test_table = test_table.sort_values(by=['difference %'])
242 test_table = test_table.reset_index(drop=True)
243 fu.log('TEST DATA\n\n%s\n' % test_table, self.out_log, self.global_log)
245 # save test data
246 if (self.io_dict["out"]["output_test_table_path"]):
247 fu.log('Saving testing data to %s' % self.io_dict["out"]["output_test_table_path"], self.out_log, self.global_log)
248 test_table.to_csv(self.io_dict["out"]["output_test_table_path"], index=False, header=True)
250 # create test plot
251 if (self.io_dict["out"]["output_plot_path"]):
252 fu.log('Saving plot to %s' % self.io_dict["out"]["output_plot_path"], self.out_log, self.global_log)
253 test_predictions = test_predictions.flatten()
254 train_predictions = model.predict(X_train).flatten()
255 plot = plotResultsReg(mf.history, y_test, test_predictions, y_train, train_predictions)
256 plot.savefig(self.io_dict["out"]["output_plot_path"], dpi=150)
258 # save model and parameters
259 vars_obj = {
260 'target': self.target,
261 'window_size': self.window_size,
262 'type': 'recurrent'
263 }
264 variables = json.dumps(vars_obj)
265 fu.log('Saving model to %s' % self.io_dict["out"]["output_model_path"], self.out_log, self.global_log)
266 with h5py.File(self.io_dict["out"]["output_model_path"], mode='w') as f:
267 hdf5_format.save_model_to_hdf5(model, f)
268 f.attrs['variables'] = variables
270 # Copy files to host
271 self.copy_to_host()
273 self.tmp_files.extend([
274 self.stage_io_dict.get("unique_dir")
275 ])
276 self.remove_tmp_files()
278 self.check_arguments(output_files_created=True, raise_exception=False)
280 return 0
283def recurrent_neural_network(input_dataset_path: str, output_model_path: str, output_test_table_path: str = None, output_plot_path: str = None, properties: dict = None, **kwargs) -> int:
284 """Execute the :class:`RecurrentNeuralNetwork <neural_networks.recurrent_neural_network.RecurrentNeuralNetwork>` class and
285 execute the :meth:`launch() <neural_networks.recurrent_neural_network.RecurrentNeuralNetwork.launch>` method."""
287 return RecurrentNeuralNetwork(input_dataset_path=input_dataset_path,
288 output_model_path=output_model_path,
289 output_test_table_path=output_test_table_path,
290 output_plot_path=output_plot_path,
291 properties=properties, **kwargs).launch()
294def main():
295 """Command line execution of this building block. Please check the command line documentation."""
296 parser = argparse.ArgumentParser(description="Wrapper of the TensorFlow Keras LSTM method.", formatter_class=lambda prog: argparse.RawTextHelpFormatter(prog, width=99999))
297 parser.add_argument('--config', required=False, help='Configuration file')
299 # Specific args of each building block
300 required_args = parser.add_argument_group('required arguments')
301 required_args.add_argument('--input_dataset_path', required=True, help='Path to the input dataset. Accepted formats: csv.')
302 required_args.add_argument('--output_model_path', required=True, help='Path to the output model file. Accepted formats: h5.')
303 parser.add_argument('--output_test_table_path', required=False, help='Path to the test table file. Accepted formats: csv.')
304 parser.add_argument('--output_plot_path', required=False, help='Loss, accuracy and MSE plots. Accepted formats: png.')
306 args = parser.parse_args()
307 args.config = args.config or "{}"
308 properties = settings.ConfReader(config=args.config).get_prop_dic()
310 # Specific call of each building block
311 recurrent_neural_network(input_dataset_path=args.input_dataset_path,
312 output_model_path=args.output_model_path,
313 output_test_table_path=args.output_test_table_path,
314 output_plot_path=args.output_plot_path,
315 properties=properties)
318if __name__ == '__main__':
319 main()