#!/usr/bin/env python3
"""Module containing the RecurrentNeuralNetwork class and the command line interface."""
import argparse
import h5py
import json
import numpy as np
import pandas as pd
from biobb_common.generic.biobb_object import BiobbObject
from tensorflow.python.keras.saving import hdf5_format
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import LSTM
from tensorflow.keras.callbacks import EarlyStopping
from biobb_common.configuration import settings
from biobb_common.tools import file_utils as fu
from biobb_common.tools.file_utils import launchlogger
from biobb_ml.neural_networks.common import check_input_path, check_output_path, getHeader, getTargetValue, split_sequence, plotResultsReg
[docs]class RecurrentNeuralNetwork(BiobbObject):
"""
| biobb_ml RecurrentNeuralNetwork
| Wrapper of the TensorFlow Keras LSTM method using Recurrent Neural Networks.
| 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.
Args:
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).
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).
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).
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).
properties (dic - Python dictionary object containing the tool parameters, not input/output files):
* **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.
* **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.
* **window_size** (*int*) - (5) [0~100|1] Number of steps for each window of training model.
* **test_size** (*int*) - (5) [0~100000|1] Represents the number of samples of the dataset to include in the test split.
* **hidden_layers** (*list*) - (None) List of dictionaries with hidden layers values. Format: [ { 'size': 50, 'activation': 'relu' } ].
* **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).
* **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
* **batch_size** (*int*) - (100) [0~1000|1] Number of samples per gradient update.
* **max_epochs** (*int*) - (100) [0~1000|1] Number of epochs to train the model. As the early stopping is enabled, this is a maximum.
* **normalize_cm** (*bool*) - (False) Whether or not to normalize the confusion matrix.
* **remove_tmp** (*bool*) - (True) [WF property] Remove temporal files.
* **restart** (*bool*) - (False) [WF property] Do not execute if output files exist.
Examples:
This is a use example of how to use the building block from Python::
from biobb_ml.neural_networks.recurrent_neural_network import recurrent_neural_network
prop = {
'target': {
'column': 'target'
},
'window_size': 5,
'validation_size': 0.2,
'test_size': 0.2,
'hidden_layers': [
{
'size': 10,
'activation': 'relu'
},
{
'size': 8,
'activation': 'relu'
}
],
'optimizer': 'Adam',
'learning_rate': 0.01,
'batch_size': 32,
'max_epochs': 150
}
recurrent_neural_network(input_dataset_path='/path/to/myDataset.csv',
output_model_path='/path/to/newModel.h5',
output_test_table_path='/path/to/newTable.csv',
output_plot_path='/path/to/newPlot.png',
properties=prop)
Info:
* wrapped_software:
* name: TensorFlow Keras LSTM
* version: >2.1.0
* license: MIT
* ontology:
* name: EDAM
* schema: http://edamontology.org/EDAM.owl
"""
def __init__(self, input_dataset_path, output_model_path,
output_test_table_path=None, output_plot_path=None, properties=None, **kwargs) -> None:
properties = properties or {}
# Call parent class constructor
super().__init__(properties)
self.locals_var_dict = locals().copy()
# Input/Output files
self.io_dict = {
"in": {"input_dataset_path": input_dataset_path},
"out": {"output_model_path": output_model_path, "output_test_table_path": output_test_table_path, "output_plot_path": output_plot_path}
}
# Properties specific for BB
self.target = properties.get('target', '')
self.validation_size = properties.get('validation_size', 0.1)
self.window_size = properties.get('window_size', 5)
self.test_size = properties.get('test_size', 5)
self.hidden_layers = properties.get('hidden_layers', [])
self.optimizer = properties.get('optimizer', 'Adam')
self.learning_rate = properties.get('learning_rate', 0.02)
self.batch_size = properties.get('batch_size', 100)
self.max_epochs = properties.get('max_epochs', 100)
self.normalize_cm = properties.get('normalize_cm', False)
self.properties = properties
# Check the properties
self.check_properties(properties)
self.check_arguments()
[docs] def check_data_params(self, out_log, err_log):
""" Checks all the input/output paths and parameters """
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__)
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__)
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__)
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__)
[docs] def build_model(self, input_shape):
""" Builds Neural network according to hidden_layers property """
# create model
model = Sequential([])
# if no hidden_layers provided, create manually a hidden layer with default values
if not self.hidden_layers:
self.hidden_layers = [{'size': 50, 'activation': 'relu'}]
# generate hidden_layers
for i, layer in enumerate(self.hidden_layers):
if i == 0:
model.add(LSTM(layer['size'], activation=layer['activation'], kernel_initializer='he_normal', input_shape=input_shape)) # 1st hidden layer
else:
model.add(Dense(layer['size'], activation=layer['activation'], kernel_initializer='he_normal'))
model.add(Dense(1)) # output layer
return model
[docs] @launchlogger
def launch(self) -> int:
"""Execute the :class:`RecurrentNeuralNetwork <neural_networks.recurrent_neural_network.RecurrentNeuralNetwork>` neural_networks.recurrent_neural_network.RecurrentNeuralNetwork object."""
# check input/output paths and parameters
self.check_data_params(self.out_log, self.err_log)
# Setup Biobb
if self.check_restart():
return 0
self.stage_files()
# load dataset
fu.log('Getting dataset from %s' % self.io_dict["in"]["input_dataset_path"], self.out_log, self.global_log)
if 'column' in self.target:
labels = getHeader(self.io_dict["in"]["input_dataset_path"])
skiprows = 1
else:
labels = None
skiprows = None
data = pd.read_csv(self.io_dict["in"]["input_dataset_path"], header=None, sep="\\s+|;|:|,|\t", engine="python", skiprows=skiprows, names=labels)
# get target column
target = data[getTargetValue(self.target)].to_numpy()
# split into samples
X, y = split_sequence(target, self.window_size)
# reshape into [samples, timesteps, features]
X = X.reshape((X.shape[0], X.shape[1], 1))
# train / test split
fu.log('Creating train and test sets', self.out_log, self.global_log)
X_train, X_test, y_train, y_test = X[:-self.test_size], X[-self.test_size:], y[:-self.test_size], y[-self.test_size:]
# build model
fu.log('Building model', self.out_log, self.global_log)
model = self.build_model((X_train.shape[1], 1))
# model summary
stringlist = []
model.summary(print_fn=lambda x: stringlist.append(x))
model_summary = "\n".join(stringlist)
fu.log('Model summary:\n\n%s\n' % model_summary, self.out_log, self.global_log)
# get optimizer
mod = __import__('tensorflow.keras.optimizers', fromlist=[self.optimizer])
opt_class = getattr(mod, self.optimizer)
opt = opt_class(lr=self.learning_rate)
# compile model
model.compile(optimizer=opt, loss='mse', metrics=['mse', 'mae'])
# fitting
fu.log('Training model', self.out_log, self.global_log)
# set an early stopping mechanism
# set patience=2, to be a bit tolerant against random validation loss increases
early_stopping = EarlyStopping(patience=2)
# fit the model
mf = model.fit(X_train,
y_train,
batch_size=self.batch_size,
epochs=self.max_epochs,
callbacks=[early_stopping],
validation_split=self.validation_size,
verbose=1)
train_metrics = pd.DataFrame()
train_metrics['metric'] = ['Train loss', ' Train MAE', 'Train MSE', 'Validation loss', 'Validation MAE', 'Validation MSE']
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]]
fu.log('Training metrics\n\nTRAINING METRICS TABLE\n\n%s\n' % train_metrics, self.out_log, self.global_log)
# testing
fu.log('Testing model', self.out_log, self.global_log)
test_loss, test_mse, test_mae = model.evaluate(X_test, y_test)
# predict data from X_test
test_predictions = model.predict(X_test)
test_predictions = np.around(test_predictions, decimals=2)
tpr = np.squeeze(np.asarray(test_predictions))
test_metrics = pd.DataFrame()
test_metrics['metric'] = ['Test loss', 'Test MAE', 'Test MSE']
test_metrics['coefficient'] = [test_loss, test_mae, test_mse]
fu.log('Testing metrics\n\nTESTING METRICS TABLE\n\n%s\n' % test_metrics, self.out_log, self.global_log)
test_table = pd.DataFrame()
test_table['prediction'] = tpr
test_table['target'] = y_test
test_table['residual'] = test_table['target'] - test_table['prediction']
test_table['difference %'] = np.absolute(test_table['residual']/test_table['target']*100)
pd.set_option('display.float_format', lambda x: '%.2f' % x)
# sort by difference in %
test_table = test_table.sort_values(by=['difference %'])
test_table = test_table.reset_index(drop=True)
fu.log('TEST DATA\n\n%s\n' % test_table, self.out_log, self.global_log)
# save test data
if (self.io_dict["out"]["output_test_table_path"]):
fu.log('Saving testing data to %s' % self.io_dict["out"]["output_test_table_path"], self.out_log, self.global_log)
test_table.to_csv(self.io_dict["out"]["output_test_table_path"], index=False, header=True)
# create test plot
if (self.io_dict["out"]["output_plot_path"]):
fu.log('Saving plot to %s' % self.io_dict["out"]["output_plot_path"], self.out_log, self.global_log)
test_predictions = test_predictions.flatten()
train_predictions = model.predict(X_train).flatten()
plot = plotResultsReg(mf.history, y_test, test_predictions, y_train, train_predictions)
plot.savefig(self.io_dict["out"]["output_plot_path"], dpi=150)
# save model and parameters
vars_obj = {
'target': self.target,
'window_size': self.window_size,
'type': 'recurrent'
}
variables = json.dumps(vars_obj)
fu.log('Saving model to %s' % self.io_dict["out"]["output_model_path"], self.out_log, self.global_log)
with h5py.File(self.io_dict["out"]["output_model_path"], mode='w') as f:
hdf5_format.save_model_to_hdf5(model, f)
f.attrs['variables'] = variables
# Copy files to host
self.copy_to_host()
self.tmp_files.extend([
self.stage_io_dict.get("unique_dir")
])
self.remove_tmp_files()
self.check_arguments(output_files_created=True, raise_exception=False)
return 0
[docs]def 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:
"""Execute the :class:`RecurrentNeuralNetwork <neural_networks.recurrent_neural_network.RecurrentNeuralNetwork>` class and
execute the :meth:`launch() <neural_networks.recurrent_neural_network.RecurrentNeuralNetwork.launch>` method."""
return RecurrentNeuralNetwork(input_dataset_path=input_dataset_path,
output_model_path=output_model_path,
output_test_table_path=output_test_table_path,
output_plot_path=output_plot_path,
properties=properties, **kwargs).launch()
[docs]def main():
"""Command line execution of this building block. Please check the command line documentation."""
parser = argparse.ArgumentParser(description="Wrapper of the TensorFlow Keras LSTM method.", formatter_class=lambda prog: argparse.RawTextHelpFormatter(prog, width=99999))
parser.add_argument('--config', required=False, help='Configuration file')
# Specific args of each building block
required_args = parser.add_argument_group('required arguments')
required_args.add_argument('--input_dataset_path', required=True, help='Path to the input dataset. Accepted formats: csv.')
required_args.add_argument('--output_model_path', required=True, help='Path to the output model file. Accepted formats: h5.')
parser.add_argument('--output_test_table_path', required=False, help='Path to the test table file. Accepted formats: csv.')
parser.add_argument('--output_plot_path', required=False, help='Loss, accuracy and MSE plots. Accepted formats: png.')
args = parser.parse_args()
args.config = args.config or "{}"
properties = settings.ConfReader(config=args.config).get_prop_dic()
# Specific call of each building block
recurrent_neural_network(input_dataset_path=args.input_dataset_path,
output_model_path=args.output_model_path,
output_test_table_path=args.output_test_table_path,
output_plot_path=args.output_plot_path,
properties=properties)
if __name__ == '__main__':
main()