def main():
"""
Loads the config file, creates the library and controller, and starts the
training loop.
"""
# Load the config file
config_filename = 'config.json'
with open(config_filename, encoding='utf-8') as f:
config = json.load(f)
config_dataset = config["dataset"] # Problem specification hyperparameters
config_training = config["training"] # Training hyperparameters
config_controller = config["controller"] # Controller hyperparameters
# Define the dataset and library
dataset = Dataset(**config_dataset)
Program.set_training_data(dataset)
Program.set_library(dataset.function_set, dataset.n_input_var)
print("Ground truth expression:\n{}".format(indent(dataset.pretty(),
'\t')))
with tf.Session() as sess:
# Instantiate the controller
controller = Controller(sess,
debug=config_training["debug"],
summary=config_training["summary"],
**config_controller)
learn(sess, controller, **config_training)
def setup(self, seed=0):
# Clear the cache, reset the compute graph, and set the seed
Program.clear_cache()
tf.reset_default_graph()
self.seed(seed) # Must be called _after_ resetting graph
self.pool = self.make_pool()
self.prior = self.make_prior()
# if self.config_task['enforce_sum']:
# n_tensors = 0
# for input in self.config_task['dataset_info']['input']:
# # count the number of tensors in the inputs, if any:
# if input in ['T1', 'T2', 'T3', 'T4', 'T5', 'T6', 'T7', 'T8', 'T9', 'T10']:
# n_tensors += 1
# else:
# n_tensors = 1
# for ii in range(n_tensors):
graph = tf.Graph()
with graph.as_default():
self.seed(seed)
self.sess = tf.Session()
new_controller = Controller(self.sess,
self.prior,
self.config_task['enforce_sum'],
seed,
**self.config_controller)
new_controller.sess.run(tf.global_variables_initializer()) # initializer should be part of the graph
self.controller = new_controller
def make_controller(self):
controller = []
for sess in self.sess:
controller.append(Controller(sess,
self.prior,
**self.config_controller))
return controller
def fit(self, X, y):
# Define the dataset and library
dataset = Dataset(X, y)
Program.clear_cache()
Program.set_training_data(dataset)
Program.set_library(
['add', 'sub', 'mul', 'div', 'sin', 'cos', 'exp', 'log'],
X.shape[1])
tf.reset_default_graph()
# Shift actual seed by checksum to ensure it's different across different benchmarks
tf.set_random_seed(0)
with tf.Session() as sess:
# Instantiate the controller
controller = Controller(sess,
debug=False,
summary=False,
**self.config_controller)
# Train the controller
result: Program = learn(
sess,
controller,
**self.config_training,
return_estimator=True) # r, base_r, expression, traversal
self.result = result
def train_dsr(name_and_seed, config_dataset, config_controller,
config_training):
"""Trains DSR and returns dict of reward, expression, and traversal"""
name, seed = name_and_seed
try:
import tensorflow as tf
from dsr.controller import Controller
from dsr.train import learn
# Ignore TensorFlow warnings
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
except:
pass
start = time.time()
# Rename the output file
config_training["output_file"] = "dsr_{}_{}.csv".format(name, seed)
# Define the dataset and library
dataset = get_dataset(name, config_dataset)
Program.clear_cache()
Program.set_training_data(dataset)
Program.set_library(dataset.function_set, dataset.n_input_var)
tf.reset_default_graph()
# Shift actual seed by checksum to ensure it's different across different benchmarks
tf.set_random_seed(seed + zlib.adler32(name.encode("utf-8")))
with tf.Session() as sess:
# Instantiate the controller
controller = Controller(sess,
debug=config_training["debug"],
summary=config_training["summary"],
**config_controller)
# Train the controller
result = learn(sess, controller,
**config_training) # r, base_r, expression, traversal
result["name"] = name
result["t"] = time.time() - start
result["seed"] = seed
return result
def make_controller(self):
controller = Controller(self.sess, self.prior,
**self.config_controller)
return controller