def get_search_configs():
"""
Defines the configurations for hyper parameter search
"""
configurations.DEFINE_string("template", None,
"Template file for hyper-param search")
configurations.DEFINE_string(
"search_algorithm", "genetic",
"Algorithm for hyper-param optimization. Select from 'genetic', 'grid_search'"
)
configurations.DEFINE_integer(
"generations", 100, "Number of generations for genetic algorithm")
configurations.DEFINE_integer("pop_size", 20,
"Population size for genetic algorithm")
configurations.DEFINE_integer("num_survivors", 10,
"Number of survivors for genetic algorithm")
configurations.DEFINE_integer(
"num_threads", 4,
"NUmber of parallel threads (Number of parallel executions)")
configurations.DEFINE_integer(
"num_gpu", 1, "Number of GPU on the machine, Use 0 if there are None")
configurations.DEFINE_integer("sleep_time", 1, "Sleep time")
configurations.DEFINE_float("mutate_rate", 0.2,
"Mutation rate for genetic algorithm")
configurations.DEFINE_string(
"init_pop", None,
"Specify starting population. Path to the pickle file")
c = configurations.ConfigValues()
return c
def main(_):
"""
The model specified command line arg --model_dir is applied to every data
point in --test_datafile and the model output is sent to --output. The unix
command 'paste' can be used to stich the input file and output together.
e.g.,
$ classifiy_data.py --config=train.conf --test_datafile=test.dat > output.dat
$ paste -d ' ' test.dat output.dat > input_and_output.dat
"""
configs.DEFINE_string('test_datafile',None,'file with test data')
configs.DEFINE_string('time_field','date','fields used for dates/time')
configs.DEFINE_string('print_start','190001','only print data on or after')
configs.DEFINE_string('print_end','999912','only print data on or before')
configs.DEFINE_integer('num_batches',None,'num_batches overrride')
config = configs.get_configs()
if config.test_datafile is None:
config.test_datafile = config.datafile
batch_size = 1
data_path = model_utils.get_data_path(config.data_dir,config.test_datafile)
# print("Loading data %s"%data_path)
dataset = BatchGenerator(data_path, config,
batch_size=batch_size,
num_unrollings=config.num_unrollings)
num_data_points = dataset.num_batches
if config.num_batches is not None:
num_data_points = config.num_batches
#print("num_batches = ", num_data_points)
tf_config = tf.ConfigProto( allow_soft_placement=True,
log_device_placement=False )
with tf.Graph().as_default(), tf.Session(config=tf_config) as session:
#print("Loading model.")
model = model_utils.get_trained_model(session, config, verbose=False)
for i in range(num_data_points):
batch = dataset.next_batch()
preds = model.step(session, batch)
seq_len = get_seq_length(batch)
key, date = get_key_and_date(batch, seq_len-1)
if (date < config.print_start or date > config.print_end):
continue
score = get_score(config, preds, seq_len-1)
target = get_target(config, batch, seq_len-1)
print("%s %s %.6f %.6f %d" % (key, date, score, target, seq_len))
def get_configs():
"""
Defines the configurations for hyper parameter search
"""
configurations.DEFINE_string("configs_fname",None,"CSV containing all the configs to run")
configurations.DEFINE_boolean("predict",True,"Run predictions after training")
configurations.DEFINE_integer("num_threads",4,"NUmber of parallel threads (Number of parallel executions)")
configurations.DEFINE_integer("num_gpu",1,"Number of GPU on the machine, Use 0 if there are None")
configurations.DEFINE_integer("sleep_time",1,"Sleep time")
configurations.DEFINE_integer("start_date",None,"First date for prediction on as YYYYMM")
configurations.DEFINE_integer("end_date",None,"Last date for prediction on as YYYYMM")
c = configurations.ConfigValues()
return c
def get_configs():
"""
Defines all configuration params passable to command line.
"""
configs.DEFINE_string("datasource", 'big_datafile',
"The source of the data.")
configs.DEFINE_string("tkrlist", "big_tkrlist.csv",
"The list of filters to use.")
configs.DEFINE_string("datafile", 'big_datafile.dat', "a datafile name.")
configs.DEFINE_string("mse_outfile", None,
"A file to write mse values during predict phase.")
configs.DEFINE_string("default_gpu", '',
"The default GPU to use e.g., /gpu:0")
configs.DEFINE_string("nn_type", 'DeepRnnModel', "Model type")
configs.DEFINE_string("active_field", 'active',
"Key column name header for active indicator")
configs.DEFINE_string("key_field", 'gvkey',
"Key column name header in datafile")
configs.DEFINE_string("target_field", 'oiadpq_ttm',
"Target column name header in datafile")
configs.DEFINE_string("scale_field", 'mrkcap',
"Feature to scale inputs by")
configs.DEFINE_string("feature_fields", '',
"shared input and target field names")
configs.DEFINE_string("aux_input_fields", None,
"non-target, input only fields")
configs.DEFINE_string("data_dir", '', "The data directory")
configs.DEFINE_string("model_dir", '', "Model directory")
configs.DEFINE_string("rnn_cell", 'gru', "lstm or gru")
configs.DEFINE_integer("num_inputs", -1, "")
configs.DEFINE_integer("num_outputs", -1, "")
configs.DEFINE_integer("target_idx", None, "")
configs.DEFINE_integer("min_unrollings", None,
"Min number of unrolling steps")
configs.DEFINE_integer("max_unrollings", None,
"Max number of unrolling steps")
# num_unrollings is being depricated by max_unrollings
configs.DEFINE_integer("num_unrollings", 4, "Number of unrolling steps")
configs.DEFINE_integer("stride", 12,
"How many steps to skip per unrolling")
configs.DEFINE_integer("forecast_n", 12,
"How many steps to forecast into the future")
configs.DEFINE_integer("batch_size", 1, "Size of each batch")
configs.DEFINE_integer("num_layers", 1, "Numer of RNN layers")
configs.DEFINE_integer("num_hidden", 10, "Number of hidden layer units")
configs.DEFINE_float("init_scale", 0.1, "Initial scale for weights")
configs.DEFINE_float("max_grad_norm", 10.0, "Gradient clipping")
configs.DEFINE_integer("start_date", None,
"First date to train on as YYYYMM")
configs.DEFINE_integer("end_date", None, "Last date to train on as YYYYMM")
configs.DEFINE_float("keep_prob", 1.0, "Keep probability for dropout")
configs.DEFINE_boolean("train", True,
"Train model otherwise inference only")
configs.DEFINE_boolean("input_dropout", False, "Do dropout on input layer")
configs.DEFINE_boolean("hidden_dropout", False,
"Do dropout on hidden layers")
configs.DEFINE_boolean("rnn_dropout", False,
"Do dropout on recurrent connections")
configs.DEFINE_boolean(
"skip_connections", False,
"Have direct connections between input and output in MLP")
configs.DEFINE_boolean(
"use_cache", True,
"Load data for logreg from cache (vs processing from batch generator)")
configs.DEFINE_boolean(
"pretty_print_preds", False,
"Print predictions in tabular format with inputs, targets, and keys")
configs.DEFINE_boolean("scale_targets", True, "")
configs.DEFINE_string("data_scaler", None,
'sklearn scaling algorithm or None if no scaling')
configs.DEFINE_string("optimizer", 'GradientDescentOptimizer',
'Any tensorflow optimizer in tf.train')
configs.DEFINE_string("optimizer_params", None,
'Additional optimizer params such as momentum')
configs.DEFINE_float("learning_rate", 0.6,
"The initial starting learning rate")
configs.DEFINE_float("lr_decay", 0.9, "Learning rate decay")
configs.DEFINE_float("validation_size", 0.0,
"Size of validation set as %, ie. .3 = 30% of data")
configs.DEFINE_float("passes", 1.0, "Passes through day per epoch")
configs.DEFINE_float("target_lambda", 0.5,
"How much to weight last step vs. all steps in loss")
configs.DEFINE_float("rnn_lambda", 0.5,
"How much to weight last step vs. all steps in loss")
configs.DEFINE_integer("max_epoch", 0, "Stop after max_epochs")
configs.DEFINE_integer("early_stop", None, "Early stop parameter")
configs.DEFINE_integer("seed", None, "Seed for deterministic training")
configs.DEFINE_integer("cache_id", None,
"A unique experiment key for traking a cahce")
c = configs.ConfigValues()
if c.min_unrollings is None:
c.min_unrollings = c.num_unrollings
if c.max_unrollings is None:
c.max_unrollings = c.num_unrollings
# optimizer_params is a string of the form "param1=value1,param2=value2,..."
# this maps it to dictionary { param1 : value1, param2 : value2, ...}
if c.optimizer_params is None:
c.optimizer_params = dict()
else:
args_list = [p.split('=') for p in c.optimizer_params.split(',')]
params = dict()
for p in args_list:
params[p[0]] = float(p[1])
c.optimizer_params = params
assert ('learning_rate' not in c.optimizer_params)
return c
def get_configs():
configs.DEFINE_string("name", 'trial1', "")
configs.DEFINE_string("datafile", 'Group2-Dataset.csv', "")
configs.DEFINE_string("predict_datafile", None, "")
configs.DEFINE_string("mse_outfile", None, "")
configs.DEFINE_string("scalesfile", None, "")
configs.DEFINE_string("default_gpu", '/gpu:0', "")
configs.DEFINE_string("nn_type", 'DeepRnnModel', "")
configs.DEFINE_string("active_field", 'active', "")
configs.DEFINE_string("date_field", 'date', "")
configs.DEFINE_string("key_field", 'gvkey', "")
configs.DEFINE_string("target_field", 'mkvaltq_ttm', "")
configs.DEFINE_string("scale_field", 'mrkcap', "")
configs.DEFINE_string("financial_fields", 'saleq_ttm-ltq_mrq', "")
configs.DEFINE_string("aux_fields", 'mom3m-mom9m', "")
configs.DEFINE_string("dont_scale", None, "")
configs.DEFINE_string("data_dir", 'datasets', "")
configs.DEFINE_string("model_dir", 'chkpts-wrds-rnn', "")
configs.DEFINE_string("rnn_cell", 'lstm', "")
configs.DEFINE_string("activation_fn", 'relu', "")
configs.DEFINE_integer("num_inputs", -1, "")
configs.DEFINE_integer("num_outputs", -1, "")
configs.DEFINE_integer("target_idx", None, "")
configs.DEFINE_integer("min_unrollings", 5, "")
configs.DEFINE_integer("max_unrollings", 5, "")
configs.DEFINE_integer("min_years", None, "")
configs.DEFINE_integer("max_years", None, "")
configs.DEFINE_integer("pls_years", None, "")
configs.DEFINE_integer("num_unrollings", 5, "")
configs.DEFINE_integer("stride", 12, "")
configs.DEFINE_integer("forecast_n", 3, "")
configs.DEFINE_integer("batch_size", 128, "")
configs.DEFINE_integer("num_layers", 5, "")
configs.DEFINE_integer("num_hidden", 128, "")
configs.DEFINE_float("training_noise", None, "")
configs.DEFINE_float("init_scale", 0.01, "")
configs.DEFINE_float("max_grad_norm", 10.0, "")
configs.DEFINE_integer("start_date", None, "")
configs.DEFINE_integer("end_date", None, "")
configs.DEFINE_integer("split_date", None, "")
configs.DEFINE_float("keep_prob", 0.75, "")
configs.DEFINE_boolean("train", False, "")
configs.DEFINE_boolean("require_targets", False, "")
configs.DEFINE_boolean("input_dropout", False, "")
configs.DEFINE_boolean("hidden_dropout", False, "")
configs.DEFINE_boolean("rnn_dropout", True, "")
configs.DEFINE_boolean("skip_connections", False, "")
configs.DEFINE_boolean("direct_connections", False, "")
configs.DEFINE_boolean("use_cache", True, "")
configs.DEFINE_boolean("pretty_print_preds", True, "")
configs.DEFINE_boolean("scale_targets", True, "")
configs.DEFINE_boolean("backfill", False, "")
configs.DEFINE_boolean("log_squasher", True, "")
configs.DEFINE_boolean("ts_smoother", False, "")
configs.DEFINE_string("data_scaler", 'RobustScaler', '')
configs.DEFINE_string("optimizer", 'AdadeltaOptimizer', '')
configs.DEFINE_string("optimizer_params", None, '')
configs.DEFINE_float("learning_rate", 0.6, "")
configs.DEFINE_float("lr_decay", 0.95, "")
configs.DEFINE_float("validation_size", 0.3, "")
configs.DEFINE_float("train_until", 0.0, "")
configs.DEFINE_float("passes", 0.2, "")
configs.DEFINE_float("target_lambda", 0.8, "")
configs.DEFINE_float("rnn_lambda", 0.2, "")
configs.DEFINE_float("l2_alpha", 0.0, "")
configs.DEFINE_integer("max_epoch", 1000, "")
configs.DEFINE_integer("early_stop", 10, "")
configs.DEFINE_integer("seed", 100, "")
configs.DEFINE_integer("cache_id", 100, "")
configs.DEFINE_string("output_file", "mkvaltq_2016.csv", "")
c = configs.ConfigValues()
if c.min_unrollings is None:
c.min_unrollings = c.num_unrollings
if c.max_unrollings is None:
c.max_unrollings = c.num_unrollings
if c.min_years is not None:
c.min_unrollings = c.min_years * (12 // c.stride)
if c.max_years is not None:
c.max_unrollings = (c.max_years) * (12 // c.stride)
elif c.pls_years is None:
c.max_unrollings = c.min_unrollings
else:
c.max_unrollings = (c.min_years + c.pls_years) * (12 // c.stride)
# optimizer_params is a string of the form "param1=value1,param2=value2,..."
# this maps it to dictionary { param1 : value1, param2 : value2, ...}
if c.optimizer_params is None:
c.optimizer_params = dict()
else:
args_list = [p.split('=') for p in c.optimizer_params.split(',')]
params = dict()
for p in args_list:
params[p[0]] = float(p[1])
c.optimizer_params = params
assert ('learning_rate' not in c.optimizer_params)
return c
def get_configs():
"""
Defines all configuration params passable to command line.
"""
configs.DEFINE_string("name", 'hpo-test', "A name for the config.")
configs.DEFINE_string("datafile", 'source-ml-data-v8-100M.dat',
"a datafile name.")
configs.DEFINE_string(
"predict_datafile", None,
"If predict_datafile is not None, use it instead of datafile for predictions"
)
configs.DEFINE_string("mse_outfile", None,
"A file to write mse values during predict phase.")
configs.DEFINE_string("scalesfile", None,
"Optional file for storing scaling params")
configs.DEFINE_string("default_gpu", '/gpu:0',
"The default GPU to use e.g., /gpu:0")
configs.DEFINE_string("nn_type", 'RNNPointEstimate', "Model type")
configs.DEFINE_string("active_field", 'active',
"Key column name header for active indicator")
configs.DEFINE_string("date_field", 'date', "Name of data column.")
configs.DEFINE_string("key_field", 'gvkey',
"Key column name header in datafile")
configs.DEFINE_string("target_field", 'oiadpq_ttm',
"Target column name header in datafile")
configs.DEFINE_string("scale_field", 'mrkcap',
"Feature to scale inputs by")
configs.DEFINE_string("financial_fields", 'saleq_ttm-ltq_mrq',
"Shared input and target field names")
configs.DEFINE_string("aux_fields", 'rel_mom1m-rel_mom9m',
"non-target, input only fields")
configs.DEFINE_string("dont_scale", None, "Names of fields to not scale")
configs.DEFINE_string("data_dir", 'datasets', "The data directory")
configs.DEFINE_string("model_dir", 'test-model', "Model directory")
configs.DEFINE_string("experiments_dir", './', "Experiments directory")
configs.DEFINE_list_string("rnn_cell", 'lstm', "lstm or gru")
configs.DEFINE_list_string("activation_fn", 'relu',
"MLP activation function in tf.nn.*")
configs.DEFINE_integer("num_inputs", -1, "")
configs.DEFINE_integer("num_outputs", -1, "")
configs.DEFINE_integer("target_idx", None, "")
configs.DEFINE_list_integer("min_unrollings", 5,
"Min number of unrolling steps")
configs.DEFINE_list_integer("max_unrollings", 5,
"Max number of unrolling steps")
configs.DEFINE_list_integer("min_years", None, "Alt to min_unrollings")
configs.DEFINE_list_integer("max_years", None, "Alt to max_unrollings")
configs.DEFINE_integer("pls_years", None,
"Alt to max_years. max_years = min_year+pls_years")
configs.DEFINE_list_integer("stride", 12,
"How many steps to skip per unrolling")
configs.DEFINE_list_integer("batch_size", 256, "Size of each batch")
configs.DEFINE_list_integer("num_layers", 2, "Numer of RNN layers")
configs.DEFINE_integer("forecast_n", 12,
"How many steps to forecast into the future")
configs.DEFINE_list_integer("num_hidden", 64,
"Number of hidden layer units")
configs.DEFINE_list_float("init_scale", 1.0, "Initial scale for weights")
configs.DEFINE_list_float("max_grad_norm", 50.0, "Gradient clipping")
configs.DEFINE_integer("start_date", 197501,
"First date to train on as YYYYMM")
configs.DEFINE_integer("end_date", 199812,
"Last date to train on as YYYYMM")
configs.DEFINE_integer("split_date", None, "Date to split train/test on.")
configs.DEFINE_boolean("train", True,
"Train model otherwise inference only")
configs.DEFINE_list_float("dropout", 0.0, "Dropout rate for hidden layers")
configs.DEFINE_list_float("recurrent_dropout", 0.3,
"Dropout rate for recurrent connections")
configs.DEFINE_boolean(
"log_squasher", True,
"Squash large normalized inputs with natural log function")
configs.DEFINE_list_string(
"data_scaler", 'RobustScaler',
'sklearn scaling algorithm or None if no scaling')
configs.DEFINE_list_string("optimizer", 'Adadelta',
'Any tensorflow optimizer in tf.train')
configs.DEFINE_list_float("learning_rate", 0.6,
"The initial starting learning rate")
configs.DEFINE_list_float("lr_decay", 0.96, "Learning rate decay")
configs.DEFINE_float("validation_size", 0.3,
"Size of validation set as %, ie. 0.3 = 30% of data")
configs.DEFINE_list_float(
"target_lambda", 0.5,
"How much to weight last step vs. all steps in loss")
configs.DEFINE_list_float(
"rnn_lambda", 0.7,
"How much to weight last step vs. all steps in loss")
configs.DEFINE_integer("max_epoch", 35, "Stop after max_epochs")
configs.DEFINE_integer("early_stop", 15, "Early stop parameter")
configs.DEFINE_integer("seed", 521, "Seed for deterministic training")
configs.DEFINE_boolean("UQ", False, "Uncertainty Quantification Mode")
configs.DEFINE_list_float("l2_alpha", 0.0,
"L2 regularization for weight parameters.")
configs.DEFINE_float("recurrent_l2_alpha", 0.0,
"L2 regularization for recurrent weight parameters.")
configs.DEFINE_list_boolean("huber_loss", False,
"Use huber loss instead of mse")
configs.DEFINE_list_float("huber_delta", 1.0, "delta for huber loss")
configs.DEFINE_integer("forecast_steps", 1,
"How many future predictions need to me made")
configs.DEFINE_string('forecast_steps_weights', '1.0',
'weights for the forecast steps')
configs.DEFINE_integer(
"logging_interval", 100,
"Number of batches for logging interval during training")
configs.DEFINE_boolean("write_inp_to_out_file", True,
"Write input sequence to the output files")
configs.DEFINE_string(
"training_type", 'fixed_dates',
'Choose between "fixed_dates" and "iterative" training')
configs.DEFINE_integer("member_id", 1, "Id of member in a population")
configs.DEFINE_boolean('load_saved_weights', False,
'Load weights saved in the checkpoint directory')
configs.DEFINE_integer(
"epoch_logging_interval", 1,
"Number of batches for logging interval during training")
configs.DEFINE_string('preds_fname', 'preds.dat',
'Name of the prediction file')
configs.DEFINE_integer("num_procs", 1,
"Total number of training/prediction processes")
# HPO related params
configs.DEFINE_integer("NPE", 1, "Number of Parallel Executions")
configs.DEFINE_string(
"search_algorithm", "genetic",
"Algorithm for hyper-param optimization. Select from 'genetic', 'grid_search', 'doe' "
)
configs.DEFINE_integer("generations", 5,
"Number of generations for genetic algorithm")
configs.DEFINE_integer("pop_size", 16,
"Population size for genetic algorithm")
configs.DEFINE_integer(
"num_gpu", 1, "Number of GPU on the machine, Use 0 if there are None")
configs.DEFINE_float("mutate_rate", 0.2,
"Mutation rate for genetic algorithm")
configs.DEFINE_string("objective", 'mse', "Select between mse or uq_loss")
configs.DEFINE_string("init_pop", None,
"Initial population to begin hyper param search")
configs.DEFINE_boolean("save_latest_pop", False,
"Save the latest population")
configs.DEFINE_string('doe_file', None, 'Design of experiments csv file')
configs.DEFINE_integer("decay_steps", 100000,
"Number of training steps between decay steps")
configs.DEFINE_string("initializer", 'GlorotUniform',
'variable initializers available in Keras')
configs.DEFINE_boolean(
"use_custom_init", True,
'Use RandomUniform initializer with init_scale values')
configs.DEFINE_boolean(
"aux_masking", False,
'Mask aux features of all time steps except the last one with 0')
configs.DEFINE_integer("max_norm", None, "Max Norm for kernel constraint")
configs.DEFINE_float("sgd_momentum", 0.0, "momentum for SGD optimizer")
configs.DEFINE_float("end_learning_rate", 0.01,
"end lr for polynomial decay")
configs.DEFINE_float(
'decay_power', 0.5,
'power to decay the learning rate with for polynomial decay')
configs.DEFINE_string('piecewise_lr_boundaries', None,
'boundaries for piecewise constant lr')
configs.DEFINE_string('piecewise_lr_values', None,
'values for piecewise constant lr')
configs.DEFINE_string('lr_schedule', 'ExponentialDecay',
'Learning rate scheduler')
c = configs.ConfigValues()
c.data_dir = os.path.join(_data_dir_path, c.data_dir)
c.forecast_steps_weights = [
float(x) for x in c.forecast_steps_weights.split(',')
]
return c
import numpy as np
import tensorflow as tf
from sklearn.linear_model import LogisticRegression
import model_utils
from model_utils import get_tabular_data
import configs
from tensorflow.python.platform import gfile
from batch_generator import BatchGenerator
"""
Entry point and main loop for train_net.py. Uses command line arguments to get
model and training specification (see config.py).
"""
configs.DEFINE_string("train_datafile", None, "Training file")
configs.DEFINE_float("lr_decay", 0.9, "Learning rate decay")
configs.DEFINE_float("initial_learning_rate", 1.0, "Initial learning rate")
configs.DEFINE_float("validation_size", 0.0, "Size of validation set as %")
configs.DEFINE_integer("passes", 1, "Passes through day per epoch")
configs.DEFINE_integer("max_epoch", 0, "Stop after max_epochs")
configs.DEFINE_integer("early_stop", None, "Early stop parameter")
configs.DEFINE_integer("seed", None, "Seed for deterministic training")
config = configs.get_configs()
datafile = config.train_datafile if config.train_datafile else config.datafile
train_path = model_utils.get_data_path(config.data_dir, datafile)
cache_path = os.path.splitext(train_path)[0] + '.cache'
def get_configs():
"""
Defines all configuration params passable to command line.
"""
configs.DEFINE_string("name", 'none', "A name for the config.")
configs.DEFINE_string("datafile", 'open_dataset.dat', "a datafile name.")
configs.DEFINE_string(
"predict_datafile", None,
"If predict_datafile is not None, use it instead of datafile for predictions"
)
configs.DEFINE_string("mse_outfile", None,
"A file to write mse values during predict phase.")
configs.DEFINE_string("scalesfile", None,
"Optional file for storing scaling params")
configs.DEFINE_string(
"mse_var_outfile", None,
"A file to write mse_var values during predict phase.")
configs.DEFINE_string("default_gpu", '',
"The default GPU to use e.g., /gpu:0")
configs.DEFINE_string("nn_type", 'DeepRnnModel', "Model type")
configs.DEFINE_string("active_field", 'active',
"Key column name header for active indicator")
configs.DEFINE_string("date_field", 'date', "Name of data column.")
configs.DEFINE_string("key_field", 'gvkey',
"Key column name header in datafile")
configs.DEFINE_string("target_field", 'oiadpq_ttm',
"Target column name header in datafile")
configs.DEFINE_string("scale_field", 'mrkcap',
"Feature to scale inputs by")
configs.DEFINE_string("financial_fields", '',
"Shared input and target field names")
configs.DEFINE_string("aux_fields", None, "non-target, input only fields")
configs.DEFINE_string("dont_scale", None, "Names of fields to not scale")
configs.DEFINE_string("data_dir", '', "The data directory")
configs.DEFINE_string("model_dir", 'chkpts',
"Model (checkpoint) directory")
configs.DEFINE_string("rnn_cell", 'gru', "lstm or gru")
configs.DEFINE_string("activation_fn", 'relu',
"MLP activation function in tf.nn.*")
configs.DEFINE_integer("num_inputs", -1, "")
configs.DEFINE_integer("num_outputs", -1, "")
configs.DEFINE_integer("target_idx", None, "")
configs.DEFINE_integer("min_unrollings", None,
"Min number of unrolling steps")
configs.DEFINE_integer("max_unrollings", None,
"Max number of unrolling steps")
configs.DEFINE_integer("min_years", None, "Alt to min_unrollings")
configs.DEFINE_integer("max_years", None, "Alt to max_unrollings")
configs.DEFINE_integer("pls_years", None,
"Alt to max_years. max_years = min_year+pls_years")
# num_unrollings is being depricated, replaced with max_unrollings
configs.DEFINE_integer("num_unrollings", 4, "Number of unrolling steps")
configs.DEFINE_integer("stride", 12,
"How many steps to skip per unrolling")
configs.DEFINE_integer("forecast_n", 12,
"How many steps to forecast into the future")
configs.DEFINE_integer("batch_size", 1, "Size of each batch")
configs.DEFINE_integer("num_layers", 1, "Numer of RNN layers")
configs.DEFINE_integer("num_hidden", 10, "Number of hidden layer units")
configs.DEFINE_float("training_noise", None,
"Level of training noise as multiple of 1-stdev")
configs.DEFINE_float("init_scale", 0.1, "Initial scale for weights")
configs.DEFINE_float("max_grad_norm", 10.0, "Gradient clipping")
configs.DEFINE_integer("start_date", None,
"First date to train on as YYYYMM")
configs.DEFINE_integer("end_date", None, "Last date to train on as YYYYMM")
configs.DEFINE_integer("split_date", None, "Date to split train/test on.")
configs.DEFINE_float("keep_prob", 1.0, "Keep probability for dropout")
configs.DEFINE_boolean("train", True,
"Train model otherwise inference only")
configs.DEFINE_boolean("require_targets", False,
"Require target values for test predictions")
configs.DEFINE_boolean("input_dropout", False, "Do dropout on input layer")
configs.DEFINE_boolean("hidden_dropout", False,
"Do dropout on hidden layers")
configs.DEFINE_boolean("rnn_dropout", False,
"Do dropout on recurrent connections")
configs.DEFINE_boolean(
"skip_connections", False,
"Have a linear fully connected weight skip hidden units in MLP")
configs.DEFINE_boolean(
"direct_connections", False,
"Have direct connections between input and output in MLP")
configs.DEFINE_boolean(
"use_cache", True,
"Load data for logreg from cache (vs processing from batch generator)")
configs.DEFINE_boolean(
"pretty_print_preds", False,
"Print predictions in tabular format with inputs, targets, and keys")
configs.DEFINE_boolean(
"print_preds", False,
"Print predictions with just date, gvkey and output values")
configs.DEFINE_string(
"df_dirname", None,
"Saves dataframes for target, output, variance/variance, mse and mse_var in df_dirname"
)
configs.DEFINE_boolean("scale_targets", True, "")
configs.DEFINE_boolean(
"backfill", False,
"Backfill seq history to max_unrollings with data in first time step")
configs.DEFINE_boolean(
"log_squasher", True,
"Squash large normalized inputs with natural log function")
configs.DEFINE_boolean("ts_smoother", False,
"Use smoother on data time series during training")
configs.DEFINE_string("data_scaler", None,
'sklearn scaling algorithm or None if no scaling')
configs.DEFINE_string("optimizer", 'GradientDescentOptimizer',
'Any tensorflow optimizer in tf.train')
configs.DEFINE_string("optimizer_params", None,
'Additional optimizer params such as momentum')
configs.DEFINE_float("learning_rate", 0.6,
"The initial starting learning rate")
configs.DEFINE_float("lr_decay", 0.9, "Learning rate decay")
configs.DEFINE_float("validation_size", 0.0,
"Size of validation set as %, ie. .3 = 30% of data")
configs.DEFINE_float("train_until", 0.0,
"Train until validation MSE is less than this value")
configs.DEFINE_float("passes", 1.0, "Passes through day per epoch")
configs.DEFINE_float("target_lambda", 0.5,
"How much to weight last step vs. all steps in loss")
configs.DEFINE_float("rnn_lambda", 0.5,
"How much to weight last step vs. all steps in loss")
configs.DEFINE_integer("max_epoch", 0, "Stop after max_epochs")
configs.DEFINE_integer("early_stop", None, "Early stop parameter")
configs.DEFINE_integer("seed", None, "Seed for deterministic training")
configs.DEFINE_integer("cache_id", None,
"A unique experiment key for traking a cahce")
configs.DEFINE_float("keep_prob_pred", 1.0,
"Keep Prob for dropout during prediction")
configs.DEFINE_boolean(
"print_normalized_outputs", False,
"Print normalized outputs. Doesn't apply to pretty print")
configs.DEFINE_boolean("UQ", False, "Uncertainty Quantification Mode")
configs.DEFINE_string("UQ_model_type", 'MVE', "Select between MVE or PIE")
configs.DEFINE_float(
"noise_lambda", 1.0,
"Weight decay for noise in the loss function. Refer to DeepBayesUQ Model"
)
configs.DEFINE_float("l2_alpha", 0.0,
"L2 regularization for weight parameters.")
configs.DEFINE_float("picp_lambda", 1.0,
"Contribution of PICP loss term for HQPI UQ model")
configs.DEFINE_float(
"smoothing_pi_check", 100,
"Smoothing parameter for calculation of PI check in HQPI UQ model")
configs.DEFINE_float(
"confidence_alpha", 0.1,
"Alpha used for calculating confidence level (= 1 - alpha)")
configs.DEFINE_boolean("huber_loss", False,
"Use huber loss instead of mse")
configs.DEFINE_float("huber_delta", 1.0, "delta for huber loss")
c = configs.ConfigValues()
if c.min_unrollings is None:
c.min_unrollings = c.num_unrollings
if c.max_unrollings is None:
c.max_unrollings = c.num_unrollings
if c.min_years is not None:
c.min_unrollings = c.min_years * (12 // c.stride)
if c.max_years is not None:
c.max_unrollings = (c.max_years) * (12 // c.stride)
elif c.pls_years is None:
c.max_unrollings = c.min_unrollings
else:
c.max_unrollings = (c.min_years + c.pls_years) * (12 // c.stride)
# optimizer_params is a string of the form "param1=value1,param2=value2,..."
# this maps it to dictionary { param1 : value1, param2 : value2, ...}
if c.optimizer_params is None:
c.optimizer_params = dict()
else:
args_list = [p.split('=') for p in c.optimizer_params.split(',')]
params = dict()
for p in args_list:
params[p[0]] = float(p[1])
c.optimizer_params = params
assert ('learning_rate' not in c.optimizer_params)
return c
def main(_):
"""
The model specified command line arg --model_dir is applied to every data
point in --test_datafile and the model output is sent to --output. The unix
command 'paste' can be used to stich the input file and output together.
e.g.,
$ classifiy_data.py --config=train.conf --test_datafile=test.dat --output=output.dat
$ paste -d ' ' test.dat output.dat > input_and_output.dat
"""
configs.DEFINE_string('test_datafile', None, 'file with test data')
configs.DEFINE_string('output', 'preds.dat', 'file for predictions')
configs.DEFINE_string('time_field', 'date', 'fields used for dates/time')
configs.DEFINE_string('print_start', '190001',
'only print data on or after')
configs.DEFINE_string('print_end', '210012',
'only print data on or before')
configs.DEFINE_integer('min_test_k', 1, 'minimum seq length classified')
configs.DEFINE_integer('num_batches', None, 'num_batches overrride')
config = configs.get_configs()
if config.test_datafile is None:
config.test_datafile = config.datafile
batch_size = 1
data_path = model_utils.get_data_path(config.data_dir,
config.test_datafile)
print("Loading data %s" % data_path)
dataset = BatchGenerator(data_path,
config,
batch_size=batch_size,
num_unrollings=config.num_unrollings)
num_data_points = dataset.num_batches
if config.num_batches is not None:
num_data_points = config.num_batches
print("num_batches = ", num_data_points)
tf_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
with tf.Graph().as_default(), tf.Session(config=tf_config) as session:
print("Loading model.")
model = model_utils.get_trained_model(session, config)
stats = dict()
key = 'ALL'
stats[key] = list()
with open(config.output, "w") as outfile:
for i in range(num_data_points):
batch = dataset.next_batch()
preds = model.step(session, batch)
seq_len = get_seq_length(batch)
start = seq_len - 1
if seq_len < config.num_unrollings:
continue
#if config.nn_type != 'rnn' and seq_len < config.num_unrollings:
# continue
#elif config.nn_type == 'rnn' and classify_entire_seq(batch):
# start = config.min_test_k - 1
for i in range(start, seq_len):
key, date = get_key_and_date(batch, i)
if (date < config.print_start or date > config.print_end):
continue
prob = get_pos_prob(config, preds, i)
target = get_target(batch, i)
outfile.write("%s %s "
"%.4f %.4f %d %d\n" %
(key, date, 1.0 - prob, prob, target, i + 1))
pred = +1.0 if prob >= 0.5 else 0.0
error = 0.0 if (pred == target) else 1.0
tpos = 1.0 if (pred == 1 and target == 1) else 0.0
tneg = 1.0 if (pred == 0 and target == 0) else 0.0
fpos = 1.0 if (pred == 1 and target == 0) else 0.0
fneg = 1.0 if (pred == 0 and target == 1) else 0.0
# print("pred=%.2f target=%.2f tp=%d tn=%d fp=%d fn=%d"%(pred,target,tp,tn,fp,fn))
curstat = {
'error': error,
'tpos': tpos,
'tneg': tneg,
'fpos': fpos,
'fneg': fneg
}
if date not in stats:
stats[date] = list()
stats[date].append(curstat)
stats['ALL'].append(curstat)
print_summary_stats(stats)
def main(_):
"""
"""
configs.DEFINE_string('test_datafile', None, 'file with test data')
configs.DEFINE_string('output', 'preds.dat', 'file for predictions')
configs.DEFINE_string('time_field', 'date', 'fields used for dates/time')
configs.DEFINE_string('print_start', '190001',
'only print data on or after')
configs.DEFINE_string('print_end', '210012',
'only print data on or before')
configs.DEFINE_string('factor_name', None,
'Name of factor if nn_type=factor')
configs.DEFINE_integer('min_test_k', 1, 'minimum seq length classified')
configs.DEFINE_integer('num_batches', None, 'num_batches overrride')
config = configs.get_configs()
factor_name = config.factor_name
assert (factor_name is not None)
if config.test_datafile is None:
config.test_datafile = config.datafile
batch_size = 1
num_unrollings = config.num_unrollings
data_path = model_utils.get_data_path(config.data_dir,
config.test_datafile)
filename = data_path
print("Loading data %s" % data_path)
if not os.path.isfile(filename):
raise RuntimeError("The data file %s does not exists" % filename)
data = pd.read_csv(filename,
sep=' ',
dtype={
config.key_field: str,
'date': str
})
if config.end_date is not None:
data = data.drop(data[data['date'] > str(config.end_date)].index)
num_data_points = len(data)
params = dict()
print("num data points = ", num_data_points)
stats = dict()
key = 'ALL'
stats[key] = list()
with open(config.output, "w") as outfile:
last_key = ''
seq_len = 0
for i in range(num_data_points):
key = get_value(data, config.key_field, i)
date = get_value(data, 'date', i)
seq_len = seq_len + 1 if key == last_key else 1
last_key = key
if (str(date) < config.print_start
or str(date) > config.print_end):
continue
if seq_len < config.min_test_k:
continue
prob = get_value(data, factor_name, i)
out = get_value(data, config.target_field, i)
target = (out + 1.0) / 2.0
k = min(seq_len, config.num_unrollings)
outfile.write("%s %s "
"%.4f %.4f %d %d\n" %
(key, date, 1.0 - prob, prob, target, k))
pred = +1.0 if prob >= 0.5 else 0.0
error = 0.0 if (pred == target) else 1.0
tpos = 1.0 if (pred == 1 and target == 1) else 0.0
tneg = 1.0 if (pred == 0 and target == 0) else 0.0
fpos = 1.0 if (pred == 1 and target == 0) else 0.0
fneg = 1.0 if (pred == 0 and target == 1) else 0.0
# print("pred=%.2f target=%.2f tp=%d tn=%d fp=%d fn=%d"%(pred,target,tp,tn,fp,fn))
rec = {
'error': error,
'tpos': tpos,
'tneg': tneg,
'fpos': fpos,
'fneg': fneg
}
if date not in stats:
stats[date] = list()
stats[date].append(rec)
stats['ALL'].append(rec)
print_summary_stats(stats)
def get_configs():
"""
Defines all configuration params passable to command line.
"""
configs.DEFINE_string("name", 'test', "A name for the config.")
configs.DEFINE_string("datafile", None, "a datafile name.")
configs.DEFINE_string("scalesfile", None,
"Optional file for storing scaling params")
configs.DEFINE_string("default_gpu", '/gpu:0',
"The default GPU to use e.g., /gpu:0")
configs.DEFINE_string("nn_type", 'RNNPointEstimate', "Model type")
configs.DEFINE_string("active_field", 'active',
"Key column name header for active indicator")
configs.DEFINE_string("date_field", 'date', "Name of data column.")
configs.DEFINE_string("key_field", 'gvkey',
"Key column name header in datafile")
configs.DEFINE_string("target_field", 'oiadpq_ttm',
"Target column name header in datafile")
configs.DEFINE_string("scale_field", 'mrkcap',
"Feature to scale inputs by")
configs.DEFINE_string("financial_fields", 'saleq_ttm-ltq_mrq',
"Shared input and target field names")
configs.DEFINE_string("aux_fields", 'rel_mom1m-rel_mom9m',
"non-target, input only fields")
configs.DEFINE_string("dont_scale_fields", None,
"Names of fields to not scale")
configs.DEFINE_string("data_dir", 'datasets', "The data directory")
configs.DEFINE_string("model_dir", 'test-model', "Model directory")
configs.DEFINE_string("experiments_dir", './', "Experiments directory")
configs.DEFINE_string("rnn_cell", 'lstm', "lstm or gru")
configs.DEFINE_string("activation_fn", 'relu',
"MLP activation function in tf.nn.*")
configs.DEFINE_integer("num_inputs", -1, "")
configs.DEFINE_integer("num_outputs", -1, "")
configs.DEFINE_integer("target_idx", None, "")
configs.DEFINE_integer("min_unrollings", 5,
"Min number of unrolling steps")
configs.DEFINE_integer("max_unrollings", 5,
"Max number of unrolling steps")
configs.DEFINE_integer("min_years", None, "Alt to min_unrollings")
configs.DEFINE_integer("max_years", None, "Alt to max_unrollings")
configs.DEFINE_integer("pls_years", None,
"Alt to max_years. max_years = min_year+pls_years")
configs.DEFINE_integer("stride", 12,
"How many steps to skip per unrolling")
configs.DEFINE_integer("batch_size", 256, "Size of each batch")
configs.DEFINE_integer("num_layers", 2, "Numer of RNN layers")
configs.DEFINE_integer("forecast_n", 12,
"How many steps to forecast into the future")
configs.DEFINE_integer("num_hidden", 64, "Number of hidden layer units")
configs.DEFINE_float("init_scale", 1.0, "Initial scale for weights")
configs.DEFINE_float("max_grad_norm", 50.0, "Gradient clipping")
configs.DEFINE_integer("start_date", 197501,
"First date to train on as YYYYMM")
configs.DEFINE_integer("end_date", 199912,
"Last date to train on as YYYYMM")
configs.DEFINE_integer("split_date", None, "Date to split train/test on.")
configs.DEFINE_boolean("train", True,
"Train model otherwise inference only")
configs.DEFINE_float("dropout", 0.0, "Dropout rate for hidden layers")
configs.DEFINE_float("recurrent_dropout", 0.0,
"Dropout rate for recurrent connections")
configs.DEFINE_boolean(
"log_squasher", True,
"Squash large normalized inputs with natural log function")
configs.DEFINE_string("data_scaler", 'RobustScaler',
'sklearn scaling algorithm or None if no scaling')
configs.DEFINE_string("optimizer", 'Adadelta',
'Any tensorflow optimizer in tf.train')
configs.DEFINE_float("learning_rate", 0.6,
"The initial starting learning rate")
configs.DEFINE_float("lr_decay", 1.0,
"Learning rate decay for exponential decay")
configs.DEFINE_float("validation_size", 0.3,
"Size of validation set as %, ie. 0.3 = 30% of data")
configs.DEFINE_float("target_lambda", 0.5,
"How much to weight last step vs. all steps in loss")
configs.DEFINE_float("rnn_lambda", 0.7,
"How much to weight last step vs. all steps in loss")
configs.DEFINE_integer("max_epoch", 1, "Stop after max_epochs")
configs.DEFINE_integer("early_stop", 1, "Early stop parameter")
configs.DEFINE_integer("seed", 521, "Seed for deterministic training")
configs.DEFINE_boolean("UQ", False, "Uncertainty Quantification Mode")
configs.DEFINE_float("l2_alpha", 0.0,
"L2 regularization for weight parameters.")
configs.DEFINE_float("recurrent_l2_alpha", 0.0,
"L2 regularization for recurrent weight parameters.")
configs.DEFINE_boolean("huber_loss", False,
"Use huber loss instead of mse")
configs.DEFINE_float("huber_delta", 1.0, "delta for huber loss")
configs.DEFINE_integer("forecast_steps", 1,
"How many future predictions need to me made")
configs.DEFINE_string('forecast_steps_weights', '1.0',
'weights for the forecast steps')
configs.DEFINE_integer(
"logging_interval", 100,
"Number of batches for logging interval during training")
configs.DEFINE_boolean("write_inp_to_out_file", True,
"Write input sequence to the output files")
configs.DEFINE_string(
"training_type", 'fixed_dates',
'Choose between "fixed_dates" and "iterative" training')
configs.DEFINE_integer("NPE", 1, "Number of Parallel Executions")
configs.DEFINE_integer("num_procs", 1,
"Total number of training/prediction processes")
configs.DEFINE_integer("num_gpu", 1, "NUmber of GPUs")
configs.DEFINE_boolean('load_saved_weights', False,
'Load weights saved in the checkpoint directory')
configs.DEFINE_integer(
"epoch_logging_interval", 1,
"Number of batches for logging interval during training")
configs.DEFINE_integer("decay_steps", 1500,
"Number of training steps between decay steps")
configs.DEFINE_string("initializer", 'GlorotUniform',
'variable initializers available in Keras')
configs.DEFINE_boolean(
"use_custom_init", True,
'Use RandomUniform initializer with init_scale values')
configs.DEFINE_boolean(
"aux_masking", False,
'Mask aux features of all time steps except the last one with 0')
configs.DEFINE_integer("max_norm", 3, "Max Norm for kernel constraint")
configs.DEFINE_float("sgd_momentum", 0.0, "momentum for SGD optimizer")
configs.DEFINE_float("end_learning_rate", 0.01,
"end lr for polynomial decay")
configs.DEFINE_float(
'decay_power', 0.5,
'power to decay the learning rate with for polynomial decay')
configs.DEFINE_string('piecewise_lr_boundaries', '4000-5500-5500',
'boundaries for piecewise constant lr')
configs.DEFINE_string('piecewise_lr_values', '0.5-0.1-0.05-0.1',
'values for piecewise constant lr')
configs.DEFINE_string('lr_schedule', 'ExponentialDecay',
'Learning rate scheduler')
configs.DEFINE_string('preds_fname', 'preds.dat',
'Name of the prediction file')
configs.DEFINE_integer("member_id", 1, "Id of member in a population")
configs.DEFINE_boolean("cdrs_inference", False,
'If the execution is for inference on CDRS data')
configs.DEFINE_string('cdrs_src_fname', 'cdrs-src.dat',
'Filename of the CDRS source file')
configs.DEFINE_string('cdrs_ml_fname', 'cdrs-ml-data.dat',
'Filename of the CDRS ML data file')
configs.DEFINE_string('model_ranking_fname', './model-ranking.dat',
'Model Ranking File Name')
configs.DEFINE_string('model_ranking_factor', 'pred_var_entval',
'Model ranking factor')
configs.DEFINE_string("cdrs_inference_date", None,
"CDRS Inference date. Format: '%Y-%m-%d' ")
c = configs.ConfigValues()
if c.min_unrollings is None:
c.min_unrollings = c.num_unrollings
if c.max_unrollings is None:
c.max_unrollings = c.num_unrollings
if c.min_years is not None:
c.min_unrollings = c.min_years * (12 // c.stride)
if c.max_years is not None:
c.max_unrollings = (c.max_years) * (12 // c.stride)
elif c.pls_years is None:
c.max_unrollings = c.min_unrollings
else:
c.max_unrollings = (c.min_years + c.pls_years) * (12 // c.stride)
c.forecast_steps_weights = [
float(x) for x in c.forecast_steps_weights.split('-')
]
c.piecewise_lr_boundaries = [
float(x) for x in c.piecewise_lr_boundaries.split('-')
]
c.piecewise_lr_values = [
float(x) for x in c.piecewise_lr_values.split('-')
]
return c
def main(_):
"""
Entry point and main loop for train_net.py. Uses command line arguments to get
model and training specification (see config.py).
"""
configs.DEFINE_string("train_datafile", None, "Training file")
configs.DEFINE_string("optimizer", 'gd',
'Optimizer to use gd, adam, adagrad, momentum')
configs.DEFINE_float("lr_decay", 0.9, "Learning rate decay")
configs.DEFINE_float("initial_learning_rate", 1.0, "Initial learning rate")
configs.DEFINE_float("validation_size", 0.0, "Size of validation set as %")
configs.DEFINE_float("passes", 1.0, "Passes through day per epoch")
configs.DEFINE_float("rnn_loss_weight", None,
"How much moret to weight kth example")
configs.DEFINE_integer("max_epoch", 0, "Stop after max_epochs")
configs.DEFINE_integer("early_stop", None, "Early stop parameter")
configs.DEFINE_integer("seed", None, "Seed for deterministic training")
config = configs.get_configs()
if config.train_datafile is None:
config.train_datafile = config.datafile
train_path = model_utils.get_data_path(config.data_dir,
config.train_datafile)
print("Loading training data ...")
train_data = BatchGenerator(train_path,
config,
config.batch_size,
config.num_unrollings,
validation_size=config.validation_size,
randomly_sample=True)
tf_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
with tf.Graph().as_default(), tf.Session(config=tf_config) as session:
if config.seed is not None:
tf.set_random_seed(config.seed)
print("Constructing model ...")
model = model_utils.get_training_model(session, config, verbose=True)
if config.early_stop is not None:
print("Training will early stop without "
"improvement after %d epochs." % config.early_stop)
train_history = list()
valid_history = list()
# This sets the initial learning rate tensor
lr = model.assign_lr(session, config.initial_learning_rate)
for i in range(config.max_epoch):
trc, tre, vdc, vde = run_epoch(session,
model,
train_data,
keep_prob=config.keep_prob,
passes=config.passes,
verbose=True)
trc = 999.0 if trc > 999.0 else trc
vdc = 999.0 if vdc > 999.0 else vdc
print(('Epoch: %d loss: %.6f %.6f'
' error: %.6f %.6f Learning rate: %.4f') %
(i + 1, trc, vdc, tre, vde, lr))
sys.stdout.flush()
train_history.append(trc)
valid_history.append(vdc)
# update learning rate
if config.optimizer == 'gd' or config.optimizer == 'momentum':
lr = model_utils.adjust_learning_rate(session, model, lr,
config.lr_decay,
train_history)
if not os.path.exists(config.model_dir):
print("Creating directory %s" % config.model_dir)
os.mkdir(config.model_dir)
chkpt_file_prefix = "training.ckpt"
if model_utils.stop_training(config, valid_history,
chkpt_file_prefix):
print("Training stopped.")
quit()
else:
checkpoint_path = os.path.join(config.model_dir,
chkpt_file_prefix)
tf.train.Saver().save(session, checkpoint_path, global_step=i)
