def _test_8x_v100_half_precision(self, num_epochs, min_accuracy,
max_accuracy):
"""Utility to benchmark ImageNet on 8xV100 GPUs. Use in your test func."""
model_dir = self.get_tmp_model_dir()
FLAGS.batch_size = 2048
FLAGS.half_precision = True
FLAGS.num_epochs = num_epochs
FLAGS.model_dir = model_dir
start_time = time.time()
train.main([])
benchmark_time = time.time() - start_time
summaries = self.read_summaries(model_dir)
# Summaries contain all the information necessary for the regression
# metrics.
wall_time, _, eval_accuracy = zip(*summaries['eval_accuracy'])
wall_time = np.array(wall_time)
sec_per_epoch = np.mean(wall_time[1:] - wall_time[:-1])
end_accuracy = eval_accuracy[-1]
# Assertions are deferred until the test finishes, so the metrics are
# always reported and benchmark success is determined based on *all*
# assertions.
self.assertBetween(end_accuracy, min_accuracy, max_accuracy)
# Use the reporting API to report single or multiple metrics/extras.
self.report_wall_time(benchmark_time)
self.report_metrics({
'sec_per_epoch': sec_per_epoch,
'accuracy': end_accuracy
})
def __call__(self, seed):
# algs, lrs, n_epochs = ['sgd'], [1e-1], [ 5000]
algs, lrs, n_epochs = ['adam'], [1e-3], [5000]
for a, l, e in zip(algs, lrs, n_epochs):
params = deepcopy(default_D2_pars)
params['net_params']['saturations'] = [0, 1e8]
params['net_params']['save_folder'] = 'out/D2/relu_avg/'
params['train_params'].update({
'optimizer_name': a,
'lr': l,
'n_epochs': e,
'stop_loss': 1e-8,
'loss_name': 'avg_d2'
})
params['test_suite']['fit_internal_representation'][
'batch_size'] = 512
for test_name in params['test_suite'].keys():
params['test_suite'][test_name]['period'] = e // 4
os.makedirs(params['net_params']['save_folder'], exist_ok=True)
with open(params['net_params']['save_folder'] + 'full_params.json',
'w+') as f:
json.dump(params, f, indent=4)
main(params, seed)
gc.collect()
def __call__(self, seed):
algs, lrs, n_epochs = ['adam'], [5e-4], [5000]
for a, l, e in zip(algs, lrs, n_epochs):
for n in [1024]:
params = deepcopy(default_D3_pars)
params['sampler_params']['decays'] = [0.8, 0.77, 0.75]
params['net_params']['saturations'] = [0, 1e8]
params['sampler_params']['batch_size'] = 4096
params['net_params']['n'] = n
params['net_params'][
'save_folder'] = 'out/D3_relu_avg/n_{}/'.format(n)
params['train_params'].update({
'optimizer_name': a,
'lr': l,
'n_epochs': e,
'loss_name': 'avg_generic'
})
params['test_suite'] = { # By default, tests only run at final step
'weight_analysis': {'period': 1000},
'sanity_check': {'T': 200, 'period': 1000},
}
os.makedirs(params['net_params']['save_folder'], exist_ok=True)
with open(
params['net_params']['save_folder'] +
'full_params.json', 'w+') as f:
json.dump(params, f, indent=4)
main(params, seed)
gc.collect()
def __call__(self, seed):
algs, n_epochs, lr_list = ['adam'], [10000], [5e-5]
train_d = True
for a, e, l in zip(algs, n_epochs, lr_list):
for train_bias in [True, False]:
for sig_slope in [50.]: #1.]:
for sig_thresh in [
.1,
]: #, .1]:
params = deepcopy(default_D3_pars)
params['sampler_params']['decays'] = [0.8, 0.75, .75]
params['sampler_params']['batch_size'] = 1024
params['sampler_params']['epoch_length'] = 10
params['train_params'].update({
'optimizer_name': a,
'lr': l,
'n_epochs': e
})
params['train_params'].update({
'train_d_only': train_d,
'stop_loss': 1e-8,
'loss_name': 'batch'
})
params['net_params']['activation_type'] = 'Sigmoid'
params['net_params']['sigmoid_threshold'] = sig_thresh
params['net_params']['sigmoid_slope'] = sig_slope
params['net_params']['sigmoid_random_bias'] = False
params['net_params']['sigmoid_train_bias'] = train_bias
n = 1000
params['net_params']['n'] = n
params['net_params'][
'save_folder'] = 'out/D3_sigmoid_batch/n_{}_slope_{}_thresh_{}_train_bias_{}/'.format(
n, sig_slope, sig_thresh, train_bias)
# if params['net_params']['sigmoid_random_bias']:
# params['net_params']['save_folder'] = params['net_params']['save_folder'] + 'randomly_biased/'
# if train_bias:
# params['net_params']['save_folder'] = params['net_params']['save_folder'] +
# print(d_scale, e_scale)
params['test_suite'] = { # By default, tests only run at final step
'weight_analysis': {'period': 2**20},
'sanity_check': {'T': 200, 'period': 2**20},
'fit_internal_representation': {'T': 200, 'period': 2**20},
}
for test_name in params['test_suite'].keys():
params['test_suite'][test_name]['period'] = 1000
os.makedirs(params['net_params']['save_folder'],
exist_ok=True)
with open(
params['net_params']['save_folder'] +
'full_params.json', 'w+') as f:
json.dump(params, f, indent=4)
main(params, seed)
gc.collect()
def test_keras(self):
"""Test training_with_keras."""
output_dir = tempfile.mkdtemp()
output_model = os.path.join(output_dir, "model.h5")
body_pp_dpkl = os.path.join(output_dir, "body_pp.dpkl")
title_pp_dpkl = os.path.join(output_dir, "body_pp.dpkl")
title_vecs = os.path.join(output_dir, "title.npy")
body_vecs = os.path.join(output_dir, "body.npy")
this_dir = os.path.dirname(__file__)
args = [
"--sample_size=100",
"--num_epochs=1",
"--input_data=" +
os.path.join(this_dir, "test_data", "github_issues_sample.csv"),
"--output_model=" + output_model,
"--output_body_preprocessor_dpkl=" + body_pp_dpkl,
"--output_title_preprocessor_dpkl=" + title_pp_dpkl,
"--output_train_title_vecs_npy=" + title_vecs,
"--output_train_body_vecs_npy=" + body_vecs,
]
train.main(args)
output_files = [
output_model, body_pp_dpkl, title_pp_dpkl, title_vecs, body_vecs
]
for f in output_files:
self.assertTrue(os.path.exists(f))
def train_language_model(data_dir, arch, extra_flags=None):
train_parser = options.get_training_parser()
train_args = options.parse_args_and_arch(
train_parser,
[
'--task',
'language_modeling',
data_dir,
'--arch',
arch,
'--optimizer',
'adam',
'--lr',
'0.0001',
'--criterion',
'adaptive_loss',
'--adaptive-softmax-cutoff',
'5,10,15',
'--max-tokens',
'500',
'--tokens-per-sample',
'500',
'--save-dir',
data_dir,
'--max-epoch',
'1',
'--no-progress-bar',
'--distributed-world-size',
'1',
'--ddp-backend',
'no_c10d',
] + (extra_flags or []),
)
train.main(train_args)
def test_8x_v100_half_precision(self):
"""Run ImageNet on 8x V100 GPUs in half precision for 2 epochs."""
model_dir = self.get_tmp_model_dir()
FLAGS.batch_size = 2048
FLAGS.half_precision = True
FLAGS.num_epochs = 2
FLAGS.model_dir = model_dir
start_time = time.time()
train.main([])
benchmark_time = time.time() - start_time
summaries = self.read_summaries(model_dir)
# Summaries contain all the information necessary for the regression
# metrics.
wall_time, _, eval_accuracy = zip(*summaries['eval_accuracy'])
wall_time = np.array(wall_time)
sec_per_epoch = np.mean(wall_time[1:] - wall_time[:-1])
end_accuracy = eval_accuracy[-1]
# Assertions are deferred until the test finishes, so the metrics are
# always reported and benchmark success is determined based on *all*
# assertions.
self.assertBetween(end_accuracy, 0.06, 0.09)
# Use the reporting API to report single or multiple metrics/extras.
self.report_wall_time(benchmark_time)
self.report_metrics({'sec_per_epoch': sec_per_epoch,
'accuracy': end_accuracy})
self.report_extras({
'description': 'Toy 8 x V100 test for ImageNet ResNet50.',
'model_name': 'resnet50',
'parameters': 'hp=true,bs=2048',
})
def __call__(self, seed):
algs, lrs, n_epochs = ['sgd'], [1e0], [
10000,
]
for a, l, e in zip(algs, lrs, n_epochs):
for T in [4]:
params = deepcopy(default_D2_pars)
params['sampler_params']['epoch_length'] = T
params['net_type'] = 'DaleNet'
params['net_params']['saturations'] = [0, 1e8]
params['net_params'][
'save_folder'] = 'out/D2/dale/{}/T_{}/'.format(a, T)
params['net_params']['inhib_proportion'] = .25
params['net_params']['l2_penalty'] = 0.
params['train_params'].update({
'optimizer_name': a,
'lr': l,
'n_epochs': e
})
for test_name in params['test_suite'].keys():
params['test_suite'][test_name]['period'] = e // 4
os.makedirs(params['net_params']['save_folder'], exist_ok=True)
with open(
params['net_params']['save_folder'] +
'full_params.json', 'w+') as f:
json.dump(params, f, indent=4)
main(params, seed)
gc.collect()
def main():
ctr = 1
while ctr:
inp = input(
f'Select: \n\t1 Configure \n\t2 Train \n\t3 Interact \n\n > ')
cls()
if inp == '1':
list_config()
change_config()
elif inp == '2':
import train
train.main()
elif inp == '3':
import interact
interact.main()
elif inp == '0':
cls()
break
cls()
ctr += 1
def test_cpu(self):
"""Run full training for MNIST CPU training."""
model_dir = self.get_tmp_model_dir()
FLAGS.model_dir = model_dir
start_time = time.time()
train.main([])
benchmark_time = time.time() - start_time
summaries = self.read_summaries(model_dir)
# Summaries contain all the information necessary for the regression
# metrics.
wall_time, _, eval_accuracy = zip(*summaries['eval_accuracy'])
wall_time = np.array(wall_time)
sec_per_epoch = np.mean(wall_time[1:] - wall_time[:-1])
end_eval_accuracy = eval_accuracy[-1]
# Assertions are deferred until the test finishes, so the metrics are
# always reported and benchmark success is determined based on *all*
# assertions.
self.assertBetween(end_eval_accuracy, 0.98, 1.0)
# Use the reporting API to report single or multiple metrics/extras.
self.report_wall_time(benchmark_time)
self.report_metrics({
'sec_per_epoch': sec_per_epoch,
'accuracy': end_eval_accuracy,
})
self.report_extras({
'model_name': 'MNIST',
'description': 'CPU test for MNIST.'
})
def main():
if parser.extract:
extract.main()
if parser.train:
train.main()
def train_translation_model(data_dir, arch, extra_flags=None, task='translation'):
train_parser = options.get_training_parser()
train_args = options.parse_args_and_arch(
train_parser,
[
'--task', task,
data_dir,
'--save-dir', data_dir,
'--arch', arch,
'--lr', '0.05',
'--max-tokens', '500',
'--max-epoch', '1',
'--no-progress-bar',
'--distributed-world-size', '1',
'--source-lang', 'in',
'--target-lang', 'out',
] + (extra_flags or []),
)
train.main(train_args)
# test validation
validate_parser = options.get_validation_parser()
validate_args = options.parse_args_and_arch(
validate_parser,
[
'--task', task,
data_dir,
'--path', os.path.join(data_dir, 'checkpoint_last.pt'),
'--valid-subset', 'valid',
'--max-tokens', '500',
'--no-progress-bar',
]
)
validate.main(validate_args)
def test_main(self, mock_gan_train, mock_define_train_ops, mock_cyclegan_loss,
mock_define_model, mock_data_provider, mock_gfile):
FLAGS.image_set_x_file_pattern = '/tmp/x/*.jpg'
FLAGS.image_set_y_file_pattern = '/tmp/y/*.jpg'
FLAGS.batch_size = 3
FLAGS.patch_size = 8
FLAGS.generator_lr = 0.02
FLAGS.discriminator_lr = 0.3
FLAGS.train_log_dir = '/tmp/foo'
FLAGS.master = 'master'
FLAGS.task = 0
FLAGS.cycle_consistency_loss_weight = 2.0
FLAGS.max_number_of_steps = 1
mock_data_provider.provide_custom_datasets.return_value = (tf.zeros(
[1, 2], dtype=tf.float32), tf.zeros([1, 2], dtype=tf.float32))
train.main(None)
mock_data_provider.provide_custom_datasets.assert_called_once_with(
['/tmp/x/*.jpg', '/tmp/y/*.jpg'], batch_size=3, patch_size=8)
mock_define_model.assert_called_once_with(mock.ANY, mock.ANY)
mock_cyclegan_loss.assert_called_once_with(
mock_define_model.return_value,
cycle_consistency_loss_weight=2.0,
tensor_pool_fn=mock.ANY)
mock_define_train_ops.assert_called_once_with(
mock_define_model.return_value, mock_cyclegan_loss.return_value)
mock_gan_train.assert_called_once_with(
mock_define_train_ops.return_value,
'/tmp/foo',
get_hooks_fn=mock.ANY,
hooks=mock.ANY,
master='master',
is_chief=True)
def __call__(self, seed):
# algs, lrs, n_epochs = ['adam', 'sgd'], [3e-4, 5e-1], [1000, 5000]
# algs, lrs, n_epochs = ['adam'], [1e-3], [1000, ]
algs, lrs, n_epochs = ['sgd'], [1e0], [
10000,
]
T = 4
for a, l, e in zip(algs, lrs, n_epochs):
for inhib_frac in [.5, .4, .25, .1]:
params = deepcopy(default_D3_pars)
params['sampler_params']['epoch_length'] = T
params['net_type'] = 'DaleNet'
params['net_params']['saturations'] = [0, 1e8]
params['net_params'][
'save_folder'] = 'out/D3/dale/inhib_frac_{}/'.format(
inhib_frac)
params['train_params'].update({
'optimizer_name': a,
'lr': l,
'n_epochs': e
})
params['net_params']['inhib_proportion'] = inhib_frac
params['net_params']['l2_penalty'] = 0
for test_name in params['test_suite'].keys():
params['test_suite'][test_name]['period'] = e // 4
os.makedirs(params['net_params']['save_folder'], exist_ok=True)
with open(
params['net_params']['save_folder'] +
'full_params.json', 'w+') as f:
json.dump(params, f, indent=4)
main(params, seed)
gc.collect()
def main():
print("Ara Voice trainer 1.0.0 in development")
print("checking system")
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
import os.path
import sys
import numpy as np
from six.moves import xrange
import tensorflow as tf
import train
import freeze
import input_data
import models
from tensorflow.python.platform import gfile
print("Stating...")
train.main()
print("Exporting to .tflite")
freeze.main()
print("done")
def run_experiment(run_config):
for rt in replay_types:
for bs in buffer_size:
for i in range(num_of_runs):
train_config = create_config(run_config, bs, rt)
print(f'Starting trainin on: {train_config.name}')
train.main(train_config)
def main():
csv_list = cvt_midi_2_csv('midi')
with open('data/input.txt', 'w+', encoding='utf-8'):
pass
for csv in csv_list:
cvt_csv_2_chrmid(csv)
train.main()
def __call__(self, seed):
algs, lrs, n_epochs = ['adam', 'sgd'], [3e-4, 5e-1], [1000, 5000]
T = 4
for a, l, e in zip(algs, lrs, n_epochs):
for ed_init in [
'random', 'support_same', 'support_disjoint',
'support_random_e', 'support_same_with_overlap'
]:
params = deepcopy(default_D2_pars)
params['sampler_params']['epoch_length'] = T
params['net_params']['saturations'] = [0, 1e8]
params['net_params']['init_vectors_type'] = ed_init
params['net_params'][
'save_folder'] = 'out/D2/relu_support/{}/{}/'.format(
ed_init, a)
params['train_params'].update({
'optimizer_name': a,
'lr': l,
'n_epochs': e
})
for test_name in params['test_suite'].keys():
params['test_suite'][test_name]['period'] = e // 4
os.makedirs(params['net_params']['save_folder'], exist_ok=True)
with open(
params['net_params']['save_folder'] +
'full_params.json', 'w+') as f:
json.dump(params, f, indent=4)
main(params, seed)
gc.collect()
def test_1x_v100(self):
"""Run Wide ResNet CIFAR10 on 1x V100 GPUs for 2 epochs."""
model_dir = tempfile.mkdtemp()
FLAGS.num_epochs = 2
FLAGS.arch = 'wrn26_10'
FLAGS.model_dir = model_dir
start_time = time.time()
train.main([])
benchmark_time = time.time() - start_time
summaries = self.read_summaries(model_dir)
# Summaries contain all the information necessary for the regression
# metrics.
wall_time, _, eval_error_rate = zip(*summaries['eval_error_rate'])
wall_time = np.array(wall_time)
sec_per_epoch = np.mean(wall_time[1:] - wall_time[:-1])
end_error_rate = eval_error_rate[-1]
# Assertions are deferred until the test finishes, so the metrics are
# always reported and benchmark success is determined based on *all*
# assertions.
self.assertBetween(sec_per_epoch, 80., 84.)
self.assertBetween(end_error_rate, 0.30, 0.36)
# Use the reporting API to report single or multiple metrics/extras.
self.report_wall_time(benchmark_time)
self.report_metrics({
'sec_per_epoch': sec_per_epoch,
'error_rate': end_error_rate
})
self.report_extra('description',
'Toy 1 x V100 test for CIFAR10 WideResNet26_10.')
def run_grid_search(run_config):
for lr in lrs:
for df in dfs:
for i in range(num_of_runs):
train_config = create_config_gs(run_config, lr, df)
print(f'Starting training on: {train_config.name}')
train.main(train_config)
def __call__(self, seed):
algs, lrs, n_epochs = ['sgd'], [5e-1], [5000]
# algs, lrs, n_epochs = ['adam', 'sgd'], [3e-4, 5e-1], [1000, 5000]
for a, l, e in zip(algs, lrs, n_epochs):
# for T in [3,10]:
for T in [3]:
params = deepcopy(default_D2_pars)
params['sampler_params']['epoch_length'] = T
params['net_params']['saturations'] = [0, 1e8]
params['net_params'][
'save_folder'] = 'out/D2/relu/{}/T_{}/'.format(a, T)
params['train_params'].update({
'optimizer_name': a,
'lr': l,
'n_epochs': e
})
params['test_suite']['fit_internal_representation'][
'batch_size'] = 512
for test_name in params['test_suite'].keys():
params['test_suite'][test_name]['period'] = e // 4
os.makedirs(params['net_params']['save_folder'], exist_ok=True)
with open(
params['net_params']['save_folder'] +
'full_params.json', 'w+') as f:
json.dump(params, f, indent=4)
main(params, seed)
gc.collect()
def train_translation_model(data_dir,
arch,
extra_flags=None,
task='translation'):
train_parser = options.get_training_parser()
train_args = options.parse_args_and_arch(
train_parser,
[
'--task',
task,
data_dir,
'--save-dir',
data_dir,
'--arch',
arch,
'--lr',
'0.05',
'--max-tokens',
'500',
'--max-epoch',
'1',
'--no-progress-bar',
'--distributed-world-size',
'1',
'--source-lang',
'in',
'--target-lang',
'out',
] + (extra_flags or []),
)
train.main(train_args)
def test_main(self, mock_gan_train, mock_define_train_ops,
mock_cyclegan_loss, mock_define_model, mock_data_provider,
mock_gfile):
FLAGS.image_set_x_file_pattern = '/tmp/x/*.jpg'
FLAGS.image_set_y_file_pattern = '/tmp/y/*.jpg'
FLAGS.batch_size = 3
FLAGS.patch_size = 8
FLAGS.generator_lr = 0.02
FLAGS.discriminator_lr = 0.3
FLAGS.train_log_dir = '/tmp/foo'
FLAGS.master = 'master'
FLAGS.task = 0
FLAGS.cycle_consistency_loss_weight = 2.0
FLAGS.max_number_of_steps = 1
mock_data_provider.provide_custom_datasets.return_value = (tf.zeros(
[1, 2], dtype=tf.float32), tf.zeros([1, 2], dtype=tf.float32))
train.main(None)
mock_data_provider.provide_custom_datasets.assert_called_once_with(
['/tmp/x/*.jpg', '/tmp/y/*.jpg'], batch_size=3, patch_size=8)
mock_define_model.assert_called_once_with(mock.ANY, mock.ANY)
mock_cyclegan_loss.assert_called_once_with(
mock_define_model.return_value,
cycle_consistency_loss_weight=2.0,
tensor_pool_fn=mock.ANY)
mock_define_train_ops.assert_called_once_with(
mock_define_model.return_value, mock_cyclegan_loss.return_value)
mock_gan_train.assert_called_once_with(
mock_define_train_ops.return_value,
'/tmp/foo',
get_hooks_fn=mock.ANY,
hooks=mock.ANY,
master='master',
is_chief=True)
def main(args):
args = parser.parse_args(args)
config = {
'name': args.name,
'n_gpu': args.gpus,
'weights_path': args.weights_path,
'arch': {
'type': args.model_type,
'args': {
'variant': args.model_layers,
'num_classes': 2,
'print_model': True
}
},
'loss': 'cross_entropy_loss',
'metrics': ['accuracy'],
'data_loader': {
'type': 'COVID_Dataset',
'args': {
'root': args.dataset,
'k_fold_idx': args.k_fold_idx,
'mode': 'ct',
'pos_neg_file': args.labels,
'splits': [0.7, 0.15, 0.15],
'replicate_channel': 1,
'batch_size': 32,
'input_size': 224,
'num_workers': 2,
'self_supervised': 0
}
},
'optimizer': {
'type': 'Adam',
'args': {
'lr': args.learning_rate,
'weight_decay': args.weight_decay,
'amsgrad': True
}
},
'lr_scheduler': {
'type': 'StepLR',
'args': {
'step_size': args.lr_step_size,
'gamma': 0.1
}
},
'trainer': {
'epochs': args.epochs,
'save_dir': args.outdir,
'save_period': 1,
'verbosity': 2,
'monitor': 'min val_loss',
'early_stop': 10,
'tensorboard': False
}
}
config = ConfigParser(config)
train.main(config)
def objective(params):
try:
bounding_box_fbeta = params["bounding_box_fbeta"]
ir_coverage = params["ir_coverage"]
crop = params["crop"]
dropout = params["dropout"]
global workspace, data_dir, preprocess_config, training_config, analysis_results
# preprocess data
pp_data_dir = os.path.join(workspace, "data")
if not os.path.exists(pp_data_dir):
os.mkdir(pp_data_dir)
f_scores = visualize_analysis_results.f_score(
analysis_results["precisions"], analysis_results["recalls"],
bounding_box_fbeta)
i = np.argmax(f_scores)
bounding_box = np.unravel_index(i, f_scores.shape) + np.array(
analysis_results["bb_range"][0])
if tuple(bounding_box) != tuple(
preprocess_config["cropping"]["bb_size"]):
preprocess_config["cropping"]["bb_size"] = bounding_box
preprocess_dataset.main(preprocess_config,
data_dir,
pp_data_dir,
do_resample=False,
do_normalize=False,
crops=[crop])
histogram = analyze_dataset.analyze(
pp_data_dir,
do_bounding_boxes=False)["histograms"][LABELS.index(crop)]
_, i_min, i_max = visualize_analysis_results.crop_histogram(
histogram, ir_coverage)
if (i_min, i_max) != preprocess_config["normalization"][f"ir_{crop}"]:
preprocess_config["normalization"][f"ir_{crop}"] = (i_min, i_max)
preprocess_dataset.main(preprocess_config,
data_dir,
pp_data_dir,
do_resample=False,
do_crop=False,
crops=[crop])
# train
if os.path.exists(training_config["workspace"]):
shutil.rmtree(training_config["workspace"])
os.mkdir(training_config["workspace"])
def get_model(input_shape):
return models.simple_net.get_model(input_shape, dropout=dropout)
train.main(training_config, custom_model_generator=get_model)
with open(os.path.join(training_config["workspace"], "summary.json"),
"r") as file:
summary = json.load(file)
return {
"loss": -summary["auc_mean"],
"status": STATUS_OK,
"epochs": summary["epochs_mean"],
}
except Exception as e:
return {"status": STATUS_FAIL, "attachments": {"exception": str(e)}}
def main():
for i in range(7, 10):
train.main()
src = Path('./model/cost.h5')
to = Path(f'./model/cost-1024x4-1000x1000x{i + 1:02d}00.h5')
copy(str(src), str(to))
def test_experiment_vhrs1s2(self):
args = self.get_common_args()
args["experiment"] = "vhrs1s2"
try:
main(**args)
except Exception as err:
self.fail(err)
def main(args):
"""
Main running script
"""
# Get the config file
config = util.get_config(args.config)
root_dir = config['ROOT_DIR']
# fill out initial folders
if not os.path.isdir('{}/metadata'.format(root_dir)):
os.mkdir('{}/metadata'.format(root_dir))
print('created metadata dir')
if not os.path.isdir('{}'.format(config['OBS_ROOT'])):
os.mkdir('{}'.format(config['OBS_ROOT']))
print('created OBS dir')
if not os.path.isdir('{}'.format(config['ESTIMATORS_ROOT'])):
os.mkdir('{}'.format(config['ESTIMATORS_ROOT']))
print('created ESTIMATORS dir')
if not os.path.isdir('{}'.format(config['PREDICTIONS_ROOT'])):
os.mkdir('{}'.format(config['PREDICTIONS_ROOT']))
print('created PREDICTIONS dir')
if not os.path.isdir('{}'.format(config['QAQC_ROOT'])):
os.mkdir('{}'.format(config['QAQC_ROOT']))
print('created QAQC dir')
if not os.path.isdir('{}'.format(config['PLOT_ROOT'])):
os.mkdir('{}'.format(config['PLOT_ROOT']))
print('created PLOT dir')
# --- download data ---
if args.clean:
clean.main(config)
else:
print('skipping database cleaning')
# --- download data ---
if args.download:
download.main(config)
else:
print('skipping download of new data')
# --- train models
if args.train:
train.main(config)
else:
print('skip training')
# --- make predictions ---
if args.predict:
predict.main(config)
else:
print('skipping download of new data')
# --- run qaqc checks ---
if args.qaqc:
qaqc.main(config)
else:
print('skipping qaqc')
# --- plot ---
if args.plot:
plot.main(config)
else:
print('skipping plots')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('mode', choices=['train', 'plot', 'eval', 'hogwild'])
# IO arguments.
parser.add_argument('--name', type=str, default='text8.10k',
help='name for model')
parser.add_argument('--vocab-dir', type=str, default='data/vocab',
help='input path for vocabulary')
parser.add_argument('--matrix-dir', type=str, default='data/cooccur',
help='input path for cooccurence matrix')
parser.add_argument('--out-dir', type=str, default='vec',
help='ouput directory to write vectors')
parser.add_argument('--model-dir', type=str, default='models',
help='directory to save model')
parser.add_argument('--log-dir', type=str, default='log',
help='directory log losses')
parser.add_argument('--vec-path', type=str, default='',
help='path to load vectors for plotting')
parser.add_argument('--gensim-format', action='store_true',
help='save vectors in gensim format')
# Model arguments.
parser.add_argument('--emb-dim', type=int, default=50,
help='dimension of vectors')
# Train arguments.
parser.add_argument('--num-updates', type=int, default=10000,
help='number of parameter updates')
parser.add_argument('--batch-size', type=int, default=512,
help='size of minibatches')
parser.add_argument('--lr', type=float, default=1e-2,
help='learning rate')
parser.add_argument('--seed', type=int, default=42,
help='random seed')
parser.add_argument('--use-schedule', action='store_true',
help='using scheduler for optimizer')
parser.add_argument('--save-every', type=int, default=1000,
help='how often to save the model parameters')
parser.add_argument('--print-every', type=int, default=100,
help='how often to print loss to screen')
# Plot arguments.
parser.add_argument('--tsne', action='store_true',
help='plot tsne')
parser.add_argument('--matrices', action='store_true',
help='plot matrices and decomposition')
args = parser.parse_args()
if args.mode == 'train':
train.main(args)
if args.mode == 'plot':
plot.main(args)
if args.mode == 'hogwild':
hogwild.main(args)
def main():
logger.info("Running main ...")
train.main()
predict.main()
logger.info("Run complete ...")
return
def main(argv):
parser = argparse.ArgumentParser()
parser.add_argument("--type", default=0, help="number of param")
args = parser.parse_args()
r.reset()
train.main(args.type)
m.main(args.type)
def main():
print "TRAINING"
train.main()
print "PREDICTING"
predict.main()
print "EVAlUATING"
evaluate.main()
def test_main(self):
FLAGS.image_file_patterns = [
os.path.join(FLAGS.test_srcdir, TESTDATA_DIR, 'black/*.jpg'),
os.path.join(FLAGS.test_srcdir, TESTDATA_DIR, 'blond/*.jpg'),
os.path.join(FLAGS.test_srcdir, TESTDATA_DIR, 'brown/*.jpg'),
]
FLAGS.max_number_of_steps = 1
FLAGS.steps_per_eval = 1
FLAGS.batch_size = 1
train.main(None, _test_generator, _test_discriminator)
def test_main(self, mock_provide_data):
FLAGS.image_file_patterns = ['/tmp/A/*.jpg', '/tmp/B/*.jpg', '/tmp/C/*.jpg']
FLAGS.max_number_of_steps = 10
FLAGS.batch_size = 2
num_domains = 3
images_shape = [FLAGS.batch_size, FLAGS.patch_size, FLAGS.patch_size, 3]
img_list = [tf.zeros(images_shape)] * num_domains
lbl_list = [tf.one_hot([0] * FLAGS.batch_size, num_domains)] * num_domains
mock_provide_data.return_value = (img_list, lbl_list)
train.main(None)
def _test_build_graph_helper(self, weight_factor):
FLAGS.max_number_of_steps = 0
FLAGS.weight_factor = weight_factor
FLAGS.batch_size = 9
FLAGS.patch_size = 32
mock_imgs = np.zeros(
[FLAGS.batch_size, FLAGS.patch_size, FLAGS.patch_size, 3],
dtype=np.float32)
with mock.patch.object(train, 'data_provider') as mock_data_provider:
mock_data_provider.provide_data.return_value = mock_imgs
train.main(None)
def testTrainingAndInferenceGraphsAreCompatible(
self, mock_provide_custom_data, unused_mock_gan_train):
# Training and inference graphs can get out of sync if changes are made
# to one but not the other. This test will keep them in sync.
# Save the training graph
train_sess = tf.Session()
FLAGS.image_set_x_file_pattern = '/tmp/x/*.jpg'
FLAGS.image_set_y_file_pattern = '/tmp/y/*.jpg'
FLAGS.batch_size = 3
FLAGS.patch_size = 128
FLAGS.generator_lr = 0.02
FLAGS.discriminator_lr = 0.3
FLAGS.train_log_dir = self._export_dir
FLAGS.master = 'master'
FLAGS.task = 0
FLAGS.cycle_consistency_loss_weight = 2.0
FLAGS.max_number_of_steps = 1
mock_provide_custom_data.return_value = (
tf.zeros([3, 4, 4, 3,]), tf.zeros([3, 4, 4, 3]))
train.main(None)
init_op = tf.global_variables_initializer()
train_sess.run(init_op)
train_saver = tf.train.Saver()
train_saver.save(train_sess, save_path=self._ckpt_path)
# Create inference graph
tf.reset_default_graph()
FLAGS.patch_dim = FLAGS.patch_size
logging.info('dir_path: %s', os.listdir(self._export_dir))
FLAGS.checkpoint_path = self._ckpt_path
FLAGS.image_set_x_glob = self._image_glob
FLAGS.image_set_y_glob = self._image_glob
FLAGS.generated_x_dir = self._genx_dir
FLAGS.generated_y_dir = self._geny_dir
inference_demo.main(None)
logging.info('gen x: %s', os.listdir(self._genx_dir))
# Check that the image names match
self.assertSetEqual(
set(_basenames_from_glob(FLAGS.image_set_x_glob)),
set(os.listdir(FLAGS.generated_y_dir)))
self.assertSetEqual(
set(_basenames_from_glob(FLAGS.image_set_y_glob)),
set(os.listdir(FLAGS.generated_x_dir)))
# Check that each image in the directory looks as expected
for directory in [FLAGS.generated_x_dir, FLAGS.generated_x_dir]:
for base_name in os.listdir(directory):
image_path = os.path.join(directory, base_name)
self.assertRealisticImage(image_path)
def test_build_graph(self, gan_type):
FLAGS.max_number_of_steps = 0
FLAGS.gan_type = gan_type
# Mock input pipeline.
mock_imgs = np.zeros([FLAGS.batch_size, 28, 28, 1], dtype=np.float32)
mock_lbls = np.concatenate(
(np.ones([FLAGS.batch_size, 1], dtype=np.int32),
np.zeros([FLAGS.batch_size, 9], dtype=np.int32)), axis=1)
with mock.patch.object(train, 'data_provider') as mock_data_provider:
mock_data_provider.provide_data.return_value = (
mock_imgs, mock_lbls, None)
train.main(None)
def test_full_flow(self, mock_data_provider):
FLAGS.eval_dir = self.get_temp_dir()
FLAGS.batch_size = 16
FLAGS.max_number_of_steps = 2
FLAGS.noise_dims = 3
# Construct mock inputs.
mock_imgs = np.zeros([FLAGS.batch_size, 28, 28, 1], dtype=np.float32)
mock_lbls = np.concatenate(
(np.ones([FLAGS.batch_size, 1], dtype=np.int32),
np.zeros([FLAGS.batch_size, 9], dtype=np.int32)), axis=1)
mock_data_provider.provide_data.return_value = (mock_imgs, mock_lbls, None)
train.main(None)
def test_build_graph(self, conditional, use_sync_replicas):
FLAGS.max_number_of_steps = 0
FLAGS.conditional = conditional
FLAGS.use_sync_replicas = use_sync_replicas
FLAGS.batch_size = 16
# Mock input pipeline.
mock_imgs = np.zeros([FLAGS.batch_size, 32, 32, 3], dtype=np.float32)
mock_lbls = np.concatenate(
(np.ones([FLAGS.batch_size, 1], dtype=np.int32),
np.zeros([FLAGS.batch_size, 9], dtype=np.int32)), axis=1)
with mock.patch.object(train, 'data_provider') as mock_data_provider:
mock_data_provider.provide_data.return_value = (
mock_imgs, mock_lbls, None, None)
train.main(None)
def test_run_one_train_step(self, mock_data_provider):
FLAGS.max_number_of_steps = 1
FLAGS.gan_type = 'unconditional'
FLAGS.batch_size = 5
FLAGS.grid_size = 1
tf.set_random_seed(1234)
# Mock input pipeline.
mock_imgs = np.zeros([FLAGS.batch_size, 28, 28, 1], dtype=np.float32)
mock_lbls = np.concatenate(
(np.ones([FLAGS.batch_size, 1], dtype=np.int32),
np.zeros([FLAGS.batch_size, 9], dtype=np.int32)), axis=1)
mock_data_provider.provide_data.return_value = (mock_imgs, mock_lbls, None)
train.main(None)
def train(query):
ans = "<NA>"
import train
ans = list(train.main(query))
#print ans
if ans:
ans={"train":ans}
#print ans
return [ans]
def train_translation_model(data_dir, arch, extra_flags=None):
train_parser = options.get_training_parser()
train_args = options.parse_args_and_arch(
train_parser,
[
'--task', 'translation',
data_dir,
'--save-dir', data_dir,
'--arch', arch,
'--optimizer', 'nag',
'--lr', '0.05',
'--max-tokens', '500',
'--max-epoch', '1',
'--no-progress-bar',
'--distributed-world-size', '1',
'--source-lang', 'in',
'--target-lang', 'out',
] + (extra_flags or []),
)
train.main(train_args)
def train_language_model(data_dir, arch):
train_parser = options.get_training_parser()
train_args = options.parse_args_and_arch(
train_parser,
[
'--task', 'language_modeling',
data_dir,
'--arch', arch,
'--optimizer', 'nag',
'--lr', '1.0',
'--criterion', 'adaptive_loss',
'--adaptive-softmax-cutoff', '5,10,15',
'--decoder-layers', '[(850, 3)] * 2 + [(1024,4)]',
'--decoder-embed-dim', '280',
'--max-tokens', '500',
'--tokens-per-sample', '500',
'--save-dir', data_dir,
'--max-epoch', '1',
'--no-progress-bar',
'--distributed-world-size', '1',
],
)
train.main(train_args)
def test_train(self):
argv = ["train.py", "-niter", "2", "-batch", "2", "tests/test_train_data.arff", "tests/test_train_model.model"]
train.main(argv)
argv = ["train.py", "-niter", "2", "-batch", "2", "tests/test_train_data.arff", "tests/test_train_model.model"]
train.main(argv)
argv = ["train.py", "-niter", "2", "-batch", "2", "tests/test_train_data.arff", "tests/test_train_model.model"]
train.main(argv)
def main(config_file):
with open(config_file) as f:
# use safe_load instead load
dataMap = yaml.safe_load(f)
filebase = os.path.splitext(os.path.basename(config_file))[0]
sig_file = os.path.join('../data', filebase + '_data.mat')
scat_file = os.path.join('../data', filebase + '_scat.mat')
labels = dataMap["data"]["labels"]
Nsig = dataMap["data"]["N"]
if (('derivative' in dataMap["data"]) and
dataMap['data']["derivative"]):
extract_derivatives(labels, Nsig, sig_file)
else:
extract_signals(labels, Nsig, sig_file)
if dataMap["features"] == 'holder':
holder_exp(sig_file, scat_file)
elif 'scattering_transfer' in dataMap["features"]:
scat = dataMap["features"]["scattering_transfer"]
scatter_transfer(sig_file, scat_file, scat["N"], scat["T"])
else:
scat = dataMap["features"]["scatter"]
fractal_scatter(sig_file, scat_file, scat["N"], scat["M"], scat["T"])
score, X, y, yhat = train.main(scat_file, dataMap["machine_learning"])
# Make the summary plot
fig = plot_results(y, yhat, score)
fig.savefig(os.path.join('../results/', filebase +'.eps'))
dataMap["results"] = str(score)
outyaml = os.path.join('../results', filebase+'.yaml')
with open(outyaml, 'w') as f:
f.write(yaml.dump(dataMap, default_flow_style=False))
theano.config.floatX = 'float32'
np.random.seed(0)
if __name__ == '__main__':
import argparse
import train
parser = argparse.ArgumentParser(description='Train/Test A Language Model.')
parser.add_argument('-data', help='path to data')
# NN architecture
parser.add_argument('--vocab', type=int, default=100000000, help='vocabulary size')
parser.add_argument('--emb', type=int, default=32, help='dimension of embeddings')
parser.add_argument('--hidden', type=int, default=32, help='dimension of hidden layer')
parser.add_argument('--layer', type=int, default=1, help='number of layers')
# Training Parameters
parser.add_argument('--batch', type=int, default=10, help='batch size')
parser.add_argument('--opt', default='adagrad', help='optimization method')
parser.add_argument('--epoch', type=int, default=100, help='number of epochs')
parser.add_argument('--lr', type=float, default=0.0001, help='learning rate')
parser.add_argument('--save', type=bool, default=False, help='parameters to be saved or not')
parser.add_argument('--init_emb', default=None, help='Initial embedding to be loaded')
parser.add_argument('--n_words', type=int, default=100, help='number of words')
train.main(parser.parse_args())
#!/usr/bin/env python
from secrets import *
from twython import TwythonStreamer
import train
train.main()
from colorama import init, Fore, Back
train.categories[0].colour = Fore.GREEN + Back.BLACK
train.categories[1].colour = Fore.RED + Back.WHITE
reset = Fore.RESET + Back.RESET
class SentimentStreamer(TwythonStreamer):
def on_success(self, data):
if 'text' in data:
mood = train.classify(data['text'])
if mood is None:
print data['text'].encode('utf-8')
print "neutral"
else:
print mood.colour + data['text'].encode('utf-8') + reset
print(mood)
def on_error(self, status_code, data):
print status_code
stream = SentimentStreamer(APP_KEY, APP_SECRET,
OAUTH_TOKEN, OAUTH_TOKEN_SECRET)
stream.statuses.filter(locations=[-6.833496, 49.894634, 2.087402, 59.623325])
def test_pipeline(self):
train.corpus_paths = ['../data/en-cs-test.txt']
train.main()
parser.add_argument('--init_emb', default=None, help='Initial embedding to be loaded')
parser.add_argument('--opt', default='adam', help='optimization method')
parser.add_argument('--lr', type=float, default=0.01, help='learning rate')
parser.add_argument('--reg', type=float, default=0.0001, help='L2 Reg rate')
parser.add_argument('--batch', type=int, default=32, help='batch size')
parser.add_argument('--epoch', type=int, default=500, help='number of epochs to train')
parser.add_argument('--no-shuffle', action='store_true', default=False, help='don\'t shuffle training data')
""" test options """
parser.add_argument('--model', default=None, help='path to model')
parser.add_argument('--arg_dict', default=None, help='path to arg dict')
parser.add_argument('--vocab_dict', default=None, help='path to vocab dict')
parser.add_argument('--emb_dict', default=None, help='path to emb dict')
argv = parser.parse_args()
print
print argv
print
if argv.mode == 'train':
import train
train.main(argv)
else:
import test
assert argv.model is not None
assert argv.arg_dict is not None
assert argv.vocab_dict is not None
assert argv.emb_dict is not None
test.main(argv)
def main():
parser = argparse.ArgumentParser(description="CliRel (Clinical Relation) \
extractor- trains a classifier able to \
determine the type of relation between \
two medical concepts in a sentence.")
# Add arguments here
parser.add_argument("--train", nargs=3,
metavar=("train_dir", "model_file", "model_type"), type=str,
help="Directory should contain three subdirs (txt, \
concept, rel) containing .txt, .con, .rel files. \
Will train a classifier on this data. \
Trained model will be written to specified model file.\n \
Current model types:[svm-spt, svm-insert, svm-suffix]",
default=None)
parser.add_argument("--predict", nargs=3,
metavar=("test_dir", "model_file", "results_dir"), type=str,
help="Directory contains concept and text files \
that the specified (or default) model will predict. \
Resulting relation files will be written to \
the specified results directory.",
default=None)
parser.add_argument("--evaluate", nargs=3,
metavar=("test_dir", "gold_dir", "eval_file"), type=str,
help="Evaluate the relation files in the test directory \
in comparison with those in the gold directory. The \
results will be written to the evaluation file.",
default=None)
parser.add_argument("--verbose", action="store_true",
default=False, help="Show debugging info.")
args = parser.parse_args()
if not args.predict and not args.train and not args.evaluate:
sys.stderr.write("ERROR: No valid flag specified.\n")
parser.print_help()
sys.exit(1)
if args.train:
checkDir(args.train[0])
checkDir(os.path.dirname(args.train[1]))
if (os.path.isdir(args.train[1])):
sys.stderr.write("ERROR: Model expected to be a file, %s is a directory\n"
% args.train[1])
sys.exit(1)
train.main(args.train[0], args.train[1], args.train[2], args.verbose)
if args.predict:
checkDir(args.predict[0])
checkFile(args.predict[1])
checkDir(args.predict[2])
predict.main(args.predict[0], args.predict[1], args.predict[2], args.verbose)
if args.evaluate:
checkDir(args.evaluate[0])
checkDir(args.evaluate[1])
checkDir(os.path.dirname(args.evaluate[2]))
if (os.path.isdir(args.evaluate[2])):
sys.stderr.write("ERROR: eval_file expected to be a file, %s is a \
directory\n" % args.evaluate[2])
sys.exit(1)
evaluate.main(args.evaluate[0], args.evaluate[1], args.evaluate[2], args.verbose)
from bionlp.utils.crf_arguments import crf_model_arguments
from train import main
if __name__=="__main__":
config_params=crf_model_arguments()
# Required for choosing CRF network models.
config_params['CRF_MODEL_ON']=True
main(config_params)
def main(run_dir="rfe_chain", start=None, start_auc=None,
verbose=None, logfile=None):
"""
Main function to run the chain.
"""
if logfile is not None:
sys.stdout = open(logfile, "w")
# load starting json
with open(start) as f:
start = json.load(f)
if start_auc is None:
startauc = 0.8
start['AUC_SCORE_PATH'] = run_dir
# have to load a list of possible features to replace with
if all("10feat" in feature for feature in start['FEATURES']):
with open("10featlist.json") as fh:
featlist = json.load(fh)['FEATURES']
else:
featlist = get_featlist()
# and possible preceding modifiers
modlist = get_modlist()
# creat list of combinations of these two lists
comblist = []
for mod in modlist:
for feature in featlist:
comblist.append('{0}_{1}_'.format(mod, feature))
# define sampled json
prevsample = copy.deepcopy(start)
# initialise auc
prevauc = startauc
first = 1
counter = 0
converged = False
# will decide what constitutes converged later
while not converged:
sample = copy.deepcopy(prevsample)
# If this isn't the first one, sample new settings
if not first:
# Sample a new hdf5 and replace existing at random
# Or, just push it in, or just drop a hdf5 at random
utils.print_verbose("===== Sampling new proposal "
"settings ======", flag=verbose)
# sample new settings
# shuffle combinations
random.shuffle(comblist)
# pop 3 features off this
added = [comblist.pop() for i in range(3)]
# add them to the settings
sample['FEATURES'] = added
utils.print_verbose("============================"
"===============", flag=verbose)
# ensure that ordering of the features is the same between jsons
sample['FEATURES'].sort()
# Then save this new json with a descriptive name
# unless it's already been generated
if first:
featurerecord = "".join(sample['FEATURES'])
else:
featurerecord = featurerecord + "".join(sample['FEATURES'])
md5name = hashlib.md5(featurerecord.encode('UTF-8')).hexdigest()
# get a list of the files in the run_dir
existingjsons = glob.glob(run_dir + "/*.json")
# check if the md5 exists
if md5name + ".json" in existingjsons:
# then load the results of that run
with open(os.path.join(run_dir, "AUC_scores.csv"), "r") as fh:
c = csv.reader(fh, delimiter="\t")
utils.print_verbose("Already ran {0},"
"reading from results.".format(md5name), flag=verbose)
for line in c:
# look for that md5sum
if md5name in line[0]:
auc_score = line[-1]
else:
# save a json with this name and run train.py on it
samplefname = os.path.join(run_dir, md5name + ".json")
utils.print_verbose("Creating new settings"
" file for {0}".format(samplefname), flag=verbose)
with open(samplefname, "w") as fh:
json.dump(sample, fh)
# call train.py
try:
if first:
auc_score_dict = train.main(samplefname, verbose=verbose,
store_models=False, store_features=True)
else:
picklefname = prevsamplefname.split(".")[0] + \
"_feature_dump.pickle"
# load the features saved in the last run
auc_score_dict = train.main(samplefname, verbose=verbose,
store_models=False, store_features=True,
load_pickled=picklefname)
prevsamplefname = samplefname
auc_score = auc_score_dict['all']
except IndexError:
print("Warning: accidentally added invalid feature.")
os.remove(samplefname)
# set auc to zero so these settings are not accepted
auc_score = 0
prevsample = sample
# can't be first anymore
first = 0
# as it may be bad manners to run infinite loops
counter += 1
if counter > 100:
converged = True
return None
from train import main
logger = logging.getLogger(__name__)
# Get the arguments
parser = argparse.ArgumentParser()
parser.add_argument(
"--proto", default="get_config_ch_topical_former_douban",
help="Prototype config to use for config")
parser.add_argument(
"--bokeh", default=False, action="store_true",
help="Use bokeh server for plotting")
parser.add_argument(
"--mode", choices=["train", "translate"], default='translate',
help="The mode to run. In the `train` mode a model is trained."
" In the `translate` mode a trained model is used to translate"
" an input file and generates tokenized translation.")
parser.add_argument(
"--test-file", default='', help="Input test file for `translate` mode")
args = parser.parse_args()
if __name__ == "__main__":
# Get configurations for model
configuration = getattr(configurations, args.proto)()
# configuration['test_set'] = args.test_file
logger.info("Model options:\n{}".format(pprint.pformat(configuration)))
# Get data streams and call main
main(args.mode, configuration, args.bokeh)
def main(mcmcdir="hdf5mcmc", start=None, start_auc=None,
verbose=True, logfile=None, discr_flag=False):
"""
Contains the main loop for this script.
Pseudo-MHMCMC to find optimal AUC scoring
combinations of HDF5s.
start - location of json file settings to begin at
"""
if logfile is not None:
sys.stdout = open(logfile, "w")
# pseudo-code for the MCMC iteration
# want it to start with the probably good features
with open(start) as f:
start = json.load(f)
if start_auc is None:
startauc = 0.8
# hardcode AUC results to the hdf5mcmc directory
start['AUC_SCORE_PATH'] = mcmcdir
# have to load a list of possible features to replace with
if all("10feat" in feature for feature in start['FEATURES']):
with open("10featlist.json") as fh:
featlist = json.load(fh)['FEATURES']
else:
featlist = get_featlist()
# and possible preceding modifiers
modlist = get_modlist()
# define sampled json
prevsample = copy.deepcopy(start)
# initialise auc
prevauc = startauc
counter = 0
converged = False
# will decide what constitutes converged later
while not converged:
sample = copy.deepcopy(prevsample)
# Sample a new hdf5 and replace existing at random
# Or, just push it in, or just drop a hdf5 at random
utils.print_verbose("===== Sampling new proposal "
"settings ======", flag=verbose)
u = np.random.rand()
if u < 0.25:
# drop an element at random
features = sample['FEATURES'][:]
random.shuffle(features)
dropped = features.pop()
sample['FEATURES'] = features
utils.print_verbose(
"Dropped feature {0}".format(dropped),
flag=verbose)
elif u > 0.25 and u < 0.5:
# keep trying to sample a new feature until we
# find one that's not in there already
while True:
# push a new feature, but don't remove an old one
newfeature = random.sample(featlist, 1)[0]
newmod = random.sample(modlist, 1)[0]
added = '{0}_{1}_'.format(newmod, newfeature)
if added not in sample['FEATURES']:
break
sample['FEATURES'].append(added)
utils.print_verbose(
"Added feature {0}".format(added),
flag=verbose)
elif u > 0.5:
# push a new feature and remove an old one
features = sample['FEATURES'][:]
random.shuffle(features)
dropped = features.pop()
# keep trying to sample a new feature until we
# find one that's not in there already
while True:
# push a new feature, but don't remove an old one
newfeature = random.sample(featlist, 1)[0]
newmod = random.sample(modlist, 1)[0]
added = '{0}_{1}_'.format(newmod, newfeature)
if added not in sample['FEATURES']:
break
features.append(added)
sample['FEATURES'] = features
utils.print_verbose("Switched feature {0} for "
"{1}".format(dropped, added), flag=verbose)
utils.print_verbose("============================"
"===============", flag=verbose)
# ensure that ordering of the features is the same between jsons
sample['FEATURES'].sort()
# Then save this new json with a descriptive name
# unless it's already been generated
md5name = hashlib.md5(
"".join(sample['FEATURES']).encode('UTF-8')).hexdigest()
# get a list of the files in the mcmcdir
existingjsons = glob.glob(mcmcdir + "/*.json")
# check if the md5 exists
if md5name + ".json" in existingjsons:
# then load the results of that run
with open(os.path.join(mcmcdir, "AUC_scores.csv"), "r") as fh:
c = csv.reader(fh, delimiter="\t")
utils.print_verbose("Already ran {0},"
"reading from results.".format(md5name), flag=verbose)
for line in c:
# look for that md5sum
if md5name in line[0]:
auc_score = line[-1]
else:
# save a json with this name and run train.py on it
samplefname = os.path.join(mcmcdir, md5name + ".json")
utils.print_verbose("Creating new settings"
" file for {0}".format(samplefname), flag=verbose)
with open(samplefname, "w") as fh:
json.dump(sample, fh)
# call train.py or discriminate.py
if discr_flag:
try:
auc_score_dict = discriminate.main(samplefname,
verbose=verbose)
# don't want to rename this variable
# even though it is no longer an AUC score
# want a low accuracy score, strangely enough
auc_score = 1 - auc_score_dict['all']
except IndexError:
print("Warning: accidentally added invalid feature.")
os.remove(samplefname)
# set auc to zero so these settings are not accepted
auc_score = 0
else:
try:
auc_score_dict = train.main(samplefname,
verbose=verbose, store_models=False)
auc_score = auc_score_dict['all'] - 0.5
except IndexError:
print("Warning: accidentally added invalid feature.")
os.remove(samplefname)
# set auc to zero so these settings are not accepted
auc_score = 0
utils.print_verbose("==== Acceptance calculation ====", flag=verbose)
# compute acceptance probability from AUC:
# r = min(1,AUC/(previous AUC))
acceptance = np.max([np.min([1, auc_score / prevauc]), 0])
u = np.random.rand()
# accept new point with probability r
if u < acceptance:
prevsample = sample
# save current auc
prevauc = auc_score
utils.print_verbose("accepting new settings with probability "
"{0}".format(acceptance), flag=verbose)
else:
utils.print_verbose("rejecting new settings with probability "
"{0}".format(1.0 - acceptance), flag=verbose)
utils.print_verbose("================================", flag=verbose)
# otherwise it will not overwrite prevsample, so continue from where it
# was
# as it may be bad manners to run infinite loops
counter += 1
if counter > 100:
converged = True
import train
import config
import sys
if __name__=="__main__":
jobid = sys.argv[1]
with open("."+jobid+".id") as f:
arrayid = int(f.readline())-1
params = config.parameters_list[arrayid]
params.append("-j " + jobid)
params = " ".join(params).split()
train.main(params)
