def _init_():
'''
Inference script for image-classification task on mxnet
Update: 2019-01-03
Author: @Northrend
Contributor:
Change log:
2019/01/03 v3.5 fix category list sort bug
2018/09/30 v3.4 support parallelized image pre-processing
2018/07/23 v3.3 fix testing bug caused by mxnet v1.0.0
2018/06/22 v3.2 support mutable images testing
2018/06/19 v3.1 support multi(3 for now) crop
2018/06/11 v3.0 code-refactoring
2018/05/31 v2.6 support log file name with parent path
2018/04/18 v2.5 support print foward fps
2017/12/29 v2.4 fix numpy truth value err bug
2017/12/11 v2.3 fix center crop bug
2017/12/07 v2.2 convert img-data to float before resizing
2017/11/29 v2.1 support center crop
2017/11/17 v2.0 support mean and std
2017/09/25 v1.3 support alternative gpu
2017/09/21 v1.2 support batch-inference & test mode
2017/07/31 v1.1 support different label file
2017/06/20 v1.0 basic functions
Usage:
mxnet_image_classifier.py <input-cfg> [--single-img=str]
mxnet_image_classifier.py -v | --version
mxnet_image_classifier.py -h | --help
Arguments:
<input-cfg> path to customized config file
Options:
-h --help show this help screen
-v --version show current version
-------------------------------------------------------
--single-img=str give path to one image file
to use single image testing
'''
# merge configuration
merge_cfg_from_file(args["<input-cfg>"])
# config logger
logger.setLevel(eval('logging.' + cfg.LOG_LEVEL))
assert cfg.LOG_PATH, logger.error('Missing LOG_PATH!')
fhandler = logging.FileHandler(cfg.LOG_PATH, mode=cfg.LOG_MODE)
logger.addHandler(fhandler)
# print arguments
logger.info('=' * 80 + '\nCalled with arguments:')
for key in sorted(args.keys()):
logger.info('{:<20}= {}'.format(key.replace('--', ''), args[key]))
logger.info('=' * 80)
# reset logger format
fhandler.setFormatter(logging.Formatter(log_format))
def _init_():
'''
Training script for image-classification task on mxnet
Update: 2019-05-14
Author: @Northrend
Contributor:
Changelog:
2019/05/14 v2.3 fix top_k bug
2019/04/08 v2.2 fix crashing caused by loading image error
2018/11/08 v2.1 fix pre-eval batch size bug
2018/11/05 v2.0 support lr fixed decay step
2018/10/29 v1.9 fix PIL loading bug
2018/10/24 v1.8 support mobilenet-v2 and se-mobilenet-v2
2018/10/23 v1.7 fix cuda oom caused by val-phase grad
2018/10/18 v1.6 fix mix-up training bug
2018/10/15 v1.5 support mix-up training trick
2018/10/10 v1.4 support pre-evaluation
2018/10/09 v1.3 support check nets mode
support resnet-v2
2018/09/26 v1.2 optimize logging info
2018/09/25 v1.1 support finetune & scratch training
support xavier initialization
2018/09/13 v1.0 basic functions
Usage:
pytorch_train.py <input-cfg> [-c|--check-nets]
pytorch_train.py -v | --version
pytorch_train.py -h | --help
Arguments:
<input-cfg> path to customized config file
Options:
-h --help show this help screen
-v --version show current version
-------------------------------------------------------------
-c --check-nets check available network arch only
'''
# merge configuration
merge_cfg_from_file(args["<input-cfg>"])
# config logger
logger.setLevel(eval('logging.' + cfg.LOG_LEVEL))
assert cfg.LOG_PATH, logger.error('Missing LOG_PATH!')
fhandler = logging.FileHandler(cfg.LOG_PATH, mode=cfg.LOG_MODE)
logger.addHandler(fhandler)
# print arguments
logger.info('=' * 80 + '\nCalled with arguments:')
for key in sorted(args.keys()):
logger.info('{:<20}= {}'.format(key.replace('--', ''), args[key]))
logger.info('=' * 80)
# reset logger format
fhandler.setFormatter(logging.Formatter(log_format))
def _init_():
'''
Inference script for image-classification task on mxnet
Update: 2018-06-22 12:09:37
Author: @Northrend
Contributor:
Change log:
2018/06/22 v3.2 support mutable images testing
2018/06/19 v3.1 support multi(3 for now) crop
2018/06/11 v3.0 code-refactoring
2018/05/31 v2.6 support log file name with parent path
2018/04/18 v2.5 support print foward fps
2017/12/29 v2.4 fix numpy truth value err bug
2017/12/11 v2.3 fix center crop bug
2017/12/07 v2.2 convert img-data to float before resizing
2017/11/29 v2.1 support center crop
2017/11/17 v2.0 support mean and std
2017/09/25 v1.3 support alternative gpu
2017/09/21 v1.2 support batch-inference & test mode
2017/07/31 v1.1 support different label file
2017/06/20 v1.0 basic functions
Usage:
mxnet_image_classifier.py <input-cfg> [--single-img=str]
mxnet_image_classifier.py -v | --version
mxnet_image_classifier.py -h | --help
Arguments:
<input-cfg> path to customized config file
Options:
-h --help show this help screen
-v --version show current version
-------------------------------------------------------
--single-img=str give path to one image file
to use single image testing
'''
# merge configuration
merge_cfg_from_file(args["<input-cfg>"])
# config logger
logger.setLevel(eval('logging.' + cfg.LOG_LEVEL))
assert cfg.LOG_PATH, logger.error('Missing LOG_PATH!')
fhandler = logging.FileHandler(cfg.LOG_PATH, mode=cfg.LOG_MODE)
logger.addHandler(fhandler)
# print arguments
logger.info('=' * 80 + '\nCalled with arguments:')
for key in sorted(args.keys()):
logger.info('{:<20}= {}'.format(key.replace('--', ''), args[key]))
logger.info('=' * 80)
# reset logger format
fhandler.setFormatter(logging.Formatter(log_format))
def _init_():
'''
Alternately training multi-image-classification networks on mxnet
Update: 2018-06-07 15:12:24
Author: @Northrend
Contributor:
Changelog:
2018/06/07 v1.2 modify config param name
2018/05/28 v1.1 fix evaluation bug
2018/05/24 v1.0 basic functions
Usage:
altet_train.py <input-cfg>
altet_train.py -v | --version
altet_train.py -h | --help
Arguments:
<input-cfg> path to customized config file
Options:
-h --help show this help screen
-v --version show current version
'''
# merge configuration
merge_cfg_from_file(args["<input-cfg>"])
# config logger
logger.setLevel(eval('logging.' + cfg.LOG_LEVEL))
assert cfg.LOG_PATH, logger.error('Missing LOG_PATH!')
fhandler = logging.FileHandler(cfg.LOG_PATH, mode=cfg.LOG_MODE)
logger.addHandler(fhandler)
# print arguments
logger.info('=' * 80 + '\nCalled with arguments:')
for key in sorted(args.keys()):
logger.info('{:<20}= {}'.format(key.replace('--', ''), args[key]))
logger.info('=' * 80)
# reset logger format
fhandler.setFormatter(logging.Formatter(log_format))
def _init_():
'''
Precision-guided training for image classification task on mxnet
Update: 2018/09/06
Author: @Northrend
Contributor:
Changelog:
2018/09/06 v1.1 support single image heatmap
2018/05/30 v1.0 basic functions
Usage:
pg_train.py <input-cfg> [-s|--single-img]
pg_train.py -v | --version
pg_train.py -h | --help
Arguments:
<input-cfg> path to customized config file
Options:
-h --help show this help screen
-v --version show current version
-s --single-img single img model, will draw heat-map simultaneously
'''
# merge configuration
merge_cfg_from_file(args["<input-cfg>"])
# config logger
logger.setLevel(eval('logging.' + cfg.LOG_LEVEL))
assert cfg.LOG_PATH, logger.error('Missing LOG_PATH!')
fhandler = logging.FileHandler(cfg.LOG_PATH, mode=cfg.LOG_MODE)
logger.addHandler(fhandler)
# print arguments
logger.info('=' * 80 + '\nCalled with arguments:')
for key in sorted(args.keys()):
logger.info('{:<20}= {}'.format(key.replace('--', ''), args[key]))
logger.info('=' * 80)
# reset logger format
fhandler.setFormatter(logging.Formatter(log_format))
def _init_():
'''
Alternately training multi-image-classification networks on mxnet
Update: 2018-06-07 15:12:24
Author: @Northrend
Contributor:
Changelog:
2018/06/07 v1.2 modify config param name
2018/05/28 v1.1 fix evaluation bug
2018/05/24 v1.0 basic functions
Usage:
altet_train.py <input-cfg>
altet_train.py -v | --version
altet_train.py -h | --help
Arguments:
<input-cfg> path to customized config file
Options:
-h --help show this help screen
-v --version show current version
'''
# merge configuration
merge_cfg_from_file(args["<input-cfg>"])
# config logger
logger.setLevel(eval('logging.' + cfg.LOG_LEVEL))
assert cfg.LOG_PATH, logger.error('Missing LOG_PATH!')
fhandler = logging.FileHandler(cfg.LOG_PATH, mode=cfg.LOG_MODE)
logger.addHandler(fhandler)
# print arguments
logger.info('=' * 80 + '\nCalled with arguments:')
for key in sorted(args.keys()):
logger.info('{:<20}= {}'.format(key.replace('--', ''), args[key]))
logger.info('=' * 80)
# reset logger format
fhandler.setFormatter(logging.Formatter(log_format))
def _init_():
'''
Training script for image-classification task on mxnet
Update: 2018-09-29
Author: @Northrend
Contributor:
Changelog:
2018/09/29 v1.0 basic functions
Usage:
pytorch_train.py <input-cfg>
pytorch_train.py -v | --version
pytorch_train.py -h | --help
Arguments:
<input-cfg> path to customized config file
Options:
-h --help show this help screen
-v --version show current version
'''
# merge configuration
merge_cfg_from_file(args["<input-cfg>"])
# config logger
logger.setLevel(eval('logging.' + cfg.LOG_LEVEL))
assert cfg.LOG_PATH, logger.error('Missing LOG_PATH!')
fhandler = logging.FileHandler(cfg.LOG_PATH, mode=cfg.LOG_MODE)
logger.addHandler(fhandler)
# print arguments
logger.info('=' * 80 + '\nCalled with arguments:')
for key in sorted(args.keys()):
logger.info('{:<20}= {}'.format(key.replace('--', ''), args[key]))
logger.info('=' * 80)
# reset logger format
fhandler.setFormatter(logging.Formatter(log_format))
import os
import unittest
import numpy as np
from scipy import signal
from mne import create_info, EpochsArray
import Offline.model as Model
import Offline.utils as util
from sklearn.linear_model import LogisticRegression
import matplotlib.pyplot as plt
from config import merge_cfg_from_file
merge_cfg_from_file('./test_stimulator.yml')
class TestOfflineUtils(unittest.TestCase):
def test_cut_epochs(self):
timestamp = [1, 1.2, 2, 3]
start, end, fs = -0.1, 0.5, 10
timestamp = list(map(lambda x: int(x * fs), timestamp))
data = np.zeros((1, fs * 4))
start_p = tuple(map(lambda x: x + int(start * fs), timestamp))
data[:, start_p] = 1
t = (start, end, fs)
epochs = util.cut_epochs(t, data, timestamp)
self.assertTrue(
epochs.shape == (4, 1, int(fs * (end - start))))
self.assertTrue(np.allclose(epochs[:, :, 0], 1))
def _init_():
'''
Training script for image-classification task on mxnet
Update: 2019-04-08
Author: @Northrend
Contributor:
Changelog:
2019/04/08 v5.8 fix DATA_TYPE config bug
2019/01/15 v5.7 support shufflenet v1 & v2
2019/01/09 v5.6 compatiable with mxnet v1.3.0 for dcnv2
2018/12/24 v5.5 support softmax label smoothing
2018/12/16 v5.4 support COSINE_DECAY learning rate
2018/12/04 v5.3 support loading gluoncv pretrained-model
2018/11/27 v5.2 fix random_resized_crop arg bug
2018/11/13 v5.1 support xception
2018/11/13 v5.0 add new dataloader preprocess supported by MXNet 1.3
2018/10/19 v4.3 support NV-DALI
2018/09/13 v4.2 fix svm bug
2018/09/12 v4.1 support directly image loading
2018/06/07 v4.0 code-refactoring
2018/03/12 v3.2 support freeze feature layer weights
2018/02/28 v3.1 support svm classifier
2018/02/11 v3.0 support customized finetune layer name
2018/02/10 v2.9 support resize dev data separately
2018/02/03 v2.8 support random resize scale
2018/01/29 v2.7 fix resume training job bugs
2017/12/27 v2.6 support se-inception-v4
2017/12/04 v2.5 support change shorter edge size
2017/11/21 v2.4 support input nomalization
2017/09/22 v2.3 support resume training job
2017/09/04 v2.2 support modify cardinality for resnext
2017/08/15 v2.1 support modify dropout argument
2017/08/09 v2.0 apply logging module
2017/08/08 v1.5 support scratch-training and more networks
2017/07/24 v1.4 support customized input shape
2017/07/13 v1.3 fix lr descend bug
2017/05/19 v1.2 support multi-metrics during training
support threshold and save-json options
modify script name from mxnet_train_finetune to mxnet_train
2017/05/10 v1.1 separate _init_ as a function
2017/05/08 v1.0 finetune task tested
resume training mode unsupported
Usage:
mxnet_train.py <input-cfg>
mxnet_train.py -v | --version
mxnet_train.py -h | --help
Arguments:
<input-cfg> path to customized config file
Options:
-h --help show this help screen
-v --version show current version
'''
# merge configuration
merge_cfg_from_file(args["<input-cfg>"])
# config logger
logger.setLevel(eval('logging.' + cfg.LOG_LEVEL))
assert cfg.LOG_PATH, logger.error('Missing LOG_PATH!')
fhandler = logging.FileHandler(cfg.LOG_PATH, mode=cfg.LOG_MODE)
logger.addHandler(fhandler)
# print arguments
logger.info('=' * 80 + '\nCalled with arguments:')
for key in sorted(args.keys()):
logger.info('{:<20}= {}'.format(key.replace('--', ''), args[key]))
logger.info('=' * 80)
# reset logger format
fhandler.setFormatter(logging.Formatter(log_format))
def online_main(screen_index, test, cfg_file=None, model_date=None):
if cfg_file is not None:
merge_cfg_from_file(cfg_file)
# q_result: controller -> screen (sending judgement result and start/quit signal)
# q_stim: screen -> controller (sending event order)
q_result = Queue()
q_stim = Queue()
# create stimulator object
stim_string = cfg.exp_config.train_string if not test else cfg.exp_config.test_string
kwargs = {
'q_stim': q_stim,
'q_result': q_result,
'screen_index': screen_index,
'stim_string': stim_string,
'amp': cfg.amp_info.amp,
'trigger_type': cfg.amp_info.trigger_type,
'stim_dir': cfg.exp_config.stim_dir
if not cfg.exp_config.bidir else None # None for bidir
}
print('Configuration finished. Start process.')
# start process
process = Process(target=Stimulator.run_exp, kwargs=kwargs)
process.start()
# main process
if test:
# testing mode
# create data_client object
n_channel = len(cfg.subj_info.montage)
if cfg.amp_info.amp == 'neuracle':
from Online import Neuracle
data_client = Neuracle.Neuracle(
n_channel=n_channel + 1, # +1 for trigger channel
samplerate=cfg.amp_info.samplerate)
else:
raise ValueError("Unexpected amplifier type")
controller = Controller.TestingController(q_stim=q_stim,
q_result=q_result,
dataclient=data_client,
model_date=model_date,
stim_string=stim_string)
else:
# training mode
controller = Controller.TrainingController(q_stim=q_stim,
q_result=q_result,
stim_string=stim_string)
# write exp config info to log file
controller.write_exp_log()
# waiting start signal
keyboard.wait('s')
# put starting signal into q_result
q_result.put(-2)
# set quit hotkey
keyboard.add_hotkey('q', quit_process)
while controller.char_cnt < len(
controller.stim_string) and not Controller.quit_flag:
controller.run()
time.sleep(0.05)
# close events file io
controller.close()
if test:
# writing down itr
print('accu: %.2f, average time: %.2f, itr: %.2f' % controller.itr())
# turn off data client thread
data_client.close()
# terminate process
if Controller.quit_flag:
process.terminate()
process.join()
# start train
for epoch_idx in range(0, cfg.SOLVER.EPOCHS):
logger.info('train epoch: {0}'.format(epoch_idx))
TrainEpoch(epoch_idx, train_loader, model, loss_funcs, optimizer,
lr_scheduler, device, logger)
# if not epoch_idx % 10:
# val_loader = MakeTrainLoader(os.path.join(cfg.DATA.ROOT_DIR, cfg.DATA.VAL_PATH), cfg.DATA.VAL_BATCH)
# Valid(val_loader, model, loss_funcs, device, logger)
def parse_args(args):
import argparse
parser = argparse.ArgumentParser(description='train')
parser.add_argument('--yaml_filepath',
type=str,
help='config filepath',
default='ConfigYaml/config.yaml')
args = parser.parse_args(args)
return args
if __name__ == '__main__':
args = parse_args(sys.argv[1:])
if os.path.exists(args.yaml_filepath):
merge_cfg_from_file(args.yaml_filepath, cfg)
else:
raise Exception('invalid config filepath: ', args.yaml_filepath)
Train(cfg)
def _init_():
'''
Training script for image-classification task on mxnet
Update: 2018-06-07 16:43:27
Author: @Northrend
Contributor:
Changelog:
2018/06/07 v4.0 code-refactoring
2018/03/12 v3.2 support freeze feature layer weights
2018/02/28 v3.1 support svm classifier
2018/02/11 v3.0 support customized finetune layer name
2018/02/10 v2.9 support resize dev data separately
2018/02/03 v2.8 support random resize scale
2018/01/29 v2.7 fix resume training job bugs
2017/12/27 v2.6 support se-inception-v4
2017/12/04 v2.5 support change shorter edge size
2017/11/21 v2.4 support input nomalization
2017/09/22 v2.3 support resume training job
2017/09/04 v2.2 support modify cardinality for resnext
2017/08/15 v2.1 support modify dropout argument
2017/08/09 v2.0 apply logging module
2017/08/08 v1.5 support scratch-training and more networks
2017/07/24 v1.4 support customized input shape
2017/07/13 v1.3 fix lr descend bug
2017/05/19 v1.2 support multi-metrics during training
support threshold and save-json options
modify script name from mxnet_train_finetune to mxnet_train
2017/05/10 v1.1 separate _init_ as a function
2017/05/08 v1.0 finetune task tested
resume training mode unsupported
Usage:
mxnet_train.py <input-cfg>
mxnet_train.py -v | --version
mxnet_train.py -h | --help
Arguments:
<input-cfg> path to customized config file
Options:
-h --help show this help screen
-v --version show current version
'''
# merge configuration
merge_cfg_from_file(args["<input-cfg>"])
# config logger
logger.setLevel(eval('logging.' + cfg.LOG_LEVEL))
assert cfg.LOG_PATH, logger.error('Missing LOG_PATH!')
fhandler = logging.FileHandler(cfg.LOG_PATH, mode=cfg.LOG_MODE)
logger.addHandler(fhandler)
# print arguments
logger.info('=' * 80 + '\nCalled with arguments:')
for key in sorted(args.keys()):
logger.info('{:<20}= {}'.format(key.replace('--', ''), args[key]))
logger.info('=' * 80)
# reset logger format
fhandler.setFormatter(logging.Formatter(log_format))
def main(args):
if args.cfg_file is not None:
merge_cfg_from_file(args.cfg_file)
# loading dataset
dataset = Dataset(subject=cfg.subj_info.subjname, date=args.date)
# print dataset information
print(dataset.data.shape)
print(len(dataset.timestamp))
print(dataset.events.shape)
print(dataset.ch_names)
# extract features
extractor = FeatExtractor(sfreq=cfg.amp_info.samplerate,
band_erp=cfg.subj_info.erp_band)
# select channels in montage
data = dataset.data[dataset.montage_indices]
ch_names = dataset.montage.copy()
print(data.shape)
model = Model(subject=cfg.subj_info.subjname)
feature = model.extract_feature(extractor, data, channel_selection=False)
X = model.raw2epoch(feature, dataset.timestamp, dataset.events)
# get target
if cfg.exp_config.bidir:
target = utils.get_label_bidir(cfg.exp_config.train_string,
cfg.exp_config.n_rep, dataset.events)
else:
target = utils.get_label(cfg.exp_config.train_string,
cfg.exp_config.n_rep)
y = target.flatten()
try:
plots_path = os.path.join(dataset.root_dir, 'plots')
os.mkdir(plots_path)
except FileExistsError:
pass
# split train/validate set and cross validating on train set
# randomly split 20%
train_ind, val_ind = utils.uniform_split(y, shuffle=True, random_state=42)
X_train = X[train_ind]
y_train = y[train_ind]
X_val = X[val_ind]
y_val = y[val_ind]
ind = utils.chan_select(X_train, y_train, cfg.off_config.k_best_channel)
print('Selected k best channel')
print(ind)
if args.p:
if cfg.subj_info.type == 'eeg':
# parameter search
C = np.logspace(-4, 2, 10)
n_components = range(1, (X_train.shape[1] // 2) + 1)
best_params, best_result = _grid_search(X_train,
y_train,
args,
ind,
C=C,
n_components=n_components)
else:
raise KeyError('Unsupported experiment type.')
for key in best_params:
print('Best %s: %.4f' % (key, best_params[key]))
for i in best_result:
print('%s: ' % i, best_result[i])
print('')
else:
if cfg.subj_info.type == 'eeg':
best_params = {'C': 1., 'n_components': 3}
else:
raise KeyError('Unsupported experiment type.')
best_result = _k_fold(X_train, y_train, args, ch_select=ind)
# train a model with selected parameters
model = _train(X_train, y_train, select=ind, **best_params)
y_pred = _prediction(X_val, y_val, model, args.n_avg)
if cfg.exp_config.bidir:
fig_roc, fig_cm, fig_pr, param_dict = utils.evaluate_multiclass(
y_pred, y_val, if_plot=True)
else:
fig_roc, fig_cm, fig_pr, param_dict = utils.evaluate_binary(
y_pred, y_val, if_plot=True)
print('Cross validation results:')
for i in best_result:
print('%s: ' % i, best_result[i])
print('')
print('Split out validation result:')
for i in param_dict:
print('%s: %.4f' % (i, param_dict[i]))
print('')
# save evaluation results
fig_roc.savefig(os.path.join(plots_path, 'roc.png'))
fig_cm.savefig(os.path.join(plots_path, 'cm.png'))
fig_pr.savefig(os.path.join(plots_path, 'pr.png'))
# dump training info
info = {
'average':
args.n_avg,
'k-fold':
args.k,
'selected channels': [
ch_names[i] +
' ERPs' if i < len(ch_names) else ch_names[i - len(ch_names)] +
' HG' for i in ind
],
}
info.update(best_params)
try:
os.mkdir(os.path.join(dataset.root_dir, 'logs'))
except FileExistsError:
pass
with open(
os.path.join(
dataset.root_dir, 'logs',
'log%s.txt' % datetime.now().strftime("%Y-%m-%d-%H-%M-%S")),
'w') as f:
# training parameters
f.write(json.dumps(info))
f.write('\n')
f.write('Cross validation result:\n')
f.write(json.dumps(best_result))
f.write('\n')
f.write('Split out validation result:\n')
f.write(json.dumps(param_dict))
f.write('\n')
f.write('\n')
f.write(json.dumps(cfg.subj_info))
f.write('\n')
f.write('\n')
f.write(json.dumps(cfg.off_config))
# training with whole dataset and save model
model = _train(X, y, select=ind, date=args.date, **best_params)
model.dump()
# show plots
plt.show()
