params=config,
)
save_yaml_file(yaml_str,
os.path.join(config.experiment_root, 'settings.yaml'))
with log_timing(log, 'training network'):
train.main_loop()
def get_default_config_path():
return os.path.join(os.path.dirname(__file__), 'train_convnet.cfg')
if __name__ == '__main__':
config = load_config(default_config=get_default_config_path(),
reset_logging=False)
if not config.get('only_extract_results', False):
train_convnet(config)
scan_for_best_performance(config.experiment_root, 'valid_y_misclass')
scan_for_best_performance(config.experiment_root,
'valid_ptrial_misclass_rate')
values = extract_results(config.experiment_root, mode='misclass')
print np.multiply(
100,
[
# 1 - values['test_y_misclass'],
# 1 - values['test_wseq_misclass_rate'],
train_mlp(trainset, testset, config);
# load model
with log_timing(log, 'loading MLP model from {}'.format(mlp_file)):
mlp = serial.load(mlp_file);
y_true = trainset.labels;
# y_pred = mlp.fprop(trainset.X);
X = mlp.get_input_space().make_theano_batch()
Y = mlp.fprop( X )
Y = T.argmax( Y, axis = 1 )
f = theano.function( [X], Y )
y_pred = f( trainset.X );
# Compute confusion matrix
print classification_report(y_true, y_pred);
print confusion_matrix(y_true, y_pred);
return mlp;
if __name__ == '__main__':
config = load_config(default_config='../train_sda.cfg', reset_logging=False);
train(config);
log.debug('job overrides: {}'.format(job.overrides))
job_config = merge_params(job_config, job.overrides)
job_config.experiment_root = os.path.join(config.output_root,
job_config.type, job.name)
log.debug('experiment root: {}'.format(job_config.experiment_root))
print job_config
# try:
if job_config.type == 'cnn':
train_convnet(job_config)
elif job_config.type == 'fftcnn':
train_convnet(job_config)
elif job_config.type == 'sda':
train_mlp(job_config)
else:
log.error('unsupported job type {}'.format(job_config.type))
# except:
# log.fatal("Unexpected error:", sys.exc_info());
if __name__ == '__main__':
default_config = os.path.join(os.path.dirname(__file__), 'batch.cfg')
config = load_config(default_config=default_config, reset_logging=False)
config = merge_params(Config(file(default_config)), config)
process_jobs(config)
params=params,
);
save_yaml_file(train_yaml_str, os.path.join(params.experiment_root, 'mlp_train.yaml'));
with log_timing(log, 'training MLP'):
train.main_loop();
log.info('done');
def get_default_config_path():
return os.path.join(os.path.dirname(__file__),'train_sda_mlp.cfg');
if __name__ == '__main__':
# config = load_config(default_config='../../train_sda.cfg', reset_logging=False);
config = load_config(default_config=get_default_config_path(), reset_logging=False);
hyper_params = {
};
params = merge_params(config, hyper_params);
if not config.get('only_extract_results', False):
train_mlp(params);
scan_for_best_performance(params.experiment_root, 'valid_y_misclass');
scan_for_best_performance(params.experiment_root, 'valid_ptrial_misclass_rate')
values = extract_results(config.experiment_root, mode='misclass');
print np.multiply(100, [
job_config.experiment_root = os.path.join(
config.output_root,
job_config.type,
job.name
);
log.debug('experiment root: {}'.format(job_config.experiment_root));
print job_config;
# try:
if job_config.type == 'cnn':
train_convnet(job_config);
elif job_config.type == 'fftcnn':
train_convnet(job_config);
elif job_config.type == 'sda':
train_mlp(job_config);
else:
log.error('unsupported job type {}'.format(job_config.type));
# except:
# log.fatal("Unexpected error:", sys.exc_info());
if __name__ == '__main__':
default_config = os.path.join(os.path.dirname(__file__), 'batch.cfg');
config = load_config(default_config=default_config, reset_logging=False);
config = merge_params(Config(file(default_config)), config);
process_jobs(config);
x**x;
log.info('done');
def run(params):
try:
log.debug('running {}'.format(params.experiment_root));
# dummy();
train_convnet(params);
# plot2(config.experiment_root);
except:
log.fatal("Unexpected error:", sys.exc_info());
if __name__ == '__main__':
config = load_config(default_config=
os.path.join(os.path.dirname(__file__),'train_fftconvnet.cfg'), reset_logging=True);
lr_values = config.get('lr_values', [0.001, 0.0033, 0.01, 0.00033, 0.033, 0.1]);
beat_patterns = config.get('beat_patterns', [10,20,30]);
bar_patterns = config.get('bar_patterns', [10,20,30]);
beat_pools = config.get('beat_pools', [1,3,5]);
bar_pools = config.get('bar_pools', [1,3,5]);
# with ThreadPoolExecutor(max_workers=config.num_processes) as executor:
for lr in lr_values:
for h1pat in bar_patterns:
for h1pool in bar_pools:
for h0pat in beat_patterns:
for h0pool in beat_pools:
# collect params
[11, 'all', 'all']
];
for col, group, subjects in groups:
hyper_params = {
'experiment_root' : os.path.join(config.experiment_root, str(group)),
'subjects' : subjects
# NOTE: layerX_content should still point to global sda/ folder
};
table_row[col:col+3] = run_experiment(config, hyper_params, random_seeds);
print table_row;
return table_row;
if __name__ == '__main__':
config = load_config(default_config=get_default_config_path());
config.save_output = True;
# FIXME: remove manual override
# config.only_extract_results = True;
#
# config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu.old/cnn/bigbatch/';
#
# config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu/fftcnn/exp05.2/'; ## <= BEST so far
# 72.2 & 72.8 & 73.7
row = run_experiments_for_table_row(config);
print config.experiment_root;
s = '';
report = analyze_reconstruction_error(datasets, minibatched_output_fn);
for line in report:
print '{:>6} : mean = {:.3f} std = {:.3f} max = {:.3f}'.format(*line);
# analyze_frames(dataset, output_fn);
if worst_frames:
analyze_worst_frames(datasets['train'], minibatched_output_fn, output_path=output_path);
# analyze_reconstruction(dataset, minibatched_output_fn, hop_size);
analyze_complex(datasets['test'], minibatched_output_fn, config.hop_size, output_path=output_path);
return report;
if __name__ == '__main__':
config = load_config(default_config=
# '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_400Hz_80-50-25-10/sda_400Hz_80-50-25-10.cfg'
# '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_400Hz_100-50-25-10/sda_400Hz_100-50-25-10.cfg'
'/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_100Hz_100-50-25-10/sda_100Hz_100-50-25-10.cfg'
);
# os.path.join(os.path.dirname(__file__), 'run', 'train_sda_mlp.cfg'), reset_logging=False);
# config.dataset_suffix = '_channels';
config.dataset_root = '/Users/sstober/work/datasets/Dan/eeg';
config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_400Hz_80-50-25-10/';
config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_400Hz_100-50-25-10/';
config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_100Hz_100-50-25-10/';
# analyze(config, subjects='all', worst_frames=True);
analyze(config, subjects=[1], worst_frames=False);
if worst_frames:
analyze_worst_frames(datasets['train'],
minibatched_output_fn,
output_path=output_path)
# analyze_reconstruction(dataset, minibatched_output_fn, hop_size);
analyze_complex(datasets['test'],
minibatched_output_fn,
config.hop_size,
output_path=output_path)
return report
if __name__ == '__main__':
config = load_config(
default_config=
# '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_400Hz_80-50-25-10/sda_400Hz_80-50-25-10.cfg'
# '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_400Hz_100-50-25-10/sda_400Hz_100-50-25-10.cfg'
'/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_100Hz_100-50-25-10/sda_100Hz_100-50-25-10.cfg'
)
# os.path.join(os.path.dirname(__file__), 'run', 'train_sda_mlp.cfg'), reset_logging=False);
# config.dataset_suffix = '_channels';
config.dataset_root = '/Users/sstober/work/datasets/Dan/eeg'
config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_400Hz_80-50-25-10/'
config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_400Hz_100-50-25-10/'
config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu/sda/batch_100Hz_100-50-25-10/'
# analyze(config, subjects='all', worst_frames=True);
analyze(config, subjects=[1], worst_frames=False)
accuracy[i] = 100 * (1 - extract_best_result(
params.experiment_root, mode='misclass', check_dataset='test')[0])
print results
print results.mean(axis=0)
print results.max(axis=1)
print accuracy
print accuracy.mean()
return results, accuracy
if __name__ == '__main__':
config = load_config(default_config=os.path.join(os.path.dirname(__file__),
'train_convnet.cfg'),
reset_logging=True)
# override settings
# config.experiment_root = os.path.join(config.experiment_root, 'cross-trial');
# config.subjects = None;
# config.subjects = 8; # subj9 0-indexed
# config.max_epochs = 5;
# FIXME: remove manual override
# config.only_extract_results = True;
# config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu/fftcnn/cross-test2/'
# config.experiment_root = '/Users/sstober/git/deepbeat/deepbeat/output/gpu/fftcnn/exp08.a-crosstrial.subj9/';
if config.get('full_cross_trial', False):
full_cross_trial_test(config)
def plot_all(root_path):
import fnmatch
# matches = []
for root, dirnames, filenames in os.walk(root_path):
for filename in fnmatch.filter(filenames, 'mlp.pkl'):
# matches.append(os.path.join(root, filename))
plot2(root)
if __name__ == '__main__':
config = load_config(
default_config=
# os.path.join(os.path.dirname(__file__), '..', 'run', 'train_convnet.cfg'), reset_logging=True);
os.path.join(os.path.dirname(__file__), '..', 'run',
'train_fftconvnet.cfg'),
reset_logging=True)
# os.path.join(os.path.dirname(__file__), '..', 'run', 'train_sda_mlp.cfg'), reset_logging=True);
# FIXME:
# root = '/Users/stober/git/deepbeat/deepbeat/output/gpu/sda/exp6.14all/';
# root = '/Users/stober/git/deepbeat/deepbeat/output/gpu/cnn/bigbatch/individual/';
# plot_batch(root);
# for i in xrange(12):
# root = '/Users/stober/git/deepbeat/deepbeat/output/gpu/cnn/bigbatch/cross-trial/pair'+str(i);
# plot(root+'/mlp.pkl', channels, root+'/plot.pdf');
# plot(root+'/mlp.pkl', less_channels, root+'/plot-less.pdf', get_color_variants(2));
# FIXME
mlp_file = config.get('mlp_file')
if not os.path.isfile(mlp_file):
train_mlp(trainset, testset, config)
# load model
with log_timing(log, 'loading MLP model from {}'.format(mlp_file)):
mlp = serial.load(mlp_file)
y_true = trainset.labels
# y_pred = mlp.fprop(trainset.X);
X = mlp.get_input_space().make_theano_batch()
Y = mlp.fprop(X)
Y = T.argmax(Y, axis=1)
f = theano.function([X], Y)
y_pred = f(trainset.X)
# Compute confusion matrix
print classification_report(y_true, y_pred)
print confusion_matrix(y_true, y_pred)
return mlp
if __name__ == '__main__':
config = load_config(default_config='../train_sda.cfg',
reset_logging=False)
train(config)
# trial_id += 1;
data = np.vstack(data);
labels = np.vstack(labels);
# trial_meta = np.vstack(trial_meta);
channel_meta = np.vstack(channel_meta);
# subject_meta = np.vstack(subject_meta);
log.debug('generated {} data points and {} labels '.format(data.shape, labels.shape));
# return data, labels, trial_meta, channel_meta;
return data, labels, channel_meta;
if __name__ == '__main__':
config = load_config(default_config='../train_sda.cfg');
DATA_ROOT = config.eeg.get('dataset_root', './');
SAMPLE_RATE = 400; # in Hz
TRIAL_LENGTH = 32; # in sec
TRIAL_LENGTH += 2; # add 2s after end of presentation
TRIAL_SAMPLE_LENGTH = SAMPLE_RATE * TRIAL_LENGTH;
log.info('using dataset at {}'.format(DATA_ROOT));
'''
Note from Dan:
All subjects should have channels 15, 16, 17 and 18 removed [...]
If you want to make them truly identical, you could remove channel 19 from
