def __init__(self,
data,
n_units,
save=False,
learning_rate=1e-4,
name='',
optimizer='rmsprop',
num_gpus=1,
testdata=None,
**kwargs):
""" initialize data and conv net """
self.save = save
self.model = n.CNN(n_units,
save,
learning_rate,
optimizer,
name,
num_gpus=num_gpus,
**kwargs)
self.name = self.model.name
self.num_gpus = num_gpus
self.data = data
self.testdata = testdata
self.learning_rate = learning_rate
if optimizer not in ['descent', 'adam', 'adagrad', 'rmsprop']:
raise ValueError('optimizer must be \'descent\', \'adagrad\', '
'\'rmsprop\', or \'adam\'')
self.optimizer = optimizer
self.params = pp.NNParams('%s/resources/nupack/parameters/%s' %
(s.BASE_DIR, s.paramfile))
self.cache = {}
self.add_all_motifs()
self.update_params()
def main():
""" test existing model on given data """
args = parse_args()
data = d.KdDataset(args.filename.split(','), reporter=args.reporter,
lowmem=args.lowmem)
# get number of units if not provided
if args.n_units is None:
m = re.search('([0-9]+x[0-9]+)units', args.restore)
try:
args.n_units = m.group(1)
except:
raise ValueError('unable to parse number of units from model name')
# get model
if args.model == 'cnn':
print 'building cnn...'
model = n.CNN(args.n_units, optimizer=args.optimizer, lowmem=args.lowmem)
elif args.model == 'rnn':
print 'building rnn...'
model = n.RNN(args.n_units, optimizer=args.optimizer, lowmem=args.lowmem)
else:
raise ValueError('model must be "cnn" or "rnn"')
model.restore(args.restore)
data.get_predictions(model)
print 'rmse: %.4f' % data.get_rmse(sterr=False)
files = '+'.join([os.path.splitext(os.path.basename(f))[0]
for f in args.filename.split(',')])
f = '%s/%s_%s.txt' % (s.RESULTS_DIR, os.path.basename(args.restore), files)
data.to_file(f)
print 'results written to %s' % f
def main():
args = parse_args()
cnn = nn.CNN(args.layers, write=True)
cnn.restore(args.filename)
# cnn.get_activation_profile(1,0, iters=1000, learning_rate=0.1)
for i, layer in enumerate(args.layers):
for j in range(layer):
cnn.get_activation_profile(i + 1, j)
def train(data, layers, restore, epochs, save, keepprob=1., **kwargs):
""" train the CNN """
cnn = nn.CNN(layers, low_mem=True, save=save, **kwargs)
if restore is not None:
cnn.restore(restore)
print 'training model %s' % cnn.name
cnn.train(data, epochs=epochs, keepprob=keepprob)
print 'saved model to %s/%s' % (s.MODELS_DIR, cnn.name)
print 'writing results to %s/%s' % (s.RESULTS_DIR, cnn.name)
if save:
cnn.save_model()
results = get_results(data, cnn, test=False)
summarize_results(results, '%s_train.txt' % cnn.name)
return cnn
def __init__(self,
data,
n_units,
save=False,
n_cores=1,
learning_rate=None,
batch_norm=False,
name='',
low_mem=False,
optimizer='rmsprop',
num_gpus=1,
testdata=None,
**kwargs):
""" initialize data and conv net """
if learning_rate is None:
if isinstance(data, e.MeltDataset):
learning_rate = 1e-6
else:
learning_rate = 1e-4
self.save = save
self.model = n.CNN(n_units,
save,
learning_rate,
optimizer,
name,
low_mem=low_mem,
**kwargs)
self.name = self.model.name
self.num_gpus = num_gpus
self.data = data
self.testdata = testdata
self.learning_rate = learning_rate
if optimizer not in ['descent', 'adam', 'adagrad', 'rmsprop']:
raise ValueError('optimizer must be \'descent\', \'adagrad\', '
'\'rmsprop\', or \'adam\'')
self.optimizer = optimizer
self.low_mem = low_mem
self.params = pp.NNParams('%s/resources/nupack/parameters/%s' %
(s.BASE_DIR, s.paramfile))
self.n_cores = n_cores
self.cache = {}
self.update_params()
######------hyperParam-----#####
epoch = 40
imgSize = 160
Neuron = 320
hidden1 = 320
hidden2 = 160
lr = 0.000005
batchSize = 1
ptName = "models/nn1.pt"
lossPng = "lossImg/loss1.png"
accPng = "accImg/acc1.png"
######------モデルの設定(損失関数も変える)-----#####
# model = NN.Net_log_softmax(num=6,inputSize=imgSize,Neuron=Neuron)
model = NN.CNN(num=6, inputSize=imgSize, hidden1=hidden1, hidden2=hidden2)
optimizer = torch.optim.Adam(params=model.parameters(), lr=lr)
######------個別推論用ディレクトリ設定-----#####
andoDir = "../ReadOnlyDataSet2/Resources/test/ando"
higashiDir = "../ReadOnlyDataSet2/Resources/test/higashi"
kataokaDir = "../ReadOnlyDataSet2/Resources/test/kataoka"
kodamaDir = "../ReadOnlyDataSet2/Resources/test/kodama"
masudaDir = "../ReadOnlyDataSet2/Resources/test/masuda"
suetomoDir = "../ReadOnlyDataSet2/Resources/test/suetomo"
######------学習用ディレクトリ設定-----#####
dirImgTrainPath = "new-dataset"
dirImgTestPath = "../ReadOnlyDataSet2/Resources" + "/test/"
def main():
"""train a neural network with given data"""
args = parse_args()
if args.seed is not None:
np.random.seed(args.seed)
tf.set_random_seed(args.seed)
# read data from file(s)
print 'reading data...'
data = d.KdDataset(args.filename.split(','), reporter=args.reporter,
lowmem=args.lowmem)
print '\t%s points in training data' % data.n
print '\tstandard deviation of train data: %.2f' \
% np.nanstd(data.data.dG_measured)
if args.testfile is not None:
testdata = d.KdDataset(args.testfile.split(','), reporter=args.reporter,
lowmem=args.lowmem)
print '\t%s points in testing data' % testdata.n
print '\tstandard deviation of test data: %.2f' \
% np.nanstd(testdata.data.dG_measured)
else:
testdata = None
# construct neural network model object
if args.model == 'cnn':
print 'building cnn...'
files = '+'.join([os.path.splitext(os.path.basename(f))[0]
for f in args.filename.split(',')])
name = '%s_%s_%sunits_%.2e' % \
(files, args.model, args.n_units, args.learning_rate)
if args.keepprob != 1.:
name += '_keepprob%.2f' % args.keepprob
model = n.CNN(args.n_units, args.write, args.learning_rate,
args.optimizer, name, batch_norm=args.batch_norm,
lowmem=args.lowmem, num_gpus=args.num_gpus)
elif args.model == 'rnn':
print 'building rnn...'
files = '+'.join([os.path.splitext(os.path.basename(f))[0]
for f in args.filename.split(',')])
name = '%s_%s_%sunits_%.2e' % \
(files, args.model, args.n_units, args.learning_rate)
if args.keepprob != 1.:
name += '_keepprob%.2f' % args.keepprob
model = n.RNN(args.n_units, args.write, args.learning_rate,
args.optimizer, name, bidirectional=args.bidirectional,
lowmem=args.lowmem, num_gpus=args.num_gpus)
else:
raise ValueError('model must be "cnn" or "rnn"')
# restore model from file if specified
if args.restore is not None:
print 'restoring model parameters from file...'
model.restore(args.restore)
# train model
print 'training model %s...' % model.name
loss = model.train(data, args.epochs, args.batch_size, args.keepprob,
testdata, args.sterr)
print 'finished training model'
# save loss to file
if not os.path.exists(s.RESULTS_DIR):
os.makedirs(s.RESULTS_DIR)
np.savetxt('%s/%s.loss' % (s.RESULTS_DIR, model.name), loss.T, delimiter='\t',
header='train\ttest', comments='')
# run train and test predictions for final model
print 'getting final predictions...'
if testdata:
testdata.get_predictions(model)
testdata.to_file('%s/%s_test.txt' % (s.RESULTS_DIR, model.name))
testrmse = testdata.get_rmse()
print 'final test: %.4f' % testrmse
print 'final test rmse: %.4f' % testdata.get_rmse(sterr=False)
data.get_predictions(model)
# save and print results
data.to_file('%s/%s_train.txt' % (s.RESULTS_DIR, model.name))
print 'final: %.4f' % data.get_rmse()
print 'final rmse: %.4f' % data.get_rmse(sterr=False)
if args.save:
print 'saving model parameters to file %s...' % model.name
model.save('_%f' % testdata.get_rmse(sterr=False))