def train_one_model(task, mdl_dir, archi_fun, x_arr, y_arr, label_arr):
"""Function run by a single process."""
x, y, label = pp.getNumpy(x_arr, y_arr, label_arr)
n_data = x.shape[0]
for i in range(task[0], task[1]):
maski = i == label
datai = {'x': x[maski], 'y': y[maski]}
net = archi_fun(datai['x'].shape[0] / float(n_data))
config = genTrainConfig(network=net, batch_size=64, test_batch_size=1024,
outdir=mdl_dir, outname='model_%d_of_%s.pt' % (i, joinNumber(net)))
trainOne(config, datai)
def main():
net = GaoNetBN([2, 10, 2])
net.eval()
y = net.eval(np.random.random(2))
print(y)
# train a mapping with those variables
net.train()
N = 50
xin = np.random.random((N, 2))
y = np.c_[0.5 * xin[:, 0], 0.6 * xin[:, 1] + 0.4]
data_dict = {'x': xin, 'y': y}
config = genTrainConfig(network=net)
trainOne(config, data_dict, net=net, scalex=False, scaley=False)
net.printWeights()
def run_the_training(args):
"""Run the MoE training without using any clustering information but let it find it on its own."""
cfg, lbl = util.get_label_cfg_by_args(args)
uid = cfg['uniqueid']
print('We are playing with %s' % uid)
data = npload(cfg['file_path'], uid)
data_feed = {'x': data[cfg['x_name']], 'y': data[cfg['y_name']]}
dimx = data_feed['x'].shape[1]
dimy = data_feed['y'].shape[1]
n_model = args.k
# create the network
net = MoMNet([dimx, 100, n_model], [[dimx, int(np.ceil(300.0 / n_model)), dimy] for _ in range(n_model)])
net.argmax = False
config = genTrainConfig(outdir='models/%s/mom' % uid, outname='mom_model.pt', overwrite=False)
if args.eval:
mdl_path = os.path.join(config['outdir'], config['outname'])
eval_fun = momLoader(mdl_path, withclus=args.prob, argmax=False)
predy = eval_fun(data_feed['x'])
return {'x': data_feed['x'], 'y': data_feed['y'], 'predy': predy}
trainOne(config, data_feed, net=net)
def train_one_classifier(task, keys, uid, lbl_name, clas, label_data, x_arr):
"""A single process that trains one classifier.
Paramters
---------
task: list of two integers, specifying which tasks we use
keys: list of keys, work with task to locate cluster strategy
clas: list, integer, neural network architecture
label_data: dict, mapping from key to label
x_arr: sharedNumpy, the feature vectors
"""
x = pp.getNumpy(x_arr)
for i in range(task[0], task[1]):
key = keys[i]
n_cluster = int(key)
clas[-1] = n_cluster
model_directory = 'models/%s/%s/%s' % (uid, lbl_name, key)
if not os.path.exists(model_directory):
os.makedirs(model_directory)
label = label_data[key]
datai = {'x': x, 'label': label, 'n_label': n_cluster}
config = genTrainConfig(network=clas, batch_size=256, test_batch_size=2048,
outdir=model_directory, outname='classifier_%d_of_%s.pt' % (n_cluster, joinNumber(clas)))
trainOne(config, datai, is_reg_task=False) # switch to classifier