k = p.name
d[k] = p.get_value(borrow=False)
np.savez(fname, **d)
def load_parameters(epoch=None):
fname = "weights.npz" if epoch is None else "weights.%d.npz" % epoch
f = np.load(fname)
for p in parameters:
p.set_value(f[p.name])
def report(d):
import json
with open("learning_curve.txt", "a") as fh:
fh.write(json.dumps(d))
fh.write("\n")
print "epoch=%-6s -- trn_cost=%0.8f tst_cost=%0.8f" % (d['_epoch'],
d['trn'], d['tst'])
if __name__ == '__main__':
# generate some simple training data
import uuid
run_name = uuid.uuid4().hex[:8]
n_samples = 10000
x_train_val, y_train_val = datasets.sinewaves(n_steps, n=n_samples)
x_test_val, y_test_val = datasets.sinewaves(
n_steps, n=20) # (n_samples, n_steps, n_x)
train(x_train_val, y_train_val, x_test_val, y_test_val)
def train(x_train_val, y_train_val):
x_train = theano.shared(x_train_val)
y_train = theano.shared(y_train_val)
n_samples = x_train_val.shape[0]
# creating training function
learn_fn = make_learner(x_train, y_train, param_updates)
# start gradient descent w/ batch_size==1
max_epochs = 1000
lr_val = 0.5
for epoch in range(max_epochs):
costs = []
for idx in range(n_samples):
sample_cost = learn_fn(idx, lr_val)
costs.append(sample_cost[0])
print "epoch=%-6s -- cost=%0.8f" % (epoch, np.mean(costs))
if __name__ == '__main__':
# generate some simple training data
n_samples = 100
timesteps = 20
x_train_val, y_train_val = datasets.sinewaves(timesteps, n=n_samples)
train(x_train_val, y_train_val)
predictor = make_predictor()
for p in parameters:
k = p.name
d[k] = p.get_value(borrow=False)
np.savez("weights.npz", **d)
def load_parameters():
f = np.load("weights.npz")
for p in parameters:
p.set_value(f[p.name])
def report(d):
import json
with open("learning_curve.txt", "a") as fh:
fh.write(json.dumps(d))
fh.write("\n")
print "epoch=%-6s -- trn_cost=%0.8f tst_cost=%0.8f" % (d["_epoch"], d["trn"], d["tst"])
if __name__ == "__main__":
# generate some simple training data
import uuid
run_name = uuid.uuid4().hex[:8]
n_samples = 1
x_train_val, y_train_val = datasets.sinewaves(n_steps, n=n_samples)
x_test_val, y_test_val = datasets.sinewaves(n_steps, n=20)
train(x_train_val, y_train_val, x_test_val, y_test_val)