updates.update(layer.updates(grads))
train = theano.function(
[sx, sy], [cost, error], updates=updates)
# --- make test function
y_pred = tt.argmax(self._propup(sx, testsize, noise=False), axis=1)
error = tt.mean(tt.neq(y_pred, sy))
test = theano.function([sx, sy], error)
return train, test
################################################################################
if __name__ == '__main__':
[train_x, train_y], [test_x, test_y] = dvs_loader.load_dataset()
assert train_x.shape[1] == train_x.shape[2]
train_x = train_x[:, None, :, :].astype(dtype)
train_y = train_y.astype(np.int32)
test_x = test_x[:, None, :, :].astype(dtype)
test_y = test_y.astype(np.int32)
batch_size = 100
n = floor_multiple(len(train_x), batch_size)
train_x, train_y = train_x[:n], train_y[:n]
batch_x = train_x.reshape(-1, batch_size, *train_x.shape[1:])
batch_y = train_y.reshape(-1, batch_size)
test_size = 2000
n = floor_multiple(len(test_x), test_size)
FullLayer(outputs=2, initW=0.01, initB=0.0, wc=0.01),
]
convfile = "convnet_2015-07-15_20-56-16.npz"
convdata = np.load(convfile)
net = Convnet((1,) + shape, layers)
propup = net.get_propup()
for i, [w, b] in enumerate(zip(convdata["weights"], convdata["biases"])):
net.layers[i].weights.set_value(w)
net.layers[i].biases.set_value(b)
if 0:
# test
_, [test_x, test_y] = dvs_loader.load_dataset()
test_x, test_y = test_x[:1000], test_y[:1000]
test_x = test_x[:, None, :, :].astype(dtype)
test_y = test_y.astype(np.int32)
_, test = net.get_train()
y = test(test_x, test_y)
print y.mean()
y = propup(test_x)
print (test_y != y).mean()
sys.exit(0)
# load data
