self.O = param((dim, self.outdim), lrmul=1.).glorotuniform()
def apply(self, inptensor):
emb = self.W(inptensor)
out = T.dot(emb, self.O)
out.output_as("out")
probs = Softmax()(out)
return probs
def run(epochs=1, dim=10, vocabsize=2000, lr=0.02, numbats=100):
lr *= numbats
data = np.arange(0, vocabsize).astype("int32")
ae = Dummy(indim=vocabsize, dim=dim)
ae = ae.train([data], data).adadelta(lr=lr).dlr_thresh().cross_entropy()\
.split_validate(5, random=True).cross_entropy().accuracy().autosave\
.train(numbats=numbats, epochs=epochs)
pdata = range(100)
pembs = ae.W.predict(pdata)
#print np.linalg.norm(pembs, axis=1)
pred = ae.predict(pdata)
print pred.shape
#print np.argmax(pred, axis=1)
#print err, verr
return pred
if __name__ == "__main__":
run(**argparsify(run))
probs = Softmax()(out)
return probs
def run(
epochs=1,
dim=10,
vocabsize=2000,
lr=0.02,
numbats=100
):
lr *= numbats
data = np.arange(0, vocabsize).astype("int32")
ae = Dummy(indim=vocabsize, dim=dim)
ae = ae.train([data], data).adadelta(lr=lr).dlr_thresh().cross_entropy()\
.split_validate(5, random=True).cross_entropy().accuracy().autosave\
.train(numbats=numbats, epochs=epochs)
pdata = range(100)
pembs = ae.W.predict(pdata)
#print np.linalg.norm(pembs, axis=1)
pred = ae.predict(pdata)
#print pred.shape
#print np.argmax(pred, axis=1)
#print err, verr
return pred
if __name__ == "__main__":
run(**argparsify(run))
def test_argparsify(self):
def testf(a=1, b="str"):
pass
self.assertEqual(argparsify(testf, test="-a 1"), {"a": 1})
def test_argparsify(self):
def testf(a=1, b="str"):
pass
self.assertEqual(argparsify(testf, test="-a 1"), {"a": 1})