def test(self, glove_vocab, glove_embed):
test1 = ["its", "a", "good", "product"]
test2 = ["This", "is", "bad", "indeed"]
print("\n\nTest 1: Its a good product", end=": ")
#test 1
self.node.c = autoTensor(torch.zeros(1, 25))
h = autoTensor(torch.zeros(1, 25))
for word in test1:
try:
index = glove_vocab.index(word)
sub_x = glove_embed[index].view(1, 25)
except:
sub_x = glove_embed[0].view(1, 25) * 0
x = autoTensor(sub_x)
h = self.node.forward(h, x, j)
z = F.sigmoid(self.out(h))
print(z)
print("\n\nTest 2: [This,is,bad,indeed]", end=": ")
#test 1
self.node.c = autoTensor(torch.zeros(1, 25))
h = autoTensor(torch.zeros(1, 25))
for word in test2:
try:
index = glove_vocab.index(word)
sub_x = glove_embed[index].view(1, 25)
except:
sub_x = glove_embed[0].view(1, 25) * 0
x = autoTensor(sub_x)
h = self.node.forward(h, x, j)
z = F.sigmoid(self.out(h))
print(z)
def forward(self,h,x,it):
it = str(it)
f = F.sigmoid(self.f(h,x))
i = F.sigmoid(self.i(h,x))
c_ = F.tanh(self.c_(h,x))
o = F.sigmoid(self.o(h,x))
self.c = f*self.c + i*c_
h.requires_grad=False
h = o*F.tanh(self.c)
f.name = "g_f"+it
i.name = "g_i"+it
c_.name = "g_c_"+it
o.name = "g_o"+it
return h
def test_der(self):
obj1 = make_autoTensor([[-3,3],[4,5]])
obj1.requires_grad = True
obj = sigmoid(obj1)
obj.backprop(make_autoTensor([[1,1],[1,1]]))
assert torch.sum(obj1.grad.value - make_autoTensor([[1,1],[1,1]]).value*obj.value *(1-obj.value)) == 0
def run(self, X, y, glove_vocab, glove_embed):
for epoch in range(115):
print("epoch: ", epoch)
epoch_loss = 0
acc = 0
for i in range(len(X)):
if i == 6:
break
self.node.c = autoTensor(torch.zeros(1, 25))
y_t = autoTensor(y[i])
h = autoTensor(torch.zeros(1, 25), requires_grad=True)
batch = X[i]
print("\t", i, len(batch[0]), end=" ")
t = time()
for j in range(len(batch[0])):
for k in range(len(batch)):
try:
word = str(batch[k][j])
index = glove_vocab.index(word)
sub_x = glove_embed[index].view(1, 25)
except:
sub_x = glove_embed[0].view(1, 25) * 0
if k == 0:
x = sub_x
else:
x = torch.cat((x, sub_x))
x = autoTensor(x)
h = self.node.forward(h, x, j)
print("t:", round(time() - t, 4), end="\t")
s = time()
z = F.sigmoid(self.out(h))
loss = Loss.BinaryCrossEntropy(z, y_t)
print(loss, (get_accuracy_value(z, y_t), loss.value.size()[0]),
end=" ")
#loss.trace_backprop()
#loss.reset_count()
loss.backward()
op = Optimizer("sgd", loss, 0.00005)
op.step()
gc.collect()
print(round(time() - s, 4))
epoch_loss += loss.single()
acc += get_accuracy_value(z, y_t)
print("\nepoch summary: loss: ", epoch_loss / 5, " ACC: ",
acc / 316)
def forward(self, X, y, glove_vocab, glove_embed):
for i, sentence in enumerate(X):
self.node.c = autoTensor(torch.zeros(1, 25))
y_t = autoTensor(y[i])
h = autoTensor(torch.zeros(1, 25), requires_grad=True)
print(i, end=" ")
for word in sentence:
try:
i = glove_vocab.index(word)
x = autoTensor(glove_embed[i].view(1, 25))
except:
x = autoTensor(glove_embed[0].view(1, 25) * 0)
h = self.node.forward(h, x)
s = time()
z = F.sigmoid(self.out(h))
loss = Loss.BinaryCrossEntropy(z, y_t)
print(loss, z.value.item(), y_t.value.item())
loss.backward()
op = Optimizer("sgd", loss, 0.0001)
op.step()
gc.collect()
print(time() - s, end=" ")
Y[i, y[i] - 1] = 1
y = Y.astype(np.float64)
X = autoTensor(X)
y = autoTensor(y)
initer = Initializer("xavier")
layer1 = Linear(400,25,initializer=initer)
layer2 = Linear(25,25,initializer=initer)
layer3 = Linear(25,10,initializer=initer)
lr = 0.0001
for x in range(1000):
l1 = layer1(X)
l2 = F.relu(layer2(l1))
l3 = F.sigmoid(layer3(l2))
loss = Loss.SquareError(l3,y)
loss.backward()
SGD = Optimizer("gdmomentum",loss,lr,beta=0.2)
SGD.step()
if x%50 == 0:
print(x,loss,get_accuracy_value(l3,y))
loss.grad_sweep()
def test_init(self):
obj1 = make_autoTensor([[-3,3],[4,5]])
obj1.requires_grad = True
obj = sigmoid(obj1)
assert obj.channels[0].autoVariable == obj1