def dycosine(query_vec, question_vec):
num = dy.transpose(query_vec) * question_vec
dem1 = dy.sqrt(dy.transpose(query_vec) * query_vec)
dem2 = dy.sqrt(dy.transpose(question_vec) * question_vec)
dem = dem1 * dem2
return dy.cdiv(num, dem)
def eval_dict_dataset(dataset, net, shortlist, proj, parsed):
ranks = []
num_batches = len(dataset)
if parsed:
dim = dataset[0][0].shape[0]
batch_size = 1
else:
dim, batch_size = dataset[0][0].shape
for batch_num, data in enumerate(dataset):
if parsed:
words, definitions, _ = data
else:
words, definitions = data
words = np.reshape(np.transpose(words), (batch_size, dim))
dy.renew_cg()
P = dy.parameter(proj)
if parsed:
outputs = net.do_parse_tree(definitions)
else:
outputs, _ = net(definitions)
outputs = P * outputs
normalised_outputs = outputs * dy.cdiv(dy.inputTensor([1]), dy.sqrt(dy.squared_norm(outputs)))
normalised_outputs = np.reshape(np.transpose(normalised_outputs.npvalue()), (batch_size, dim))
for output, word in zip(normalised_outputs, words):
target_similarity = np.dot(word, output)
similarities = np.dot(shortlist, output)
rank = (similarities > target_similarity).sum()
ranks.append(rank)
total = len(ranks)
accuracy10 = float(sum(int(r <= 10) for r in ranks))/total
accuracy100 = float(sum(int(r <= 100) for r in ranks))/total
return np.median(ranks), accuracy10, accuracy100
def l2_normalize(vector):
square_sum = dy.sqrt(
dy.bmax(
dy.sum_elems(dy.square(vector)),
np.finfo(float).eps * dy.ones((1))[0],
))
return dy.cdiv(vector, square_sum)
def transduce(self, src: ExpressionSequence) -> ExpressionSequence:
src = src.as_tensor()
src_height = src.dim()[0][0]
src_width = 1
batch_size = src.dim()[1]
W = dy.parameter(self.pW)
b = dy.parameter(self.pb)
src = dy.reshape(src, (src_height, src_width), batch_size=batch_size) # ((276, 80, 3), 1)
# convolution and pooling layers
l1 = (W*src)+b
output = dy.cdiv(l1,dy.sqrt(dy.squared_norm(l1)))
return ExpressionSequence(expr_tensor=output)
def eval(self, mini_batch):
words, pos_tags, chars, langs, signs, masks = mini_batch
h_out = self.rnn_mlp(mini_batch, False)[-1]
t_out = dy.transpose(
dy.reshape(h_out, (h_out.dim()[0][0], h_out.dim()[1])))
sims = []
for i in range(len(langs)):
for j in range(i + 1, len(langs)):
sims.append(dy.sqrt(dy.squared_distance(t_out[i], t_out[j])))
sim = dy.esum(sims)
sim.forward()
sim_value = sim.value() / len(sims)
dy.renew_cg()
return sim_value
def transduce(self, src):
src = src.as_tensor()
src_height = src.dim()[0][0]
src_width = src.dim()[0][1]
src_channels = 1
batch_size = src.dim()[1]
src = dy.reshape(src, (src_height, src_width, src_channels),
batch_size=batch_size) # ((276, 80, 3), 1)
# print(self.filters1)
# convolution and pooling layers
l1 = dy.rectify(
dy.conv2d(src,
dy.parameter(self.filters1),
stride=[self.stride[0], self.stride[0]],
is_valid=True))
pool1 = dy.maxpooling2d(l1, (1, 4), (1, 2), is_valid=True)
l2 = dy.rectify(
dy.conv2d(pool1,
dy.parameter(self.filters2),
stride=[self.stride[1], self.stride[1]],
is_valid=True))
pool2 = dy.maxpooling2d(l2, (1, 4), (1, 2), is_valid=True)
l3 = dy.rectify(
dy.conv2d(pool2,
dy.parameter(self.filters3),
stride=[self.stride[2], self.stride[2]],
is_valid=True))
pool3 = dy.kmax_pooling(l3, 1, d=1)
# print(pool3.dim())
output = dy.cdiv(pool3, dy.sqrt(dy.squared_norm(pool3)))
output = dy.reshape(output, (self.num_filters[2], ),
batch_size=batch_size)
# print("my dim: ", output.dim())
return ExpressionSequence(expr_tensor=output)
def cosine_similarity(a, b):
# FIXME do I really need to do this for scalar division? :(
norm = dy.cdiv(dy.inputTensor([1]), dy.sqrt(dy.squared_norm(a)))
return dy.dot_product(a, b) * norm
def score_constituent(self, constituent):
w_score = dy.parameter(self.w_score)
w_score = dy.cdiv(w_score, dy.sqrt(dy.squared_norm(w_score)))
h = dy.cdiv(constituent.h, dy.sqrt(dy.squared_norm(constituent.h)))
return dy.dot_product(w_score, h)*self.inv_temp
def l2_normalize(x):
square_sum = dynet.sqrt(dynet.bmax(dynet.sum_elems(dynet.square(x)), np.finfo(float).eps * dynet.ones((1))[0]))
return dynet.cdiv(x, square_sum)
def l2_normalize(x):
epsilon = np.finfo(float).eps * dy.ones(pred.dim()[0])
norm = dy.sqrt(dy.sum_elems(dy.square(x)))
sign = dy.cdiv(x, dy.bmax(dy.abs(x), epsilon))
return dy.cdiv(dy.cmult(sign, dy.bmax(dy.abs(x), epsilon)), dy.bmax(norm, epsilon[0]))