def on_calc_additional_loss(self, translator_loss):
if not self.learn_segmentation or self.segment_decisions is None:
return None
reward = -translator_loss["mle"]
if not self.log_reward:
reward = dy.exp(reward)
reward = dy.nobackprop(reward)
# Make sure that reward is not scalar, but rather based on the each batch item
assert reward.dim()[1] == len(self.src_sent)
# Mask
enc_mask = self.enc_mask.get_active_one_mask().transpose() if self.enc_mask is not None else None
# Compose the lose
ret = LossBuilder()
## Length prior
alpha = self.length_prior_alpha.value() if self.length_prior_alpha is not None else 0
if alpha > 0:
reward += self.segment_length_prior * alpha
# reward z-score normalization
if self.z_normalization:
reward = dy.cdiv(reward-dy.mean_batches(reward), dy.std_batches(reward) + EPS)
## Baseline Loss
if self.use_baseline:
baseline_loss = []
for i, baseline in enumerate(self.bs):
loss = dy.squared_distance(reward, baseline)
if enc_mask is not None:
loss = dy.cmult(dy.inputTensor(enc_mask[i], batched=True), loss)
baseline_loss.append(loss)
ret.add_loss("Baseline", dy.esum(baseline_loss))
if self.print_sample:
print(dy.exp(self.segment_logsoftmaxes[i]).npvalue().transpose()[0])
## Reinforce Loss
lmbd = self.lmbd.value()
if lmbd > 0.0:
reinforce_loss = []
# Calculating the loss of the baseline and reinforce
for i in range(len(self.segment_decisions)):
ll = dy.pick_batch(self.segment_logsoftmaxes[i], self.segment_decisions[i])
if self.use_baseline:
r_i = reward - dy.nobackprop(self.bs[i])
else:
r_i = reward
if enc_mask is not None:
ll = dy.cmult(dy.inputTensor(enc_mask[i], batched=True), ll)
reinforce_loss.append(r_i * -ll)
loss = dy.esum(reinforce_loss) * lmbd
ret.add_loss("Reinforce", loss)
if self.confidence_penalty:
ls_loss = self.confidence_penalty(self.segment_logsoftmaxes, enc_mask)
ret.add_loss("Confidence Penalty", ls_loss)
# Total Loss
return ret
def reparameterize(mu, logvar):
# Get z by reparameterization.
d = mu.dim()[0][0]
eps = dy.random_normal(d)
std = dy.exp(logvar * 0.5)
return mu + dy.cmult(std, eps)
def reparameterize(self, mu, logvar):
if self.training:
std = dy.exp(logvar * 0.5)
eps = dy.random_normal(dim=std.dim()[0], mean=0.0, stddev=1.0)
return dy.cmult(eps, std) + mu
else:
return mu
def log_sum_exp(scores, n_tags):
npval = scores.npvalue()
argmax_score = np.argmax(npval)
max_score_expr = dy.pick(scores, argmax_score)
max_score_expr_broadcast = dy.concatenate([max_score_expr] * n_tags)
return max_score_expr + dy.log(
dy.sum_cols(dy.transpose(dy.exp(scores - max_score_expr_broadcast))))
def reparameterize(mu, logvar):
# Get z by reparameterization.
d = mu.dim()[0][0]
eps = dy.random_normal(d)
std = dy.exp(logvar * 0.5)
return mu + dy.cmult(std, eps)
def test_item(model, sentence):
seq = [
model.wlookup[int(model.w2i.get(entry, 0))]
for entry in sentence.preprocessed_sentence
]
if len(seq) > 0:
encoded_sequence = encode_sequence(model, seq, model.sentence_rnn)
global_max = max_pooling(encoded_sequence)
global_min = average_pooling(encoded_sequence)
if len(encoded_sequence) > 0:
att_mlp_outputs = []
for e in encoded_sequence:
mlp_out = (model.attention_w * e) + model.attention_b
att_mlp_outputs.append(mlp_out)
lst = []
for o in att_mlp_outputs:
lst.append(dy.exp(dy.sum_elems(dy.cmult(o,
model.att_context))))
sum_all = dy.esum(lst)
probs = [dy.cdiv(e, sum_all) for e in lst]
att_context = dy.esum(
[dy.cmult(p, h) for p, h in zip(probs, encoded_sequence)])
context = dy.concatenate([att_context, global_max, global_min])
y_pred = dy.logistic((model.mlp_w * context) + model.mlp_b)
sentence.prediction_result = y_pred.scalar_value()
dy.renew_cg()
return sentence.prediction_result
return 0
def softmax(x):
"""
Compute the softmax function in tensorflow.
You might find the tensorflow functions tf.exp, tf.reduce_max,
tf.reduce_sum, tf.expand_dims useful. (Many solutions are possible, so you may
not need to use all of these functions). Recall also that many common
tensorflow operations are sugared (e.g. x * y does a tensor multiplication
if x and y are both tensors). Make sure to implement the numerical stability
fixes as in the previous homework!
Args:
x: tf.Tensor with shape (n_samples, n_features). Note feature vectors are
represented by row-vectors. (For simplicity, no need to handle 1-d
input as in the previous homework)
Returns:
out: tf.Tensor with shape (n_sample, n_features). You need to construct this
tensor in this problem.
"""
### YOUR CODE HERE
x_max = dy.max_dim(x, 1)
x_sub = dy.colwise_add(x, -x_max)
x_exp = dy.exp(x_sub)
sum_exp = dy.colwise_add(dy.zeroes(x.dim()[0]), dy.sum_cols(x_exp))
out = dy.cdiv(x_exp, sum_exp)
### END YOUR CODE
return out
def span_parser(self, sentence, is_train, elmo_embeddings, cur_word_index, gold=None):
if gold is not None:
assert isinstance(gold, ParseNode)
lstm_outputs = self._featurize_sentence(sentence, is_train=is_train,
elmo_embeddings=elmo_embeddings,
cur_word_index=cur_word_index)
encodings = []
span_to_index = {}
for start in range(0, len(sentence)):
for end in range(start + 1, len(sentence) + 1):
span_to_index[(start, end)] = len(encodings)
encodings.append(self._get_span_encoding(start, end, lstm_outputs))
label_log_probabilities = self._encodings_to_label_log_probabilities(encodings)
total_loss = dy.zeros(1)
if is_train:
for start in range(0, len(sentence)):
for end in range(start + 1, len(sentence) + 1):
gold_label = gold.oracle_label(start, end)
gold_label_index = self.label_vocab.index(gold_label)
index = span_to_index[(start, end)]
total_loss -= label_log_probabilities[gold_label_index][index]
return None, total_loss
else:
label_log_probabilities_np = label_log_probabilities.npvalue()
tree, additional_info = optimal_parser(label_log_probabilities_np,
span_to_index,
sentence,
self.empty_label_index,
self.label_vocab,
gold)
return tree, additional_info, dy.exp(label_log_probabilities).npvalue()
def train(self, mini_batch, num_train, k):
words, pos_tags, chars, langs, signs, masks = mini_batch
# Getting the last hidden layer from BiLSTM.
rnn_out = self.rnn_mlp(mini_batch, True)
h_out = rnn_out[-1]
t_out_d = dy.reshape(h_out, (h_out.dim()[0][0], h_out.dim()[1]))
t_out = dy.transpose(t_out_d)
# Calculating the kq values for NCE.
kq = dy.scalarInput(float(k) / num_train)
lkq = dy.log(kq)
loss_values = []
for i in range(len(langs)):
for j in range(i + 1, len(langs)):
if (langs[i] != langs[j]) and (signs[i] == 1 or signs[j] == 1):
lu = -dy.squared_distance(t_out[i], t_out[j])
denom = dy.log(dy.exp(lu) + kq)
if signs[i] == signs[j]: # both one
nom = lu
else:
nom = lkq
loss_values.append(denom - nom)
err_value = 0
if len(loss_values) > 0:
err = dy.esum(loss_values) / len(loss_values)
err.forward()
err_value = err.value()
err.backward()
self.trainer.update()
dy.renew_cg()
return err_value
def _policy_shape_probs(self,
prob_dist):
# TODO: this is specific to Alchemy
num_actions = len(self.output_action_vocabulary) - 1
num_locations = len(self.output_location_vocabulary) - 1
num_arguments = len(self.output_argument_vocabulary) - 1
new_probdist = dy.zeros(prob_dist.dim()[0])
zeroes = numpy.zeros(num_locations * num_arguments)
ones = numpy.ones(num_locations * num_arguments)
eos_prob = prob_dist[self._all_output_vocabulary.lookup_index((EOS, NO_ARG, NO_ARG))]
action_idx = 0
for action in self.output_action_vocabulary:
masks = numpy.concatenate(
(numpy.repeat(zeroes, action_idx),
ones,
numpy.repeat(zeroes, num_actions - action_idx - 1)))
actions_masks = dy.reshape(dy.inputTensor(masks),
(num_actions * num_locations * num_arguments, 1))
if action == EOS:
new_probdist += dy.cmult(actions_masks, prob_dist) / 2.
elif action == "push":
new_probdist += dy.cmult(actions_masks, prob_dist) + eos_prob / (2. * 56.)
elif action == "pop":
new_probdist += dy.cmult(actions_masks, prob_dist)
if self.args.syntax_restricted:
return dy.exp(dy.log_softmax(dy.cmult(new_probdist, prob_dist),
restrict = self._valid_action_indices))
else:
return dy.softmax(dy.cmult(new_probdist, prob_dist))
def intra_sent_attend(self, vecs):
numVecs = len(vecs)
fVecs = [dt.tanh(self.SelIntraFW * v) for v in vecs]
expE = []
for i, fq in enumerate(fVecs):
row = []
for j, fc in enumerate(fVecs):
row.append(
dt.exp(
dt.dot_product(fq, fc) +
self.SelIntraBias[i - j +
int(config.d["DIST_BIAS_DIM"] / 2)]))
expE.append(row)
invSumExpE = []
for i in xrange(numVecs):
invSumExpE.append(dt.pow(dt.esum(expE[i]), dt.scalarInput(-1)))
alpha = []
for i in xrange(numVecs):
s = dt.esum([vecs[j] * expE[i][j] for j in xrange(numVecs)])
alpha.append(s * invSumExpE[i])
return [
dt.tanh(self.SelIntraHW * dt.concatenate([v, a]))
for v, a in zip(vecs, alpha)
]
def reparameterize(self, mu, logvar):
if self.training:
std = dy.exp(logvar * 0.5)
eps = dy.random_normal(dim=std.dim()[0], mean=0.0, stddev=1.0)
return dy.cmult(eps, std) + mu
else:
return mu
def calc_loss(sent):
dy.renew_cg()
# Transduce all batch elements with an LSTM
src = sent[0]
trg = sent[1]
# initialize the LSTM
init_state_src = LSTM_SRC_BUILDER.initial_state()
# get the output of the first LSTM
src_output = init_state_src.add_inputs([LOOKUP_SRC[x]
for x in src])[-1].output()
# Now compute mean and standard deviation of source hidden state.
W_mean = dy.parameter(W_mean_p)
V_mean = dy.parameter(V_mean_p)
b_mean = dy.parameter(b_mean_p)
W_var = dy.parameter(W_var_p)
V_var = dy.parameter(V_var_p)
b_var = dy.parameter(b_var_p)
# The mean vector from the encoder.
mu = mlp(src_output, W_mean, V_mean, b_mean)
# This is the diagonal vector of the log co-variance matrix from the encoder
# (regard this as log variance is easier for furture implementation)
log_var = mlp(src_output, W_var, V_var, b_var)
# Compute KL[N(u(x), sigma(x)) || N(0, I)]
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
kl_loss = -0.5 * dy.sum_elems(1 + log_var -
dy.pow(mu, dy.inputVector([2])) -
dy.exp(log_var))
z = reparameterize(mu, log_var)
# now step through the output sentence
all_losses = []
current_state = LSTM_TRG_BUILDER.initial_state().set_s([z, dy.tanh(z)])
prev_word = trg[0]
W_sm = dy.parameter(W_sm_p)
b_sm = dy.parameter(b_sm_p)
for next_word in trg[1:]:
# feed the current state into the
current_state = current_state.add_input(LOOKUP_TRG[prev_word])
output_embedding = current_state.output()
s = dy.affine_transform([b_sm, W_sm, output_embedding])
all_losses.append(dy.pickneglogsoftmax(s, next_word))
prev_word = next_word
softmax_loss = dy.esum(all_losses)
return kl_loss, softmax_loss
def decomp_attend(self, vecsA, vecsB):
# Fq^T Fc -> need to expedite using native matrix/tensor multiplication
Fq = vecsA # the original word vector, not yet passing a NN as in Eq.1, # need a function F
Fc = vecsB # need a function F
expE = []
for fq in Fq:
row = []
for fc in Fc:
row.append(dt.exp(dt.dot_product(fq, fc)))
expE.append(row)
#print ("debug: expE", expE[0][0].value())
invSumExpEi = []
for i in xrange(len(Fq)):
invSumExpEi.append(dt.pow(dt.esum(expE[i]), dt.scalarInput(-1)))
invSumExpEj = []
for j in xrange(len(Fc)):
invSumExpEj.append(
dt.pow(dt.esum([expE[i][j] for i in xrange(len(Fq))]),
dt.scalarInput(-1)))
beta = []
for i in xrange(len(Fq)):
s = dt.esum([Fc[j] * expE[i][j] for j in xrange(len(Fc))])
beta.append(s * invSumExpEi[i])
#print("debug: beta", beta[0].value())
alpha = []
for j in xrange(len(Fc)):
s = dt.esum([Fc[j] * expE[i][j] for i in xrange(len(Fq))])
alpha.append(s * invSumExpEj[j])
#print("debug: alpha", alpha[0].value())
# Compare
v1i = [
dt.logistic(dt.concatenate([Fq[i], beta[i]]))
for i in xrange(len(Fq))
] # need a function G
v2j = [
dt.logistic(dt.concatenate([Fc[j], alpha[j]]))
for j in xrange(len(Fc))
] # need a function G
#print ("debug: v1i", v1i[0].value())
#print ("debug: v2j", v2j[0].value())
# Aggregate
v1 = dt.esum(v1i)
v2 = dt.esum(v2j)
#print ("debug: v1.value()", v1.value())
#print ("debug: v2.value()", v2.value())
#colScore = dt.logistic(dt.dot_product(self.SelHW, dt.concatenate([v1,v2])))
return dt.dot_product(v1, v2)
def log_sum_exp(scores):
npval = scores.npvalue()
argmax_score = np.argmax(npval)
max_score_expr = dy.pick(scores, argmax_score)
max_score_expr_broadcast = dy.concatenate([max_score_expr] *
self.dim_output)
return max_score_expr + dy.log(
dy.sum_elems(
dy.transpose(dy.exp(scores - max_score_expr_broadcast))))
def log_sum_exp(tag_score_arr):
argmax = np.argmax(tag_score_arr.value())
max_score = tag_score_arr[argmax]
score = max_score
max_arr = dynet.concatenate(
[max_score for i in range(len(self.pos) + 2)])
score += dynet.log(
dynet.sum_dim(dynet.exp(tag_score_arr - max_arr), [0]))
return score
def loss_function(recon_x, x, mu, logvar):
BCE = dy.binary_log_loss(recon_x, x) # equiv to torch.nn.functional.binary_cross_entropy(?,?, size_average=False)
# see Appendix B from VAE paper:
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
# https://arxiv.org/abs/1312.6114
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
KLD = -0.5 * dy.sum_elems(1 + logvar - dy.pow(mu, dy.scalarInput(2)) - dy.exp(logvar))
return BCE + KLD
def log_sum_exp(scores):
npval = scores.npvalue()
argmax_score = np.argmax(npval)
max_score_expr = dy.pick(scores, argmax_score)
max_score_expr_broadcast = dy.concatenate([max_score_expr] *
self.tagset_size)
return max_score_expr + dy.log(
dy.sum_dim(
dy.transpose(dy.exp(scores - max_score_expr_broadcast)),
[1]))
def calc_loss(self, policy):
if self.weight < 1e-8:
return None
neg_entropy = []
for i, ll in enumerate(policy):
if self.valid_pos is not None:
ll = dy.pick_batch_elems(ll, self.valid_pos[i])
loss = dy.sum_batches(dy.sum_elems(dy.cmult(dy.exp(ll), ll)))
neg_entropy.append(dy.sum_batches(loss))
return self.weight * dy.esum(neg_entropy)
def log_sum_exp(scores):
npval = scores.npvalue()
argmax_score = np.argmax(npval)
max_score_expr = dynet.pick(scores, argmax_score)
max_score_expr_broadcast = dynet.concatenate([max_score_expr] *
(self.n_tags + 2))
return max_score_expr + dynet.log(
dynet.sum_cols(
dynet.transpose(
dynet.exp(scores - max_score_expr_broadcast))))
def calc_loss(sent):
dy.renew_cg()
# Transduce all batch elements with an LSTM
src = sent[0]
trg = sent[1]
# initialize the LSTM
init_state_src = LSTM_SRC_BUILDER.initial_state()
# get the output of the first LSTM
src_output = init_state_src.add_inputs([LOOKUP_SRC[x] for x in src])[-1].output()
# Now compute mean and standard deviation of source hidden state.
W_mean = dy.parameter(W_mean_p)
V_mean = dy.parameter(V_mean_p)
b_mean = dy.parameter(b_mean_p)
W_var = dy.parameter(W_var_p)
V_var = dy.parameter(V_var_p)
b_var = dy.parameter(b_var_p)
# The mean vector from the encoder.
mu = mlp(src_output, W_mean, V_mean, b_mean)
# This is the diagonal vector of the log co-variance matrix from the encoder
# (regard this as log variance is easier for furture implementation)
log_var = mlp(src_output, W_var, V_var, b_var)
# Compute KL[N(u(x), sigma(x)) || N(0, I)]
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
kl_loss = -0.5 * dy.sum_elems(1 + log_var - dy.pow(mu, dy.inputVector([2])) - dy.exp(log_var))
z = reparameterize(mu, log_var)
# now step through the output sentence
all_losses = []
current_state = LSTM_TRG_BUILDER.initial_state().set_s([z, dy.tanh(z)])
prev_word = trg[0]
W_sm = dy.parameter(W_sm_p)
b_sm = dy.parameter(b_sm_p)
for next_word in trg[1:]:
# feed the current state into the
current_state = current_state.add_input(LOOKUP_TRG[prev_word])
output_embedding = current_state.output()
s = dy.affine_transform([b_sm, W_sm, output_embedding])
all_losses.append(dy.pickneglogsoftmax(s, next_word))
prev_word = next_word
softmax_loss = dy.esum(all_losses)
return kl_loss, softmax_loss
def softmax(x):
### YOUR CODE HERE
x_max = dy.max_dim(x, 1)
x_sub = dy.colwise_add(x, -x_max)
x_exp = dy.exp(x_sub)
x_sum = dy.sum_cols(x_exp)
x_tmp = dy.zeroes(x.dim()[0])
x_tmp = dy.colwise_add(x_tmp, x_sum)
out = dy.cdiv(x_exp, x_tmp)
### END YOUR CODE
return out
def selu(x):
""" :type x: dn.Expression
:rtype: dn.Expression """
positive = dn.rectify(x)
positive_indicator = dn.rectify(dn.cdiv(positive, positive + epsilon))
negative = -dn.rectify(-x)
exp_negative = dn.exp(negative) - positive_indicator
exp_negative_minus_alpha = exp_negative * alpha - alpha + positive_indicator * alpha
# x>0: x=x * scale; x<0: x = (alpha * exp(x) - alpha) * scale
ret = (positive + exp_negative_minus_alpha) * scale
return ret
def marginals(self,
inside_chart,
outside_chart,
lognormalizer,
semiring=LogProbSemiring):
marginals = {}
for node in inside_chart:
marginals[node] = dy.exp(
semiring.division(
semiring.product(inside_chart[node], outside_chart[node]),
lognormalizer))
return marginals
def loss_function(recon_x, x, mu, logvar):
BCE = dy.binary_log_loss(
recon_x, x
) # equiv to torch.nn.functional.binary_cross_entropy(?,?, size_average=False)
# see Appendix B from VAE paper:
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
# https://arxiv.org/abs/1312.6114
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
KLD = -0.5 * dy.sum_elems(1 + logvar - dy.pow(mu, dy.scalarInput(2)) -
dy.exp(logvar))
return BCE + KLD
def __call__(self, logsoftmaxes, mask):
strength = self.strength.value()
if strength == 0:
return 0
neg_entropy = []
for i, logsoftmax in enumerate(logsoftmaxes):
loss = dy.cmult(dy.exp(logsoftmax), logsoftmax)
if mask is not None:
loss = dy.cmult(dy.inputTensor(mask[i], batched=True), loss)
neg_entropy.append(loss)
return strength * dy.sum_elems(dy.esum(neg_entropy))
def calc_reinforce_loss(words, tags, delta):
dy.renew_cg()
# Transduce all batch elements with an LSTM
word_reps = LSTM.transduce([LOOKUP[x] for x in words])
# Softmax scores
W = dy.parameter(W_sm)
b = dy.parameter(b_sm)
#calculate the probability distribution
scores = [dy.affine_transform([b, W, x]) for x in word_reps]
losses = [
dy.pickneglogsoftmax(score, tag) for score, tag in zip(scores, tags)
]
probs = [-dy.exp(loss).as_array() for loss in losses]
#then take samples from the probability distribution
samples = [np.random.choice(range(len(x)), p=x) for x in probs]
#calculate accuracy=reward
correct = [sample == tag for sample, tag in zip(samples, tags)]
r_i = float(sum(correct)) / len(correct)
r = dy.constant((1), r_i)
# Reward baseline for each word
W_bl = dy.parameter(W_bl_p)
b_bl = dy.parameter(b_bl_p)
r_b = [
dy.affine_transform([b_bl, W_bl, dy.nobackprop(x)]) for x in word_reps
]
#we need to take the value in order to break the computation graph
#as the reward portion is trained seperatley and not backpropogated through during the overall score
rewards_over_baseline = [(r - dy.nobackprop(x)) for x in r_b]
#the scores for training the baseline
baseline_scores = [dy.square(r - x) for x in r_b]
#then calculate the reinforce scores using reinforce
reinforce_scores = [
r_s * score for r_s, score in zip(rewards_over_baseline, scores)
]
#we want the first len(sent)-delta scores from xent then delta scores from reinforce
#for mixer
if len(scores) > delta:
mixer_scores = scores[:len(scores) - delta] + reinforce_scores[delta -
1:]
else:
mixer_scores = reinforce_scores
return dy.esum(mixer_scores), dy.esum(baseline_scores)
def log_sum_exp_dim_0(x):
# numerically stable log_sum_exp
dims = x.dim()
max_score = dy.max_dim(x, 0) # (dim_1, batch_size)
if len(dims[0]) == 1:
max_score_extend = max_score
else:
max_score_reshape = dy.reshape(max_score, (1, dims[0][1]), batch_size=dims[1])
max_score_extend = dy.concatenate([max_score_reshape] * dims[0][0])
x = x - max_score_extend
exp_x = dy.exp(x)
# (dim_1, batch_size), if no dim_1, return ((1,), batch_size)
log_sum_exp_x = dy.log(dy.mean_dim(exp_x, d=[0], b=False) * dims[0][0])
return log_sum_exp_x + max_score
def log_sum_exp(self, scores):
"""
:param scores: observation scores for all possible tag sequences
:return: \log (\sum(exp(S(y))))
"""
scores_val = scores.npvalue()
max_idx = np.argmax(scores_val)
# introduce max_scores to avoid underflow
# if not, the results will be INF or -INF
# dynet expression of maximum scores
max_score = dy.pick(scores, max_idx)
max_score_broadcast = dy.concatenate([max_score] * (self.dim_ts_y + 2))
# shift the center of exponential sum to (scores - max)
return max_score + dy.log(dy.sum_elems(dy.transpose(dy.exp(scores - max_score_broadcast))))
def backward(self, word_vectors, label):
dy.renew_cg()
x = dy.inputTensor(word_vectors)
y = dy.inputTensor(label)
logit = self.build_graph(x)
# q表示对正样本的加权
# 公式见https://www.tensorflow.org/api_docs/python/tf/nn/weighted_cross_entropy_with_logits
q = 15
l = 1 + (q - 1) * y
loss = (1 - y) * logit + l * (dy.log(1 + dy.exp(-dy.abs(logit))) +
dy.rectify(-logit))
res = loss.value()
loss.backward()
return res
def pz(self, eq):
"""
Gumbel softmax on distribution over z.
"""
W = dy.parameter(self.W)
prob = dy.softmax(W * eq)
gumbel = dy.random_gumbel(self.num_clusters)
y = []
denom = []
for z in range(self.num_clusters):
pi_i = prob[z]
g_i = gumbel[z]
val = dy.exp((dy.log(pi_i) + g_i) / self.temp)
denom.append(val)
denom = dy.esum(denom)
for z in range(self.num_clusters):
pi_i = prob[z]
g_i = gumbel[z]
numerator = dy.exp((dy.log(pi_i) + g_i) / self.temp)
y.append(dy.cdiv(numerator, denom))
logits = dy.concatenate(y)
return logits
def calc_loss_basic(self, frames, label):
# Renew the computation graph
dy.renew_cg()
# Initialize LSTM
init_state_src = self.lstm_builder.initial_state()
# Instantiate the params
W_mean = dy.parameter(self.W_mean_p)
V_mean = dy.parameter(self.V_mean_p)
b_mean = dy.parameter(self.b_mean_p)
W_var = dy.parameter(self.W_var_p)
V_var = dy.parameter(self.V_var_p)
b_var = dy.parameter(self.b_var_p)
input_frames = dy.inputTensor(frames)
output_label = label
# Get the LSTM embeddings
src_output = init_state_src.add_inputs(
[frame for frame in input_frames])[-1].output()
# Get the mean and diagonal log covariance from the encoder
mu = self.mlp(src_output, W_mean, V_mean, b_mean)
log_var = self.mlp(src_output, W_mean, V_mean, b_mean)
# Compute the KL Divergence loss
kl_loss = -0.5 * dy.sum_elems(1 + log_var -
dy.pow(mu, dy.inputVector([2])) -
dy.exp(log_var))
# Reparameterize
z = self.reparameterize(mu, log_var)
W_sm = dy.parameter(self.W_sm_p)
b_sm = dy.parameter(self.b_sm_p)
# Calculate the reconstruction loss
pred = dy.affine_transform([b_sm, W_sm, z])
label_embedding = self.lookup[label]
#print label, label_embedding
recons_loss = dy.pickneglogsoftmax(pred, label)
return kl_loss, recons_loss
def max_margin_weighting(instance, pred_states, pred_scores_v):
pred_scores_v = np.array(pred_scores_v)
assert len(pred_states) == len(pred_scores_v)
# assert len(instance.states) == len(pred_states[0])
correct_denotations = []
for i, states in enumerate(pred_states):
if states[-1] == instance.states[-1]:
correct_denotations.append(i)
if not correct_denotations:
weights = np.zeros_like(pred_scores_v)
else:
weights = dy.exp(
dy.log_softmax(dy.inputVector(pred_scores_v),
correct_denotations)).npvalue()
return weights, correct_denotations
def calc_reinforce_loss(words, tags, delta):
dy.renew_cg()
# Transduce all batch elements with an LSTM
word_reps = LSTM.transduce([LOOKUP[x] for x in words])
# Softmax scores
W = dy.parameter(W_sm)
b = dy.parameter(b_sm)
#calculate the probability distribution
scores = [dy.affine_transform([b, W, x]) for x in word_reps]
losses = [dy.pickneglogsoftmax(score, tag) for score, tag in zip(scores, tags)]
probs = [-dy.exp(loss).as_array() for loss in losses]
#then take samples from the probability distribution
samples = [np.random.choice(range(len(x)), p=x) for x in probs]
#calculate accuracy=reward
correct = [sample == tag for sample, tag in zip(samples, tags)]
r_i = float(sum(correct))/len(correct)
r = dy.constant((1), r_i)
# Reward baseline for each word
W_bl = dy.parameter(W_bl_p)
b_bl = dy.parameter(b_bl_p)
r_b = [dy.affine_transform([b_bl, W_bl, dy.nobackprop(x)]) for x in word_reps]
#we need to take the value in order to break the computation graph
#as the reward portion is trained seperatley and not backpropogated through during the overall score
rewards_over_baseline = [(r - dy.nobackprop(x)) for x in r_b]
#the scores for training the baseline
baseline_scores = [dy.square(r - x) for x in r_b]
#then calculate the reinforce scores using reinforce
reinforce_scores = [r_s*score for r_s, score in zip(rewards_over_baseline, scores)]
#we want the first len(sent)-delta scores from xent then delta scores from reinforce
#for mixer
if len(scores) > delta:
mixer_scores = scores[:len(scores)-delta] + reinforce_scores[delta-1:]
else:
mixer_scores = reinforce_scores
return dy.esum(mixer_scores), dy.esum(baseline_scores)
def sample(self, x: dy.Expression, n: numbers.Integral, temperature: numbers.Real=1.0):
assert temperature != 0.0
scores_expr = self.calc_log_probs(x)
if temperature != 1.0:
scores_expr *= 1.0 / temperature
scores = dy.softmax(scores_expr).npvalue()
else:
scores = dy.exp(scores_expr).npvalue()
# Numpy is very picky. If the sum is off even by 1e-8 it complains.
scores /= sum(scores)
a = range(scores.shape[0])
samples = np.random.choice(a, (n,), replace=True, p=scores)
r = []
for word in samples:
r.append((word, dy.pick(scores_expr, word)))
return r