def attend(self, input_mat, state, w1dt, w2, v, coverage):
w2dt = w2 * dy.concatenate(list(state.s()))
if coverage:
w1dt = w1dt + self.w_cov * dy.transpose(coverage)
a_t = dy.transpose(v * dy.tanh(dy.colwise_add(w1dt, w2dt)))
a_t = dy.softmax(a_t)
return a_t, (input_mat * a_t)
def __call__(self, h, s):
# hT -> ((L, h_dim), B), s -> ((s_dim, L), B)
if len(h.dim()[0]) == 2:
L = h.dim()[0][1]
if self.h_bias:
s = dy.concatenate(
[s, dy.inputTensor(np.ones((1, L), dtype=np.float32))])
if self.s_bias:
h = dy.concatenate(
[h, dy.inputTensor(np.ones((1, L), dtype=np.float32))])
else:
if self.h_bias:
s = dy.concatenate(
[s, dy.inputTensor(np.ones((1, ), dtype=np.float32))])
if self.s_bias:
h = dy.concatenate(
[h, dy.inputTensor(np.ones((1, ), dtype=np.float32))])
hT = dy.transpose(h)
lin = self.U * s # ((h_dim*n_label, L), B)
if self.n_label > 1:
lin = dy.reshape(lin, (self.h_dim, self.n_label))
blin = hT * lin
if self.n_label == 1:
return blin
else:
return dy.transpose(blin)
def get_score(self,word,context):
## Get the loss given word, context pair and perform negative sampling
objective = dy.logistic(((dy.transpose(self.context_embeddings[context]))*self.word_embeddings[word]))
negative_sample = np.random.choice(self.context_size, self.num_sampled, replace=False, p=self.context_fre)
for context_prime in negative_sample:
objective *= dy.logistic(-((dy.transpose(self.context_embeddings[context_prime]))*self.word_embeddings[word]))
loss = -dy.log(objective)
return loss
def __call__(self, h, s):
# hT -> ((L, h_dim), B), s -> ((s_dim, L), B)
hT = dy.transpose(h)
lin = self.U * s # ((h_dim*n_label, L), B)
if self.n_label > 1:
lin = dy.reshape(lin, (self.h_dim, self.n_label))
blin = hT * lin
if self.n_label == 1:
return blin + (hT * self.B if self.bias else 0)
else:
return dy.transpose(blin) + (self.V * dy.concatenate([h, s]) +
self.B if self.bias else 0)
def __call__(self, h, s):
if self.h_bias:
if len(h.dim()[0]) == 2:
h = dy.concatenate([
h,
dy.inputTensor(
np.ones((1, h.dim()[0][1]), dtype=np.float32))
])
else:
h = dy.concatenate(
[h, dy.inputTensor(np.ones((1, ), dtype=np.float32))])
if self.s_bias:
if len(s.dim()[0]) == 2:
s = dy.concatenate([
s,
dy.inputTensor(
np.ones((1, s.dim()[0][1]), dtype=np.float32))
])
else:
s = dy.concatenate(
[s, dy.inputTensor(np.ones((1, ), dtype=np.float32))])
lin = self.U * s
if self.n_label > 1:
lin = dy.reshape(lin, (self.h_dim, self.n_label))
blin = dy.transpose(h) * lin
return blin
def selection_by_tree(self, tree, mode, idx=0):
input_layers, pairs = self._select_by_tree(tree, mode, True)
if len(pairs) == 0:
if not self.opt['allow_partial']:
input_layers, pairs = self._select_by_tree(tree, mode, False)
else:
print 'early stop! discard {} / {}.'.format(
len(tree.V), len(tree.terms))
return None, None
W1_rl = dy.parameter(self.model_parameters['W1_rl'])
b1_rl = dy.parameter(self.model_parameters['b1_rl'])
if not self.opt['one_layer']:
W2_rl = dy.parameter(self.model_parameters['W2_rl'])
b2_rl = dy.parameter(self.model_parameters['b2_rl'])
# pr = W2_rl * dy.rectify(W1_rl * dy.concatenate_to_batch(input_layers) + b1_rl) + b2_rl
# (V x N)x160 160x50 50x60 60x1
input_layers = dy.concatenate_cols(input_layers)
input_layers = dy.transpose(input_layers)
if not self.opt['one_layer']:
if self.opt['use_history']:
pr = input_layers * dy.rectify(W2_rl * dy.rectify(
W1_rl * self.history[idx].output() + b1_rl) + b2_rl)
else:
pr = dy.rectify(input_layers * W2_rl + b2_rl) * W1_rl + b1_rl
else:
if self.opt['use_history']:
pr = input_layers * dy.rectify(
W1_rl * self.history[idx].output() + b1_rl)
else:
pr = input_layers * W1_rl + b1_rl
# (#actions, )
pr = dy.reshape(pr, (len(pairs), ))
return dy.softmax(pr), pairs
def predict(self,
feature_vector,
task_ids,
train=False,
soft_labels=False,
temperature=None,
dropout_rate=0.0,
orthogonality_weight=0.0,
domain_id=None):
dynet.renew_cg() # new graph
feature_vector = feature_vector.toarray()
feature_vector = np.squeeze(feature_vector, axis=0)
# self.input = dynet.vecInput(self.vocab_size)
# self.input.set(feature_vector)
# TODO this takes too long; can we speed this up somehow?
input = dynet.inputVector(feature_vector)
for i in range(self.h_layers):
if train: # add some noise
input = dynet.noise(input, self.noise_sigma)
input = dynet.dropout(input, dropout_rate)
input = self.layers[i](input)
outputs = []
for task_id in task_ids:
output = self.output_layers_dict[task_id](input,
soft_labels=soft_labels,
temperature=temperature)
outputs.append(output)
constraint, adv_loss = 0, 0
if orthogonality_weight != 0:
# put the orthogonality constraint either directly on the
# output layer or on the hidden layer if it's an MLP
F0_layer = self.output_layers_dict["F0"]
F1_layer = self.output_layers_dict["F1"]
F0_param = F0_layer.W_mlp if self.add_hidden else F0_layer.W
F1_param = F1_layer.W_mlp if self.add_hidden else F1_layer.W
F0_W = dynet.parameter(F0_param)
F1_W = dynet.parameter(F1_param)
# calculate the matrix product of the task matrix with both others
matrix_product = dynet.transpose(F0_W) * F1_W
# take the squared Frobenius norm by squaring
# every element and then summing them
squared_frobenius_norm = dynet.sum_elems(
dynet.square(matrix_product))
constraint += squared_frobenius_norm
# print('Constraint with first matrix:', squared_frobenius_norm.value())
if domain_id is not None:
# flip the gradient when back-propagating through here
adv_input = dynet.flip_gradient(input) # last state
adv_output = self.adv_layer(adv_input)
adv_loss = self.pick_neg_log(adv_output, domain_id)
# print('Adversarial loss:', avg_adv_loss.value())
return outputs, constraint, adv_loss
def set_initial_states(self, x):
self.xt_embs = [dy.lookup(self.F, x_t) for x_t in x]
if self.encoder_type == 'bow':
self.W_enc = self.W * dy.average(self.xt_embs)
elif self.encoder_type == 'attention':
self.xb = dy.concatenate([
dy.esum(self.xt_embs[max(i - self.q, 0
):min(len(x) - 1 + 1, i + self.q +
1)]) / self.q
for i in range(len(x))
],
d=1)
self.xt = dy.transpose(dy.concatenate(self.xt_embs, d=1))
def __call__(self, x, h_matrix, noprob=False):
s_t = x
for i in range(self.layers - 1):
e_t = self.V[i] * dy.tanh(self.W1[i] * h_matrix + self.W2[i] * s_t)
a_t = dy.softmax(dy.transpose(e_t))
c_t = h_matrix * a_t
s_t = dy.concatenate([x, c_t])
e_t = self.V[-1] * dy.tanh(self.W1[-1] * h_matrix + self.W2[-1] *
s_t) + self.B1 * h_matrix + self.B2 * s_t
if len(h_matrix.dim()[0]) > 1:
e_t = dy.reshape(e_t,
(self.V[-1].dim()[0][0] * h_matrix.dim()[0][1], ))
if not noprob:
p_t = dy.softmax(e_t)
return p_t
else:
return e_t
def main():
parser = argparse.ArgumentParser(
description=
'Convolutional Neural Networks for Sentence Classification in DyNet')
parser.add_argument('--gpu',
type=int,
default=0,
help='GPU ID to use. For cpu, set -1 [default: 0]')
parser.add_argument(
'--train_x_path',
type=str,
default='./data/train_x.txt',
help='File path of train x data [default: `./data/train_x.txt`]')
parser.add_argument(
'--train_y_path',
type=str,
default='./data/train_y.txt',
help='File path of train y data [default: `./data/train_x.txt`]')
parser.add_argument(
'--valid_x_path',
type=str,
default='./data/valid_x.txt',
help='File path of valid x data [default: `./data/valid_x.txt`]')
parser.add_argument(
'--valid_y_path',
type=str,
default='./data/valid_y.txt',
help='File path of valid y data [default: `./data/valid_y.txt`]')
parser.add_argument('--n_epochs',
type=int,
default=10,
help='Number of epochs [default: 10]')
parser.add_argument('--batch_size',
type=int,
default=64,
help='Mini batch size [default: 64]')
parser.add_argument('--win_sizes',
type=int,
nargs='*',
default=[3, 4, 5],
help='Window sizes of filters [default: [3, 4, 5]]')
parser.add_argument(
'--num_fil',
type=int,
default=100,
help='Number of filters in each window size [default: 100]')
parser.add_argument('--s',
type=float,
default=3.0,
help='L2 norm constraint on w [default: 3.0]')
parser.add_argument('--dropout_prob',
type=float,
default=0.5,
help='Dropout probability [default: 0.5]')
parser.add_argument(
'--v_strategy',
type=str,
default='static',
help=
'Embedding strategy. rand: Random initialization. static: Load pretrained embeddings and do not update during the training. non-static: Load pretrained embeddings and update during the training. [default: static]'
)
parser.add_argument(
'--alloc_mem',
type=int,
default=4096,
help='Amount of memory to allocate [mb] [default: 4096]')
args = parser.parse_args()
print(args)
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
N_EPOCHS = args.n_epochs
WIN_SIZES = args.win_sizes
BATCH_SIZE = args.batch_size
EMB_DIM = 300
OUT_DIM = 1
L2_NORM_LIM = args.s
NUM_FIL = args.num_fil
DROPOUT_PROB = args.dropout_prob
V_STRATEGY = args.v_strategy
ALLOC_MEM = args.alloc_mem
if V_STRATEGY in ['rand', 'static', 'non-static']:
NUM_CHA = 1
else:
NUM_CHA = 2
# FILE paths
W2V_PATH = './GoogleNews-vectors-negative300.bin'
TRAIN_X_PATH = args.train_x_path
TRAIN_Y_PATH = args.train_y_path
VALID_X_PATH = args.valid_x_path
VALID_Y_PATH = args.valid_y_path
# DyNet setting
dyparams = dy.DynetParams()
dyparams.set_random_seed(RANDOM_SEED)
dyparams.set_mem(ALLOC_MEM)
dyparams.init()
# Load pretrained embeddings
pretrained_model = gensim.models.KeyedVectors.load_word2vec_format(
W2V_PATH, binary=True)
vocab = pretrained_model.wv.vocab.keys()
w2v = pretrained_model.wv
# Build dataset =======================================================================================================
w2c = build_w2c(TRAIN_X_PATH, vocab=vocab)
w2i, i2w = build_w2i(TRAIN_X_PATH, w2c, unk='unk')
train_x, train_y = build_dataset(TRAIN_X_PATH,
TRAIN_Y_PATH,
w2i,
unk='unk')
valid_x, valid_y = build_dataset(VALID_X_PATH,
VALID_Y_PATH,
w2i,
unk='unk')
train_x, train_y = sort_data_by_length(train_x, train_y)
valid_x, valid_y = sort_data_by_length(valid_x, valid_y)
VOCAB_SIZE = len(w2i)
print('VOCAB_SIZE:', VOCAB_SIZE)
V_init = init_V(w2v, w2i)
with open(os.path.join(RESULTS_DIR, './w2i.dump'),
'wb') as f_w2i, open(os.path.join(RESULTS_DIR, './i2w.dump'),
'wb') as f_i2w:
pickle.dump(w2i, f_w2i)
pickle.dump(i2w, f_i2w)
# Build model =================================================================================
model = dy.Model()
trainer = dy.AdamTrainer(model)
# V1
V1 = model.add_lookup_parameters((VOCAB_SIZE, EMB_DIM))
if V_STRATEGY in ['static', 'non-static', 'multichannel']:
V1.init_from_array(V_init)
if V_STRATEGY in ['static', 'multichannel']:
V1_UPDATE = False
else: # 'rand', 'non-static'
V1_UPDATE = True
make_emb_zero(V1, [w2i['<s>'], w2i['</s>']], EMB_DIM)
# V2
if V_STRATEGY == 'multichannel':
V2 = model.add_lookup_parameters((VOCAB_SIZE, EMB_DIM))
V2.init_from_array(V_init)
V2_UPDATE = True
make_emb_zero(V2, [w2i['<s>'], w2i['</s>']], EMB_DIM)
layers = [
CNNText(model, EMB_DIM, WIN_SIZES, NUM_CHA, NUM_FIL, dy.tanh,
DROPOUT_PROB),
Dense(model, 3 * NUM_FIL, OUT_DIM, dy.logistic)
]
# Train model ================================================================================
n_batches_train = math.ceil(len(train_x) / BATCH_SIZE)
n_batches_valid = math.ceil(len(valid_x) / BATCH_SIZE)
start_time = time.time()
for epoch in range(N_EPOCHS):
# Train
loss_all_train = []
pred_all_train = []
for i in tqdm(range(n_batches_train)):
# Create a new computation graph
dy.renew_cg()
associate_parameters(layers)
# Create a mini batch
start = i * BATCH_SIZE
end = start + BATCH_SIZE
x = build_batch(train_x[start:end], w2i, max(WIN_SIZES)).T
t = np.array(train_y[start:end])
sen_len = x.shape[0]
if V_STRATEGY in ['rand', 'static', 'non-static']:
x_embs = dy.concatenate_cols(
[dy.lookup_batch(V1, x_t, update=V1_UPDATE) for x_t in x])
x_embs = dy.transpose(x_embs)
x_embs = dy.reshape(x_embs, (sen_len, EMB_DIM, 1))
else: # multichannel
x_embs1 = dy.concatenate_cols(
[dy.lookup_batch(V1, x_t, update=V1_UPDATE) for x_t in x])
x_embs2 = dy.concatenate_cols(
[dy.lookup_batch(V2, x_t, update=V2_UPDATE) for x_t in x])
x_embs1 = dy.transpose(x_embs1)
x_embs2 = dy.transpose(x_embs2)
x_embs = dy.concatenate([x_embs1, x_embs2], d=2)
t = dy.inputTensor(t, batched=True)
y = forwards(layers, x_embs, test=False)
mb_loss = dy.mean_batches(dy.binary_log_loss(y, t))
# Forward prop
loss_all_train.append(mb_loss.value())
pred_all_train.extend(list(binary_pred(y.npvalue().flatten())))
# Backward prop
mb_loss.backward()
trainer.update()
# L2 norm constraint
layers[1].scale_W(L2_NORM_LIM)
# Make padding embs zero
if V_STRATEGY in ['rand', 'non-static']:
make_emb_zero(V1, [w2i['<s>'], w2i['</s>']], EMB_DIM)
elif V_STRATEGY in ['multichannel']:
make_emb_zero(V2, [w2i['<s>'], w2i['</s>']], EMB_DIM)
# Valid
loss_all_valid = []
pred_all_valid = []
for i in range(n_batches_valid):
# Create a new computation graph
dy.renew_cg()
associate_parameters(layers)
# Create a mini batch
start = i * BATCH_SIZE
end = start + BATCH_SIZE
x = build_batch(valid_x[start:end], w2i, max(WIN_SIZES)).T
t = np.array(valid_y[start:end])
sen_len = x.shape[0]
if V_STRATEGY in ['rand', 'static', 'non-static']:
x_embs = dy.concatenate_cols(
[dy.lookup_batch(V1, x_t, update=V1_UPDATE) for x_t in x])
x_embs = dy.transpose(x_embs)
x_embs = dy.reshape(x_embs, (sen_len, EMB_DIM, 1))
else: # multichannel
x_embs1 = dy.concatenate_cols(
[dy.lookup_batch(V1, x_t, update=V1_UPDATE) for x_t in x])
x_embs2 = dy.concatenate_cols(
[dy.lookup_batch(V2, x_t, update=V2_UPDATE) for x_t in x])
x_embs1 = dy.transpose(x_embs1)
x_embs2 = dy.transpose(x_embs2)
x_embs = dy.concatenate([x_embs1, x_embs2], d=2)
t = dy.inputTensor(t, batched=True)
y = forwards(layers, x_embs, test=True)
mb_loss = dy.mean_batches(dy.binary_log_loss(y, t))
# Forward prop
loss_all_valid.append(mb_loss.value())
pred_all_valid.extend(list(binary_pred(y.npvalue().flatten())))
print(
'EPOCH: %d, Train Loss:: %.3f (F1:: %.3f, Acc:: %.3f), Valid Loss:: %.3f (F1:: %.3f, Acc:: %.3f), Time:: %.3f[s]'
% (
epoch + 1,
np.mean(loss_all_train),
f1_score(train_y, pred_all_train),
accuracy_score(train_y, pred_all_train),
np.mean(loss_all_valid),
f1_score(valid_y, pred_all_valid),
accuracy_score(valid_y, pred_all_valid),
time.time() - start_time,
))
# Save model =========================================================================================================================
if V_STRATEGY in ['rand', 'static', 'non-static']:
dy.save(os.path.join(RESULTS_DIR, './model_e' + str(epoch + 1)),
[V1] + layers)
else:
dy.save(os.path.join(RESULTS_DIR, './model_e' + str(epoch + 1)),
[V1, V2] + layers)
def __call__(self, x=None, t=None, test=False):
if test:
tt_embs = [dy.lookup(self.E, t_t) for t_t in t]
if self.encoder_type == 'bow':
# Neural language model
tt_c = dy.concatenate(tt_embs)
h = dy.tanh(self.U * tt_c)
# Output with softmax
y_t = dy.softmax(self.V * h + self.W_enc)
elif self.encoder_type == 'attention':
ttp_embs = [dy.lookup(self.G, t_t) for t_t in t]
# Neural language model
tt_c = dy.concatenate(tt_embs)
h = dy.tanh(self.U * tt_c)
# Attention
ttp_c = dy.concatenate(ttp_embs)
p = dy.softmax(self.xt * self.P * ttp_c) # Attention weight
enc = self.xb * p # Context vector
# Output with softmax
y_t = dy.softmax(self.V * h + self.W * enc)
return y_t
else:
xt_embs = [dy.lookup(self.F, x_t) for x_t in x]
tt_embs = [dy.lookup(self.E, t_t) for t_t in t]
y = []
if self.encoder_type == 'bow':
# BoW
enc = dy.average(xt_embs)
W_enc = self.W * enc
for i in range(len(t) - self.c + 1):
# Neural language model
tt_c = dy.concatenate(tt_embs[i:i + self.c])
h = dy.tanh(self.U * tt_c)
# Output without softmax
y_t = self.V * h + W_enc
y.append(y_t)
elif self.encoder_type == 'attention':
xb = dy.concatenate([
dy.esum(xt_embs[max(i - self.q, 0
):min(len(x) - 1 + 1, i + self.q + 1)])
/ self.q for i in range(len(x))
],
d=1)
xt = dy.transpose(dy.concatenate(xt_embs, d=1))
ttp_embs = [dy.lookup(self.G, t_t) for t_t in t]
for i in range(len(t) - self.c + 1):
# Neural language model
tt_c = dy.concatenate(tt_embs[i:i + self.c])
h = dy.tanh(self.U * tt_c)
# Attention
ttp_c = dy.concatenate(
ttp_embs[i:i + self.c]) # Window-sized embedding
p = dy.softmax(xt * self.P * ttp_c) # Attention weight
enc = xb * p # Context vector
# Output without softmax
y_t = self.V * h + self.W * enc
y.append(y_t)
return y
def __call__(self, x, tm1s=None, test=False):
if test:
# Initial states
s_tm1 = tm1s[0]
c_tm1 = tm1s[1]
w_tm1 = x
# GRU
s_t = self.GRUBuilder.initial_state().set_s([s_tm1]).add_input(
dy.concatenate([w_tm1, c_tm1])).output()
# Attention
e_t = dy.pick(
self.va *
dy.tanh(dy.colwise_add(self.Ua * self.hp, self.Wa * s_tm1)), 0)
a_t = dy.softmax(e_t)
c_t = dy.esum([
dy.cmult(a_t_i, h_i)
for a_t_i, h_i in zip(a_t, dy.transpose(self.hp))
])
#c_t = self.hp*a_t # memory error?
# Output
r_t = dy.concatenate_cols([
Wr_j * w_tm1 + Ur_j * c_t + Vr_j * s_t
for Wr_j, Ur_j, Vr_j in zip(self.Wr, self.Ur, self.Vr)
]) # Maxout
m_t = dy.max_dim(r_t, d=1)
y_t = dy.softmax(self.Wo * m_t)
return s_t, c_t, y_t
else:
w_embs = x
# Initial states
s_tm1 = self.s_0
c_tm1 = self.c_0
GRU = self.GRUBuilder.initial_state().set_s([s_tm1])
y = []
for w_tm1 in w_embs:
# GRU
GRU = GRU.add_input(dy.concatenate([w_tm1, c_tm1]))
s_t = GRU.output()
# Attention
e_t = dy.pick(
self.va * dy.tanh(
dy.colwise_add(self.Ua * self.hp, self.Wa * s_tm1)), 0)
a_t = dy.softmax(e_t)
c_t = dy.esum([
dy.cmult(a_t_i, h_i)
for a_t_i, h_i in zip(a_t, dy.transpose(self.hp))
])
#c_t = self.hp*a_t # memory error?
# Output
r_t = dy.concatenate_cols([
Wr_j * w_tm1 + Ur_j * c_t + Vr_j * s_t
for Wr_j, Ur_j, Vr_j in zip(self.Wr, self.Ur, self.Vr)
]) # Maxout
m_t = dy.max_dim(r_t, d=1)
y_t = self.Wo * m_t
y.append(y_t)
# t -> tm1
s_tm1 = s_t
c_tm1 = c_t
return y
def main():
parser = argparse.ArgumentParser(description='Convolutional Neural Networks for Sentence Classification in DyNet')
parser.add_argument('--gpu', type=int, default=-1, help='GPU ID to use. For cpu, set -1 [default: -1]')
parser.add_argument('--model_file', type=str, default='./model', help='Model to use for prediction [default: ./model]')
parser.add_argument('--input_file', type=str, default='./data/valid_x.txt', help='Input file path [default: ./data/valid_x.txt]')
parser.add_argument('--output_file', type=str, default='./pred_y.txt', help='Output file path [default: ./pred_y.txt]')
parser.add_argument('--w2i_file', type=str, default='./w2i.dump', help='Word2Index file path [default: ./w2i.dump]')
parser.add_argument('--i2w_file', type=str, default='./i2w.dump', help='Index2Word file path [default: ./i2w.dump]')
parser.add_argument('--alloc_mem', type=int, default=1024, help='Amount of memory to allocate [mb] [default: 1024]')
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
MODEL_FILE = args.model_file
INPUT_FILE = args.input_file
OUTPUT_FILE = args.output_file
W2I_FILE = args.w2i_file
I2W_FILE = args.i2w_file
ALLOC_MEM = args.alloc_mem
# DyNet setting
dyparams = dy.DynetParams()
dyparams.set_mem(ALLOC_MEM)
dyparams.init()
# Load model
model = dy.Model()
pretrained_model = dy.load(MODEL_FILE, model)
if len(pretrained_model) == 3:
V1, layers = pretrained_model[0], pretrained_model[1:]
MULTICHANNEL = False
else:
V1, V2, layers = pretrained_model[0], pretrained_model[1], pretrained_model[2:]
MULTICHANNEL = True
EMB_DIM = V1.shape()[0]
WIN_SIZES = layers[0].win_sizes
# Load test data
with open(W2I_FILE, 'rb') as f_w2i, open(I2W_FILE, 'rb') as f_i2w:
w2i = pickle.load(f_w2i)
i2w = pickle.load(f_i2w)
max_win = max(WIN_SIZES)
test_X, _, _ = build_dataset(INPUT_FILE, w2i=w2i, unksym='unk')
test_X = [[0]*max_win + instance_x + [0]*max_win for instance_x in test_X]
# Pred
pred_y = []
for instance_x in tqdm(test_X):
# Create a new computation graph
dy.renew_cg()
associate_parameters(layers)
sen_len = len(instance_x)
if MULTICHANNEL:
x_embs1 = dy.concatenate([dy.lookup(V1, x_t, update=False) for x_t in instance_x], d=1)
x_embs2 = dy.concatenate([dy.lookup(V2, x_t, update=False) for x_t in instance_x], d=1)
x_embs1 = dy.transpose(x_embs1)
x_embs2 = dy.transpose(x_embs2)
x_embs = dy.concatenate([x_embs1, x_embs2], d=2)
else:
x_embs = dy.concatenate([dy.lookup(V1, x_t, update=False) for x_t in instance_x], d=1)
x_embs = dy.transpose(x_embs)
x_embs = dy.reshape(x_embs, (sen_len, EMB_DIM, 1))
y = f_props(layers, x_embs, train=False)
pred_y.append(str(int(binary_pred(y.value()))))
with open(OUTPUT_FILE, 'w') as f:
f.write('\n'.join(pred_y))
def __call__(self, s_t, h_matrix):
e_t = self.v * dy.tanh(self.W1*h_matrix + self.W2 * s_t)
a_t = dy.softmax(dy.transpose(e_t))
c_t = h_matrix * a_t
return c_t
def predict(self,
word_indices,
char_indices,
task_id,
train=False,
soft_labels=False,
temperature=None,
orthogonality_weight=0.0,
domain_id=None):
"""
predict tags for a sentence represented as char+word embeddings
:param domain_id: Predict adversarial loss if domain id is provided.
"""
dynet.renew_cg() # new graph
char_emb = []
rev_char_emb = []
wfeatures = [self.wembeds[w] for w in word_indices]
if self.c_in_dim > 0:
# get representation for words
for chars_of_token in char_indices:
char_feats = [self.cembeds[c] for c in chars_of_token]
# use last state as word representation
f_char, b_char = self.char_rnn.predict_sequence(
char_feats, char_feats)
last_state = f_char[-1]
rev_last_state = b_char[-1]
char_emb.append(last_state)
rev_char_emb.append(rev_last_state)
features = [
dynet.concatenate([w, c, rev_c])
for w, c, rev_c in zip(wfeatures, char_emb, rev_char_emb)
]
else:
features = wfeatures
if train: # only do at training time
features = [dynet.noise(fe, self.noise_sigma) for fe in features]
output_expected_at_layer = self.h_layers
output_expected_at_layer -= 1
# go through layers
prev = features
prev_rev = features
num_layers = self.h_layers
constraint = 0
adv_loss = 0
for i in range(0, num_layers):
predictor = self.predictors["inner"][i]
forward_sequence, backward_sequence = predictor.predict_sequence(
prev, prev_rev)
if i > 0 and self.activation:
# activation between LSTM layers
forward_sequence = [
self.activation(s) for s in forward_sequence
]
backward_sequence = [
self.activation(s) for s in backward_sequence
]
if i == output_expected_at_layer:
concat_layer = [
dynet.concatenate([f, b]) for f, b in zip(
forward_sequence, reversed(backward_sequence))
]
if train and self.noise_sigma > 0.0:
concat_layer = [
dynet.noise(fe, self.noise_sigma)
for fe in concat_layer
]
if task_id not in ["src", "trg"]:
output_predictor = self.predictors["output_layers_dict"][
task_id]
output = output_predictor.predict_sequence(
concat_layer,
soft_labels=soft_labels,
temperature=temperature)
else:
# one src example for all three outputs
output = [] # in this case it is a list
for t_id in self.task_ids:
output_predictor = self.predictors[
"output_layers_dict"][t_id]
output_t = output_predictor.predict_sequence(
concat_layer,
soft_labels=soft_labels,
temperature=temperature)
output.append(output_t)
if orthogonality_weight != 0 and task_id != "Ft":
# put the orthogonality constraint either directly on the
# output layer or on the hidden layer if it's an MLP
# use orthogonality_weight only between F0 and F1
builder = self.predictors["output_layers_dict"][
"F0"].network_builder
task_param = builder.W_mlp if self.add_hidden else builder.W
task_W = dynet.parameter(task_param)
builder = self.predictors["output_layers_dict"][
"F1"].network_builder
other_param = builder.W_mlp if self.add_hidden else builder.W
other_task_W = dynet.parameter(other_param)
# calculate the matrix product of the task matrix with the other
matrix_product_1 = dynet.transpose(task_W) * other_task_W
# take the squared Frobenius norm by squaring
# every element and then summing them
squared_frobenius_norm = dynet.sum_elems(
dynet.square(matrix_product_1))
constraint = squared_frobenius_norm
#print('Constraint with first matrix:', squared_frobenius_norm.value())
if domain_id is not None:
# flip the gradient when back-propagating through here
adv_input = dynet.flip_gradient(
concat_layer[-1]) # last state
adv_output = self.adv_layer(adv_input)
adv_loss = self.pick_neg_log(adv_output, domain_id)
#print('Adversarial loss:', avg_adv_loss.value())
# output is list if task_id = 'src'
return output, constraint, adv_loss
prev = forward_sequence
prev_rev = backward_sequence
raise Exception("oops should not be here")
return None
