def init_params(options, preemb=None):
"""
Initialize all parameters
"""
params = OrderedDict()
# Word embedding
if preemb == None:
params['Wemb'] = norm_weight(options['n_words'], options['dim_word'])
else:
print('using preemb')
params['Wemb'] = preemb
# Encoder
params = get_layer(options['encoder'])[0](options, params, prefix='encoder',
nin=options['dim_word'], dim=options['dim'])
# Decoder: next sentence
params = get_layer(options['decoder'])[0](options, params, prefix='decoder_f',
nin=options['dim_word'], dim=options['dim'])
# Decoder: previous sentence
params = get_layer(options['decoder'])[0](options, params, prefix='decoder_b',
nin=options['dim_word'], dim=options['dim'])
# Output layer
params = get_layer('ff')[0](options, params, prefix='ff_logit', nin=options['dim'], nout=options['n_words'])
return params
def build_model(tparams, options):
"""
Computation graph for the model
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
x = tensor.matrix('x', dtype='int64')
mask = tensor.matrix('mask', dtype='float32')
im = tensor.matrix('im', dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# Encode sentences (source)
if options['encoder'] == 'bow':
sents = (emb * mask[:,:,None]).sum(0)
else:
proj = get_layer(options['encoder'])[1](tparams, emb, None, options,
prefix='encoder',
mask=mask)
sents = proj[0][-1]
sents = l2norm(sents)
# Encode images (source)
images = get_layer('ff')[1](tparams, im, options, prefix='ff_image', activ='linear')
# Compute loss
cost = contrastive_loss(options['margin'], images, sents)
return trng, [x, mask, im], cost
def init_params(options, preemb=None):
"""
Initialize all parameters
"""
params = OrderedDict()
# Word embedding
if preemb == None:
params['Wemb'] = norm_weight(options['n_words'], options['dim_word'])
else:
params['Wemb'] = preemb
# init state
params = get_layer('ff')[0](options, params, prefix='ff_state', nin=options['dimctx'], nout=options['dim'])
# Decoder
params = get_layer(options['decoder'])[0](options, params, prefix='decoder',
nin=options['dim_word'], dim=options['dim'])
# Output layer
if options['doutput']:
params = get_layer('ff')[0](options, params, prefix='ff_hid', nin=options['dim'], nout=options['dim_word'])
params = get_layer('ff')[0](options, params, prefix='ff_logit', nin=options['dim_word'], nout=options['n_words'])
else:
params = get_layer('ff')[0](options, params, prefix='ff_logit', nin=options['dim'], nout=options['n_words'])
return params
def build_model(tparams, options):
"""
Computation graph for the model
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
x = tensor.matrix('x', dtype='int64')
mask = tensor.matrix('mask', dtype='float32')
ctx = tensor.matrix('ctx', dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Index into the word embedding matrix, shift it forward in time
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# make eos the first word (it has no prev word context),
# throw out last word (it doesn't predict anything)
emb_shifted = tensor.zeros_like(emb)
emb_shifted = tensor.set_subtensor(emb_shifted[1:], emb[:-1])
emb = emb_shifted
# Init state
init_state = get_layer('ff')[1](tparams, ctx, options, prefix='ff_state', activ='tanh')
# Decoder
proj = get_layer(options['decoder'])[1](tparams, emb, init_state, options,
prefix='decoder',
mask=mask)
# Compute word probabilities
if options['doutput']:
hid = get_layer('ff')[1](tparams, proj[0], options, prefix='ff_hid', activ='tanh')
logit = get_layer('ff')[1](tparams, hid, options, prefix='ff_logit', activ='linear')
else:
logit = get_layer('ff')[1](tparams, proj[0], options, prefix='ff_logit', activ='linear')
logit_shp = logit.shape
probs = tensor.nnet.softmax(logit.reshape([logit_shp[0]*logit_shp[1], logit_shp[2]]))
# Cost
x_flat = x.flatten()
p_flat = probs.flatten()
cost = -tensor.log(p_flat[tensor.arange(x_flat.shape[0])*probs.shape[1]+x_flat]+1e-8)
cost = cost.reshape([x.shape[0], x.shape[1]])
cost = (cost * mask).sum(0)
cost = cost.sum()
return trng, [x, mask, ctx], cost
def init_params(options):
"""
Initialize all parameters
"""
params = OrderedDict()
# Word embedding
# params['Wemb'] = norm_weight(options['n_words'], options['dim_word'])
# We load the embeddings that we saved earlier.
if options['dataset'] == 'amazon':
embeddings = numpy.load(
'/home/shunan/Code/skip-thoughts/experiments/amazon/word2vec_embeds.npy'
)
elif options['dataset'] == 'imdb':
embeddings = numpy.load(
'/home/shunan/Code/skip-thoughts/experiments/imdb/skip_thought_word2vec_embeds.npy'
)
params['Wemb'] = embeddings.astype('float32')
# Encoder
params = get_layer(options['encoder'])[0](options,
params,
prefix='encoder',
nin=options['dim_word'],
dim=options['dim'])
# Decoder: next sentence
params = get_layer(options['decoder'])[0](options,
params,
prefix='decoder_f',
nin=options['dim_word'],
dim=options['dim'])
# Decoder: previous sentence
params = get_layer(options['decoder'])[0](options,
params,
prefix='decoder_b',
nin=options['dim_word'],
dim=options['dim'])
# Output layer
params = get_layer('ff')[0](options,
params,
prefix='ff_logit',
nin=options['dim'],
nout=options['n_words'])
return params
def init_params(options, rng=None):
"""
Initialize all network parameters and constrains.
All parameters and their corresponding constraints are stored in an OrderedDict.
"""
params = OrderedDict()
constraints = OrderedDict()
input_size = 2 # number of player utilities
n_hidden = [input_size] + options['hidden_units']
for i in xrange(1, len(n_hidden)):
params = get_layer('hid')[0](
options,
params,
prefix='hidden%02d' % i,
nin=n_hidden[i - 1] *
(3 if options['pooling'] else 1), # 3 x parameters if pooling used
nout=n_hidden[i],
rng=rng,
b_offset=1.)
params = get_layer('softmax')[0](options,
params,
nin=n_hidden[-1],
rng=rng)
ar_layers = options['ar_layers']
for i in range(ar_layers):
for p in range(2):
if i == ar_layers - 1 and p == 1:
# don't build ar layer for pl 2 in the last layer because it is not used
continue
params, constraints = get_layer('ar')[0](options,
params,
prefix='p%d_ar%d' %
(p, i),
nin=ar_layers,
level=i,
rng=rng,
constraints=constraints)
params, constraints = get_layer('output')[0](options,
params,
constraints,
rng=rng,
nin=ar_layers)
return params, constraints
def build_model(tparams, options):
"""
Computation graph for the model
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
x = tensor.matrix('x', dtype='int64')
mask = tensor.matrix('mask', dtype='float32')
ctx = tensor.matrix('ctx', dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Index into the word embedding matrix, shift it forward in time
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
emb_shifted = tensor.zeros_like(emb)
emb_shifted = tensor.set_subtensor(emb_shifted[1:], emb[:-1])
emb = emb_shifted
# Init state
init_state = get_layer('ff')[1](tparams, ctx, options, prefix='ff_state', activ='tanh')
# Decoder
proj = get_layer(options['decoder'])[1](tparams, emb, init_state, options,
prefix='decoder',
mask=mask)
# Compute word probabilities
if options['doutput']:
hid = get_layer('ff')[1](tparams, proj[0], options, prefix='ff_hid', activ='tanh')
logit = get_layer('ff')[1](tparams, hid, options, prefix='ff_logit', activ='linear')
else:
logit = get_layer('ff')[1](tparams, proj[0], options, prefix='ff_logit', activ='linear')
logit_shp = logit.shape
probs = tensor.nnet.softmax(logit.reshape([logit_shp[0]*logit_shp[1], logit_shp[2]]))
# Cost
x_flat = x.flatten()
p_flat = probs.flatten()
cost = -tensor.log(p_flat[tensor.arange(x_flat.shape[0])*probs.shape[1]+x_flat]+1e-8)
cost = cost.reshape([x.shape[0], x.shape[1]])
cost = (cost * mask).sum(0)
cost = cost.sum()
return trng, [x, mask, ctx], cost
def build_sampler(tparams, options, trng):
# x: 1 x 1
y = tensor.vector('y_sampler', dtype='int64')
init_state = tensor.matrix('init_state', dtype='float32')
# if it's the first word, emb should be all zero
emb = tensor.switch(y[:, None] < 0,
tensor.alloc(0., 1, tparams['Wemb'].shape[1]),
tparams['Wemb'][y])
# apply one step of gru layer
proj = get_layer(options['encoder'])[1](tparams,
emb,
options,
prefix='encoder',
mask=None,
one_step=True,
init_state=init_state)
next_state = proj[0]
# compute the output probability dist and sample
logit_lstm = get_layer('ff')[1](tparams,
next_state,
options,
prefix='ff_logit_lstm',
activ='linear')
logit_prev = get_layer('ff')[1](tparams,
emb,
options,
prefix='ff_logit_prev',
activ='linear')
logit = tensor.tanh(logit_lstm + logit_prev)
logit = get_layer('ff')[1](tparams,
logit,
options,
prefix='ff_logit',
activ='linear')
next_probs = tensor.nnet.softmax(logit)
next_sample = trng.multinomial(pvals=next_probs).argmax(1)
# next word probability
print 'Building f_next..',
inps = [y, init_state]
outs = [next_probs, next_sample, next_state]
f_next = theano.function(inps, outs, name='f_next', profile=profile)
print 'Done'
return f_next
def build_encoder(tparams, options):
"""
Computation graph, encoder only
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
x = tensor.matrix('x', dtype='int64')
x_mask = tensor.matrix('x_mask', dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# encoder
proj = get_layer(options['encoder'])[1](tparams, emb, None, options,
prefix='encoder',
mask=x_mask)
ctx = proj[0][-1]
return trng, x, x_mask, ctx, emb
def build_image_encoder(tparams, options):
"""
Encoder only, for images
"""
opt_ret = dict()
trng = RandomStreams(1234)
# image features
im = tensor.matrix('im', dtype='float32')
# Encode images
images_mm = get_layer('ff')[1](tparams,
im,
options,
prefix='ff_image_mm',
activ='linear')
if not 'attention_type' in options or options['attention_type'] == 'dot':
images_mm = l2norm(images_mm)
if options['use_dropout']:
images_mm *= shared_dropout_layer(
(n_samples, options['dim_multimodal']), use_noise, trng,
retain_probability_hidden)
return trng, [im], images_mm
def init_mlp_params(params,
dim=1,
dim_v=4096,
init='glorot',
gain=1.,
**kwargs):
"""
Initialize parameters for the scoring function, an MLP.
:param params:
:param dim:
:param init: glorot initialization or uniform or normal
:param gain: parameter for glorot initializer
:param kwargs:
:return:
"""
logger.warn('MLP - init: {} gain: {}'.format(init, gain))
dim_in = 2 * dim # we concatenate the forward and backward RNN
dim_out = dim_v
# The first layer goes from the bi-directional concatenation
# to the visual feature vector dimensionality. e.g. 2 x 1000 -> 4,096
params = get_layer('ff')[0](params,
prefix='mlp',
nin=dim_in,
nout=dim_out,
ortho=False,
init=init,
gain=gain)
return params
def build_sentence_encoder(tparams, options):
"""
Encoder only, for sentences
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
x = tensor.matrix('x', dtype='int64')
mask = tensor.matrix('x_mask', dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Word embedding
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# Encode sentences
if options['encoder'] == 'bow':
sents = (emb * mask[:, :, None]).sum(0)
else:
proj = get_layer(options['encoder'])[1](tparams, emb, None, options,
prefix='encoder',
mask=mask)
sents = proj[0][-1]
sents = l2norm(sents)
return trng, [x, mask], sents
def build_encoder(tparams, options):
"""
Computation graph, encoder only
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
x = tensor.matrix('x', dtype='int64')
x_mask = tensor.matrix('x_mask', dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape(
[n_timesteps, n_samples, options['dim_word']])
# encoder
proj = get_layer(options['encoder'])[1](tparams,
emb,
None,
options,
prefix='encoder',
mask=x_mask)
ctx = proj[0][-1]
return trng, x, x_mask, ctx, emb
def encode_images(tparams, options, im):
im_emb = get_layer('ff')[1](tparams, im, options, prefix='ff_image', activ='linear')
im_emb = l2norm(im_emb)
if options['abs']:
im_emb = abs(im_emb)
return im_emb
def build_sentence_encoder(tparams, options):
"""
Encoder only, for sentences
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
x = tensor.matrix('x', dtype='int64')
mask = tensor.matrix('x_mask', dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Word embedding
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# Encode sentences
if options['encoder'] == 'bow':
sents = (emb * mask[:,:,None]).sum(0)
else:
proj = get_layer(options['encoder'])[1](tparams, emb, None, options,
prefix='encoder',
mask=mask)
sents = proj[0][-1]
sents = l2norm(sents)
return trng, [x, mask], sents
def encode_images(tparams, options, im):
im_emb = get_layer('ff')[1](tparams, im, options, prefix='ff_image', activ='linear')
im_emb = l2norm(im_emb)
if options['abs']:
im_emb = abs(im_emb)
return im_emb
def init_params(options):
"""
Initialize all parameters
"""
params = OrderedDict()
# Word embedding
params['Wemb'] = norm_weight(options['n_words'], options['dim_word'])
# Sentence encoder
params = get_layer(options['encoder'])[0](options, params, prefix='encoder',
nin=options['dim_word'], dim=options['dim'])
# Image encoder
params = get_layer('ff')[0](options, params, prefix='ff_image', nin=options['dim_image'], nout=options['dim'])
return params
def init_params(options, preemb=None):
"""
Initialize all parameters
"""
params = OrderedDict()
# Word embedding
if preemb == None:
params['Wemb'] = norm_weight(options['n_words'], options['dim_word'])
else:
params['Wemb'] = preemb
# init state
params = get_layer('ff')[0](options,
params,
prefix='ff_state',
nin=options['dimctx'],
nout=options['dim'])
# Decoder
params = get_layer(options['decoder'])[0](options,
params,
prefix='decoder',
nin=options['dim_word'],
dim=options['dim'])
# Output layer
if options['doutput']:
params = get_layer('ff')[0](options,
params,
prefix='ff_hid',
nin=options['dim'],
nout=options['dim_word'])
params = get_layer('ff')[0](options,
params,
prefix='ff_logit',
nin=options['dim_word'],
nout=options['n_words'])
else:
params = get_layer('ff')[0](options,
params,
prefix='ff_logit',
nin=options['dim'],
nout=options['n_words'])
return params
def build_sampler(tparams, options, trng):
"""
Forward sampling
"""
ctx = tensor.matrix('ctx', dtype='float32')
ctx0 = ctx
print 'Building f_init...',
init_state = get_layer('ff')[1](tparams, ctx, options, prefix='ff_state', activ='tanh')
f_init = theano.function([ctx], init_state, name='f_init', profile=False)
# x: 1 x 1
y = tensor.vector('y_sampler', dtype='int64')
init_state = tensor.matrix('init_state', dtype='float32')
# if it's the first word, emb should be all zero
emb = tensor.switch(y[:,None] < 0, tensor.alloc(0., 1, tparams['Wemb'].shape[1]),
tparams['Wemb'][y])
# decoder
proj = get_layer(options['decoder'])[1](tparams, emb, init_state, options,
prefix='decoder',
mask=None,
one_step=True)
next_state = proj[0]
if next_state.ndim > 2:
next_state = next_state[0]
# output
if options['doutput']:
hid = get_layer('ff')[1](tparams, next_state, options, prefix='ff_hid', activ='tanh')
logit = get_layer('ff')[1](tparams, hid, options, prefix='ff_logit', activ='linear')
else:
logit = get_layer('ff')[1](tparams, next_state, options, prefix='ff_logit', activ='linear')
next_probs = tensor.nnet.softmax(logit)
next_sample = trng.multinomial(pvals=next_probs).argmax(1)
# next word probability
print 'Building f_next..',
inps = [y, init_state]
outs = [next_probs, next_sample, next_state]
f_next = theano.function(inps, outs, name='f_next', profile=False)
print 'Done'
return f_init, f_next
def init_params(options):
"""
Initialize all parameters
"""
params = OrderedDict()
# Word embedding
params['Wemb'] = norm_weight(options['n_words'], options['dim_word'])
# Encoder
params = get_layer(options['encoder'])[0](options,
params,
prefix='encoder',
nin=options['dim_word'],
dim=options['dim'])
# Variational layer
params = get_layer(options['vae'])[0](options,
params,
prefix='vae',
nhid=options['vae_nhid'],
nlatent=options['vae_nlatent'],
ndim=options['dim'])
# Decoder: next sentence
params = get_layer(options['decoder'])[0](options,
params,
prefix='decoder_f',
nin=options['dim_word'],
dim=options['dim'])
# Decoder: previous sentence
params = get_layer(options['decoder'])[0](options,
params,
prefix='decoder_b',
nin=options['dim_word'],
dim=options['dim'])
# Output layer
params = get_layer('ff')[0](options,
params,
prefix='ff_logit',
nin=options['dim'],
nout=options['n_words'])
return params
def init_params(options):
"""
Initialize all parameters
"""
params = OrderedDict()
# Word embedding
params['Wemb'] = norm_weight(options['n_words'], options['dim_word'])
# Sentence encoder
if options['encoder'] != 'bow':
params = get_layer(options['encoder'])[0](options, params, prefix='encoder',
nin=options['dim_word'], dim=options['dim'])
# Image encoder
params = get_layer('ff')[0](options, params, prefix='ff_image', nin=options['dim_image'], nout=options['dim'])
return params
def build_sampler(tparams, options, trng):
"""
Forward sampling
"""
ctx = tensor.matrix('ctx', dtype='float32')
ctx0 = ctx
print 'Building f_init...',
init_state = get_layer('ff')[1](tparams, ctx, options, prefix='ff_state', activ='tanh')
f_init = theano.function([ctx], init_state, name='f_init', profile=False)
# x: 1 x 1
y = tensor.vector('y_sampler', dtype='int64')
init_state = tensor.matrix('init_state', dtype='float32')
# if it's the first word, emb should be all zero
emb = tensor.switch(y[:,None] < 0, tensor.alloc(0., 1, tparams['Wemb'].shape[1]),
tparams['Wemb'][y])
# decoder
proj = get_layer(options['decoder'])[1](tparams, emb, init_state, options,
prefix='decoder',
mask=None,
one_step=True)
next_state = proj[0]
# output
if options['doutput']:
hid = get_layer('ff')[1](tparams, next_state, options, prefix='ff_hid', activ='tanh')
logit = get_layer('ff')[1](tparams, hid, options, prefix='ff_logit', activ='linear')
else:
logit = get_layer('ff')[1](tparams, next_state, options, prefix='ff_logit', activ='linear')
next_probs = tensor.nnet.softmax(logit)
next_sample = trng.multinomial(pvals=next_probs).argmax(1)
# next word probability
print 'Building f_next..',
inps = [y, init_state]
outs = [next_probs, next_sample, next_state]
f_next = theano.function(inps, outs, name='f_next', profile=False)
print 'Done'
return f_init, f_next
def build_model(tparams, options):
"""
Computation graph for the model
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
x = tensor.matrix('x', dtype='int64')
mask = tensor.matrix('mask', dtype='float32')
im = tensor.matrix('im', dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape(
[n_timesteps, n_samples, options['dim_word']])
# Encode sentences (source)
if options['encoder'] == 'bow':
sents = (emb * mask[:, :, None]).sum(0)
else:
proj = get_layer(options['encoder'])[1](tparams,
emb,
None,
options,
prefix='encoder',
mask=mask)
sents = proj[0][-1]
sents = l2norm(sents)
# Encode images (source)
images = get_layer('ff')[1](tparams,
im,
options,
prefix='ff_image',
activ='linear')
# Compute loss
cost = contrastive_loss(options['margin'], images, sents)
return trng, [x, mask, im], cost
def encode_sentences_with_topicvector(tparams, options, x, mask, topics):
n_timesteps = x.shape[0]
n_samples = x.shape[1]
t2gru_emb = get_layer('ff')[1](tparams, topics, options, prefix='ff_topic_vector1_emb_gru', activ='linear')
# Word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# Encode sentences (source)
proj = get_layer(options['encoder'])[1](tparams, emb, t2gru_emb * 0.1, options,
prefix='encoder',
mask=mask)
s = proj[0][-1]
#if options['v_norm'] == 'l2' :
s = l2norm(s)
#s = maxnorm2(s)
if options['abs']:
#s = abs(s)
s = tensor.maximum(s, 0)
return s
def encode_images(tparams, options, im):
im_emb = get_layer('ff')[1](tparams, im, options, prefix='ff_image', activ='linear')
#if options['v_norm'] == 'l2' :
im_emb = l2norm(im_emb)
#im_emb = maxnorm2(im_emb)
if options['abs']:
#im_emb = abs(im_emb)
im_emb = tensor.maximum(im_emb, 0)
return im_emb
def encode_topic_vector2(tparams, options, topics):
t_emb = get_layer('ff')[1](tparams, topics, options, prefix='ff_topic_vector2', activ='linear')
t_emb = l2norm(t_emb)
#t_emb = maxnorm2(t_emb)
if options['abs']:
#im_emb = abs(im_emb)
t_emb = tensor.maximum(t_emb, 0)
return t_emb
def init_params(options, preemb=None):
"""
Initialize all parameters
"""
params = OrderedDict()
# Word embedding
if preemb == None:
params['Wemb'] = norm_weight(options['n_words'], options['dim_word'])
else:
print('using preemb')
params['Wemb'] = preemb
# Encoder
params = get_layer(options['encoder'])[0](options,
params,
prefix='encoder',
nin=options['dim_word'],
dim=options['dim'])
# Decoder: next sentence
params = get_layer(options['decoder'])[0](options,
params,
prefix='decoder_f',
nin=options['dim_word'],
dim=options['dim'])
# Decoder: previous sentence
params = get_layer(options['decoder'])[0](options,
params,
prefix='decoder_b',
nin=options['dim_word'],
dim=options['dim'])
# Output layer
params = get_layer('ff')[0](options,
params,
prefix='ff_logit',
nin=options['dim'],
nout=options['n_words'])
return params
def build_model(tparams, options):
"""
Computation graph for the model
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
x = tensor.matrix('x', dtype='int64')
mask = tensor.matrix('mask', dtype='float32')
im = tensor.matrix('im', dtype='float32')
con = tensor.matrix('con', dtype='int64')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# Encode sentences (source)
proj = get_layer(options['encoder'])[1](tparams, emb, None, options,
prefix='encoder',
mask=mask)
sents = proj[0][-1]
sents = l2norm(sents)
# Encode images (source)
images = get_layer('ff')[1](tparams, im, options, prefix='ff_image', activ='linear')
# Compute loss
cost, updates = theano.scan(_step,
sequences=con,
outputs_info=tensor.alloc(0.),
non_sequences = [sents, images, options['margin']],
n_steps=con.shape[0],
profile=False,
strict=True)
cost = cost[-1]
return trng, [x, mask, im, con], cost
def init_params(options):
"""
Initialize all parameters
"""
params = OrderedDict()
# if using bidirectional RNN,
# forward and backward embeddings are half the final MM embedding size because
# they will be concatenated to form the sentence embedding
sent_dim = int(options['dim'])//2 if options['bidirectional_enc'] else int(options['dim'])
langs = options['langs']
for lang in langs:
# word embeddings
params['Wemb_%s'%lang] = norm_weight(options['n_words_%s'%lang], options['dim_word'])
# encoder type (currently 'bow', 'gru' or 'lstm')
if options['encoder_%s'%lang] != 'bow':
for i in range(int(options['n_enc_hidden_layers'])):
layer_name_prefix='encoder_%s_%i'%(lang,i)
# first hidden layer has input word embeddings, next layers have input (hidden) sentence embeddings
nin=options['dim_word'] if i==0 else sent_dim
params = get_layer(options['encoder_%s'%lang])[0](options, params, prefix=layer_name_prefix,
nin=nin, dim=sent_dim)
if options['bidirectional_enc']:
for i in range(int(options['n_enc_hidden_layers'])):
layer_name_prefix='encoder_%s_r_%i'%(lang,i)
# first hidden layer has input word embeddings, next layers have input (hidden) sentence embeddings
nin=options['dim_word'] if i==0 else sent_dim
params = get_layer(options['encoder_%s'%lang])[0](options, params, prefix=layer_name_prefix,
nin=nin, dim=sent_dim)
# Image encoder
params = get_layer('ff')[0](options, params, prefix='ff_image', nin=options['dim_image'], nout=options['dim'])
return params
def encode_sentences(tparams, options, x, mask):
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# Encode sentences (source)
proj = get_layer(options['encoder'])[1](tparams, emb, None, options,
prefix='encoder',
mask=mask)
s = l2norm(proj[0][-1])
if options['abs']:
s = abs(s)
return s
def encode_sentences(tparams, options, x, mask):
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# Encode sentences (source)
proj = get_layer(options['encoder'])[1](tparams, emb, None, options,
prefix='encoder',
mask=mask)
s = l2norm(proj[0][-1])
if options['abs']:
s = abs(s)
return s
def build_image_encoder(tparams, options):
"""
Encoder only, for images
"""
opt_ret = dict()
trng = RandomStreams(1234)
# image features
im = tensor.matrix('im', dtype='float32')
# Encode images
images = get_layer('ff')[1](tparams, im, options, prefix='ff_image', activ='linear')
images = l2norm(images)
return trng, [im], images
def build_image_encoder(tparams, options):
"""
Encoder only, for images
"""
opt_ret = dict()
trng = RandomStreams(1234)
# image features
im = tensor.matrix('im', dtype='float32')
# Encode images
images = get_layer('ff')[1](tparams, im, options, prefix='ff_image', activ='linear')
images = l2norm(images)
return trng, [im], images
def build_ar_layers(x, tparams, options, features, hiddens):
u1, u2 = (x[:, 0, :, :], x[:, 1, :, :].transpose(0, 2, 1))
h1, h2 = hiddens
# concatinate the payoff matrix onto the final layer hidden units
utility = (tensor.concatenate((u1.reshape(
(u1.shape[0], 1, u1.shape[1], u1.shape[2])), h1),
axis=1),
tensor.concatenate((u2.reshape(
(u2.shape[0], 1, u2.shape[1], u2.shape[2])), h2),
axis=1))
ar_layers = options['ar_layers']
ar_lists = ([], [])
opp = [None, None]
weighted_feature_list = ([], [])
br_list = ([], [])
for i in range(ar_layers):
for p in range(2):
if i == (ar_layers - 1) and p == 1:
continue # don't build ar layer for pl 2 in the last layer
feat = features[p]
ar, weighted_features, br = get_layer('ar')[1](tparams,
feat,
options,
payoff=utility[p],
prefix='p%d_ar%d' %
(p, i),
opposition=opp[p],
level=i)
n, d = ar.shape
ar = ar.reshape((n, 1, d)) # make space to concat ar layers
weighted_feature_list[p].append(weighted_features)
if i == 0:
ar_lists[p].append(ar)
else:
ar_lists[p].append(
tensor.concatenate((ar_lists[p][i - 1], ar), axis=1))
br_list[p].append(br)
# append each layer then update the opposition variable...
if i < ar_layers - 1:
for p in range(2):
opp[1 - p] = ar_lists[p][i]
# return ar_lists[0][ar_layers-1]
return ar_lists, weighted_feature_list, br_list
def build_encoder_w2v(tparams, options):
"""
Computation graph for encoder, given pre-trained word embeddings
"""
opt_ret = dict()
trng = RandomStreams(1234)
# word embedding (source)
embedding = tensor.tensor3('embedding', dtype='float32')
x_mask = tensor.matrix('x_mask', dtype='float32')
# encoder
proj = get_layer(options['encoder'])[1](tparams,
embedding,
None,
options,
prefix='encoder',
mask=x_mask)
ctx = proj[0][-1]
return trng, embedding, x_mask, ctx
def __init__(self, net_config, data_cache):
#net_config_str = json.dumps(net_config, sort_keys=True, indent=4, separators=(',', ':'))
#LOGGER.info("\n" + net_config_str)
self.data_cache = data_cache
self.inputs = net_config.get("inputs", [])
self.check_net_config(net_config)
self.outputs = net_config.get("outputs", [])
# layers
self.layers = []
self.name2layer = {}
model_disk_base = net_config.get("model_cache",
{}).get("disk_base", None)
for layer_config in net_config["layers"]:
layer = get_layer(layer_config, self.data_cache)
layer.model_disk_base = model_disk_base
self.layers.append(layer)
self.name2layer[layer.name] = layer
def build_encoder_w2v(tparams, options):
"""
Computation graph for encoder, given pre-trained word embeddings
"""
opt_ret = dict()
trng = RandomStreams(1234)
# word embedding (source)
embedding = tensor.tensor3('embedding', dtype='float32')
x_mask = tensor.matrix('x_mask', dtype='float32')
# encoder
proj = get_layer(options['encoder'])[1](tparams, embedding, None, options,
prefix='encoder',
mask=x_mask)
ctx = proj[0][-1]
return trng, embedding, x_mask, ctx
def build_mlp_predict(self,
enc_states,
dim_emb=0,
dim=0,
activation_mlp='relu',
**kwargs):
"""
Builds an MLP scoring function for use during prediction / test time.
We want this to predict a single 4096d vector for each input.
:param dim_emb:
:param dim:
:param activation_mlp:
:param kwargs:
:return:
TODO: Redefine the predict function so it predicts the embedding of
the sentence and the image embedding.
"""
# set MLP activation function
assert activation_mlp in ('relu', 'tanh'), \
'MLP activation function must be tanh or relu'
activation_mlp = 'lambda x: tensor.nnet.relu(x)' \
if activation_mlp == 'relu' else 'lambda x: tensor.tanh(x)'
logger.warn('Using MLP activation function: {}'.format(activation_mlp))
theano_params = self.theano_params
# The input to the MLP will be the mean value of the hidden states for
# each instance in the minibatch.
if kwargs['verbose']:
logger.warn(states_mean.tag.test_value)
# train a single layer MLP to do everything
output = get_layer('ff')[1](theano_params,
enc_states,
prefix='mlp',
activ=activation_mlp)
return output
def build_model(tparams, options, rng=None):
"""
Computation graph for the model
"""
if rng is None:
rng = numpy.random.RandomState(123)
trng = RandomStreams(rng.randint(1000000))
use_noise = theano.shared(numpy.float32(0.))
x = tensor.tensor4('x')
own_features, hidden1 = build_features(x, tparams, options, use_noise,
trng)
opp_features, hidden2 = build_features(
x.transpose(
(0, 1, 3, 2))[:, [1, 0], :, :], # transpose to get player 2 model
tparams,
options,
use_noise,
trng)
ar, weighted_feature_list, br_list = build_ar_layers(
x, tparams, options, (own_features, opp_features),
(hidden1[-1], hidden2[-1]))
ar_layers = options['ar_layers']
out = get_layer('output')[1](tparams, ar[0][ar_layers - 1], options)
intermediate_fns = {
'ar': ar,
'own_features': own_features,
'opp_features': opp_features,
'hidden1': hidden1,
'hidden2': hidden2,
'weighted_feature_list': weighted_feature_list,
'br_list': br_list
}
if not options['debug']:
return trng, use_noise, x, out
else:
return trng, use_noise, x, out, intermediate_fns
def build_model(tparams, options):
"""
Computation graph for the model
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
# x: current sentence
# y: next sentence
# z: previous sentence
x = tensor.matrix('x', dtype='int64')
x_mask = tensor.matrix('x_mask', dtype='float32')
y = tensor.matrix('y', dtype='int64')
y_mask = tensor.matrix('y_mask', dtype='float32')
z = tensor.matrix('z', dtype='int64')
z_mask = tensor.matrix('z_mask', dtype='float32')
n_timesteps = x.shape[0]
n_timesteps_f = y.shape[0]
n_timesteps_b = z.shape[0]
n_samples = x.shape[1]
# Word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape([n_timesteps, n_samples, options['dim_word']])
# encoder
proj = get_layer(options['encoder'])[1](tparams, emb, None, options,
prefix='encoder',
mask=x_mask)
ctx = proj[0][-1]
dec_ctx = ctx
# Word embedding (ahead)
embf = tparams['Wemb'][y.flatten()].reshape([n_timesteps_f, n_samples, options['dim_word']])
embf_shifted = tensor.zeros_like(embf)
embf_shifted = tensor.set_subtensor(embf_shifted[1:], embf[:-1])
embf = embf_shifted
# Word embedding (behind)
embb = tparams['Wemb'][z.flatten()].reshape([n_timesteps_b, n_samples, options['dim_word']])
embb_shifted = tensor.zeros_like(embb)
embb_shifted = tensor.set_subtensor(embb_shifted[1:], embb[:-1])
embb = embb_shifted
# decoder (ahead)
projf = get_layer(options['decoder'])[1](tparams, embf, dec_ctx, options,
prefix='decoder_f',
mask=y_mask)
# decoder (behind)
projb = get_layer(options['decoder'])[1](tparams, embb, dec_ctx, options,
prefix='decoder_b',
mask=z_mask)
# compute word probabilities (ahead)
logit = get_layer('ff')[1](tparams, projf[0], options, prefix='ff_logit', activ='linear')
logit_shp = logit.shape
probs = tensor.nnet.softmax(logit.reshape([logit_shp[0]*logit_shp[1], logit_shp[2]]))
# cost (ahead)
y_flat = y.flatten()
y_flat_idx = tensor.arange(y_flat.shape[0]) * options['n_words'] + y_flat
costf = -tensor.log(probs.flatten()[y_flat_idx]+1e-8)
costf = costf.reshape([y.shape[0],y.shape[1]])
costf = (costf * y_mask).sum(0)
costf = costf.sum()
# compute word probabilities (behind)
logit = get_layer('ff')[1](tparams, projb[0], options, prefix='ff_logit', activ='linear')
logit_shp = logit.shape
probs = tensor.nnet.softmax(logit.reshape([logit_shp[0]*logit_shp[1], logit_shp[2]]))
# cost (behind)
z_flat = z.flatten()
z_flat_idx = tensor.arange(z_flat.shape[0]) * options['n_words'] + z_flat
costb = -tensor.log(probs.flatten()[z_flat_idx]+1e-8)
costb = costb.reshape([z.shape[0],z.shape[1]])
costb = (costb * z_mask).sum(0)
costb = costb.sum()
# total cost
cost = costf + costb
return trng, x, x_mask, y, y_mask, z, z_mask, opt_ret, cost
def build_model(tparams, options):
"""
Computation graph for the model
"""
opt_ret = dict()
trng = RandomStreams(1234)
# description string: #words x #samples
# x: current sentence
# y: next sentence
# z: previous sentence
x = tensor.matrix('x', dtype='int64')
x_mask = tensor.matrix('x_mask', dtype='float32')
y = tensor.matrix('y', dtype='int64')
y_mask = tensor.matrix('y_mask', dtype='float32')
z = tensor.matrix('z', dtype='int64')
z_mask = tensor.matrix('z_mask', dtype='float32')
n_timesteps = x.shape[0]
n_timesteps_f = y.shape[0]
n_timesteps_b = z.shape[0]
n_samples = x.shape[1]
# Word embedding (source)
emb = tparams['Wemb'][x.flatten()].reshape(
[n_timesteps, n_samples, options['dim_word']])
# encoder
proj = get_layer(options['encoder'])[1](tparams,
emb,
None,
options,
prefix='encoder',
mask=x_mask)
ctx = proj[0][-1]
dec_ctx = ctx
# Word embedding (ahead)
embf = tparams['Wemb'][y.flatten()].reshape(
[n_timesteps_f, n_samples, options['dim_word']])
embf_shifted = tensor.zeros_like(embf)
embf_shifted = tensor.set_subtensor(embf_shifted[1:], embf[:-1])
embf = embf_shifted
# Word embedding (behind)
embb = tparams['Wemb'][z.flatten()].reshape(
[n_timesteps_b, n_samples, options['dim_word']])
embb_shifted = tensor.zeros_like(embb)
embb_shifted = tensor.set_subtensor(embb_shifted[1:], embb[:-1])
embb = embb_shifted
# decoder (ahead)
projf = get_layer(options['decoder'])[1](tparams,
embf,
dec_ctx,
options,
prefix='decoder_f',
mask=y_mask)
# decoder (behind)
projb = get_layer(options['decoder'])[1](tparams,
embb,
dec_ctx,
options,
prefix='decoder_b',
mask=z_mask)
# compute word probabilities (ahead)
logit = get_layer('ff')[1](tparams,
projf[0],
options,
prefix='ff_logit',
activ='linear')
logit_shp = logit.shape
probs = tensor.nnet.softmax(
logit.reshape([logit_shp[0] * logit_shp[1], logit_shp[2]]))
# cost (ahead)
y_flat = y.flatten()
y_flat_idx = tensor.arange(y_flat.shape[0]) * options['n_words'] + y_flat
costf = -tensor.log(probs.flatten()[y_flat_idx] + 1e-8)
costf = costf.reshape([y.shape[0], y.shape[1]])
costf = (costf * y_mask).sum(0)
costf = costf.sum()
# compute word probabilities (behind)
logit = get_layer('ff')[1](tparams,
projb[0],
options,
prefix='ff_logit',
activ='linear')
logit_shp = logit.shape
probs = tensor.nnet.softmax(
logit.reshape([logit_shp[0] * logit_shp[1], logit_shp[2]]))
# cost (behind)
z_flat = z.flatten()
z_flat_idx = tensor.arange(z_flat.shape[0]) * options['n_words'] + z_flat
costb = -tensor.log(probs.flatten()[z_flat_idx] + 1e-8)
costb = costb.reshape([z.shape[0], z.shape[1]])
costb = (costb * z_mask).sum(0)
costb = costb.sum()
# total cost
cost = costf + costb
return trng, x, x_mask, y, y_mask, z, z_mask, opt_ret, cost
def init_params(options):
"""
Initialize all parameters
"""
params = OrderedDict()
# if using bidirectional RNN,
# forward and backward embeddings are half the final MM embedding size because
# they will be concatenated to form the sentence embedding
#sent_dim = int(options['dim'])//2 if options['bidirectional_enc'] else int(options['dim'])
sent_dim = options['dim']
ctx_dim = options['dim'] # context vector in case of mono RNN encoder
langs = options['langs']
for idx, lang in enumerate(langs):
# word embeddings
params['Wemb_%s' % lang] = norm_weight(options['n_words_%s' % lang],
options['dim_word'])
# encoder type (currently 'bow', 'gru' or 'lstm')
if options['encoder_%s' % lang] != 'bow':
for i in range(int(options['n_enc_hidden_layers'])):
layer_name_prefix = 'encoder_%s_%i' % (lang, i)
# first hidden layer has input word embeddings, next layers have input (hidden) sentence embeddings
nin = options['dim_word'] if i == 0 else sent_dim
params = get_layer(options['encoder_%s' % lang])[0](
options,
params,
prefix=layer_name_prefix,
nin=nin,
dim=sent_dim)
if options['bidirectional_enc']:
ctx_dim = 2 * options[
'dim'] # context vector in case of biRNN encoder
for i in range(int(options['n_enc_hidden_layers'])):
layer_name_prefix = 'encoder_%s_r_%i' % (lang, i)
# first hidden layer has input word embeddings, next layers have input (hidden) sentence embeddings
nin = options['dim_word'] if i == 0 else sent_dim
params = get_layer(options['encoder_%s' % lang])[0](
options,
params,
prefix=layer_name_prefix,
nin=nin,
dim=sent_dim)
# if using general attention, create matrices for each possible
# sentence-image and sentence-sentence pairs
if 'attention_type' in options and options[
'attention_type'] == 'general':
# sentence_lang-image mapping
params['image_sentence_%i_mapping'%idx] = \
norm_weight(options['dim_multimodal'], options['dim_multimodal'], ortho=False)
# sentence_lang1 - sentence_langN mappings
for idx1, lang1 in enumerate(langs):
if idx == idx1 or idx1 <= idx: continue
params['sentence_%i_sentence_%i_mapping'%(idx,idx1)] = \
norm_weight(options['dim_multimodal'], options['dim_multimodal'], ortho=False)
# Sentence-multimodal projection
params = get_layer('ff')[0](options,
params,
prefix='ff_sentence_mm',
nin=ctx_dim,
nout=options['dim_multimodal'])
# Image-multimodal projection
params = get_layer('ff')[0](options,
params,
prefix='ff_image_mm',
nin=options['dim_image'],
nout=options['dim_multimodal'])
return params
def build_model(tparams, options):
"""
Computation graph for the model
"""
opt_ret = dict()
use_noise = theano.shared(numpy.asarray(1., dtype=theano.config.floatX))
try:
trng = RandomStreams(1234, use_cuda=True)
except:
print "Could not apply use_cuda==True in RandonStreams ..."
trng = RandomStreams(1234)
xs = []
xmasks = []
langs = options['langs']
for lang in langs:
# description string: #words x #samples
x_lang = tensor.matrix('x_%s' % lang, dtype='int64')
mask_lang = tensor.matrix('mask_%s' % lang, dtype='float32')
xs.append(x_lang)
xmasks.append(mask_lang)
xs_r = []
xmasks_r = []
if options['bidirectional_enc']:
for i, lang in enumerate(langs):
x_lang = xs[i]
mask_lang = xmasks[i]
# reverse
x_lang_r = x_lang[::-1]
mask_lang_r = mask_lang[::-1]
xs_r.append(x_lang_r)
xmasks_r.append(mask_lang_r)
sents_all = []
im = tensor.matrix('im', dtype='float32')
n_samples = im.shape[0]
for i, lang in enumerate(langs):
x_lang = xs[i]
mask_lang = xmasks[i]
n_timesteps_lang = x_lang.shape[0]
n_samples_lang = x_lang.shape[1]
if options['use_dropout']:
# dropout probs for the word embeddings
retain_probability_emb = 1 - options['dropout_embedding']
# dropout probs for the RNN hidden states
retain_probability_hidden = 1 - options['dropout_hidden']
# dropout probs for the source words
retain_probability_source = 1 - options['dropout_source']
# hidden states
rec_dropout = shared_dropout_layer(
(2, n_samples_lang, options['dim']), use_noise, trng,
retain_probability_hidden)
rec_dropout_r = shared_dropout_layer(
(2, n_samples_lang, options['dim']), use_noise, trng,
retain_probability_hidden)
# word embeddings
emb_dropout = shared_dropout_layer(
(2, n_samples_lang, options['dim_word']), use_noise, trng,
retain_probability_emb)
emb_dropout_r = shared_dropout_layer(
(2, n_samples_lang, options['dim_word']), use_noise, trng,
retain_probability_emb)
# source words
source_dropout = shared_dropout_layer(
(n_timesteps_lang, n_samples_lang, 1), use_noise, trng,
retain_probability_source)
source_dropout = tensor.tile(source_dropout,
(1, 1, options['dim_word']))
else:
# hidden states
rec_dropout = theano.shared(numpy.array([1.] * 2, dtype='float32'))
rec_dropout_r = theano.shared(
numpy.array([1.] * 2, dtype='float32'))
# word embeddings
emb_dropout = theano.shared(numpy.array([1.] * 2, dtype='float32'))
emb_dropout_r = theano.shared(
numpy.array([1.] * 2, dtype='float32'))
# Word embedding (for a particular language `lang`)
# forward
emb_lang = tparams['Wemb_%s' % lang][x_lang.flatten()]
emb_lang = emb_lang.reshape(
[n_timesteps_lang, n_samples_lang, options['dim_word']])
if options['use_dropout']:
emb_lang *= source_dropout
if options['bidirectional_enc']:
x_lang_r = xs_r[i]
mask_lang_r = xmasks_r[i]
# backward lang encoder
emb_lang_r = tparams['Wemb_%s' % lang][x_lang_r.flatten()]
emb_lang_r = emb_lang_r.reshape(
[n_timesteps_lang, n_samples_lang, options['dim_word']])
if options['use_dropout']:
emb_lang_r *= source_dropout[::-1]
# Encode sentence in language `lang`
if options['encoder_%s' % lang] == 'bow':
sents_lang = (emb_lang * mask_lang[:, :, None]).sum(0)
else:
# iteratively push input from first hidden layer until the last
for i in range(int(options['n_enc_hidden_layers'])):
layer_name_prefix = 'encoder_%s_%i' % (lang, i)
# if first hidden layer use wembs, otherwise output of previous hidden layer
layer_below = emb_lang if i == 0 else layer_below[0]
# do not apply dropout on word embeddings layer
#if options['use_dropout'] and i>0:
# layer_below = dropout_layer(layer_below, use_noise, trng, prob=options['dropout_prob'])
layer_below = get_layer(options['encoder_%s' % lang])[1](
tparams,
layer_below,
options,
None,
prefix=layer_name_prefix,
mask=mask_lang,
emb_dropout=emb_dropout,
rec_dropout=rec_dropout)
if i == int(options['n_enc_hidden_layers']) - 1:
# sentence embeddings (projections) are the output of the last hidden layer
proj_lang = layer_below
# apply forward and backward steps and concatenate both
if options['bidirectional_enc']:
# concatenate forward and backward pass RNNs
# iteratively push input from first hidden layer until the last
for i in range(int(options['n_enc_hidden_layers'])):
layer_name_prefix = 'encoder_%s_r_%i' % (lang, i)
# if first hidden layer use wembs, else output of prev hidden layer
layer_below = emb_lang_r if i == 0 else layer_below[0]
# do not apply dropout on word embeddings layer
#if options['use_dropout'] and i>0:
# layer_below = dropout_layer(layer_below, use_noise, trng, prob=options['dropout_prob'])
layer_below = get_layer(options['encoder_%s' % lang])[1](
tparams,
layer_below,
options,
None,
prefix=layer_name_prefix,
mask=mask_lang_r,
emb_dropout=emb_dropout_r,
rec_dropout=rec_dropout_r)
if i == int(options['n_enc_hidden_layers']) - 1:
# sentence embeddings (projections) are the output of the last hidden layer
proj_lang_r = layer_below
# use the last state of forward + backward encoder rnns
sents_lang = concatenate(
[proj_lang[0][-1], proj_lang_r[0][-1]],
axis=proj_lang[0].ndim - 2)
else:
sents_lang = proj_lang[0][-1]
if options['use_dropout']:
sents_lang *= shared_dropout_layer(
(n_samples_lang, options['dim']), use_noise, trng,
retain_probability_hidden)
# project sentences into multimodal space
sents_mm = get_layer('ff')[1](tparams,
sents_lang,
options,
prefix='ff_sentence_mm',
activ='linear')
if options['attention_type'] == 'dot':
sents_mm = l2norm(sents_mm)
if options['use_dropout']:
sents_mm *= shared_dropout_layer(
(n_samples_lang, options['dim_multimodal']), use_noise, trng,
retain_probability_hidden)
sents_all.append(sents_mm)
# Encode images
images = get_layer('ff')[1](tparams,
im,
options,
prefix='ff_image_mm',
activ='linear')
if options['attention_type'] == 'dot':
images = l2norm(images)
if options['use_dropout']:
images *= shared_dropout_layer((n_samples, options['dim_multimodal']),
use_noise, trng,
retain_probability_hidden)
# Compute loss
lambda_img_sent = options['lambda_img_sent']
lambda_sent_sent = options['lambda_sent_sent']
if options['use_all_costs']:
cost = contrastive_loss_all(tparams, options, images, sents_all,
lambda_img_sent, lambda_sent_sent)
else:
cost = contrastive_loss(tparams, options, images, sents_all)
# return flattened inputs
inps = []
inps.extend(xs)
inps.extend(xmasks)
inps.append(im)
return trng, inps, cost
def build_sentence_encoders(tparams, options):
"""
Sentence encoder only to be used at test time
"""
opt_ret = dict()
trng = RandomStreams(1234)
#xs, masks, sents_all = [], [], []
in_outs = []
langs = options['langs']
for lang in langs:
# description string: #words x #samples
# forward
x = tensor.matrix('x_%s' % lang, dtype='int64')
mask = tensor.matrix('x_mask_%s' % lang, dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
# Word embedding (forward)
emb = tparams['Wemb_%s' % lang][x.flatten()].reshape(
[n_timesteps, n_samples, options['dim_word']])
if options['bidirectional_enc']:
# backward RNN
x_r = x[::-1]
mask_r = mask[::-1]
emb_r = tparams['Wemb_%s' % lang][x_r.flatten()].reshape(
[n_timesteps, n_samples, options['dim_word']])
if options['use_dropout']:
retain_probability_emb = 1 - options['dropout_embedding']
retain_probability_hidden = 1 - options['dropout_hidden']
retain_probability_source = 1 - options['dropout_source']
rec_dropout = theano.shared(
numpy.array([retain_probability_hidden] * 2, dtype='float32'))
rec_dropout_r = theano.shared(
numpy.array([retain_probability_hidden] * 2, dtype='float32'))
emb_dropout = theano.shared(
numpy.array([retain_probability_emb] * 2, dtype='float32'))
emb_dropout_r = theano.shared(
numpy.array([retain_probability_emb] * 2, dtype='float32'))
source_dropout = theano.shared(
numpy.float32(retain_probability_source))
emb *= source_dropout
if options['bidirectional_enc']:
embr *= source_dropout
else:
rec_dropout = theano.shared(numpy.array([1.] * 2, dtype='float32'))
rec_dropout_r = theano.shared(
numpy.array([1.] * 2, dtype='float32'))
emb_dropout = theano.shared(numpy.array([1.] * 2, dtype='float32'))
emb_dropout_r = theano.shared(
numpy.array([1.] * 2, dtype='float32'))
# Encode sentences
if options['encoder_%s' % lang] == 'bow':
sents = (emb * mask[:, :, None]).sum(0)
else:
# iteratively push input from first hidden layer until the last
for i in range(int(options['n_enc_hidden_layers'])):
layer_name_prefix = 'encoder_%s_%i' % (lang, i)
# if first layer input are wembs, otherwise input will be output of last hidden layer
layer_below = emb if i == 0 else layer_below[0]
layer_below = get_layer(options['encoder_%s' % lang])[1](
tparams,
layer_below,
options,
None,
prefix=layer_name_prefix,
mask=mask,
emb_dropout=emb_dropout,
rec_dropout=rec_dropout)
if i == int(options['n_enc_hidden_layers']) - 1:
# sentence embeddings (projections) are the output of the last hidden layer
proj = layer_below
if options['bidirectional_enc']:
for i in range(int(options['n_enc_hidden_layers'])):
layer_name_prefix = 'encoder_%s_r_%i' % (lang, i)
# if first layer input are wembs, otherwise input will be output of last hidden layer
layer_below = emb_r if i == 0 else layer_below[0]
layer_below = get_layer(options['encoder_%s' % lang])[1](
tparams,
layer_below,
options,
None,
prefix=layer_name_prefix,
mask=mask_r,
emb_dropout=emb_dropout_r,
rec_dropout=rec_dropout_r)
if i == int(options['n_enc_hidden_layers']) - 1:
# sentence embeddings (projections) are the output of the last hidden layer
proj_r = layer_below
# use last hidden state of forward and backward RNNs
sents = concatenate([proj[0][-1], proj_r[0][-1]],
axis=proj[0].ndim - 2)
else:
sents = proj[0][-1]
if options['use_dropout']:
sents *= shared_dropout_layer((n_samples, options['dim']),
use_noise, trng,
retain_probability_hidden)
# project sentences into multimodal space
sents_mm = get_layer('ff')[1](tparams,
sents,
options,
prefix='ff_sentence_mm',
activ='linear')
if not 'attention_type' in options or options[
'attention_type'] == 'dot':
sents_mm = l2norm(sents_mm)
if options['use_dropout']:
sents_mm *= shared_dropout_layer(
(n_samples, options['dim_multimodal']), use_noise, trng,
retain_probability_hidden)
# outputs per language
in_outs.append(([x, mask], sents_mm))
return trng, in_outs
