def _build(self):
print 'Building model...'
# build the sampling functions and model
self.trng = RandomStreams(1234)
use_noise = theano.shared(numpy.float32(0.), name='use_noise')
params = capgen.init_params(self.options)
params = capgen.load_params(self.model, params)
self.tparams = capgen.init_tparams(params)
# word index
self.f_init, self.f_next = capgen.build_sampler(self.tparams, self.options, use_noise, self.trng)
self.trng, use_noise, inps, \
alphas, alphas_samples, cost, opt_outs = capgen.build_model(self.tparams, self.options)
# get the alphas and selector value [called \beta in the paper]
# create update rules for the stochastic attention
hard_attn_updates = []
if self.options['attn_type'] == 'stochastic':
baseline_time = theano.shared(numpy.float32(0.), name='baseline_time')
hard_attn_updates += [(baseline_time, baseline_time * 0.9 + 0.1 * opt_outs['masked_cost'].mean())]
hard_attn_updates += opt_outs['attn_updates']
self.f_alpha = theano.function(inps, alphas, name='f_alpha', updates=hard_attn_updates)
if self.options['selector']:
self.f_sels = theano.function(inps, opt_outs['selector'], name='f_sels', updates=hard_attn_updates)
print 'Done'
def load_model(model_path):
# print 'loading model'
model = model_path
options = load_pkl(model + '.pkl')
# build the sampling functions and model
trng = RandomStreams(1234)
use_noise = theano.shared(numpy.float32(0.), name='use_noise')
params = capgen.init_params(options)
params = capgen.load_params(model, params)
tparams = capgen.init_tparams(params)
f_init, f_next = capgen.build_sampler(tparams, options, use_noise, trng)
trng, use_noise, inps, alphas, alphas_samples, cost, opt_outs = capgen.build_model(tparams, options)
# print 'done'
return tparams, f_init, f_next, options, trng
def train(dim_word=100, # word vector dimensionality
ctx_dim=512, # context vector dimensionality
dim=1000, # the number of LSTM units
attn_type='stochastic', # [see section 4 from paper]
n_layers_att=1, # number of layers used to compute the attention weights
n_layers_out=1, # number of layers used to compute logit
n_layers_lstm=1, # number of lstm layers
n_layers_init=1, # number of layers to initialize LSTM at time 0
lstm_encoder=False, # if True, run bidirectional LSTM on input units
prev2out=False, # Feed previous word into logit
ctx2out=False, # Feed attention weighted ctx into logit
alpha_entropy_c=0.002, # hard attn param
RL_sumCost=True, # hard attn param
semi_sampling_p=0.5, # hard attn param
temperature=1., # hard attn param
patience=10,
max_epochs=5000,
dispFreq=100,
decay_c=0., # weight decay coeff
alpha_c=0., # doubly stochastic coeff
lrate=0.01, # used only for SGD
selector=False, # selector (see paper)
n_words=10000, # vocab size
maxlen=100, # maximum length of the description
optimizer='rmsprop',
batch_size = 16,
valid_batch_size = 16,
saveto='model.npz', # relative path of saved model file
validFreq=1000,
saveFreq=1000, # save the parameters after every saveFreq updates
sampleFreq=100, # generate some samples after every sampleFreq updates
data_path='./data', # path to find data
dataset='flickr8k',
dictionary=None, # word dictionary
use_dropout=False, # setting this true turns on dropout at various points
use_dropout_lstm=False, # dropout on lstm gates
reload_=False,
save_per_epoch=False): # this saves down the model every epoch
# hyperparam dict
model_options = locals().copy()
model_options = validate_options(model_options)
# reload options
if reload_ and os.path.exists(saveto):
print "Reloading options"
with open('%s.pkl'%saveto, 'rb') as f:
model_options = pkl.load(f)
print "Using the following parameters:"
print model_options
print 'Loading data'
load_data, prepare_data = get_dataset(dataset)
train, valid, test, worddict = load_data(path=data_path)
if dataset == 'coco':
valid, _ = valid # the second one contains all the validation data
# index 0 and 1 always code for the end of sentence and unknown token
word_idict = dict()
for kk, vv in worddict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
# Initialize (or reload) the parameters using 'model_options'
# then build the Theano graph
print 'Building model'
params = init_params(model_options)
if reload_ and os.path.exists(saveto):
print "Reloading model"
params = load_params(saveto, params)
# numpy arrays -> theano shared variables
tparams = init_tparams(params)
# In order, we get:
# 1) trng - theano random number generator
# 2) use_noise - flag that turns on dropout
# 3) inps - inputs for f_grad_shared
# 4) cost - log likelihood for each sentence
# 5) opts_out - optional outputs (e.g selector)
trng, use_noise, \
inps, alphas, alphas_sample,\
cost, \
opt_outs = \
build_model(tparams, model_options)
# To sample, we use beam search: 1) f_init is a function that initializes
# the LSTM at time 0 [see top right of page 4], 2) f_next returns the distribution over
# words and also the new "initial state/memory" see equation
print 'Building sampler'
f_init, f_next = build_sampler(tparams, model_options, use_noise, trng)
# we want the cost without any the regularizers
# define the log probability
f_log_probs = theano.function(inps, -cost, profile=False,
updates=opt_outs['attn_updates']
if model_options['attn_type']=='stochastic'
else None, allow_input_downcast=True)
# Define the cost function + Regularization
cost = cost.mean()
# add L2 regularization costs
if decay_c > 0.:
decay_c = theano.shared(numpy.float32(decay_c), name='decay_c')
weight_decay = 0.
for kk, vv in tparams.iteritems():
weight_decay += (vv ** 2).sum()
weight_decay *= decay_c
cost += weight_decay
# Doubly stochastic regularization
if alpha_c > 0.:
alpha_c = theano.shared(numpy.float32(alpha_c), name='alpha_c')
alpha_reg = alpha_c * ((1.-alphas.sum(0))**2).sum(0).mean()
cost += alpha_reg
hard_attn_updates = []
# Backprop!
if model_options['attn_type'] == 'deterministic':
grads = tensor.grad(cost, wrt=itemlist(tparams))
else:
# shared variables for hard attention
baseline_time = theano.shared(numpy.float32(0.), name='baseline_time')
opt_outs['baseline_time'] = baseline_time
alpha_entropy_c = theano.shared(numpy.float32(alpha_entropy_c), name='alpha_entropy_c')
alpha_entropy_reg = alpha_entropy_c * (alphas*tensor.log(alphas)).mean()
# [see Section 4.1: Stochastic "Hard" Attention for derivation of this learning rule]
if model_options['RL_sumCost']:
grads = tensor.grad(cost, wrt=itemlist(tparams),
disconnected_inputs='raise',
known_grads={alphas:(baseline_time-opt_outs['masked_cost'].mean(0))[None,:,None]/10.*
(-alphas_sample/alphas) + alpha_entropy_c*(tensor.log(alphas) + 1)})
else:
grads = tensor.grad(cost, wrt=itemlist(tparams),
disconnected_inputs='raise',
known_grads={alphas:opt_outs['masked_cost'][:,:,None]/10.*
(alphas_sample/alphas) + alpha_entropy_c*(tensor.log(alphas) + 1)})
# [equation on bottom left of page 5]
hard_attn_updates += [(baseline_time, baseline_time * 0.9 + 0.1 * opt_outs['masked_cost'].mean())]
# updates from scan
hard_attn_updates += opt_outs['attn_updates']
# to getthe cost after regularization or the gradients, use this
# f_cost = theano.function([x, mask, ctx], cost, profile=False)
# f_grad = theano.function([x, mask, ctx], grads, profile=False)
# f_grad_shared computes the cost and updates adaptive learning rate variables
# f_update updates the weights of the model
lr = tensor.scalar(name='lr')
f_grad_shared, f_update = eval(optimizer)(lr, tparams, grads, inps, cost, hard_attn_updates)
print 'Optimization'
# [See note in section 4.3 of paper]
train_iter = HomogeneousData(train, batch_size=batch_size, maxlen=maxlen)
if valid:
kf_valid = KFold(len(valid[0]), n_folds=len(valid[0])/valid_batch_size, shuffle=False)
if test:
kf_test = KFold(len(test[0]), n_folds=len(test[0])/valid_batch_size, shuffle=False)
# history_errs is a bare-bones training log that holds the validation and test error
history_errs = []
# reload history
if reload_ and os.path.exists(saveto):
history_errs = numpy.load(saveto)['history_errs'].tolist()
best_p = None
bad_counter = 0
if validFreq == -1:
validFreq = len(train[0])/batch_size
if saveFreq == -1:
saveFreq = len(train[0])/batch_size
if sampleFreq == -1:
sampleFreq = len(train[0])/batch_size
uidx = 0
estop = False
for eidx in xrange(max_epochs):
n_samples = 0
print 'Epoch ', eidx
for caps in train_iter:
n_samples += len(caps)
uidx += 1
# turn on dropout
use_noise.set_value(1.)
# preprocess the caption, recording the
# time spent to help detect bottlenecks
pd_start = time.time()
x, mask, ctx = prepare_data(caps,
train[1],
worddict,
maxlen=maxlen,
n_words=n_words)
pd_duration = time.time() - pd_start
if x is None:
print 'Minibatch with zero sample under length ', maxlen
continue
# get the cost for the minibatch, and update the weights
ud_start = time.time()
cost = f_grad_shared(x, mask, ctx)
f_update(lrate)
ud_duration = time.time() - ud_start # some monitoring for each mini-batch
# Numerical stability check
if numpy.isnan(cost) or numpy.isinf(cost):
print 'NaN detected'
return 1., 1., 1.
if numpy.mod(uidx, dispFreq) == 0:
print 'Epoch ', eidx, 'Update ', uidx, 'Cost ', cost, 'PD ', pd_duration, 'UD ', ud_duration
# Checkpoint
if numpy.mod(uidx, saveFreq) == 0:
print 'Saving...',
if best_p is not None:
params = copy.copy(best_p)
else:
params = unzip(tparams)
numpy.savez(saveto, history_errs=history_errs, **params)
pkl.dump(model_options, open('%s.pkl'%saveto, 'wb'))
print 'Done'
# Print a generated sample as a sanity check
if numpy.mod(uidx, sampleFreq) == 0:
# turn off dropout first
use_noise.set_value(0.)
x_s = x
mask_s = mask
ctx_s = ctx
# generate and decode the a subset of the current training batch
for jj in xrange(numpy.minimum(10, len(caps))):
sample, score = gen_sample(tparams, f_init, f_next, ctx_s[jj], model_options,
trng=trng, k=5, maxlen=30, stochastic=False)
# Decode the sample from encoding back to words
print 'Truth ',jj,': ',
for vv in x_s[:,jj]:
if vv == 0:
break
if vv in word_idict:
print word_idict[vv],
else:
print 'UNK',
print
for kk, ss in enumerate([sample[0]]):
print 'Sample (', kk,') ', jj, ': ',
for vv in ss:
if vv == 0:
break
if vv in word_idict:
print word_idict[vv],
else:
print 'UNK',
print
# Log validation loss + checkpoint the model with the best validation log likelihood
if numpy.mod(uidx, validFreq) == 0:
use_noise.set_value(0.)
train_err = 0
valid_err = 0
test_err = 0
if valid:
valid_err = -pred_probs(f_log_probs, model_options, worddict, prepare_data, valid, kf_valid).mean()
if test:
test_err = -pred_probs(f_log_probs, model_options, worddict, prepare_data, test, kf_test).mean()
history_errs.append([valid_err, test_err])
# the model with the best validation long likelihood is saved seperately with a different name
if uidx == 0 or valid_err <= numpy.array(history_errs)[:,0].min():
best_p = unzip(tparams)
print 'Saving model with best validation ll'
params = copy.copy(best_p)
params = unzip(tparams)
numpy.savez(saveto+'_bestll', history_errs=history_errs, **params)
bad_counter = 0
# abort training if perplexity has been increasing for too long
if eidx > patience and len(history_errs) > patience and valid_err >= numpy.array(history_errs)[:-patience,0].min():
bad_counter += 1
if bad_counter > patience:
print 'Early Stop!'
estop = True
break
print 'Train ', train_err, 'Valid ', valid_err, 'Test ', test_err
print 'Seen %d samples' % n_samples
if estop:
break
if save_per_epoch:
numpy.savez(saveto + '_epoch_' + str(eidx + 1), history_errs=history_errs, **unzip(tparams))
# use the best nll parameters for final checkpoint (if they exist)
if best_p is not None:
zipp(best_p, tparams)
use_noise.set_value(0.)
train_err = 0
valid_err = 0
test_err = 0
if valid:
valid_err = -pred_probs(f_log_probs, model_options, worddict, prepare_data, valid, kf_valid)
if test:
test_err = -pred_probs(f_log_probs, model_options, worddict, prepare_data, test, kf_test)
print 'Train ', train_err, 'Valid ', valid_err, 'Test ', test_err
params = copy.copy(best_p)
numpy.savez(saveto, zipped_params=best_p, train_err=train_err,
valid_err=valid_err, test_err=test_err, history_errs=history_errs,
**params)
return train_err, valid_err, test_err
def train(dim_word=100, # word vector dimensionality
ctx_dim=512, # context vector dimensionality
dim=1000, # the number of LSTM units
attn_type='deterministic', # [see section 4 from paper]
n_layers_att=1, # number of layers used to compute the attention weights
n_layers_out=1, # number of layers used to compute logit
n_layers_lstm=1, # number of lstm layers
n_layers_init=1, # number of layers to initialize LSTM at time 0
lstm_encoder=False, # if True, run bidirectional LSTM on input units
prev2out=False, # Feed previous word into logit
ctx2out=False, # Feed attention weighted ctx into logit
alpha_entropy_c=0.002, # hard attn param
RL_sumCost=False, # hard attn param
semi_sampling_p=0.5, # hard attn param
temperature=1., # hard attn param
patience=10,
max_epochs=5000,
dispFreq=100,
decay_c=0., # weight decay coeff
alpha_c=0., # doubly stochastic coeff
lrate=0.01, # used only for SGD
selector=False, # selector (see paper)
n_words=10000, # vocab size
maxlen=100, # maximum length of the description
optimizer='rmsprop',
batch_size = 16,
valid_batch_size = 2,#change from 16
saveto='model.npz', # relative path of saved model file
validFreq=1000,
saveFreq=1000, # save the parameters after every saveFreq updates
sampleFreq=5, # generate some samples after every sampleFreq updates
data_path='./data', # path to find data
dataset='flickr30k',
dictionary=None, # word dictionary
use_dropout=False, # setting this true turns on dropout at various points
use_dropout_lstm=False, # dropout on lstm gates
reload_=False,
save_per_epoch=False): # this saves down the model every epoch
# hyperparam dict
model_options = locals().copy()
model_options = validate_options(model_options)
# reload options
if reload_ and os.path.exists(saveto):
print "Reloading options"
with open('%s.pkl'%saveto, 'rb') as f:
model_options = pkl.load(f)
print "Using the following parameters:"
print model_options
print 'Loading data'
load_data, prepare_data = get_dataset(dataset)
train, valid, test, worddict = load_data(path=data_path)
# index 0 and 1 always code for the end of sentence and unknown token
word_idict = dict()
for kk, vv in worddict.iteritems():
word_idict[vv] = kk
word_idict[0] = '<eos>'
word_idict[1] = 'UNK'
# Initialize (or reload) the parameters using 'model_options'
# then build the Theano graph
print 'Building model'
params = init_params(model_options)
if reload_ and os.path.exists(saveto):
print "Reloading model"
params = load_params(saveto, params)
# numpy arrays -> theano shared variables
tparams = init_tparams(params)
# In order, we get:
# 1) trng - theano random number generator
# 2) use_noise - flag that turns on dropout
# 3) inps - inputs for f_grad_shared
# 4) cost - log likelihood for each sentence
# 5) opts_out - optional outputs (e.g selector)
trng, use_noise, \
inps, alphas, alphas_sample,\
cost, \
opt_outs = \
build_model(tparams, model_options)
# To sample, we use beam search: 1) f_init is a function that initializes
# the LSTM at time 0 [see top right of page 4], 2) f_next returns the distribution over
# words and also the new "initial state/memory" see equation
print 'Building sampler'
f_init, f_next = build_sampler(tparams, model_options, use_noise, trng)
# we want the cost without any the regularizers
# define the log probability
f_log_probs = theano.function(inps, -cost, profile=False,
updates=opt_outs['attn_updates']
if model_options['attn_type']=='stochastic'
else None, allow_input_downcast=True)
# Define the cost function + Regularization
cost = cost.mean()
# add L2 regularization costs
if decay_c > 0.:
decay_c = theano.shared(numpy.float32(decay_c), name='decay_c')
weight_decay = 0.
for kk, vv in tparams.iteritems():
weight_decay += (vv ** 2).sum()
weight_decay *= decay_c
cost += weight_decay
# Doubly stochastic regularization
if alpha_c > 0.:
alpha_c = theano.shared(numpy.float32(alpha_c), name='alpha_c')
alpha_reg = alpha_c * ((1.-alphas.sum(0))**2).sum(0).mean()
cost += alpha_reg
hard_attn_updates = []
# Backprop!
if model_options['attn_type'] == 'deterministic':
grads = tensor.grad(cost, wrt=itemlist(tparams))
else:
# shared variables for hard attention
baseline_time = theano.shared(numpy.float32(0.), name='baseline_time')
opt_outs['baseline_time'] = baseline_time
alpha_entropy_c = theano.shared(numpy.float32(alpha_entropy_c), name='alpha_entropy_c')
alpha_entropy_reg = alpha_entropy_c * (alphas*tensor.log(alphas)).mean()
# [see Section 4.1: Stochastic "Hard" Attention for derivation of this learning rule]
if model_options['RL_sumCost']:
grads = tensor.grad(cost, wrt=itemlist(tparams),
disconnected_inputs='raise',
known_grads={alphas:(baseline_time-opt_outs['masked_cost'].mean(0))[None,:,None]/10.*
(-alphas_sample/alphas) + alpha_entropy_c*(tensor.log(alphas) + 1)})
else:
grads = tensor.grad(cost, wrt=itemlist(tparams),
disconnected_inputs='raise',
known_grads={alphas:opt_outs['masked_cost'][:,:,None]/10.*
(alphas_sample/alphas) + alpha_entropy_c*(tensor.log(alphas) + 1)})
# [equation on bottom left of page 5]
hard_attn_updates += [(baseline_time, baseline_time * 0.9 + 0.1 * opt_outs['masked_cost'].mean())]
# updates from scan
hard_attn_updates += opt_outs['attn_updates']
# to getthe cost after regularization or the gradients, use this
# f_cost = theano.function([x, mask, ctx], cost, profile=False)
# f_grad = theano.function([x, mask, ctx], grads, profile=False)
# f_grad_shared computes the cost and updates adaptive learning rate variables
# f_update updates the weights of the model
lr = tensor.scalar(name='lr')
f_grad_shared, f_update = eval(optimizer)(lr, tparams, grads, inps, cost, hard_attn_updates)
print 'Optimization'
# [See note in section 4.3 of paper]
train_iter = HomogeneousData(train, batch_size=batch_size, maxlen=maxlen)
if valid:
kf_valid = KFold(len(valid[0]), n_folds=len(valid[0])/valid_batch_size, shuffle=False)
if test:
kf_test = KFold(len(test[0]), n_folds=len(test[0])/valid_batch_size, shuffle=False)
# history_errs is a bare-bones training log that holds the validation and test error
history_errs = []
# reload history
if reload_ and os.path.exists(saveto):
history_errs = numpy.load(saveto)['history_errs'].tolist()
best_p = None
bad_counter = 0
if validFreq == -1:
validFreq = len(train[0])/batch_size
if saveFreq == -1:
saveFreq = len(train[0])/batch_size
if sampleFreq == -1:
sampleFreq = len(train[0])/batch_size
uidx = 0
estop = False
for eidx in xrange(max_epochs):
n_samples = 0
print 'Epoch ', eidx
for caps in train_iter:
n_samples += len(caps)
uidx += 1
# turn on dropout
use_noise.set_value(1.)
# preprocess the caption, recording the
# time spent to help detect bottlenecks
pd_start = time.time()
x, mask, ctx = prepare_data(caps,
train[1],
worddict,
maxlen=maxlen,
n_words=n_words)
pd_duration = time.time() - pd_start
if x is None:
print 'Minibatch with zero sample under length ', maxlen
continue
# get the cost for the minibatch, and update the weights
ud_start = time.time()
cost = f_grad_shared(x, mask, ctx)
f_update(lrate)
ud_duration = time.time() - ud_start # some monitoring for each mini-batch
# Numerical stability check
if numpy.isnan(cost) or numpy.isinf(cost):
print 'NaN detected'
return 1., 1., 1.
if numpy.mod(uidx, dispFreq) == 0:
print 'Epoch ', eidx, 'Update ', uidx, 'Cost ', cost, 'PD ', pd_duration, 'UD ', ud_duration
# Checkpoint
if numpy.mod(uidx, saveFreq) == 0:
print 'Saving...',
if best_p is not None:
params = copy.copy(best_p)
else:
params = unzip(tparams)
numpy.savez(saveto, history_errs=history_errs, **params)
pkl.dump(model_options, open('%s.pkl'%saveto, 'wb'))
print 'Done'
# Print a generated sample as a sanity check
if numpy.mod(uidx, sampleFreq) == 0:
# turn off dropout first
use_noise.set_value(0.)
x_s = x
mask_s = mask
ctx_s = ctx
# generate and decode the a subset of the current training batch
for jj in xrange(numpy.minimum(10, len(caps))):
sample, score = gen_sample(tparams, f_init, f_next, ctx_s[jj], model_options,
trng=trng, k=5, maxlen=30, stochastic=False)
# Decode the sample from encoding back to words
print 'Truth ',jj,': ',
for vv in x_s[:,jj]:
if vv == 0:
break
if vv in word_idict:
print word_idict[vv],
else:
print 'UNK',
print
for kk, ss in enumerate([sample[0]]):
print 'Sample (', kk,') ', jj, ': ',
for vv in ss:
if vv == 0:
break
if vv in word_idict:
print word_idict[vv],
else:
print 'UNK',
print
# Log validation loss + checkpoint the model with the best validation log likelihood
if numpy.mod(uidx, validFreq) == 0:
use_noise.set_value(0.)
train_err = 0
valid_err = 0
test_err = 0
if valid:
valid_err = -pred_probs(f_log_probs, model_options, worddict, prepare_data, valid, kf_valid).mean()
if test:
test_err = -pred_probs(f_log_probs, model_options, worddict, prepare_data, test, kf_test).mean()
history_errs.append([valid_err, test_err])
# the model with the best validation long likelihood is saved seperately with a different name
if uidx == 0 or valid_err <= numpy.array(history_errs)[:,0].min():
best_p = unzip(tparams)
print 'Saving model with best validation ll'
params = copy.copy(best_p)
params = unzip(tparams)
numpy.savez(saveto+'_bestll', history_errs=history_errs, **params)
bad_counter = 0
# abort training if perplexity has been increasing for too long
if eidx > patience and len(history_errs) > patience and valid_err >= numpy.array(history_errs)[:-patience,0].min():
bad_counter += 1
if bad_counter > patience:
print 'Early Stop!'
estop = True
break
print 'Train ', train_err, 'Valid ', valid_err, 'Test ', test_err
print 'Seen %d samples' % n_samples
if estop:
break
if save_per_epoch:
numpy.savez(saveto + '_epoch_' + str(eidx + 1), history_errs=history_errs, **unzip(tparams))
# use the best nll parameters for final checkpoint (if they exist)
if best_p is not None:
zipp(best_p, tparams)
use_noise.set_value(0.)
train_err = 0
valid_err = 0
test_err = 0
if valid:
valid_err = -pred_probs(f_log_probs, model_options, worddict, prepare_data, valid, kf_valid)
if test:
test_err = -pred_probs(f_log_probs, model_options, worddict, prepare_data, test, kf_test)
print 'Train ', train_err, 'Valid ', valid_err, 'Test ', test_err
params = copy.copy(best_p)
numpy.savez(saveto, zipped_params=best_p, train_err=train_err,
valid_err=valid_err, test_err=test_err, history_errs=history_errs,
**params)
return train_err, valid_err, test_err
# build the sampling functions and model
trng = RandomStreams(1234)
use_noise = theano.shared(numpy.float32(0.), name='use_noise')
params = capgen.init_params(options)
params = capgen.load_params(model, params)
tparams = capgen.init_tparams(params)
# word index
f_init, f_next = capgen.build_sampler(tparams, options, use_noise, trng)
# In[43]:
trng,use_noise,inps, alphas, alphas_samples,cost, opt_outs = capgen.build_model(tparams, options)
# In[44]:
# get the alphas and selector value [called \beta in the paper]
# create update rules for the stochastic attention
hard_attn_updates = []
if options['attn_type'] == 'stochastic':
baseline_time = theano.shared(numpy.float32(0.), name='baseline_time')
hard_attn_updates += [(baseline_time, baseline_time * 0.9 + 0.1 * opt_outs['masked_cost'].mean())]
hard_attn_updates += opt_outs['attn_updates']
f_alpha = theano.function(inps, alphas, name='f_alpha', updates=hard_attn_updates)
if options['selector']:
def train(
dim_word=300, # word vector dimensionality
ctx_dim=300, # context vector dimensionality
semantic_dim=300,
dim=1000, # the number of LSTM units
cnn_dim=4096, # CNN feature dimension
n_layers_att=1, # number of layers used to compute the attention weights
n_layers_out=1, # number of layers used to compute logit
n_layers_lstm=1, # number of lstm layers
n_layers_init=1, # number of layers to initialize LSTM at time 0
lstm_encoder=True, # if True, run bidirectional LSTM on input units
prev2out=False, # Feed previous word into logit
ctx2out=False, # Feed attention weighted ctx into logit
cutoff=10,
patience=5,
max_epochs=30,
dispFreq=500,
decay_c=0., # weight decay coeff
alpha_c=0., # doubly stochastic coeff
lrate=1e-4, # used only for SGD
selector=False, # selector (see paper)
maxlen=30, # maximum length of the description
optimizer='rmsprop',
pretrained='',
batch_size=256,
saveto='model', # relative path of saved model file
saveFreq=1000, # save the parameters after every saveFreq updates
sampleFreq=100, # generate some samples after every sampleFreq updates
embedding='../Data/GloVe/vocab_glove.pkl',
cnn_type='vgg',
prefix='../Data', # path to find data
dataset='coco',
criterion='Bleu_4',
switch_test_val=False,
use_cnninit=True,
use_dropout=True, # setting this true turns on dropout at various points
use_dropout_lstm=False, # dropout on lstm gates
save_per_epoch=False): # this saves down the model every epoch
# hyperparam dict
model_options = locals().copy()
model_options = validate_options(model_options)
# reload options
if os.path.exists('%s.pkl' % saveto):
print "Reloading options"
with open('%s.pkl' % saveto, 'rb') as f:
model_options = pkl.load(f)
print "Using the following parameters:"
print model_options
print 'Loading data'
load_data, prepare_data = get_dataset(model_options['dataset'])
# Load data from data path
if 'switch_test_val' in model_options and model_options['switch_test_val']:
train, valid, worddict = load_data(path=osp.join(
model_options['prefix'], model_options['dataset']),
options=model_options,
load_train=True,
load_test=True)
else:
train, valid, worddict = load_data(path=osp.join(
model_options['prefix'], model_options['dataset']),
options=model_options,
load_train=True,
load_val=True)
# Automatically calculate the update frequency
validFreq = len(train[0]) / model_options['batch_size']
print "Validation frequency is %d" % validFreq
word_idict = {vv: kk for kk, vv in worddict.iteritems()}
model_options['n_words'] = len(worddict)
# Initialize (or reload) the parameters using 'model_options'
# then build the Theano graph
print 'Building model'
params = init_params(model_options)
# Initialize it with glove
if 'VCemb' in params:
params['VCemb'] = read_pkl(
model_options['embedding']).astype('float32')
# If there is a same experiment, don't use pretrained weights
if os.path.exists('%s.npz' % saveto):
print "Reloading model"
params = load_params('%s.npz' % saveto, params)
elif pretrained != '':
params = load_params(pretrained, params,
False) # Only pretrain the Language model
# numpy arrays -> theano shared variables
tparams = init_tparams(params)
# In order, we get:
# 1) trng - theano random number generator
# 2) use_noise - flag that turns on dropout
# 3) inps - inputs for f_grad_shared
# 4) cost - log likelihood for each sentence
# 5) opts_out - optional outputs (e.g selector)
trng, use_noise, \
inps, alphas,\
cost, \
opt_outs = \
build_model(tparams, model_options)
# Load evaluator to calculate bleu score
evaluator = cocoEvaluation(model_options['dataset'])
# To sample, we use beam search: 1) f_init is a function that initializes
# the LSTM at time 0 [see top right of page 4], 2) f_next returns the distribution over
# words and also the new "initial state/memory" see equation
print 'Building sampler'
f_init, f_next = build_sampler(tparams, model_options, use_noise, trng)
# we want the cost without any the regularizers
# define the log probability
f_log_probs = theano.function(inps,
-cost,
profile=False,
updates=None,
allow_input_downcast=True)
# Define the cost function + Regularization
cost = cost.mean()
# add L2 regularization costs
if decay_c > 0.:
decay_c = theano.shared(numpy.float32(decay_c), name='decay_c')
weight_decay = 0.
for kk, vv in tparams.iteritems():
weight_decay += (vv**2).sum()
weight_decay *= decay_c
cost += weight_decay
# Doubly stochastic regularization
if alpha_c > 0.:
alpha_c = theano.shared(numpy.float32(alpha_c), name='alpha_c')
alpha_reg = sum([
alpha_c * ((1. - alpha.sum(0))**2).sum(0).mean()
for alpha in alphas
])
cost += alpha_reg
# Backprop!
grads = tensor.grad(cost, wrt=itemlist(tparams))
# to getthe cost after regularization or the gradients, use this
# f_grad_shared computes the cost and updates adaptive learning rate variables
# f_update updates the weights of the model
lr = tensor.scalar(name='lr')
f_grad_shared, f_update = eval(model_options['optimizer'])(lr, tparams,
grads, inps,
cost)
print 'Optimization'
train_iter = HomogeneousData(train,
batch_size=batch_size,
maxlen=model_options['maxlen'])
# history_bleu is a bare-bones training log, reload history
history_bleu = []
if os.path.exists('%s.npz' % saveto):
history_bleu = numpy.load('%s.npz' % saveto)['history_bleu'].tolist()
start_epochs = len(history_bleu)
best_p = None
bad_counter = 0
if validFreq == -1:
validFreq = len(train[0]) / batch_size
if saveFreq == -1:
saveFreq = len(train[0]) / batch_size
if sampleFreq == -1:
sampleFreq = len(train[0]) / batch_size
uidx = 0
estop = False
for eidx in xrange(start_epochs, model_options['max_epochs']):
n_samples = 0
print 'Epoch ', eidx
for caps in train_iter:
n_samples += len(caps)
uidx += 1
# turn on dropout
use_noise.set_value(1.)
# preprocess the caption, recording the
# time spent to help detect bottlenecks
pd_start = time.time()
x, mask, ctx, cnn_feats = prepare_data(caps, train[1], train[2],
worddict, model_options)
pd_duration = time.time() - pd_start
if x is None:
print 'Minibatch with zero sample under length ', model_options[
'maxlen']
continue
# get the cost for the minibatch, and update the weights
ud_start = time.time()
cost = f_grad_shared(x, mask, ctx, cnn_feats)
print "Epoch %d, Updates: %d, Cost is: %f" % (eidx, uidx, cost)
f_update(model_options['lrate'])
ud_duration = time.time(
) - ud_start # some monitoring for each mini-batch
# Numerical stability check
if numpy.isnan(cost) or numpy.isinf(cost):
print 'NaN detected'
return 1., 1., 1.
if numpy.mod(uidx, dispFreq) == 0:
print 'Epoch ', eidx, 'Update ', uidx, 'Cost ', cost, 'PD ', pd_duration, 'UD ', ud_duration
# Print a generated sample as a sanity check
if numpy.mod(uidx, model_options['sampleFreq']) == 0:
# turn off dropout first
use_noise.set_value(0.)
x_s = x
mask_s = mask
ctx_s = ctx
# generate and decode the a subset of the current training batch
for jj in xrange(numpy.minimum(10, len(caps))):
sample, score, alphas = gen_sample(
f_init,
f_next,
ctx_s[jj],
cnn_feats[jj],
model_options,
trng=trng,
maxlen=model_options['maxlen'])
# Decode the sample from encoding back to words
print 'Truth ', jj, ': ',
print seqs2words(x_s[:, jj], word_idict)
for kk, ss in enumerate([sample[0]]):
print 'Sample (', kk, ') ', jj, ': ',
print seqs2words(ss, word_idict)
# Log validation loss + checkpoint the model with the best validation log likelihood
if numpy.mod(uidx, validFreq) == 0:
use_noise.set_value(0.)
# Do evaluation on validation set
imgid = collapse([elem[-1] for elem in valid[0]])
caps = process_examples([f_init], [f_next], imgid, valid[1],
valid[2], word_idict, model_options)
folder = osp.join('../output', '%s_%s' % (saveto, 'val'))
if not osp.exists(folder):
os.mkdir(folder)
with open(osp.join(folder, 'captions_val2014_results.json'),
'w') as f:
json.dump(caps, f)
eva_result = evaluator.evaluate(folder, False)
if model_options['criterion'] == 'combine':
history_bleu.append(eva_result['Bleu_4'] +
eva_result['CIDEr'])
else:
history_bleu.append(eva_result[model_options['criterion']])
# the model with the best validation long likelihood is saved seperately with a different name
if uidx == 0 or history_bleu[-1] == max(history_bleu):
best_p = unzip(tparams)
print 'Saving model with best validation ll'
params = copy.copy(best_p)
params = unzip(tparams)
numpy.savez(saveto + '_bestll',
history_bleu=history_bleu,
**params)
bad_counter = 0
# abort training if perplexity has been increasing for too long
if len(history_bleu) > model_options[
'patience'] and history_bleu[-1] <= max(
history_bleu[:-model_options['patience']]):
bad_counter += 1
if bad_counter > model_options['patience']:
print 'Early Stop!'
estop = True
break
print ' BLEU-4 score ', history_bleu[-1]
# Checkpoint
if numpy.mod(uidx, model_options['saveFreq']) == 0:
print 'Saving...',
if best_p is not None:
params = copy.copy(best_p)
else:
params = unzip(tparams)
numpy.savez(saveto, history_bleu=history_bleu, **params)
pkl.dump(model_options, open('%s.pkl' % saveto, 'wb'))
print 'Done'
print 'Seen %d samples' % n_samples
if estop:
break
if model_options['save_per_epoch']:
numpy.savez(saveto + '_epoch_' + str(eidx + 1),
history_bleu=history_bleu,
**unzip(tparams))
# use the best nll parameters for final checkpoint (if they exist)
if best_p is not None:
zipp(best_p, tparams)
params = copy.copy(best_p)
numpy.savez(saveto,
zipped_params=best_p,
history_bleu=history_bleu,
**params)