def param_init_fflayer(options, params, prefix='ff', nin=None, nout=None, ortho=True):
if nin is None:
nin = options['dim_proj']
if nout is None:
nout = options['dim_proj']
params[prfx(prefix,'W')] = norm_weight(nin, nout, scale=0.01, ortho=ortho)
params[prfx(prefix,'b')] = numpy.zeros((nout,)).astype('float32')
return params
def gru_layer(tparams, state_below, options, prefix='gru', mask=None, **kwargs):
nsteps = state_below.shape[0]
if state_below.ndim == 3:
n_samples = state_below.shape[1]
else:
n_samples = 1
dim = tparams[prfx(prefix,'Ux')].shape[1]
if mask == None:
mask = tensor.alloc(1., state_below.shape[0], 1)
def _slice(_x, n, dim):
if _x.ndim == 3:
return _x[:, :, n*dim:(n+1)*dim]
return _x[:, n*dim:(n+1)*dim]
state_below_ = tensor.dot(state_below, tparams[prfx(prefix, 'W')]) + tparams[prfx(prefix, 'b')]
state_belowx = tensor.dot(state_below, tparams[prfx(prefix, 'Wx')]) + tparams[prfx(prefix, 'bx')]
U = tparams[prfx(prefix, 'U')]
Ux = tparams[prfx(prefix, 'Ux')]
def _step_slice(m_, x_, xx_, h_, U, Ux):
preact = tensor.dot(h_, U)
preact += x_
r = tensor.nnet.sigmoid(_slice(preact, 0, dim))
u = tensor.nnet.sigmoid(_slice(preact, 1, dim))
preactx = tensor.dot(h_, Ux)
preactx = preactx * r
preactx = preactx + xx_
h = tensor.tanh(preactx)
h = u * h_ + (1. - u) * h
h = m_[:,None] * h + (1. - m_)[:,None] * h_
return h
seqs = [mask, state_below_, state_belowx]
_step = _step_slice
rval, updates = theano.scan(_step,
sequences=seqs,
outputs_info = [tensor.alloc(0., n_samples, dim)],
non_sequences = [tparams[prfx(prefix, 'U')],
tparams[prfx(prefix, 'Ux')]],
name=prfx(prefix, '_layers'),
n_steps=nsteps,
profile=profile,
strict=True)
rval = [rval]
return rval
def param_init_gru_cond(options, params, prefix='gru_cond',
nin=None, dim=None, dimctx=None):
if nin is None:
nin = options['dim']
if dim is None:
dim = options['dim']
if dimctx is None:
dimctx = options['dim']
params = param_init_gru(options, params, prefix, nin=nin, dim=dim)
# context to LSTM
Wc = norm_weight(dimctx,dim*2)
params[prfx(prefix,'Wc')] = Wc
Wcx = norm_weight(dimctx,dim)
params[prfx(prefix,'Wcx')] = Wcx
# attention: prev -> hidden
Wi_att = norm_weight(nin,dimctx)
params[prfx(prefix,'Wi_att')] = Wi_att
# attention: context -> hidden
Wc_att = norm_weight(dimctx)
params[prfx(prefix,'Wc_att')] = Wc_att
# attention: LSTM -> hidden
Wd_att = norm_weight(dim,dimctx)
params[prfx(prefix,'Wd_att')] = Wd_att
# attention: hidden bias
b_att = numpy.zeros((dimctx,)).astype('float32')
params[prfx(prefix,'b_att')] = b_att
# attention:
U_att = norm_weight(dimctx,1)
params[prfx(prefix,'U_att')] = U_att
c_att = numpy.zeros((1,)).astype('float32')
params[prfx(prefix, 'c_tt')] = c_att
return params
def param_init_gru(options, params, prefix='gru', nin=None, dim=None, hiero=False):
if nin == None:
nin = options['dim_proj']
if dim == None:
dim = options['dim_proj']
if not hiero:
W = numpy.concatenate([norm_weight(nin,dim),
norm_weight(nin,dim)], axis=1)
params[prfx(prefix,'W')] = W
params[prfx(prefix,'b')] = numpy.zeros((2 * dim,)).astype('float32')
U = numpy.concatenate([ortho_weight(dim),
ortho_weight(dim)], axis=1)
params[prfx(prefix,'U')] = U
Wx = norm_weight(nin, dim)
params[prfx(prefix,'Wx')] = Wx
Ux = ortho_weight(dim)
params[prfx(prefix,'Ux')] = Ux
params[prfx(prefix,'bx')] = numpy.zeros((dim,)).astype('float32')
return params
def train(
dim_word_desc=400, # word vector dimensionality
dim_word_q=400,
dim_word_ans=600,
dim_proj=300,
dim=400, # the number of LSTM units
encoder_desc='lstm',
encoder_desc_word='lstm',
encoder_desc_sent='lstm',
use_dq_sims=False,
eyem=None,
learn_h0=False,
use_desc_skip_c_g=False,
debug=False,
encoder_q='lstm',
patience=10,
max_epochs=5000,
dispFreq=100,
decay_c=0.,
alpha_c=0.,
clip_c=-1.,
lrate=0.01,
n_words_q=49145,
n_words_desc=115425,
n_words_ans=409,
pkl_train_files=None,
pkl_valid_files=None,
maxlen=2000, # maximum length of the description
optimizer='rmsprop',
batch_size=2,
vocab=None,
valid_batch_size=16,
use_elu_g=False,
saveto='model.npz',
model_dir=None,
ms_nlayers=3,
validFreq=1000,
saveFreq=1000, # save the parameters after every saveFreq updates
datasets=[None],
truncate=400,
momentum=0.9,
use_bidir=False,
cost_mask=None,
valid_datasets=[
'/u/yyu/stor/caglar/rc-data/cnn/cnn_test_data.h5',
'/u/yyu/stor/caglar/rc-data/cnn/cnn_valid_data.h5'
],
dropout_rate=0.5,
use_dropout=True,
reload_=True,
**opt_ds):
ensure_dir_exists(model_dir)
mpath = os.path.join(model_dir, saveto)
mpath_best = os.path.join(model_dir, prfx("best", saveto))
mpath_last = os.path.join(model_dir, prfx("last", saveto))
mpath_stats = os.path.join(model_dir, prfx("stats", saveto))
# Model options
model_options = locals().copy()
model_options['use_sent_reps'] = opt_ds['use_sent_reps']
stats = defaultdict(list)
del model_options['eyem']
del model_options['cost_mask']
if cost_mask is not None:
cost_mask = sharedX(cost_mask)
# reload options and parameters
if reload_:
print "Reloading the model."
if os.path.exists(mpath_best):
print "Reloading the best model from %s." % mpath_best
with open(os.path.join(mpath_best, '%s.pkl' % mpath_best),
'rb') as f:
models_options = pkl.load(f)
params = init_params(model_options)
params = load_params(mpath_best, params)
elif os.path.exists(mpath):
print "Reloading the model from %s." % mpath
with open(os.path.join(mpath, '%s.pkl' % mpath), 'rb') as f:
models_options = pkl.load(f)
params = init_params(model_options)
params = load_params(mpath, params)
else:
raise IOError("Couldn't open the file.")
else:
print "Couldn't reload the models initializing from scratch."
params = init_params(model_options)
if datasets[0]:
print "Short dataset", datasets[0]
print 'Loading data'
print 'Building model'
if pkl_train_files is None or pkl_valid_files is None:
train, valid, test = load_data(path=datasets[0],
valid_path=valid_datasets[0],
test_path=valid_datasets[1],
batch_size=batch_size,
**opt_ds)
else:
train, valid, test = load_pkl_data(train_file_paths=pkl_train_files,
valid_file_paths=pkl_valid_files,
batch_size=batch_size,
vocab=vocab,
eyem=eyem,
**opt_ds)
tparams = init_tparams(params)
trng, use_noise, inps_d, \
opt_ret, \
cost, errors, ent_errors, ent_derrors, probs = \
build_model(tparams,
model_options,
prepare_data if not opt_ds['use_sent_reps'] \
else prepare_data_sents,
valid,
cost_mask=cost_mask)
alphas = opt_ret['dec_alphas']
if opt_ds['use_sent_reps']:
inps = [inps_d["desc"], \
inps_d["word_mask"], \
inps_d["q"], \
inps_d['q_mask'], \
inps_d['ans'], \
inps_d['wlen'],
inps_d['slen'], inps_d['qlen'],\
inps_d['ent_mask']
]
else:
inps = [inps_d["desc"], \
inps_d["word_mask"], \
inps_d["q"], \
inps_d['q_mask'], \
inps_d['ans'], \
inps_d['wlen'], \
inps_d['qlen'], \
inps_d['ent_mask']]
outs = [cost, errors, probs, alphas]
if ent_errors:
outs += [ent_errors]
if ent_derrors:
outs += [ent_derrors]
# before any regularizer
print 'Building f_log_probs...',
f_log_probs = theano.function(inps, outs, profile=profile)
print 'Done'
# Apply weight decay on the feed-forward connections
if decay_c > 0.:
decay_c = theano.shared(numpy.float32(decay_c), name='decay_c')
weight_decay = 0.
for kk, vv in tparams.iteritems():
if "logit" in kk or "ff" in kk:
weight_decay += (vv**2).sum()
weight_decay *= decay_c
cost += weight_decay
# after any regularizer
print 'Computing gradient...',
grads = safe_grad(cost, itemlist(tparams))
print 'Done'
# Gradient clipping:
if clip_c > 0.:
g2 = get_norms(grads)
for p, g in grads.iteritems():
grads[p] = tensor.switch(g2 > (clip_c**2),
(g / tensor.sqrt(g2 + 1e-8)) * clip_c, g)
inps.pop()
if optimizer.lower() == "adasecant":
learning_rule = Adasecant(delta_clip=25.0,
use_adagrad=True,
grad_clip=0.25,
gamma_clip=0.)
elif optimizer.lower() == "rmsprop":
learning_rule = RMSPropMomentum(init_momentum=momentum)
elif optimizer.lower() == "adam":
learning_rule = Adam()
elif optimizer.lower() == "adadelta":
learning_rule = AdaDelta()
lr = tensor.scalar(name='lr')
print 'Building optimizers...',
learning_rule = None
if learning_rule:
f_grad_shared, f_update = learning_rule.get_funcs(learning_rate=lr,
grads=grads,
inp=inps,
cost=cost,
errors=errors)
else:
f_grad_shared, f_update = eval(optimizer)(lr, tparams, grads, inps,
cost, errors)
print 'Done'
print 'Optimization'
history_errs = []
# reload history
if reload_ and os.path.exists(mpath):
history_errs = list(numpy.load(mpath)['history_errs'])
best_p = None
bad_count = 0
if validFreq == -1:
validFreq = len(train[0]) / batch_size
if saveFreq == -1:
saveFreq = len(train[0]) / batch_size
best_found = False
uidx = 0
estop = False
train_cost_ave, train_err_ave, \
train_gnorm_ave = reset_train_vals()
for eidx in xrange(max_epochs):
n_samples = 0
if train.done:
train.reset()
for d_, q_, a, em in train:
n_samples += len(a)
uidx += 1
use_noise.set_value(1.)
if opt_ds['use_sent_reps']:
# To mask the description and the question.
d, d_mask, q, q_mask, dlen, slen, qlen = prepare_data_sents(
d_, q_)
if d is None:
print 'Minibatch with zero sample under length ', maxlen
uidx -= 1
continue
ud_start = time.time()
cost, errors, gnorm, pnorm = f_grad_shared(
d, d_mask, q, q_mask, a, dlen, slen, qlen)
else:
d, d_mask, q, q_mask, dlen, qlen = prepare_data(d_, q_)
if d is None:
print 'Minibatch with zero sample under length ', maxlen
uidx -= 1
continue
ud_start = time.time()
cost, errors, gnorm, pnorm = f_grad_shared(
d, d_mask, q, q_mask, a, dlen, qlen)
upnorm = f_update(lrate)
ud = time.time() - ud_start
# Collect the running ave train stats.
train_cost_ave = running_ave(train_cost_ave, cost)
train_err_ave = running_ave(train_err_ave, errors)
train_gnorm_ave = running_ave(train_gnorm_ave, gnorm)
if numpy.isnan(cost) or numpy.isinf(cost):
print 'NaN detected'
import ipdb
ipdb.set_trace()
if numpy.mod(uidx, dispFreq) == 0:
print 'Epoch ', eidx, ' Update ', uidx, \
' Cost ', cost, ' UD ', ud, \
' UpNorm ', upnorm[0].tolist(), \
' GNorm ', gnorm, \
' Pnorm ', pnorm, 'Terrors ', errors
if numpy.mod(uidx, saveFreq) == 0:
print 'Saving...',
if best_p is not None and best_found:
numpy.savez(mpath_best,
history_errs=history_errs,
**best_p)
pkl.dump(model_options, open('%s.pkl' % mpath_best, 'wb'))
else:
params = unzip(tparams)
numpy.savez(mpath, history_errs=history_errs, **params)
pkl.dump(model_options, open('%s.pkl' % mpath, 'wb'))
pkl.dump(stats, open("%s.pkl" % mpath_stats, 'wb'))
print 'Done'
print_param_norms(tparams)
if numpy.mod(uidx, validFreq) == 0:
use_noise.set_value(0.)
if valid.done:
valid.reset()
valid_costs, valid_errs, valid_probs, \
valid_alphas, error_ent, error_dent = eval_model(f_log_probs,
prepare_data if not opt_ds['use_sent_reps'] \
else prepare_data_sents,
model_options,
valid,
use_sent_rep=opt_ds['use_sent_reps'])
valid_alphas_ = numpy.concatenate(
[va.argmax(0) for va in valid_alphas.tolist()], axis=0)
valid_err = valid_errs.mean()
valid_cost = valid_costs.mean()
valid_alpha_ent = -negentropy(valid_alphas)
mean_valid_alphas = valid_alphas_.mean()
std_valid_alphas = valid_alphas_.std()
mean_valid_probs = valid_probs.argmax(1).mean()
std_valid_probs = valid_probs.argmax(1).std()
history_errs.append([valid_cost, valid_err])
stats['train_err_ave'].append(train_err_ave)
stats['train_cost_ave'].append(train_cost_ave)
stats['train_gnorm_ave'].append(train_gnorm_ave)
stats['valid_errs'].append(valid_err)
stats['valid_costs'].append(valid_cost)
stats['valid_err_ent'].append(error_ent)
stats['valid_err_desc_ent'].append(error_dent)
stats['valid_alphas_mean'].append(mean_valid_alphas)
stats['valid_alphas_std'].append(std_valid_alphas)
stats['valid_alphas_ent'].append(valid_alpha_ent)
stats['valid_probs_mean'].append(mean_valid_probs)
stats['valid_probs_std'].append(std_valid_probs)
if uidx == 0 or valid_err <= numpy.array(
history_errs)[:, 1].min():
best_p = unzip(tparams)
bad_counter = 0
best_found = True
else:
bst_found = False
if numpy.isnan(valid_err):
import ipdb
ipdb.set_trace()
print "============================"
print '\t>>>Valid error: ', valid_err, \
' Valid cost: ', valid_cost
print '\t>>>Valid pred mean: ', mean_valid_probs, \
' Valid pred std: ', std_valid_probs
print '\t>>>Valid alphas mean: ', mean_valid_alphas, \
' Valid alphas std: ', std_valid_alphas, \
' Valid alpha negent: ', valid_alpha_ent, \
' Valid error ent: ', error_ent, \
' Valid error desc ent: ', error_dent
print "============================"
print "Running average train stats "
print '\t>>>Train error: ', train_err_ave, \
' Train cost: ', train_cost_ave, \
' Train grad norm: ', train_gnorm_ave
print "============================"
train_cost_ave, train_err_ave, \
train_gnorm_ave = reset_train_vals()
print 'Seen %d samples' % n_samples
if estop:
break
if best_p is not None:
zipp(best_p, tparams)
use_noise.set_value(0.)
valid.reset()
valid_cost, valid_error, valid_probs, \
valid_alphas, error_ent = eval_model(f_log_probs,
prepare_data if not opt_ds['use_sent_reps'] \
else prepare_data_sents,
model_options, valid,
use_sent_rep=opt_ds['use_sent_rep'])
print " Final eval resuts: "
print 'Valid error: ', valid_error.mean()
print 'Valid cost: ', valid_cost.mean()
print '\t>>>Valid pred mean: ', valid_probs.mean(), \
' Valid pred std: ', valid_probs.std(), \
' Valid error ent: ', error_ent
params = copy.copy(best_p)
numpy.savez(mpath_last,
zipped_params=best_p,
history_errs=history_errs,
**params)
return valid_err, valid_cost
def train(dim_word_desc=400,# word vector dimensionality
dim_word_q=400,
dim_word_ans=600,
dim_proj=300,
dim=400,# the number of LSTM units
encoder_desc='lstm',
encoder_desc_word='lstm',
encoder_desc_sent='lstm',
use_dq_sims=False,
eyem=None,
learn_h0=False,
use_desc_skip_c_g=False,
debug=False,
encoder_q='lstm',
patience=10,
max_epochs=5000,
dispFreq=100,
decay_c=0.,
alpha_c=0.,
clip_c=-1.,
lrate=0.01,
n_words_q=49145,
n_words_desc=115425,
n_words_ans=409,
pkl_train_files=None,
pkl_valid_files=None,
maxlen=2000, # maximum length of the description
optimizer='rmsprop',
batch_size=2,
vocab=None,
valid_batch_size=16,
use_elu_g=False,
saveto='model.npz',
model_dir=None,
ms_nlayers=3,
validFreq=1000,
saveFreq=1000, # save the parameters after every saveFreq updates
datasets=[None],
truncate=400,
momentum=0.9,
use_bidir=False,
cost_mask=None,
valid_datasets=['/u/yyu/stor/caglar/rc-data/cnn/cnn_test_data.h5',
'/u/yyu/stor/caglar/rc-data/cnn/cnn_valid_data.h5'],
dropout_rate=0.5,
use_dropout=True,
reload_=True,
**opt_ds):
ensure_dir_exists(model_dir)
mpath = os.path.join(model_dir, saveto)
mpath_best = os.path.join(model_dir, prfx("best", saveto))
mpath_last = os.path.join(model_dir, prfx("last", saveto))
mpath_stats = os.path.join(model_dir, prfx("stats", saveto))
# Model options
model_options = locals().copy()
model_options['use_sent_reps'] = opt_ds['use_sent_reps']
stats = defaultdict(list)
del model_options['eyem']
del model_options['cost_mask']
if cost_mask is not None:
cost_mask = sharedX(cost_mask)
# reload options and parameters
if reload_:
print "Reloading the model."
if os.path.exists(mpath_best):
print "Reloading the best model from %s." % mpath_best
with open(os.path.join(mpath_best, '%s.pkl' % mpath_best), 'rb') as f:
models_options = pkl.load(f)
params = init_params(model_options)
params = load_params(mpath_best, params)
elif os.path.exists(mpath):
print "Reloading the model from %s." % mpath
with open(os.path.join(mpath, '%s.pkl' % mpath), 'rb') as f:
models_options = pkl.load(f)
params = init_params(model_options)
params = load_params(mpath, params)
else:
raise IOError("Couldn't open the file.")
else:
print "Couldn't reload the models initializing from scratch."
params = init_params(model_options)
if datasets[0]:
print "Short dataset", datasets[0]
print 'Loading data'
print 'Building model'
if pkl_train_files is None or pkl_valid_files is None:
train, valid, test = load_data(path=datasets[0],
valid_path=valid_datasets[0],
test_path=valid_datasets[1],
batch_size=batch_size,
**opt_ds)
else:
train, valid, test = load_pkl_data(train_file_paths=pkl_train_files,
valid_file_paths=pkl_valid_files,
batch_size=batch_size,
vocab=vocab,
eyem=eyem,
**opt_ds)
tparams = init_tparams(params)
trng, use_noise, inps_d, \
opt_ret, \
cost, errors, ent_errors, ent_derrors, probs = \
build_model(tparams,
model_options,
prepare_data if not opt_ds['use_sent_reps'] \
else prepare_data_sents,
valid,
cost_mask=cost_mask)
alphas = opt_ret['dec_alphas']
if opt_ds['use_sent_reps']:
inps = [inps_d["desc"], \
inps_d["word_mask"], \
inps_d["q"], \
inps_d['q_mask'], \
inps_d['ans'], \
inps_d['wlen'],
inps_d['slen'], inps_d['qlen'],\
inps_d['ent_mask']
]
else:
inps = [inps_d["desc"], \
inps_d["word_mask"], \
inps_d["q"], \
inps_d['q_mask'], \
inps_d['ans'], \
inps_d['wlen'], \
inps_d['qlen'], \
inps_d['ent_mask']]
outs = [cost, errors, probs, alphas]
if ent_errors:
outs += [ent_errors]
if ent_derrors:
outs += [ent_derrors]
# before any regularizer
print 'Building f_log_probs...',
f_log_probs = theano.function(inps, outs, profile=profile)
print 'Done'
# Apply weight decay on the feed-forward connections
if decay_c > 0.:
decay_c = theano.shared(numpy.float32(decay_c), name='decay_c')
weight_decay = 0.
for kk, vv in tparams.iteritems():
if "logit" in kk or "ff" in kk:
weight_decay += (vv ** 2).sum()
weight_decay *= decay_c
cost += weight_decay
# after any regularizer
print 'Computing gradient...',
grads = safe_grad(cost, itemlist(tparams))
print 'Done'
# Gradient clipping:
if clip_c > 0.:
g2 = get_norms(grads)
for p, g in grads.iteritems():
grads[p] = tensor.switch(g2 > (clip_c**2),
(g / tensor.sqrt(g2 + 1e-8)) * clip_c,
g)
inps.pop()
if optimizer.lower() == "adasecant":
learning_rule = Adasecant(delta_clip=25.0,
use_adagrad=True,
grad_clip=0.25,
gamma_clip=0.)
elif optimizer.lower() == "rmsprop":
learning_rule = RMSPropMomentum(init_momentum=momentum)
elif optimizer.lower() == "adam":
learning_rule = Adam()
elif optimizer.lower() == "adadelta":
learning_rule = AdaDelta()
lr = tensor.scalar(name='lr')
print 'Building optimizers...',
learning_rule = None
if learning_rule:
f_grad_shared, f_update = learning_rule.get_funcs(learning_rate=lr,
grads=grads,
inp=inps,
cost=cost,
errors=errors)
else:
f_grad_shared, f_update = eval(optimizer)(lr,
tparams,
grads,
inps,
cost,
errors)
print 'Done'
print 'Optimization'
history_errs = []
# reload history
if reload_ and os.path.exists(mpath):
history_errs = list(numpy.load(mpath)['history_errs'])
best_p = None
bad_count = 0
if validFreq == -1:
validFreq = len(train[0]) / batch_size
if saveFreq == -1:
saveFreq = len(train[0]) / batch_size
best_found = False
uidx = 0
estop = False
train_cost_ave, train_err_ave, \
train_gnorm_ave = reset_train_vals()
for eidx in xrange(max_epochs):
n_samples = 0
if train.done:
train.reset()
for d_, q_, a, em in train:
n_samples += len(a)
uidx += 1
use_noise.set_value(1.)
if opt_ds['use_sent_reps']:
# To mask the description and the question.
d, d_mask, q, q_mask, dlen, slen, qlen = prepare_data_sents(d_,
q_)
if d is None:
print 'Minibatch with zero sample under length ', maxlen
uidx -= 1
continue
ud_start = time.time()
cost, errors, gnorm, pnorm = f_grad_shared(d,
d_mask,
q,
q_mask,
a,
dlen,
slen,
qlen)
else:
d, d_mask, q, q_mask, dlen, qlen = prepare_data(d_, q_)
if d is None:
print 'Minibatch with zero sample under length ', maxlen
uidx -= 1
continue
ud_start = time.time()
cost, errors, gnorm, pnorm = f_grad_shared(d, d_mask,
q, q_mask,
a,
dlen,
qlen)
upnorm = f_update(lrate)
ud = time.time() - ud_start
# Collect the running ave train stats.
train_cost_ave = running_ave(train_cost_ave,
cost)
train_err_ave = running_ave(train_err_ave,
errors)
train_gnorm_ave = running_ave(train_gnorm_ave,
gnorm)
if numpy.isnan(cost) or numpy.isinf(cost):
print 'NaN detected'
import ipdb; ipdb.set_trace()
if numpy.mod(uidx, dispFreq) == 0:
print 'Epoch ', eidx, ' Update ', uidx, \
' Cost ', cost, ' UD ', ud, \
' UpNorm ', upnorm[0].tolist(), \
' GNorm ', gnorm, \
' Pnorm ', pnorm, 'Terrors ', errors
if numpy.mod(uidx, saveFreq) == 0:
print 'Saving...',
if best_p is not None and best_found:
numpy.savez(mpath_best, history_errs=history_errs, **best_p)
pkl.dump(model_options, open('%s.pkl' % mpath_best, 'wb'))
else:
params = unzip(tparams)
numpy.savez(mpath, history_errs=history_errs, **params)
pkl.dump(model_options, open('%s.pkl' % mpath, 'wb'))
pkl.dump(stats, open("%s.pkl" % mpath_stats, 'wb'))
print 'Done'
print_param_norms(tparams)
if numpy.mod(uidx, validFreq) == 0:
use_noise.set_value(0.)
if valid.done:
valid.reset()
valid_costs, valid_errs, valid_probs, \
valid_alphas, error_ent, error_dent = eval_model(f_log_probs,
prepare_data if not opt_ds['use_sent_reps'] \
else prepare_data_sents,
model_options,
valid,
use_sent_rep=opt_ds['use_sent_reps'])
valid_alphas_ = numpy.concatenate([va.argmax(0) for va in valid_alphas.tolist()], axis=0)
valid_err = valid_errs.mean()
valid_cost = valid_costs.mean()
valid_alpha_ent = -negentropy(valid_alphas)
mean_valid_alphas = valid_alphas_.mean()
std_valid_alphas = valid_alphas_.std()
mean_valid_probs = valid_probs.argmax(1).mean()
std_valid_probs = valid_probs.argmax(1).std()
history_errs.append([valid_cost, valid_err])
stats['train_err_ave'].append(train_err_ave)
stats['train_cost_ave'].append(train_cost_ave)
stats['train_gnorm_ave'].append(train_gnorm_ave)
stats['valid_errs'].append(valid_err)
stats['valid_costs'].append(valid_cost)
stats['valid_err_ent'].append(error_ent)
stats['valid_err_desc_ent'].append(error_dent)
stats['valid_alphas_mean'].append(mean_valid_alphas)
stats['valid_alphas_std'].append(std_valid_alphas)
stats['valid_alphas_ent'].append(valid_alpha_ent)
stats['valid_probs_mean'].append(mean_valid_probs)
stats['valid_probs_std'].append(std_valid_probs)
if uidx == 0 or valid_err <= numpy.array(history_errs)[:, 1].min():
best_p = unzip(tparams)
bad_counter = 0
best_found = True
else:
bst_found = False
if numpy.isnan(valid_err):
import ipdb; ipdb.set_trace()
print "============================"
print '\t>>>Valid error: ', valid_err, \
' Valid cost: ', valid_cost
print '\t>>>Valid pred mean: ', mean_valid_probs, \
' Valid pred std: ', std_valid_probs
print '\t>>>Valid alphas mean: ', mean_valid_alphas, \
' Valid alphas std: ', std_valid_alphas, \
' Valid alpha negent: ', valid_alpha_ent, \
' Valid error ent: ', error_ent, \
' Valid error desc ent: ', error_dent
print "============================"
print "Running average train stats "
print '\t>>>Train error: ', train_err_ave, \
' Train cost: ', train_cost_ave, \
' Train grad norm: ', train_gnorm_ave
print "============================"
train_cost_ave, train_err_ave, \
train_gnorm_ave = reset_train_vals()
print 'Seen %d samples' % n_samples
if estop:
break
if best_p is not None:
zipp(best_p, tparams)
use_noise.set_value(0.)
valid.reset()
valid_cost, valid_error, valid_probs, \
valid_alphas, error_ent = eval_model(f_log_probs,
prepare_data if not opt_ds['use_sent_reps'] \
else prepare_data_sents,
model_options, valid,
use_sent_rep=opt_ds['use_sent_rep'])
print " Final eval resuts: "
print 'Valid error: ', valid_error.mean()
print 'Valid cost: ', valid_cost.mean()
print '\t>>>Valid pred mean: ', valid_probs.mean(), \
' Valid pred std: ', valid_probs.std(), \
' Valid error ent: ', error_ent
params = copy.copy(best_p)
numpy.savez(mpath_last,
zipped_params=best_p,
history_errs=history_errs,
**params)
return valid_err, valid_cost
def fflayer(tparams, state_below, options, prefix='rconv', activ='lambda x: tensor.tanh(x)', **kwargs):
return eval(activ)(tensor.dot(state_below, tparams[prfx(prefix,'W')]) + tparams[prfx(prefix,'b')])
def gru_cond_layer(tparams,
state_below,
options,
prefix='gru',
mask=None,
context=None,
one_step=False,
init_memory=None,
init_state=None,
context_mask=None,
**kwargs):
assert context, 'Context must be provided'
if one_step:
assert init_state, 'previous state must be provided'
nsteps = state_below.shape[0]
if state_below.ndim == 3:
n_samples = state_below.shape[1]
else:
n_samples = 1
# mask
if mask == None:
mask = tensor.alloc(1., state_below.shape[0], 1)
dim = tparams[prfx(prefix, 'Wcx')].shape[1]
# initial/previous state
if init_state == None:
init_state = tensor.alloc(0., n_samples, dim)
# projected context
assert context.ndim == 3, 'Context must be 3-d: #annotation x #sample x dim'
pctx_ = tensor.dot(context, tparams[prfx(prefix,'Wc_att')]) + tparams[prfx(prefix,'b_att')]
def _slice(_x, n, dim):
if _x.ndim == 3:
return _x[:, :, n*dim:(n+1)*dim]
return _x[:, n*dim:(n+1)*dim]
# projected x
state_belowx = tensor.dot(state_below, tparams[prfx(prefix, 'Wx')]) + tparams[prfx(prefix, 'bx')]
state_below_ = tensor.dot(state_below, tparams[prfx(prefix, 'W')]) + tparams[prfx(prefix, 'b')]
state_belowc = tensor.dot(state_below, tparams[prfx(prefix, 'Wi_att')])
def _step_slice(m_, x_, xx_, xc_, h_, ctx_, alpha_, pctx_, cc_,
U, Wc, Wd_att, U_att, c_tt, Ux, Wcx):
# attention
pstate_ = tensor.dot(h_, Wd_att)
pctx__ = pctx_ + pstate_[None,:,:]
pctx__ += xc_
pctx__ = tensor.tanh(pctx__)
alpha = tensor.dot(pctx__, U_att)+c_tt
alpha = alpha.reshape([alpha.shape[0], alpha.shape[1]])
alpha = tensor.exp(alpha)
if context_mask:
alpha = alpha * context_mask
alpha = alpha / alpha.sum(0, keepdims=True)
ctx_ = (cc_ * alpha[:,:,None]).sum(0) # current context
preact = tensor.dot(h_, U)
preact += x_
preact += tensor.dot(ctx_, Wc)
preact = tensor.nnet.sigmoid(preact)
r = _slice(preact, 0, dim)
u = _slice(preact, 1, dim)
preactx = tensor.dot(h_, Ux)
preactx *= r
preactx += xx_
preactx += tensor.dot(ctx_, Wcx)
h = tensor.tanh(preactx)
h = u * h_ + (1. - u) * h
h = m_[:,None] * h + (1. - m_)[:,None] * h_
return h, ctx_, alpha.T
seqs = [mask, state_below_, state_belowx, state_belowc]
_step = _step_slice
shared_vars = [tparams[prfx(prefix, 'U')],
tparams[prfx(prefix, 'Wc')],
tparams[prfx(prefix,'Wd_att')],
tparams[prfx(prefix,'U_att')],
tparams[prfx(prefix, 'c_tt')],
tparams[prfx(prefix, 'Ux')],
tparams[prfx(prefix, 'Wcx')]]
if one_step:
rval = _step(*(seqs+[init_state, None, None, pctx_, context]+shared_vars))
else:
rval, updates = theano.scan(_step,
sequences=seqs,
outputs_info = [init_state,
tensor.alloc(0., n_samples, context.shape[2]),
tensor.alloc(0., n_samples, context.shape[0])],
non_sequences=[pctx_,
context]+shared_vars,
name=prfx(prefix, '_layers'),
n_steps=nsteps,
profile=profile,
strict=True)
return rval
