def sgd(lr, tparams, grads, inp, cost):
gshared = [
theano.shared(p.get_value() * 0., name='%s_grad' % k)
for k, p in six.iteritems(tparams)
]
gsup = [(gs, g) for gs, g in zip(gshared, grads)]
f_grad_shared = theano.function(inp, cost, updates=gsup, name='sgd')
pup = [(p, p - lr * g) for p, g in zip(itemlist(tparams), gshared)]
f_update = theano.function([lr], [], updates=pup)
return f_grad_shared, f_update
def rmsprop(lr, tparams, grads, inp, cost):
zipped_grads = [
theano.shared(p.get_value() * np.float32(0.), name='%s_grad' % k)
for k, p in six.iteritems(tparams)
]
running_grads = [
theano.shared(p.get_value() * np.float32(0.), name='%s_rgrad' % k)
for k, p in six.iteritems(tparams)
]
running_grads2 = [
theano.shared(p.get_value() * np.float32(0.), name='%s_rgrad2' % k)
for k, p in six.iteritems(tparams)
]
zgup = [(zg, g) for zg, g in zip(zipped_grads, grads)]
rgup = [(rg, 0.95 * rg + 0.05 * g) for rg, g in zip(running_grads, grads)]
rg2up = [(rg2, 0.95 * rg2 + 0.05 * (g**2))
for rg2, g in zip(running_grads2, grads)]
f_grad_shared = theano.function(inp,
cost,
updates=zgup + rgup + rg2up,
name='rmsprop')
updir = [
theano.shared(p.get_value() * np.float32(0.), name='%s_updir' % k)
for k, p in six.iteritems(tparams)
]
updir_new = [(ud, 0.9 * ud - 1e-4 * zg / tensor.sqrt(rg2 - rg**2 + 1e-4))
for ud, zg, rg, rg2 in zip(updir, zipped_grads, running_grads,
running_grads2)]
param_up = [(p, p + udn[1])
for p, udn in zip(itemlist(tparams), updir_new)]
f_update = theano.function([lr], [],
updates=updir_new + param_up,
on_unused_input='ignore')
return f_grad_shared, f_update
def adadelta(lr, tparams, grads, inp, cost):
zipped_grads = [
theano.shared(p.get_value() * np.float32(0.), name='%s_grad' % k)
for k, p in six.iteritems(tparams)
]
running_up2 = [
theano.shared(p.get_value() * np.float32(0.), name='%s_rup2' % k)
for k, p in six.iteritems(tparams)
]
running_grads2 = [
theano.shared(p.get_value() * np.float32(0.), name='%s_rgrad2' % k)
for k, p in six.iteritems(tparams)
]
zgup = [(zg, g) for zg, g in zip(zipped_grads, grads)]
rg2up = [(rg2, 0.95 * rg2 + 0.05 * (g**2))
for rg2, g in zip(running_grads2, grads)]
f_grad_shared = theano.function(inp,
cost,
updates=zgup + rg2up,
name='adadelta')
updir = [
-tensor.sqrt(ru2 + 1e-6) / tensor.sqrt(rg2 + 1e-6) * zg
for zg, ru2, rg2 in zip(zipped_grads, running_up2, running_grads2)
]
ru2up = [(ru2, 0.95 * ru2 + 0.05 * (ud**2))
for ru2, ud in zip(running_up2, updir)]
param_up = [(p, p + ud) for p, ud in zip(itemlist(tparams), updir)]
f_update = theano.function([lr], [],
updates=ru2up + param_up,
on_unused_input='ignore')
return f_grad_shared, f_update
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
def train(args, model_args):
#model_id = '/data/lisatmp4/lambalex/lsun_walkback/walkback_'
model_id = '/data/lisatmp4/anirudhg/cifar_walk_back/walkback_'
model_dir = create_log_dir(args, model_id)
model_id2 = 'logs/walkback_'
model_dir2 = create_log_dir(args, model_id2)
print model_dir
print model_dir2 + '/' + 'log.jsonl.gz'
logger = mimir.Logger(filename=model_dir2 + '/log.jsonl.gz',
formatter=None)
# TODO batches_per_epoch should not be hard coded
lrate = args.lr
import sys
sys.setrecursionlimit(10000000)
args, model_args = parse_args()
#trng = RandomStreams(1234)
if args.resume_file is not None:
print "Resuming training from " + args.resume_file
from blocks.scripts import continue_training
continue_training(args.resume_file)
## load the training data
if args.dataset == 'MNIST':
print 'loading MNIST'
from fuel.datasets import MNIST
dataset_train = MNIST(['train'], sources=('features', ))
dataset_test = MNIST(['test'], sources=('features', ))
n_colors = 1
spatial_width = 28
elif args.dataset == 'CIFAR10':
from fuel.datasets import CIFAR10
dataset_train = CIFAR10(['train'], sources=('features', ))
dataset_test = CIFAR10(['test'], sources=('features', ))
n_colors = 3
spatial_width = 32
elif args.dataset == "lsun" or args.dataset == "lsunsmall":
print "loading lsun class!"
from load_lsun import load_lsun
print "loading lsun data!"
if args.dataset == "lsunsmall":
dataset_train, dataset_test = load_lsun(args.batch_size,
downsample=True)
spatial_width = 32
else:
dataset_train, dataset_test = load_lsun(args.batch_size,
downsample=False)
spatial_width = 64
n_colors = 3
elif args.dataset == "celeba":
print "loading celeba data"
from fuel.datasets.celeba import CelebA
dataset_train = CelebA(which_sets=['train'],
which_format="64",
sources=('features', ),
load_in_memory=False)
dataset_test = CelebA(which_sets=['test'],
which_format="64",
sources=('features', ),
load_in_memory=False)
spatial_width = 64
n_colors = 3
tr_scheme = SequentialScheme(examples=dataset_train.num_examples,
batch_size=args.batch_size)
ts_scheme = SequentialScheme(examples=dataset_test.num_examples,
batch_size=args.batch_size)
train_stream = DataStream.default_stream(dataset_train,
iteration_scheme=tr_scheme)
test_stream = DataStream.default_stream(dataset_test,
iteration_scheme=ts_scheme)
dataset_train = train_stream
dataset_test = test_stream
#epoch_it = train_stream.get_epoch_iterator()
elif args.dataset == 'Spiral':
print 'loading SPIRAL'
train_set = Spiral(num_examples=100000,
classes=1,
cycles=2.,
noise=0.01,
sources=('features', ))
dataset_train = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(train_set.num_examples,
args.batch_size))
else:
raise ValueError("Unknown dataset %s." % args.dataset)
model_options = locals().copy()
if args.dataset != 'lsun' and args.dataset != 'celeba':
train_stream = Flatten(
DataStream.default_stream(
dataset_train,
iteration_scheme=ShuffledScheme(
examples=dataset_train.num_examples -
(dataset_train.num_examples % args.batch_size),
batch_size=args.batch_size)))
else:
train_stream = dataset_train
test_stream = dataset_test
print "Width", WIDTH, spatial_width
shp = next(train_stream.get_epoch_iterator())[0].shape
print "got epoch iterator"
# make the training data 0 mean and variance 1
# TODO compute mean and variance on full dataset, not minibatch
Xbatch = next(train_stream.get_epoch_iterator())[0]
scl = 1. / np.sqrt(np.mean((Xbatch - np.mean(Xbatch))**2))
shft = -np.mean(Xbatch * scl)
# scale is applied before shift
#train_stream = ScaleAndShift(train_stream, scl, shft)
#test_stream = ScaleAndShift(test_stream, scl, shft)
print 'Building model'
params = init_params(model_options)
if args.reload_:
print "Trying to reload parameters"
if os.path.exists(args.saveto_filename):
print 'Reloading Parameters'
print args.saveto_filename
params = load_params(args.saveto_filename, params)
tparams = init_tparams(params)
print tparams
'''
x = T.matrix('x', dtype='float32')
temp = T.scalar('temp', dtype='float32')
f=transition_operator(tparams, model_options, x, temp)
for data in train_stream.get_epoch_iterator():
print data[0]
a = f([data[0], 1.0, 1])
#ipdb.set_trace()
'''
x, cost, start_temperature = build_model(tparams, model_options)
inps = [x, start_temperature]
x_Data = T.matrix('x_Data', dtype='float32')
temperature = T.scalar('temperature', dtype='float32')
forward_diffusion = one_step_diffusion(x_Data, model_options, tparams,
temperature)
#print 'Building f_cost...',
#f_cost = theano.function(inps, cost)
#print 'Done'
print tparams
grads = T.grad(cost, wrt=itemlist(tparams))
#get_grads = theano.function(inps, grads)
for j in range(0, len(grads)):
grads[j] = T.switch(T.isnan(grads[j]), T.zeros_like(grads[j]),
grads[j])
# compile the optimizer, the actual computational graph is compiled here
lr = T.scalar(name='lr')
print 'Building optimizers...',
optimizer = args.optimizer
f_grad_shared, f_update = getattr(optimizers, optimizer)(lr, tparams,
grads, inps, cost)
print 'Done'
for param in tparams:
print param
print tparams[param].get_value().shape
print 'Buiding Sampler....'
f_sample = sample(tparams, model_options)
print 'Done'
uidx = 0
estop = False
bad_counter = 0
max_epochs = 4000
batch_index = 1
print 'Number of steps....'
print args.num_steps
print "Number of metasteps...."
print args.meta_steps
print 'Done'
count_sample = 1
for eidx in xrange(max_epochs):
if eidx % 20 == 0:
params = unzip(tparams)
save_params(params,
model_dir + '/' + 'params_' + str(eidx) + '.npz')
n_samples = 0
print 'Starting Next Epoch ', eidx
for data in train_stream.get_epoch_iterator():
if args.dataset == 'CIFAR10':
if data[0].shape[0] == args.batch_size:
data_use = (data[0].reshape(args.batch_size,
3 * 32 * 32), )
else:
continue
t0 = time.time()
batch_index += 1
n_samples += len(data_use[0])
uidx += 1
if data_use[0] is None:
print 'No data '
uidx -= 1
continue
ud_start = time.time()
t1 = time.time()
data_run = data_use[0]
temperature_forward = args.temperature
meta_cost = []
for meta_step in range(0, args.meta_steps):
meta_cost.append(f_grad_shared(data_run, temperature_forward))
f_update(lrate)
if args.meta_steps > 1:
data_run, sigma, _, _ = forward_diffusion(
[data_run, temperature_forward, 1])
temperature_forward *= args.temperature_factor
cost = sum(meta_cost) / len(meta_cost)
ud = time.time() - ud_start
#gradient_updates_ = get_grads(data_use[0],args.temperature)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
return 1.
t1 = time.time()
#print time.time() - t1, "time to get grads"
t1 = time.time()
logger.log({
'epoch': eidx,
'batch_index': batch_index,
'uidx': uidx,
'training_error': cost
})
#'Norm_1': np.linalg.norm(gradient_updates_[0]),
#'Norm_2': np.linalg.norm(gradient_updates_[1]),
#'Norm_3': np.linalg.norm(gradient_updates_[2]),
#'Norm_4': np.linalg.norm(gradient_updates_[3])})
#print time.time() - t1, "time to log"
#print time.time() - t0, "total time in batch"
t5 = time.time()
if batch_index % 20 == 0:
print batch_index, "cost", cost
if batch_index % 200 == 0:
count_sample += 1
temperature = args.temperature * (args.temperature_factor**(
args.num_steps * args.meta_steps - 1))
temperature_forward = args.temperature
for num_step in range(args.num_steps * args.meta_steps):
print "Forward temperature", temperature_forward
if num_step == 0:
x_data, sampled, sampled_activation, sampled_preactivation = forward_diffusion(
[data_use[0], temperature_forward, 1])
x_data = np.asarray(x_data).astype('float32').reshape(
args.batch_size, INPUT_SIZE)
x_temp = x_data.reshape(args.batch_size, n_colors,
WIDTH, WIDTH)
plot_images(
x_temp, model_dir + '/' + "batch_" +
str(batch_index) + '_corrupted' + 'epoch_' +
str(count_sample) + '_time_step_' + str(num_step))
else:
x_data, sampled, sampled_activation, sampled_preactivation = forward_diffusion(
[x_data, temperature_forward, 1])
x_data = np.asarray(x_data).astype('float32').reshape(
args.batch_size, INPUT_SIZE)
x_temp = x_data.reshape(args.batch_size, n_colors,
WIDTH, WIDTH)
plot_images(
x_temp, model_dir + '/batch_' + str(batch_index) +
'_corrupted' + '_epoch_' + str(count_sample) +
'_time_step_' + str(num_step))
temperature_forward = temperature_forward * args.temperature_factor
x_temp2 = data_use[0].reshape(args.batch_size, n_colors, WIDTH,
WIDTH)
plot_images(
x_temp2, model_dir + '/' + 'orig_' + 'epoch_' + str(eidx) +
'_batch_index_' + str(batch_index))
temperature = args.temperature * (args.temperature_factor**(
args.num_steps * args.meta_steps - 1))
for i in range(args.num_steps * args.meta_steps +
args.extra_steps):
x_data, sampled, sampled_activation, sampled_preactivation = f_sample(
[x_data, temperature, 0])
print 'On backward step number, using temperature', i, temperature
reverse_time(
scl, shft, x_data, model_dir + '/' + "batch_" +
str(batch_index) + '_samples_backward_' + 'epoch_' +
str(count_sample) + '_time_step_' + str(i))
x_data = np.asarray(x_data).astype('float32')
x_data = x_data.reshape(args.batch_size, INPUT_SIZE)
if temperature == args.temperature:
temperature = temperature
else:
temperature /= args.temperature_factor
if args.noise == "gaussian":
x_sampled = np.random.normal(
0.5, 2.0,
size=(args.batch_size, INPUT_SIZE)).clip(0.0, 1.0)
else:
s = np.random.binomial(1, 0.5, INPUT_SIZE)
temperature = args.temperature * (args.temperature_factor**(
args.num_steps * args.meta_steps - 1))
x_data = np.asarray(x_sampled).astype('float32')
for i in range(args.num_steps * args.meta_steps +
args.extra_steps):
x_data, sampled, sampled_activation, sampled_preactivation = f_sample(
[x_data, temperature, 0])
print 'On step number, using temperature', i, temperature
reverse_time(
scl, shft, x_data, model_dir + '/batch_index_' +
str(batch_index) + '_inference_' + 'epoch_' +
str(count_sample) + '_step_' + str(i))
x_data = np.asarray(x_data).astype('float32')
x_data = x_data.reshape(args.batch_size, INPUT_SIZE)
if temperature == args.temperature:
temperature = temperature
else:
temperature /= args.temperature_factor
ipdb.set_trace()
def train(args, model_args):
model_id = '/data/lisatmp4/anirudhg/spiral_walk_back/walkback_'
model_dir = create_log_dir(args, model_id)
model_id2 = 'logs/walkback_'
model_dir2 = create_log_dir(args, model_id2)
print model_dir
print model_dir2 + '/' + 'log.jsonl.gz'
logger = mimir.Logger(filename=model_dir2 + '/log.jsonl.gz',
formatter=None)
# TODO batches_per_epoch should not be hard coded
lrate = args.lr
import sys
sys.setrecursionlimit(10000000)
args, model_args = parse_args()
#trng = RandomStreams(1234)
if args.resume_file is not None:
print "Resuming training from " + args.resume_file
from blocks.scripts import continue_training
continue_training(args.resume_file)
## load the training data
if args.dataset == 'MNIST':
print 'loading MNIST'
from fuel.datasets import MNIST
dataset_train = MNIST(['train'], sources=('features', ))
dataset_test = MNIST(['test'], sources=('features', ))
n_colors = 1
spatial_width = 28
elif args.dataset == 'CIFAR10':
from fuel.datasets import CIFAR10
dataset_train = CIFAR10(['train'], sources=('features', ))
dataset_test = CIFAR10(['test'], sources=('features', ))
n_colors = 3
spatial_width = 32
elif args.dataset == "lsun" or args.dataset == "lsunsmall":
print "loading lsun class!"
from load_lsun import load_lsun
print "loading lsun data!"
if args.dataset == "lsunsmall":
dataset_train, dataset_test = load_lsun(args.batch_size,
downsample=True)
spatial_width = 32
else:
dataset_train, dataset_test = load_lsun(args.batch_size,
downsample=False)
spatial_width = 64
n_colors = 3
elif args.dataset == "celeba":
print "loading celeba data"
from fuel.datasets.celeba import CelebA
dataset_train = CelebA(which_sets=['train'],
which_format="64",
sources=('features', ),
load_in_memory=False)
dataset_test = CelebA(which_sets=['test'],
which_format="64",
sources=('features', ),
load_in_memory=False)
spatial_width = 64
n_colors = 3
tr_scheme = SequentialScheme(examples=dataset_train.num_examples,
batch_size=args.batch_size)
ts_scheme = SequentialScheme(examples=dataset_test.num_examples,
batch_size=args.batch_size)
train_stream = DataStream.default_stream(dataset_train,
iteration_scheme=tr_scheme)
test_stream = DataStream.default_stream(dataset_test,
iteration_scheme=ts_scheme)
dataset_train = train_stream
dataset_test = test_stream
#epoch_it = train_stream.get_epoch_iterator()
elif args.dataset == 'Spiral':
print 'loading SPIRAL'
train_set = Spiral(num_examples=20000,
classes=1,
cycles=1.,
noise=0.01,
sources=('features', ))
dataset_train = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(train_set.num_examples,
args.batch_size))
elif args.dataset == 'Circle':
print 'loading Circle'
train_set = Circle(num_examples=20000,
classes=1,
cycles=1.,
noise=0.0,
sources=('features', ))
dataset_train = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(train_set.num_examples,
args.batch_size))
iter_per_epoch = train_set.num_examples
else:
raise ValueError("Unknown dataset %s." % args.dataset)
model_options = locals().copy()
train_stream = dataset_train
shp = next(train_stream.get_epoch_iterator())[0].shape
print "got epoch iterator"
# make the training data 0 mean and variance 1
# TODO compute mean and variance on full dataset, not minibatch
Xbatch = next(train_stream.get_epoch_iterator())[0]
scl = 1. / np.sqrt(np.mean((Xbatch - np.mean(Xbatch))**2))
shft = -np.mean(Xbatch * scl)
# scale is applied before shift
#train_stream = ScaleAndShift(train_stream, scl, shft)
#test_stream = ScaleAndShift(test_stream, scl, shft)
print 'Building model'
params = init_params(model_options)
if args.reload_:
print "Trying to reload parameters"
if os.path.exists(args.saveto_filename):
print 'Reloading Parameters'
print args.saveto_filename
params = load_params(args.saveto_filename, params)
tparams = init_tparams(params)
print tparams
x, cost, start_temperature = build_model(tparams, model_options)
inps = [x, start_temperature]
x_Data = T.matrix('x_Data', dtype='float32')
temperature = T.scalar('temperature', dtype='float32')
forward_diffusion = one_step_diffusion(x_Data, model_options, tparams,
temperature)
#print 'Building f_cost...',
#f_cost = theano.function(inps, cost)
#print 'Done'
print tparams
grads = T.grad(cost, wrt=itemlist(tparams))
#get_grads = theano.function(inps, grads)
for j in range(0, len(grads)):
grads[j] = T.switch(T.isnan(grads[j]), T.zeros_like(grads[j]),
grads[j])
# compile the optimizer, the actual computational graph is compiled here
lr = T.scalar(name='lr')
print 'Building optimizers...',
optimizer = args.optimizer
f_grad_shared, f_update = getattr(optimizers, optimizer)(lr, tparams,
grads, inps, cost)
print 'Done'
print 'Buiding Sampler....'
f_sample = sample(tparams, model_options)
print 'Done'
uidx = 0
estop = False
bad_counter = 0
max_epochs = 4000
batch_index = 0
print 'Number of steps....', args.num_steps
print 'Done'
count_sample = 1
batch_index = 0
for eidx in xrange(max_epochs):
if eidx % 20 == 0:
params = unzip(tparams)
save_params(params,
model_dir + '/' + 'params_' + str(eidx) + '.npz')
if eidx == 30:
ipdb.set_trace()
n_samples = 0
print 'Starting Next Epoch ', eidx
for data in train_stream.get_epoch_iterator():
batch_index += 1
n_samples += len(data[0])
uidx += 1
if data[0] is None:
print 'No data '
uidx -= 1
continue
data_run = data[0]
temperature_forward = args.temperature
meta_cost = []
for meta_step in range(0, args.meta_steps):
meta_cost.append(f_grad_shared(data_run, temperature_forward))
f_update(lrate)
if args.meta_steps > 1:
data_run, sigma, _, _ = forward_diffusion(
data_run, temperature_forward)
temperature_forward *= args.temperature_factor
cost = sum(meta_cost) / len(meta_cost)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
return 1.
logger.log({
'epoch': eidx,
'batch_index': batch_index,
'uidx': uidx,
'training_error': cost
})
empty = []
spiral_x = [empty for i in range(args.num_steps)]
spiral_corrupted = []
spiral_sampled = []
grad_forward = []
grad_back = []
x_data_time = []
x_tilt_time = []
if batch_index % 8 == 0:
count_sample += 1
temperature = args.temperature * (args.temperature_factor
**(args.num_steps - 1))
temperature_forward = args.temperature
for num_step in range(args.num_steps):
if num_step == 0:
x_data_time.append(data[0])
plot_images(
data[0], model_dir + '/' + 'orig_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
x_data, mu_data, _, _ = forward_diffusion(
data[0], temperature_forward)
plot_images(
x_data, model_dir + '/' + 'corrupted_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index) +
'_time_step_' + str(num_step))
x_data_time.append(x_data)
temp_grad = np.concatenate(
(x_data_time[-2], x_data_time[-1]), axis=1)
grad_forward.append(temp_grad)
x_data = np.asarray(x_data).astype('float32').reshape(
args.batch_size, INPUT_SIZE)
spiral_corrupted.append(x_data)
mu_data = np.asarray(mu_data).astype(
'float32').reshape(args.batch_size, INPUT_SIZE)
mu_data = mu_data.reshape(args.batch_size, 2)
else:
x_data_time.append(x_data)
x_data, mu_data, _, _ = forward_diffusion(
x_data, temperature_forward)
plot_images(
x_data, model_dir + '/' + 'corrupted_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index) +
'_time_step_' + str(num_step))
x_data = np.asarray(x_data).astype('float32').reshape(
args.batch_size, INPUT_SIZE)
spiral_corrupted.append(x_data)
mu_data = np.asarray(mu_data).astype(
'float32').reshape(args.batch_size, INPUT_SIZE)
mu_data = mu_data.reshape(args.batch_size, 2)
x_data_time.append(x_data)
temp_grad = np.concatenate(
(x_data_time[-2], x_data_time[-1]), axis=1)
grad_forward.append(temp_grad)
temperature_forward = temperature_forward * args.temperature_factor
mean_sampled = x_data.mean()
var_sampled = x_data.var()
x_temp2 = data[0].reshape(args.batch_size, 2)
plot_2D(
spiral_corrupted, args.num_steps,
model_dir + '/' + 'corrupted_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
plot_2D(
x_temp2, 1, model_dir + '/' + 'orig_' + 'epoch_' +
str(count_sample) + '_batch_index_' + str(batch_index))
plot_grad(
grad_forward,
model_dir + '/' + 'grad_forward_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
for i in range(args.num_steps + args.extra_steps):
x_tilt_time.append(x_data)
x_data, sampled_mean = f_sample(x_data, temperature)
plot_images(
x_data, model_dir + '/' + 'sampled_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index) +
'_time_step_' + str(i))
x_tilt_time.append(x_data)
temp_grad = np.concatenate(
(x_tilt_time[-2], x_tilt_time[-1]), axis=1)
grad_back.append(temp_grad)
###print 'Recons, On step number, using temperature', i, temperature
x_data = np.asarray(x_data).astype('float32')
x_data = x_data.reshape(args.batch_size, INPUT_SIZE)
if temperature == args.temperature:
temperature = temperature
else:
temperature /= args.temperature_factor
plot_grad(
grad_back, model_dir + '/' + 'grad_back_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
plot_2D(
x_tilt_time, args.num_steps,
model_dir + '/' + 'sampled_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
s = np.random.normal(mean_sampled, var_sampled,
[args.batch_size, 2])
x_sampled = s
temperature = args.temperature * (args.temperature_factor
**(args.num_steps - 1))
x_data = np.asarray(x_sampled).astype('float32')
for i in range(args.num_steps + args.extra_steps):
x_data, sampled_mean = f_sample(x_data, temperature)
spiral_sampled.append(x_data)
x_data = np.asarray(x_data).astype('float32')
x_data = x_data.reshape(args.batch_size, INPUT_SIZE)
if temperature == args.temperature:
temperature = temperature
else:
temperature /= args.temperature_factor
plot_2D(
spiral_sampled, args.num_steps,
model_dir + '/' + 'inference_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
ipdb.set_trace()
def build_model(options,tparams):
"""Build up the whole computation graph
Input is the features extracted from googleNet.
"""
last_n = options['last_n'];
actionNum=options['actions'];
decay_c=options['decay_c'];
use_dropout=options['use_dropout'];
use_wta=options['use_wta'];
featureMaps=options['featureMaps'];
videosize=options['videosize'];
trng = RandomStreams(1234);
use_noise = theano.shared(np.float32(0.));
"""combine model"""
x=T.ftensor3('x');
n_steps=x.shape[0]; n_samples=x.shape[1];
mask = T.fmatrix('mask');
y=T.ftensor3('y'); # one hot vector,n_steps*n_samples*actionNum
_x=x.reshape([n_steps*n_samples,3,videosize,videosize]);
_x=_x+use_noise*trng.normal(_x.shape,avg=0,std=0.1,dtype=_x.dtype);
feature=CNN_block(_x);
feature=feature.reshape([n_steps,n_samples,featureMaps]);
if use_dropout: feature = dropout_layer(feature, use_noise, trng);
f0=ff_build(tparams,feature,prefix="recog",name='fc0',active="tanh");
f1=ff_build(tparams,f0,prefix="recog",name='fc1',active="tanh");
# if use_dropout: f0 = dropout_layer(f0, use_noise, trng);
lin=ff_build(tparams,f1,prefix="recog",name='output',active="linear"); # n_steps*n_samples*actionNum
probs=T.nnet.softmax(lin.reshape([-1,actionNum]));
probs=probs.reshape([n_steps,n_samples,actionNum]);
"""compute cost"""
cost=0;
# cross entropy
entropy_cost=-y*T.log(probs);
entropy_cost = (entropy_cost.sum(2)*mask).mean(0).sum();
cost+=entropy_cost;
# weight decay
weight_decay = 0.;
if decay_c > 0.:
decay_c = theano.shared(np.float32(decay_c), name='decay_c');
for vv in itemlist(tparams):
# if (vv.ndim==1):
weight_decay += (vv ** 2).sum()
# else:
# weight_decay+=(vv.sum()-vv.shape[0])**2;
# weight_decay += (vv ** 2).sum()
weight_decay *= decay_c;
cost+=weight_decay;
"""Predictions"""
preds = T.sum(probs[-last_n:,:,:],axis=0);
preds = T.argmax(preds,axis=1); # n_samples
# preds=T.argmax(probs[-last_n:,:,:],axis=2);
return cost,preds,[],[x,mask,y],use_noise;
def train(args, model_args, lrate):
model_id = '/data/lisatmp4/anirudhg/minst_walk_back/walkback_'
model_dir = create_log_dir(args, model_id)
model_id2 = 'walkback_'
model_dir2 = create_log_dir(args, model_id2)
print model_dir
logger = mimir.Logger(filename=model_dir2 + '/' + model_id2 +
'log.jsonl.gz',
formatter=None)
# TODO batches_per_epoch should not be hard coded
lrate = args.lr
import sys
sys.setrecursionlimit(10000000)
args, model_args = parse_args()
#trng = RandomStreams(1234)
if args.resume_file is not None:
print "Resuming training from " + args.resume_file
from blocks.scripts import continue_training
continue_training(args.resume_file)
## load the training data
if args.dataset == 'MNIST':
print 'loading MNIST'
from fuel.datasets import MNIST
dataset_train = MNIST(['train'], sources=('features', ))
dataset_test = MNIST(['test'], sources=('features', ))
n_colors = 1
spatial_width = 28
elif args.dataset == 'CIFAR10':
from fuel.datasets import CIFAR10
dataset_train = CIFAR10(['train'], sources=('features', ))
dataset_test = CIFAR10(['test'], sources=('features', ))
n_colors = 3
spatial_width = 32
elif args.dataset == 'Spiral':
print 'loading SPIRAL'
train_set = Spiral(num_examples=100000,
classes=1,
cycles=2.,
noise=0.01,
sources=('features', ))
dataset_train = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(train_set.num_examples,
args.batch_size))
else:
raise ValueError("Unknown dataset %s." % args.dataset)
model_options = locals().copy()
train_stream = Flatten(
DataStream.default_stream(dataset_train,
iteration_scheme=ShuffledScheme(
examples=dataset_train.num_examples,
batch_size=args.batch_size)))
shp = next(train_stream.get_epoch_iterator())[0].shape
# make the training data 0 mean and variance 1
# TODO compute mean and variance on full dataset, not minibatch
Xbatch = next(train_stream.get_epoch_iterator())[0]
scl = 1. / np.sqrt(np.mean((Xbatch - np.mean(Xbatch))**2))
shft = -np.mean(Xbatch * scl)
# scale is applied before shift
#train_stream = ScaleAndShift(train_stream, scl, shft)
#test_stream = ScaleAndShift(test_stream, scl, shft)
print 'Building model'
params = init_params(model_options)
if args.reload_ and os.path.exists(args.saveto_filename):
print 'Reloading Parameters'
print args.saveto_filename
params = load_params(args.saveto_filename, params)
tparams = init_tparams(params)
'''
x = T.matrix('x', dtype='float32')
f=transition_operator(tparams, model_options, x, 1)
for data in train_stream.get_epoch_iterator():
print data[0]
a = f(data[0])
print a
ipdb.set_trace()
'''
x, cost = build_model(tparams, model_options)
inps = [x]
x_Data = T.matrix('x_Data', dtype='float32')
temperature = T.scalar('temperature', dtype='float32')
forward_diffusion = one_step_diffusion(x_Data, model_options, tparams,
temperature)
print 'Building f_cost...',
f_cost = theano.function(inps, cost)
print 'Done'
print tparams
grads = T.grad(cost, wrt=itemlist(tparams))
get_grads = theano.function(inps, grads)
for j in range(0, len(grads)):
grads[j] = T.switch(T.isnan(grads[j]), T.zeros_like(grads[j]),
grads[j])
# compile the optimizer, the actual computational graph is compiled here
lr = T.scalar(name='lr')
print 'Building optimizers...',
optimizer = args.optimizer
f_grad_shared, f_update = getattr(optimizers, optimizer)(lr, tparams,
grads, inps, cost)
print 'Done'
print 'Buiding Sampler....'
f_sample = sample(tparams, model_options)
print 'Done'
uidx = 0
estop = False
bad_counter = 0
max_epochs = 4000
batch_index = 0
print 'Number of steps....'
print args.num_steps
print 'Done'
count_sample = 1
for eidx in xrange(max_epochs):
n_samples = 0
print 'Starting Next Epoch ', eidx
for data in train_stream.get_epoch_iterator():
batch_index += 1
n_samples += len(data[0])
uidx += 1
if data[0] is None:
print 'No data '
uidx -= 1
continue
ud_start = time.time()
cost = f_grad_shared(data[0])
f_update(lrate)
ud = time.time() - ud_start
if batch_index % 1 == 0:
print 'Cost is this', cost
count_sample += 1
from impainting import change_image, inpainting
train_temp = data[0]
print data[0].shape
change_image(train_temp.reshape(args.batch_size, 1, 28, 28), 3)
train_temp = train_temp.reshape(args.batch_size, 784)
output = inpainting(train_temp)
change_image(output.reshape(args.batch_size, 1, 28, 28), 1)
reverse_time(
scl, shft, output,
model_dir + '/' + 'impainting_orig_' + 'epoch_' +
str(count_sample) + '_batch_index_' + str(batch_index))
x_data = np.asarray(output).astype('float32')
temperature = args.temperature * (args.temperature_factor
**(args.num_steps - 1))
temperature = args.temperature #* (args.temperature_factor ** (args.num_steps -1 ))
orig_impainted_data = np.asarray(data[0]).astype('float32')
for i in range(args.num_steps + args.extra_steps + 5):
x_data, sampled, sampled_activation, sampled_preactivation = f_sample(
x_data, temperature)
print 'Impainting using temperature', i, temperature
x_data = do_half_image(x_data, orig_impainted_data)
reverse_time(
scl, shft, x_data, model_dir + '/' +
'impainting_orig_' + 'epoch_' + str(count_sample) +
'_batch_index_' + str(batch_index) + 'step_' + str(i))
x_data = np.asarray(x_data).astype('float32')
x_data = x_data.reshape(args.batch_size, INPUT_SIZE)
if temperature == args.temperature:
temperature = temperature
else:
temperature = temperature
#temperature /= args.temperature_factor
ipdb.set_trace()
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)
def train():
model = 'theano'
if model == 'theano':
model_options = locals().copy()
params = init_params(model_options)
tparams = init_tparams(params)
x, u, m, v, q = build_model(tparams, model_options)
reward = T.matrix('rewards')
gamma_t = T.scalar('reward')
cost = T.sqr(q - (reward + gamma_t * v))
inps = [x, u, q, reward, gamma_t, v]
inps_cost = [q, reward, gamma_t, v]
_mu = theano.function([x], m)
mu = lambda x: _mu(x)
model = theano.function([x, u], q)
params_target = init_params_target(model_options)
tparams_target = init_tparams(params_target)
x, u, m, v, q = build_model_target(tparams_target, model_options)
_V = theano.function([x], v)
V = lambda x: _V(x)
lr = T.scalar(name='lr')
f_cost = theano.function(inps_cost, cost, on_unused_input='warn')
grads = T.grad(cost.mean(), wrt=itemlist(tparams))
optimizer = 'adam'
f_grad_shared, f_update = getattr(optimizers,
optimizer)(lr, tparams, grads, inps,
cost)
for key, key2 in zip(tparams_target.items(), tparams.items()):
a, b = key
c, d = key2
params_target[a] = params[c]
prestates = []
actions = []
rewards = []
poststates = []
terminals = []
total_reward = 0
lrate = optimizer_lr
for i_episode in xrange(episodes):
observation = env.reset()
print "initial state:", observation, max_timesteps
episode_reward = 0
for t in xrange(max_timesteps):
if display:
env.render()
x = np.array([observation])
u = mu(x)
if noise_decay == 'linear':
noise = 1. / (i_episode + 1)
elif noise_decay == 'exp':
noise = 10**-i_episode
elif noise_decay == 'fixed':
noise = fixed_noise
else:
assert False
action = u[0] + np.random.randn(num_actuators) * noise
prestates.append(observation)
actions.append(action)
observation, reward, done, info = env.step(action)
episode_reward += reward
rewards.append(reward)
poststates.append(observation)
terminals.append(done)
if len(prestates) > min_train:
loss = 0
for k in xrange(train_repeat):
if len(prestates) > batch_size:
indexes = np.random.choice(len(prestates),
size=batch_size)
else:
indexes = range(len(prestates))
v = V(np.array(poststates)[indexes])
y = np.array(rewards)[indexes] + gamma * np.squeeze(v)
q = model(
np.array(prestates)[indexes],
np.array(actions)[indexes])
out = f_cost(
q,
np.array(rewards)[indexes].reshape(
(y.shape[0], 1)).astype('float32'), gamma,
np.squeeze(v).reshape((y.shape[0], 1)))
loss += out.mean()
f_grad_shared(
np.array(prestates)[indexes],
np.array(actions)[indexes], q,
np.array(rewards)[indexes].reshape(
(y.shape[0], 1)).astype('float32'), gamma,
np.squeeze(v).reshape((y.shape[0], 1)))
f_update(lrate)
for key, key2 in zip(tparams_target.items(),
tparams.items()):
a, b = key
c, d = key2
params_target[a] = params[c] * tau + params_target[
a] * (1. - tau)
if done:
break
episode_reward = episode_reward / float(t + 1)
print "Episode {} finished after {} timesteps, average reward {}".format(
i_episode + 1, t + 1, episode_reward)
total_reward += episode_reward
print "Average reward per episode {}".format(total_reward / episodes)
def train(args, model_args):
model_id = '/data/lisatmp3/anirudhg/spiral_walk_back/walkback_'
model_dir = create_log_dir(args, model_id)
model_id2 = 'walkback_'
model_dir2 = create_log_dir(args, model_id2)
print model_dir, model_dir2
logger = mimir.Logger(filename=model_dir2 + '/' + model_id2 +
'log.jsonl.gz',
formatter=None)
# TODO batches_per_epoch should not be hard coded
lrate = args.lr
import sys
sys.setrecursionlimit(10000000)
args, model_args = parse_args()
#trng = RandomStreams(1234)
if args.resume_file is not None:
print "Resuming training from " + args.resume_file
from blocks.scripts import continue_training
continue_training(args.resume_file)
## load the training data
if args.dataset == 'MNIST':
print 'loading MNIST'
from fuel.datasets import MNIST
dataset_train = MNIST(['train'], sources=('features', ))
dataset_test = MNIST(['test'], sources=('features', ))
n_colors = 1
spatial_width = 28
elif args.dataset == 'Spiral':
print 'loading SPIRAL'
train_set = Spiral(num_examples=20000,
classes=1,
cycles=1.,
noise=0.0,
sources=('features', ))
dataset_train = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(train_set.num_examples,
args.batch_size))
iter_per_epoch = train_set.num_examples
print iter_per_epoch
elif args.dataset == 'Circle':
print 'loading SPIRAL'
train_set = Circle(num_examples=20000,
classes=1,
cycles=1.,
noise=0.0,
sources=('features', ))
dataset_train = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(train_set.num_examples,
args.batch_size))
iter_per_epoch = train_set.num_examples
print iter_per_epoch
else:
raise ValueError("Unknown dataset %s." % args.dataset)
model_options = locals().copy()
train_stream = dataset_train
shp = next(train_stream.get_epoch_iterator())[0].shape
# make the training data 0 mean and variance 1
# TODO compute mean and variance on full dataset, not minibatch
Xbatch = next(train_stream.get_epoch_iterator())[0]
scl = 1. / np.sqrt(np.mean((Xbatch - np.mean(Xbatch))**2))
shft = -np.mean(Xbatch * scl)
print 'Building model'
params = init_params(model_options)
if args.reload_ and os.path.exists(args.saveto_filename):
print 'Reloading Parameters'
print args.saveto_filename
params = load_params(args.saveto_filename, params)
tparams = init_tparams(params)
x = T.matrix('x', dtype='float32')
x, cost, cost_each_step = build_model(tparams, model_options)
inps = [x]
x_Data = T.matrix('x_Data', dtype='float32')
temperature = T.scalar('temperature', dtype='float32')
forward_diffusion, debug_function = one_step_diffusion(
x_Data, model_options, tparams, temperature)
print 'Building f_cost...',
f_cost = theano.function(inps, cost)
f_cost_each_step = theano.function(inps, cost_each_step)
print 'Done'
grads = T.grad(cost, wrt=itemlist(tparams))
get_grads = theano.function(inps, grads)
for j in range(0, len(grads)):
grads[j] = T.switch(T.isnan(grads[j]), T.zeros_like(grads[j]),
grads[j])
# compile the optimizer, the actual computational graph is compiled here
lr = T.scalar(name='lr')
print 'Building optimizers...',
optimizer = args.optimizer
f_grad_shared, f_update = getattr(optimizers, optimizer)(lr, tparams,
grads, inps, cost)
print 'Done'
print 'Buiding Sampler....'
f_sample = sample(tparams, model_options)
print 'Done'
uidx = 0
estop = False
bad_counter = 0
max_epochs = 4000
batch_index = 0
print 'Number of steps....', args.num_steps
print 'Done'
count_sample = 1
for eidx in xrange(max_epochs):
if eidx % 20 == 0:
params = unzip(tparams)
save_params(params,
model_dir + '/' + 'params_' + str(eidx) + '.npz')
if eidx == 30:
ipdb.set_trace()
n_samples = 0
batch_index = 0
print 'Starting Next Epoch ', eidx
for data in train_stream.get_epoch_iterator():
batch_index += 1
n_samples += len(data[0])
uidx += 1
if data[0] is None:
print 'No data '
uidx -= 1
continue
plt.plot(*zip(*data[0]))
plt.show()
ud_start = time.time()
cost = f_grad_shared(data[0])
output_cost = f_cost_each_step(data[0])
output_cost = np.asarray(output_cost)
#cost_time_step_1, cost_time_step_2 = (output_cost[0,:].sum() *1.0) /args.batch_size, (output_cost[1,:].sum() * 1.0 )/args.batch_size
print 'Loss is', cost
f_update(lrate)
ud = time.time() - ud_start
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
return 1.
gradient_updates_ = get_grads(data[0])
logger.log({
'epoch': eidx,
'batch_index': batch_index,
'uidx': uidx,
'training_error': cost
})
empty = []
spiral_x = [empty for i in range(args.num_steps)]
spiral_corrupted = []
spiral_sampled = []
grad_forward = []
grad_back = []
x_data_time = []
x_tilt_time = []
if eidx % 10 == 0:
count_sample += 1
temperature = args.temperature * (args.temperature_factor
**(args.num_steps - 1))
temperature_forward = args.temperature
for num_step in range(args.num_steps):
if num_step == 0:
x_data_time.append(data[0])
plot_images(
data[0], model_dir + '/' + 'orig_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
x_data, mu_data = forward_diffusion(
data[0], temperature_forward)
plot_images(
x_data, model_dir + '/' + 'corrupted_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index) +
'_time_step_' + str(num_step))
x_data_time.append(x_data)
temp_grad = np.concatenate(
(x_data_time[-2], x_data_time[-1]), axis=1)
grad_forward.append(temp_grad)
x_data = np.asarray(x_data).astype('float32').reshape(
args.batch_size, INPUT_SIZE)
spiral_corrupted.append(x_data)
mu_data = np.asarray(mu_data).astype(
'float32').reshape(args.batch_size, INPUT_SIZE)
mu_data = mu_data.reshape(args.batch_size, 2)
else:
x_data_time.append(x_data)
x_data, mu_data = forward_diffusion(
x_data, temperature_forward)
plot_images(
x_data, model_dir + '/' + 'corrupted_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index) +
'_time_step_' + str(num_step))
x_data = np.asarray(x_data).astype('float32').reshape(
args.batch_size, INPUT_SIZE)
spiral_corrupted.append(x_data)
mu_data = np.asarray(mu_data).astype(
'float32').reshape(args.batch_size, INPUT_SIZE)
mu_data = mu_data.reshape(args.batch_size, 2)
x_data_time.append(x_data)
temp_grad = np.concatenate(
(x_data_time[-2], x_data_time[-1]), axis=1)
grad_forward.append(temp_grad)
temperature_forward = temperature_forward * args.temperature_factor
x_temp2 = data[0].reshape(args.batch_size, 2)
plot_2D(
spiral_corrupted, args.num_steps,
model_dir + '/' + 'corrupted_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
plot_2D(
x_temp2, 1, model_dir + '/' + 'orig_' + 'epoch_' +
str(count_sample) + '_batch_index_' + str(batch_index))
plot_grad(
grad_forward,
model_dir + '/' + 'grad_forward_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
for i in range(args.num_steps + args.extra_steps):
x_tilt_time.append(x_data)
x_data, sampled_mean = f_sample(x_data, temperature)
plot_images(
x_data, model_dir + '/' + 'sampled_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index) +
'_time_step_' + str(i))
x_tilt_time.append(x_data)
temp_grad = np.concatenate(
(x_tilt_time[-2], x_tilt_time[-1]), axis=1)
grad_back.append(temp_grad)
###print 'Recons, On step number, using temperature', i, temperature
x_data = np.asarray(x_data).astype('float32')
x_data = x_data.reshape(args.batch_size, INPUT_SIZE)
if temperature == args.temperature:
temperature = temperature
else:
temperature /= args.temperature_factor
plot_grad(
grad_back, model_dir + '/' + 'grad_back_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
np.savez('grad_backward.npz', grad_back)
plot_2D(
x_tilt_time, args.num_steps,
model_dir + '/' + 'sampled_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
s = np.random.normal(0, 1, [args.batch_size, 2])
x_sampled = s
temperature = args.temperature * (args.temperature_factor
**(args.num_steps - 1))
x_data = np.asarray(x_sampled).astype('float32')
for i in range(args.num_steps + args.extra_steps):
x_data, sampled_mean = f_sample(x_data, temperature)
spiral_sampled.append(x_data)
x_data = np.asarray(x_data).astype('float32')
x_data = x_data.reshape(args.batch_size, INPUT_SIZE)
if temperature == args.temperature:
temperature = temperature
else:
temperature /= args.temperature_factor
plot_2D(
spiral_sampled, args.num_steps,
model_dir + '/' + 'inference_' + 'epoch_' +
str(count_sample) + '_batch_' + str(batch_index))
ipdb.set_trace()
def train(args, model_args):
#model_id = '/data/lisatmp4/lambalex/lsun_walkback/walkback_'
model_id = '/data/lisatmp4/anirudhg/cifar_walk_back/walkback_'
model_dir = create_log_dir(args, model_id)
model_id2 = 'logs/walkback_'
model_dir2 = create_log_dir(args, model_id2)
print model_dir
print model_dir2 + '/' + 'log.jsonl.gz'
logger = mimir.Logger(filename=model_dir2 + '/log.jsonl.gz',
formatter=None)
# TODO batches_per_epoch should not be hard coded
lrate = args.lr
import sys
sys.setrecursionlimit(10000000)
args, model_args = parse_args()
#trng = RandomStreams(1234)
if args.resume_file is not None:
print "Resuming training from " + args.resume_file
from blocks.scripts import continue_training
continue_training(args.resume_file)
## load the training data
if args.dataset == 'MNIST':
print 'loading MNIST'
from fuel.datasets import MNIST
dataset_train = MNIST(['train'], sources=('features', ))
dataset_test = MNIST(['test'], sources=('features', ))
n_colors = 1
spatial_width = 28
elif args.dataset == 'CIFAR10':
from fuel.datasets import CIFAR10
dataset_train = CIFAR10(['train'], sources=('features', ))
dataset_test = CIFAR10(['test'], sources=('features', ))
n_colors = 3
spatial_width = 32
elif args.dataset == "lsun" or args.dataset == "lsunsmall":
print "loading lsun class!"
from load_lsun import load_lsun
print "loading lsun data!"
if args.dataset == "lsunsmall":
dataset_train, dataset_test = load_lsun(args.batch_size,
downsample=True)
spatial_width = 32
else:
dataset_train, dataset_test = load_lsun(args.batch_size,
downsample=False)
spatial_width = 64
n_colors = 3
elif args.dataset == "celeba":
print "loading celeba data"
from fuel.datasets.celeba import CelebA
dataset_train = CelebA(which_sets=['train'],
which_format="64",
sources=('features', ),
load_in_memory=False)
dataset_test = CelebA(which_sets=['test'],
which_format="64",
sources=('features', ),
load_in_memory=False)
spatial_width = 64
n_colors = 3
tr_scheme = SequentialScheme(examples=dataset_train.num_examples,
batch_size=args.batch_size)
ts_scheme = SequentialScheme(examples=dataset_test.num_examples,
batch_size=args.batch_size)
train_stream = DataStream.default_stream(dataset_train,
iteration_scheme=tr_scheme)
test_stream = DataStream.default_stream(dataset_test,
iteration_scheme=ts_scheme)
dataset_train = train_stream
dataset_test = test_stream
#epoch_it = train_stream.get_epoch_iterator()
elif args.dataset == 'Spiral':
print 'loading SPIRAL'
train_set = Spiral(num_examples=100000,
classes=1,
cycles=2.,
noise=0.01,
sources=('features', ))
dataset_train = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(train_set.num_examples,
args.batch_size))
else:
raise ValueError("Unknown dataset %s." % args.dataset)
model_options = locals().copy()
if args.dataset != 'lsun' and args.dataset != 'celeba':
train_stream = Flatten(
DataStream.default_stream(
dataset_train,
iteration_scheme=ShuffledScheme(
examples=dataset_train.num_examples -
(dataset_train.num_examples % args.batch_size),
batch_size=args.batch_size)))
else:
train_stream = dataset_train
test_stream = dataset_test
print "Width", WIDTH, spatial_width
shp = next(train_stream.get_epoch_iterator())[0].shape
print "got epoch iterator"
Xbatch = next(train_stream.get_epoch_iterator())[0]
scl = 1. / np.sqrt(np.mean((Xbatch - np.mean(Xbatch))**2))
shft = -np.mean(Xbatch * scl)
print 'Building model'
params = init_params(model_options)
if args.reload_:
print "Trying to reload parameters"
if os.path.exists(args.saveto_filename):
print 'Reloading Parameters'
print args.saveto_filename
params = load_params(args.saveto_filename, params)
tparams = init_tparams(params)
print tparams
x, cost, start_temperature, step_chain = build_model(
tparams, model_options)
inps = [x.astype('float32'), start_temperature, step_chain]
x_Data = T.matrix('x_Data', dtype='float32')
temperature = T.scalar('temperature', dtype='float32')
step_chain_part = T.scalar('step_chain_part', dtype='int32')
forward_diffusion = one_step_diffusion(x_Data, model_options, tparams,
temperature, step_chain_part)
print tparams
grads = T.grad(cost, wrt=itemlist(tparams))
#get_grads = theano.function(inps, grads)
for j in range(0, len(grads)):
grads[j] = T.switch(T.isnan(grads[j]), T.zeros_like(grads[j]),
grads[j])
# compile the optimizer, the actual computational graph is compiled here
lr = T.scalar(name='lr')
print 'Building optimizers...',
optimizer = args.optimizer
f_grad_shared, f_update = getattr(optimizers, optimizer)(lr, tparams,
grads, inps, cost)
print 'Done'
#for param in tparams:
# print param
# print tparams[param].get_value().shape
print 'Buiding Sampler....'
f_sample = sample(tparams, model_options)
print 'Done'
uidx = 0
estop = False
bad_counter = 0
max_epochs = 4000
batch_index = 1
print 'Number of steps....'
print args.num_steps
print "Number of metasteps...."
print args.meta_steps
print 'Done'
count_sample = 1
save_data = []
for eidx in xrange(1):
print 'Starting Next Epoch ', eidx
for data_ in range(500): #train_stream.get_epoch_iterator():
if args.noise == "gaussian":
x_sampled = np.random.normal(0.5,
2.0,
size=(args.batch_size,
INPUT_SIZE)).clip(0.0, 1.0)
else:
s = np.random.binomial(1, 0.5, INPUT_SIZE)
temperature = args.temperature * (args.temperature_factor**(
args.num_steps * args.meta_steps - 1))
x_data = np.asarray(x_sampled).astype('float32')
for i in range(args.num_steps * args.meta_steps +
args.extra_steps):
x_data, sampled, sampled_activation, sampled_preactivation = f_sample(
x_data.astype('float32'), temperature,
args.num_steps * args.meta_steps - i - 1)
print 'On step number, using temperature', i, temperature
#reverse_time(scl, shft, x_data, model_dir + '/batch_index_' + str(batch_index) + '_inference_' + 'epoch_' + str(count_sample) + '_step_' + str(i))
x_data = np.asarray(x_data).astype('float32')
x_data = x_data.reshape(args.batch_size, INPUT_SIZE)
if temperature == args.temperature:
temperature = temperature
else:
temperature /= args.temperature_factor
count_sample = count_sample + 1
save_data.append(x_data)
fname = model_dir + '/batch_index_' + str(
batch_index) + '_inference_' + 'epoch_' + str(
count_sample) + '_step_' + str(i)
np.savez(fname + '.npz', x_data)
save2_data = np.asarray(save_data).astype('float32')
fname = model_dir + '/generted_images_50000' #+ args.saveto_filename
np.savez(fname + '.npz', save_data)
ipdb.set_trace()