rewards_all /= 10
print('Epoch {:>3} Batch {:>4}/{} - Actor Loss: {:>6.4f} - Critic Loss: {:>6.4f} - Train Reward: {:>6.4f} - Valid Reward: {:>6.4f}'
.format(epoch_i, batch_i, len(source_int_text) // batch_size, loss_actor, loss_critic, reward_batch, rewards_all))
loss_list.append((count, loss_actor))
loss_list_critic.append((count, loss_critic))
reward_list.append((count, reward_batch))
reward_valid_list.append((count, rewards_all))
with train_sess.as_default():
with train_sess.graph.as_default():
# Save Model
saver_actor = tf.train.Saver()
saver_actor.save(train_sess, save_path)
with critic_sess.as_default():
with critic_sess.graph.as_default():
# Save Model
saver_critic = tf.train.Saver()
saver_critic.save(critic_sess, save_path_critic)
# plot
plot_loss(loss_list, "actor loss", "img/actor_loss.jpg")
plot_loss(loss_list_critic, "critic loss", "img/critic_loss.jpg")
plot_loss(reward_valid_list, "training reward", "img/training_reward.jpg")
#save model
util.save_params(save_path, 'params_actor_reinforce.p')
util.save_params(save_path_critic, 'params_critic_sup.p')
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 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()
print('reward:', rewards_c.shape)
# print(len(rewards))
reward_batch = np.sum(rewards)
with critic_sess.as_default():
with critic_sess.graph.as_default():
# critic_sess.run(tf.global_variables_initializer())
baselines, _, loss_, m = critic_sess.run(
[training_logits_, train_op_, l_, masks_], {
input_data_: target_batch,
targets_: translate_logits,
rewards_: rewards_c,
lr_: learning_rate,
target_sequence_length_: lens,
source_sequence_length_: targets_lengths,
keep_prob_: keep_probability
})
# print(m)
print('baseline:', baselines)
print(baselines.shape)
print('critic loss:', loss_)
# print(baselines.shape)
loss_list.append((count, loss_))
saver.save(critic_sess, save_path_critic)
print('Model Trained and Saved')
util.save_params(save_path_critic, 'params_critic_sup.p')
plot_loss(loss_list)
def train(m,
story_buckets,
bucket_sizes,
len_answers,
output_format,
num_updates,
outputdir,
start=0,
batch_size=BATCH_SIZE,
validation_buckets=None,
validation_bucket_sizes=None,
stop_at_accuracy=None,
stop_at_loss=None,
stop_at_overfitting=None,
save_params=1000,
validation_interval=1000,
batch_auto_adjust=None,
interrupt_file=None):
with GracefulInterruptHandler() as interrupt_h:
for i in range(start + 1, num_updates + 1):
exit_with = None
cur_bucket, cur_bucket_size = random.choice(
list(zip(story_buckets, bucket_sizes)))
cur_batch_size = adj_size(m, cur_bucket_size, batch_size,
batch_auto_adjust)
sampled_batch = sample_batch(cur_bucket, cur_batch_size,
len_answers, output_format)
loss, info = m.train(*sampled_batch)
if np.any(np.isnan(loss)):
print("Loss at timestep {} was nan! Aborting".format(i))
return TrainExitStatus.nan_loss # Don't bother saving
with open(os.path.join(outputdir, 'data.csv'), 'a') as f:
if i == 1:
f.seek(0)
f.truncate()
keylist = "iter, loss, " + ", ".join(
k for k, v in sorted(info.items())) + "\n"
f.write(keylist)
if validation_buckets is not None:
with open(os.path.join(outputdir, 'valid.csv'),
'w') as f2:
f2.write(keylist)
f.write("{}, {},".format(i, loss) +
", ".join(str(v)
for k, v in sorted(info.items())) + "\n")
if i % 1 == 0:
print("update {}: {}\n{}".format(i, loss, pformat(info)))
if i % validation_interval == 0:
if validation_buckets is not None:
cur_bucket, cur_bucket_size = random.choice(
list(zip(validation_buckets, validation_bucket_sizes)))
cur_batch_size = adj_size(m, cur_bucket_size, batch_size,
batch_auto_adjust)
sampled_batch = sample_batch(cur_bucket, cur_batch_size,
len_answers, output_format)
valid_loss, valid_info = m.eval(*sampled_batch)
print("validation at {}: {}\n{}".format(
i, valid_loss, pformat(valid_info)))
with open(os.path.join(outputdir, 'valid.csv'), 'a') as f:
f.write("{}, {}, ".format(i, valid_loss) + ", ".join(
str(v)
for k, v in sorted(valid_info.items())) + "\n")
valid_accuracy = test_accuracy(m, validation_buckets,
validation_bucket_sizes,
len_answers, output_format,
batch_size,
batch_auto_adjust,
(not m.train_with_query))
print("Best-choice accuracy at {}: {}".format(
i, valid_accuracy))
with open(os.path.join(outputdir, 'valid_acc.csv'),
'a') as f:
f.write("{}, {}\n".format(i, valid_accuracy))
if stop_at_accuracy is not None and valid_accuracy >= stop_at_accuracy:
print("Accuracy reached threshold! Stopping training")
exit_with = TrainExitStatus.success
if stop_at_loss is not None and valid_loss <= stop_at_loss:
print("Loss reached threshold! Stopping training")
exit_with = TrainExitStatus.success
if stop_at_overfitting is not None and valid_loss / loss > stop_at_overfitting:
print(
"Model appears to be overfitting! Stopping training"
)
exit_with = TrainExitStatus.overfitting
if exit_with is None and (interrupt_h.interrupted or
(interrupt_file is not None
and os.path.isfile(interrupt_file))):
exit_with = TrainExitStatus.interrupted
if (save_params is not None and i % save_params
== 0) or (exit_with is not None) or (i == num_updates):
util.save_params(
m.params,
open(os.path.join(outputdir, 'params{}.p'.format(i)),
'wb'))
if exit_with is not None:
return exit_with
return TrainExitStatus.reached_update_limit
def train(opts):
run = u.DTS()
logging.info("run %s", run)
# only run wandb stuff if it's configured, and only on primary host
wandb_enabled = (opts.group is not None) and u.primary_host()
if wandb_enabled:
wandb.init(project='evolved_channel_selection',
group=opts.group,
name=run,
reinit=True)
# save group again explicitly to work around sync bug that drops
# group when 'wandb off'
wandb.config.group = opts.group
wandb.config.seed = opts.seed
wandb.config.learning_rate = opts.learning_rate
wandb.config.batch_size = opts.batch_size
wandb.config.input_size = opts.input_size
wandb.config.dropout_channels = opts.dropout_channels
else:
logging.info("not using wandb and/or not primary host")
host_rng = jax.random.PRNGKey(opts.seed ^ jax.host_id())
pod_rng = jax.random.PRNGKey(opts.seed - 1)
pod_rng, init_key = jax.random.split(pod_rng)
representative_input = jnp.zeros((1, opts.input_size, opts.input_size, 13))
if opts.model_type == 'single':
model = models.construct_single_trunk_model()
params = model.init(init_key, representative_input)
elif opts.model_type == 'multi-res':
model = models.construct_multires_model()
representative_channel_selection = jnp.zeros(13, )
params = model.init(init_key, representative_input,
representative_channel_selection)
else:
raise Exception(opts.model_type)
# logging.debug("params %s", u.shapes_of(params))
def calc_logits(params, x, dropout_key):
if opts.model_type == 'single':
# TODO: move channel masking from data pipeline to here to
# more consistent compared to multi-res
return model.apply(params, x)
else: # multi-res
# TODO: handle --fixed-channel-selection &
# --random-channel-selection here
channel_selection = jax.random.randint(dropout_key,
minval=0,
maxval=5,
shape=(13, ))
return model.apply(params, x, channel_selection)
@jit
def mean_cross_entropy(params, x, y_true, dropout_key):
logits = calc_logits(params, x, dropout_key)
return jnp.mean(u.softmax_cross_entropy(logits, y_true))
opt = optax.adam(opts.learning_rate)
opt_state = opt.init(params)
@jit
def update(params, opt_state, x, y_true, dropout_key):
grads = grad(mean_cross_entropy)(params, x, y_true, dropout_key)
updates, opt_state = opt.update(grads, opt_state, params)
params = optax.apply_updates(params, updates)
return params, opt_state
best_validation_accuracy = 0
best_validation_epoch = None
for epoch in range(opts.epochs):
# make one pass through training set
dropout_key = None
if opts.dropout_channels:
host_rng, dropout_key = jax.random.split(host_rng)
train_dataset = d.dataset(split='train',
batch_size=opts.batch_size,
input_size=opts.input_size,
dropout_key=dropout_key)
for x, y_true in train_dataset:
host_rng, dropout_key = jax.random.split(host_rng)
params, opt_state = update(params, opt_state, x, y_true,
dropout_key)
# just report loss for final batch (note: this is _post_ the grad update)
mean_last_batch_loss = mean_cross_entropy(params, x, y_true,
dropout_key)
# calculate validation loss
validate_dataset = d.dataset(split='validate',
batch_size=opts.batch_size,
input_size=opts.input_size)
def calc_logits_for_validation(x):
if opts.model_type == 'single':
return model.apply(params, x)
else: # multi-res
# TODO: this favours validation stop based only on x64.
just_select_x64 = jnp.array([0] * 13)
return model.apply(params, x, just_select_x64)
accuracy, _mean_loss = u.accuracy_mean_loss(calc_logits_for_validation,
validate_dataset)
if accuracy > best_validation_accuracy:
best_validation_accuracy = accuracy
best_validation_epoch = epoch
u.save_params(run, epoch, params)
stats = {
'loss': float(mean_last_batch_loss),
'validate_accuracy': accuracy
}
logging.info("epoch %d stats %s", epoch, stats)
if wandb_enabled:
wandb.log(stats, step=epoch)
final_stats = {
'best_validation_accuracy': best_validation_accuracy,
'best_validation_epoch': best_validation_epoch
}
logging.info("final_stats %s", final_stats)
if wandb_enabled:
wandb.log(final_stats, step=epoch)
wandb.join()
logging.info("best params params/%s/%s.pkl", run, best_validation_epoch)
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()
source_sequence_length: sources_lengths,
target_sequence_length: targets_lengths,
keep_prob: 1.0
})
batch_valid_logits = sess.run(
inference_logits, {
input_data: valid_sources_batch,
source_sequence_length: valid_sources_lengths,
target_sequence_length: valid_targets_lengths,
keep_prob: 1.0
})
train_acc = get_accuracy(target_batch, batch_train_logits)
valid_acc = get_accuracy(valid_targets_batch,
batch_valid_logits)
print(
'Epoch {:>3} Batch {:>4}/{} - Train Accuracy: {:>6.4f}, Validation Accuracy: {:>6.4f}, Loss: {:>6.4f}'
.format(epoch_i, batch_i,
len(source_int_text) // batch_size, train_acc,
valid_acc, loss))
# Save Model
saver = tf.train.Saver()
saver.save(sess, save_path)
print('Model Trained and Saved')
util.save_params(save_path)
if batch_i % display_step == 0 and batch_i > 0:
rewards_all = 0
for batch_j, (valid_sources_batch, valid_targets_batch, valid_sources_lengths, valid_targets_lengths) in enumerate(
util.get_batches(valid_source, valid_target, batch_size,
source_vocab_to_int['<PAD>'],
target_vocab_to_int['<PAD>'])):
batch_valid_logits = sess.run(
inferencing_logits,
{input_data: valid_sources_batch,
source_sequence_length: valid_sources_lengths,
target_sequence_length: valid_targets_lengths,
keep_prob: 1.0})
rewards_v, pre_logits = util.get_bleu(valid_targets_batch, batch_valid_logits)
rewards_v = np.sum(rewards_v) / batch_size
rewards_all += rewards_v
rewards_all /= 10
print('Epoch {:>3} Batch {:>4}/{} - Loss: {:>6.4f} - Valid bleu: {:>6.4f}'
.format(epoch_i, batch_i, len(source_int_text) // batch_size, loss, rewards_all))
# Save Model
saver = tf.train.Saver()
saver.save(sess, save_path)
print('Model Trained and Saved')
plot_loss(loss_list, "pretrain loss", "img/pretrain_loss.jpg")
# Save parameters for checkpoint# Save p
util.save_params(save_path, 'params_actor_sup.p')
batches = get_batches(int_text, batch_size, seq_length)
with tf.Session(graph=train_graph) as sess:
sess.run(tf.global_variables_initializer())
for epoch_i in range(num_epochs):
state = sess.run(initial_state, {input_text: batches[0][0]})
for batch_i, (x, y) in enumerate(batches):
feed = {
input_text: x,
targets: y,
initial_state: state,
lr: learning_rate
}
train_loss, state, _ = sess.run([cost, final_state, train_op],
feed)
# Show every <show_every_n_batches> batches
if (epoch_i * len(batches) + batch_i) % show_every_n_batches == 0:
print(
'Epoch {:>3} Batch {:>4}/{} train_loss = {:.3f}'.format(
epoch_i, batch_i, len(batches), train_loss))
# Save Model
saver = tf.train.Saver()
saver.save(sess, save_dir)
print('Model Trained and Saved')
util.save_params((seq_length, save_dir))