def main():
# parse the arguments
arg_parser = ArgumentParser()
spt.register_config_arguments(config, arg_parser, title='Model options')
spt.register_config_arguments(spt.settings,
arg_parser,
prefix='tfsnippet',
title='TFSnippet options')
arg_parser.parse_args(sys.argv[1:])
# print the config
print_with_title('Configurations', pformat(config.to_dict()), after='\n')
# open the result object and prepare for result directories
results = MLResults(config.result_dir)
results.save_config(config) # save experiment settings for review
results.make_dirs('plotting', exist_ok=True)
results.make_dirs('train_summary', exist_ok=True)
# input placeholders
input_x = tf.placeholder(dtype=tf.int32,
shape=(None, config.x_dim),
name='input_x')
learning_rate = spt.AnnealingVariable('learning_rate', config.initial_lr,
config.lr_anneal_factor)
# build the posterior flow
with tf.variable_scope('posterior_flow'):
flows = []
for i in range(config.n_flows):
flows.append(spt.layers.ActNorm())
flows.append(
spt.layers.CouplingLayer(tf.make_template(
'coupling',
coupling_layer_shift_and_scale,
create_scope_now_=True),
scale_type='exp'))
flows.append(spt.layers.InvertibleDense())
posterior_flow = spt.layers.SequentialFlow(flows=flows)
# derive the initialization op
with tf.name_scope('initialization'), \
arg_scope([spt.layers.act_norm], initializing=True), \
spt.utils.scoped_set_config(spt.settings, auto_histogram=False):
init_q_net = q_net(input_x, posterior_flow)
init_chain = init_q_net.chain(p_net, observed={'x': input_x})
init_loss = tf.reduce_mean(init_chain.vi.training.sgvb())
# derive the loss and lower-bound for training
with tf.name_scope('training'):
train_q_net = q_net(input_x, posterior_flow)
train_chain = train_q_net.chain(p_net, observed={'x': input_x})
vae_loss = tf.reduce_mean(train_chain.vi.training.sgvb())
loss = vae_loss + tf.losses.get_regularization_loss()
# derive the nll and logits output for testing
with tf.name_scope('testing'):
test_q_net = q_net(input_x, posterior_flow, n_z=config.test_n_z)
test_chain = test_q_net.chain(p_net,
latent_axis=0,
observed={'x': input_x})
test_nll = -tf.reduce_mean(test_chain.vi.evaluation.is_loglikelihood())
test_lb = tf.reduce_mean(test_chain.vi.lower_bound.elbo())
# derive the optimizer
with tf.name_scope('optimizing'):
optimizer = tf.train.AdamOptimizer(learning_rate)
params = tf.trainable_variables()
grads = optimizer.compute_gradients(loss, var_list=params)
with tf.control_dependencies(tf.get_collection(
tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.apply_gradients(grads)
# derive the plotting function
with tf.name_scope('plotting'):
plot_p_net = p_net(n_z=100)
x_plots = tf.reshape(bernoulli_as_pixel(plot_p_net['x']), (-1, 28, 28))
def plot_samples(loop):
with loop.timeit('plot_time'):
images = session.run(x_plots)
save_images_collection(images=images,
filename='plotting/{}.png'.format(
loop.epoch),
grid_size=(10, 10))
# prepare for training and testing data
(x_train, y_train), (x_test, y_test) = \
spt.datasets.load_mnist(x_shape=[784])
train_flow = bernoulli_flow(x_train,
config.batch_size,
shuffle=True,
skip_incomplete=True)
test_flow = bernoulli_flow(x_test, config.test_batch_size, sample_now=True)
with spt.utils.create_session().as_default() as session, \
train_flow.threaded(5) as train_flow:
# initialize the network
spt.utils.ensure_variables_initialized()
for [batch_x] in train_flow:
print('Network initialization loss: {:.6g}'.format(
session.run(init_loss, {input_x: batch_x})))
print('')
break
# train the network
with spt.TrainLoop(params,
var_groups=['p_net', 'q_net', 'posterior_flow'],
max_epoch=config.max_epoch,
max_step=config.max_step,
summary_dir=(results.system_path('train_summary')
if config.write_summary else None),
summary_graph=tf.get_default_graph(),
early_stopping=False) as loop:
trainer = spt.Trainer(loop,
train_op, [input_x],
train_flow,
metrics={'loss': loss},
summaries=tf.summary.merge_all(
spt.GraphKeys.AUTO_HISTOGRAM))
trainer.anneal_after(learning_rate,
epochs=config.lr_anneal_epoch_freq,
steps=config.lr_anneal_step_freq)
evaluator = spt.Evaluator(loop,
metrics={
'test_nll': test_nll,
'test_lb': test_lb
},
inputs=[input_x],
data_flow=test_flow,
time_metric_name='test_time')
evaluator.events.on(
spt.EventKeys.AFTER_EXECUTION,
lambda e: results.update_metrics(evaluator.last_metrics_dict))
trainer.evaluate_after_epochs(evaluator, freq=10)
trainer.evaluate_after_epochs(functools.partial(
plot_samples, loop),
freq=10)
trainer.log_after_epochs(freq=1)
trainer.run()
# print the final metrics and close the results object
print_with_title('Results', results.format_metrics(), before='\n')
results.close()
def main():
# parse the arguments
arg_parser = ArgumentParser()
spt.register_config_arguments(config, arg_parser, title='Model options')
spt.register_config_arguments(spt.settings,
arg_parser,
prefix='tfsnippet',
title='TFSnippet options')
arg_parser.parse_args(sys.argv[1:])
# print the config
print_with_title('Configurations', pformat(config.to_dict()), after='\n')
# open the result object and prepare for result directories
results = MLResults(config.result_dir)
results.save_config(config) # save experiment settings for review
results.make_dirs('plotting', exist_ok=True)
results.make_dirs('train_summary', exist_ok=True)
# input placeholders
input_x = tf.placeholder(dtype=tf.int32,
shape=(None, ) + config.x_shape,
name='input_x')
learning_rate = spt.AnnealingVariable('learning_rate', config.initial_lr,
config.lr_anneal_factor)
# derive the loss for initializing
with tf.name_scope('initialization'), \
arg_scope([p_net, q_net], is_initializing=True), \
spt.utils.scoped_set_config(spt.settings, auto_histogram=False):
init_q_net = q_net(input_x)
init_chain = init_q_net.chain(p_net, observed={'x': input_x})
init_loss = tf.reduce_mean(init_chain.vi.training.sgvb())
# derive the loss and lower-bound for training
with tf.name_scope('training'), \
arg_scope([p_net, q_net], is_training=True):
train_q_net = q_net(input_x)
train_chain = train_q_net.chain(p_net, observed={'x': input_x})
train_loss = (tf.reduce_mean(train_chain.vi.training.sgvb()) +
tf.losses.get_regularization_loss())
# derive the nll and logits output for testing
with tf.name_scope('testing'):
input_original_x = tf.placeholder(dtype=tf.float32,
shape=(None, ) + config.x_shape,
name='input_original_x')
test_q_net = q_net(input_x, n_z=config.test_n_z)
test_chain = test_q_net.chain(p_net,
latent_axis=0,
observed={'x': input_x})
test_nll = -tf.reduce_mean(test_chain.vi.evaluation.is_loglikelihood())
test_lb = tf.reduce_mean(test_chain.vi.lower_bound.elbo())
# bernoulli test
logits = test_chain.model['x'].distribution.logits
logits = tf.clip_by_value(logits,
clip_value_max=1 - 1e-7,
clip_value_min=1e-7)
labels = tf.tile(
tf.expand_dims(tf.to_float(test_chain.model['x']), axis=0),
tf.concat([[config.test_n_z],
[1] * spt.utils.get_rank(test_chain.model['x'])],
axis=0))
log_px_given_z = tf.reduce_sum(
-tf.nn.sigmoid_cross_entropy_with_logits(labels=labels,
logits=logits),
axis=list(range(-len(config.x_shape), 0)))
vi = spt.VariationalInference(
log_joint=log_px_given_z + test_chain.model['z'].log_prob(),
latent_log_probs=[test_q_net['z'].log_prob()],
axis=0)
adv_test_nll = -tf.reduce_mean(vi.evaluation.is_loglikelihood())
adv_test_lb = tf.reduce_mean(vi.lower_bound.elbo())
# derive the optimizer
with tf.name_scope('optimizing'):
params = tf.trainable_variables()
optimizer = tf.train.AdamOptimizer(learning_rate)
grads = optimizer.compute_gradients(train_loss, params)
with tf.control_dependencies(tf.get_collection(
tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.apply_gradients(grads)
# derive the plotting function
with tf.name_scope('plotting'):
x_plots = tf.reshape(bernoulli_as_pixel(p_net(n_z=100)['x']),
(-1, ) + config.x_shape)
def plot_samples(loop):
with loop.timeit('plot_time'):
images = session.run(x_plots)
save_images_collection(
images=images,
filename='plotting/{}.png'.format(loop.epoch),
grid_size=(10, 10),
results=results,
channels_last=config.channels_last,
)
# prepare for training and testing data
(x_train, y_train), (x_test, y_test) = \
spt.datasets.load_mnist(x_shape=config.x_shape)
train_flow = bernoulli_flow(x_train,
config.batch_size,
shuffle=True,
skip_incomplete=True)
test_flow = bernoulli_flow(x_test, config.test_batch_size, sample_now=True)
'''
bernoulli_test_flow = spt.DataFlow.arrays(
[x_test], config.test_batch_size)
'''
bernoulli_test_flow = spt.DataFlow.arrays([x_test, y_test],
config.test_batch_size)
bernoulli_test_flow = spt.DataFlow.gather([test_flow, bernoulli_test_flow])
with spt.utils.create_session().as_default() as session, \
train_flow.threaded(5) as train_flow:
spt.utils.ensure_variables_initialized()
# initialize the network
for [x] in train_flow:
print('Network initialized, first-batch loss is {:.6g}.\n'.format(
session.run(init_loss, feed_dict={input_x: x})))
break
# train the network
with spt.TrainLoop(params,
var_groups=['q_net', 'p_net'],
max_epoch=config.max_epoch,
max_step=config.max_step,
summary_dir=(results.system_path('train_summary')
if config.write_summary else None),
summary_graph=tf.get_default_graph(),
early_stopping=False) as loop:
trainer = spt.Trainer(loop,
train_op, [input_x],
train_flow,
metrics={'loss': train_loss},
summaries=tf.summary.merge_all(
spt.GraphKeys.AUTO_HISTOGRAM))
trainer.anneal_after(learning_rate,
epochs=config.lr_anneal_epoch_freq,
steps=config.lr_anneal_step_freq)
evaluator = spt.Evaluator(loop,
metrics={
'test_nll': test_nll,
'test_lb': test_lb
},
inputs=[input_x],
data_flow=test_flow,
time_metric_name='test_time')
evaluator.events.on(
spt.EventKeys.AFTER_EXECUTION,
lambda e: results.update_metrics(evaluator.last_metrics_dict))
'''
bernoulli_evaluator = spt.Evaluator(
loop,
metrics={'adv_test_nll': adv_test_nll, 'adv_test_lb': adv_test_lb, 'log_px_given_z': tf.reduce_mean(log_px_given_z)},
inputs=[input_x],
data_flow=bernoulli_test_flow,
time_metric_name='test_time'
)
'''
bernoulli_evaluator = spt.Evaluator(
loop,
metrics={
'adv_test_nll': adv_test_nll,
'adv_test_lb': adv_test_lb
},
inputs=[input_x, input_original_x],
data_flow=bernoulli_test_flow,
time_metric_name='test_time')
bernoulli_evaluator.events.on(
spt.EventKeys.AFTER_EXECUTION, lambda e: results.
update_metrics(bernoulli_evaluator.last_metrics_dict))
trainer.evaluate_after_epochs(evaluator, freq=1)
trainer.evaluate_after_epochs(bernoulli_evaluator, freq=1)
trainer.evaluate_after_epochs(functools.partial(
plot_samples, loop),
freq=1)
trainer.log_after_epochs(freq=1)
trainer.run()
# print the final metrics and close the results object
print_with_title('Results', results.format_metrics(), before='\n')
results.close()
def main():
# parse the arguments
arg_parser = ArgumentParser()
spt.register_config_arguments(config, arg_parser)
arg_parser.parse_args(sys.argv[1:])
# print the config
print_with_title('Configurations', pformat(config.to_dict()), after='\n')
# open the result object and prepare for result directories
results = MLResults(config.result_dir)
results.save_config(config) # save experiment settings for review
results.make_dirs('train_summary', exist_ok=True)
# input placeholders
input_x = tf.placeholder(dtype=tf.float32,
shape=(None, config.x_dim),
name='input_x')
input_y = tf.placeholder(dtype=tf.int32, shape=[None], name='input_y')
learning_rate = spt.AnnealingVariable('learning_rate', config.initial_lr,
config.lr_anneal_factor)
# derive the loss, output and accuracy
logits = model(input_x)
cls_loss = tf.losses.sparse_softmax_cross_entropy(input_y, logits)
loss = cls_loss + tf.losses.get_regularization_loss()
y = spt.ops.softmax_classification_output(logits)
acc = spt.ops.classification_accuracy(y, input_y)
# derive the optimizer
optimizer = tf.train.AdamOptimizer(learning_rate)
params = tf.trainable_variables()
grads = optimizer.compute_gradients(loss, var_list=params)
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.apply_gradients(grads)
# prepare for training and testing data
(x_train, y_train), (x_test, y_test) = \
spt.datasets.load_mnist(normalize_x=True)
train_flow = spt.DataFlow.arrays([x_train, y_train],
config.batch_size,
shuffle=True,
skip_incomplete=True)
test_flow = spt.DataFlow.arrays([x_test, y_test], config.test_batch_size)
with spt.utils.create_session().as_default(), \
train_flow.threaded(5) as train_flow:
# train the network
with spt.TrainLoop(params,
max_epoch=config.max_epoch,
max_step=config.max_step,
summary_dir=(results.system_path('train_summary')
if config.write_summary else None),
summary_graph=tf.get_default_graph(),
early_stopping=False) as loop:
trainer = spt.Trainer(loop,
train_op, [input_x, input_y],
train_flow,
metrics={
'loss': loss,
'acc': acc
})
trainer.anneal_after(learning_rate,
epochs=config.lr_anneal_epoch_freq,
steps=config.lr_anneal_step_freq)
evaluator = spt.Evaluator(loop,
metrics={'test_acc': acc},
inputs=[input_x, input_y],
data_flow=test_flow,
time_metric_name='test_time')
evaluator.after_run.add_hook(
lambda: results.update_metrics(evaluator.last_metrics_dict))
trainer.evaluate_after_epochs(evaluator, freq=5)
trainer.log_after_epochs(freq=1)
trainer.run()
# print the final metrics and close the results object
print_with_title('Results', results.format_metrics(), before='\n')
results.close()
def main():
# parse the arguments
arg_parser = ArgumentParser()
spt.register_config_arguments(config, arg_parser)
arg_parser.parse_args(sys.argv[1:])
# print the config
print_with_title('Configurations', pformat(config.to_dict()), after='\n')
# open the result object and prepare for result directories
results = MLResults(config.result_dir)
results.save_config(config) # save experiment settings for review
results.make_dirs('plotting', exist_ok=True)
results.make_dirs('train_summary', exist_ok=True)
# input placeholders
input_x = tf.placeholder(dtype=tf.int32,
shape=(None, config.x_dim),
name='input_x')
learning_rate = spt.AnnealingVariable('learning_rate', config.initial_lr,
config.lr_anneal_factor)
# derive the output for initialization
with tf.name_scope('initialization'):
init_q_net = q_net(input_x, is_initializing=True)
init_chain = init_q_net.chain(p_net,
latent_axis=0,
observed={'x': input_x},
is_initializing=True)
init_lb = tf.reduce_mean(init_chain.vi.lower_bound.elbo())
# derive the loss and lower-bound for training
with tf.name_scope('training'):
train_q_net = q_net(input_x)
train_chain = train_q_net.chain(p_net,
latent_axis=0,
observed={'x': input_x})
vae_loss = tf.reduce_mean(train_chain.vi.training.sgvb())
loss = vae_loss + tf.losses.get_regularization_loss()
# derive the nll and logits output for testing
with tf.name_scope('testing'):
test_q_net = q_net(input_x, n_z=config.test_n_z)
test_chain = test_q_net.chain(p_net,
latent_axis=0,
observed={'x': input_x})
test_nll = -tf.reduce_mean(test_chain.vi.evaluation.is_loglikelihood())
test_lb = tf.reduce_mean(test_chain.vi.lower_bound.elbo())
# derive the optimizer
with tf.name_scope('optimizing'):
optimizer = tf.train.AdamOptimizer(learning_rate)
params = tf.trainable_variables()
grads = optimizer.compute_gradients(loss, var_list=params)
with tf.control_dependencies(tf.get_collection(
tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.apply_gradients(grads)
# derive the plotting function
with tf.name_scope('plotting'):
plot_p_net = p_net(n_z=100)
x_plots = tf.reshape(bernoulli_as_pixel(plot_p_net['x']), (-1, 28, 28))
def plot_samples(loop):
with loop.timeit('plot_time'):
images = session.run(x_plots)
save_images_collection(images=images,
filename='plotting/{}.png'.format(
loop.epoch),
grid_size=(10, 10),
results=results)
# prepare for training and testing data
(x_train, y_train), (x_test, y_test) = spt.datasets.load_mnist()
train_flow = bernoulli_flow(x_train,
config.batch_size,
shuffle=True,
skip_incomplete=True)
test_flow = bernoulli_flow(x_test, config.test_batch_size, sample_now=True)
with spt.utils.create_session().as_default() as session, \
train_flow.threaded(5) as train_flow:
spt.utils.ensure_variables_initialized()
# initialize the network
for [x] in train_flow:
lb_out = session.run(init_lb, feed_dict={input_x: x})
print('Network initialized, first-batch loss is {:.2g}.\n'.format(
lb_out))
break
# train the network
with spt.TrainLoop(params,
var_groups=['q_net', 'p_net'],
max_epoch=config.max_epoch,
max_step=config.max_step,
summary_dir=(results.system_path('train_summary')
if config.write_summary else None),
summary_graph=tf.get_default_graph(),
early_stopping=False) as loop:
trainer = spt.Trainer(loop,
train_op, [input_x],
train_flow,
metrics={'loss': loss})
trainer.anneal_after(learning_rate,
epochs=config.lr_anneal_epoch_freq,
steps=config.lr_anneal_step_freq)
evaluator = spt.Evaluator(loop,
metrics={
'test_nll': test_nll,
'test_lb': test_lb
},
inputs=[input_x],
data_flow=test_flow,
time_metric_name='test_time')
evaluator.after_run.add_hook(
lambda: results.update_metrics(evaluator.last_metrics_dict))
trainer.evaluate_after_epochs(evaluator, freq=10)
trainer.evaluate_after_epochs(functools.partial(
plot_samples, loop),
freq=10)
trainer.log_after_epochs(freq=1)
trainer.run()
# print the final metrics and close the results object
print_with_title('Results', results.format_metrics(), before='\n')
results.close()
def main(trainpath, normalpath, abnormalpath, outputpath):
if config.debug_level == -1:
spt.utils.set_assertion_enabled(False)
elif config.debug_level == 1:
spt.utils.set_check_numerics(True)
#spt.utils.set_assertion_enabled(False)
# print the config
print_with_title('Configurations', config.format_config(), after='\n')
# input and output file
train_file = trainpath
normal_file = normalpath
abnormal_file = abnormalpath
output_file = os.path.join('webankdata',
'{}_{}.csv'.format(config.flow_type or 'vae',
outputpath))
valid_file = os.path.join('webankdata',
'v{}_{}.csv'.format(config.flow_type or 'vae',
outputpath))
# you can change it by yourself
# read data
(x_train, y_train), (x_test, y_test), flows_test = \
get_data_vae(train_file, normal_file, abnormal_file)
config.x_dim = x_train.shape[1]
#config.z_dim = get_z_dim(x_train.shape[1])
all_len = x_train.shape[0]
print('origin data: %s' % all_len)
for i in range(30):
print(list(x_train[i]))
valid_rate = 0.1
x_train, x_valid = train_test_split(x_train, test_size=valid_rate)
# x_valid = x_train
print('%s for validation, %s for training v2' % (x_valid.shape[0], x_train.shape[0]))
print('%s for test' % x_test.shape[0])
print('x_dim: %s z_dim: %s' % (config.x_dim, config.z_dim))
# change it by yourself
# input placeholders
input_x = tf.placeholder(
dtype=tf.float32, shape=(None, config.x_dim), name='input_x')
learning_rate = spt.AnnealingVariable(
'learning_rate', config.initial_lr, config.lr_anneal_factor)
# build the posterior flow
if config.flow_type is None:
posterior_flow = None
elif config.flow_type == 'planar_nf':
posterior_flow = \
spt.layers.planar_normalizing_flows(config.n_planar_nf_layers)
else:
assert(config.flow_type == 'rnvp')
with tf.variable_scope('posterior_flow'):
flows = []
for i in range(config.n_rnvp_layers):
flows.append(spt.layers.ActNorm())
flows.append(spt.layers.CouplingLayer(
tf.make_template(
'coupling',
coupling_layer_shift_and_scale,
create_scope_now_=True
),
scale_type='sigmoid'
))
flows.append(spt.layers.InvertibleDense(strict_invertible=True))
posterior_flow = spt.layers.SequentialFlow(flows=flows)
# derive the initialization op
with tf.name_scope('initialization'), \
arg_scope([spt.layers.act_norm], initializing=True):
init_q_net = q_net(input_x, posterior_flow)
init_chain = init_q_net.chain(
p_net, latent_axis=0, observed={'x': input_x})
init_loss = tf.reduce_mean(init_chain.vi.training.sgvb())
# derive the loss and lower-bound for training
with tf.name_scope('training'):
train_q_net = q_net(input_x, posterior_flow)
train_chain = train_q_net.chain(
p_net, latent_axis=0, observed={'x': input_x})
vae_loss = tf.reduce_mean(train_chain.vi.training.sgvb())
loss = vae_loss + tf.losses.get_regularization_loss()
# derive the nll and logits output for testing
with tf.name_scope('testing'):
test_q_net = q_net(input_x, posterior_flow, n_z=config.test_n_z)
test_chain = test_q_net.chain(
p_net, latent_axis=0, observed={'x': input_x})
test_logp = test_chain.vi.evaluation.is_loglikelihood()
test_nll = -tf.reduce_mean(test_logp)
test_lb = tf.reduce_mean(test_chain.vi.lower_bound.elbo())
# derive the optimizer
with tf.name_scope('optimizing'):
optimizer = tf.train.AdamOptimizer(learning_rate)
params = tf.trainable_variables()
grads = optimizer.compute_gradients(loss, var_list=params)
cliped_grad = []
for grad, var in grads:
if grad is not None and var is not None:
if config.norm_clip is not None:
grad = tf.clip_by_norm(grad, config.norm_clip)
cliped_grad.append((grad, var))
with tf.control_dependencies(
tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.apply_gradients(cliped_grad)
train_flow = spt.DataFlow.arrays([x_train],
config.batch_size,
shuffle=True,
skip_incomplete=True)
valid_flow = spt.DataFlow.arrays([x_valid],
config.test_batch_size)
test_flow = spt.DataFlow.arrays([x_test],
config.test_batch_size)
# model_file
#model_name = ''
model_name = os.path.join(
'webankdata',
'md_{}_{}.model'.format(
config.flow_type or 'vae',
outputpath.split('.')[0]
)
)
with spt.utils.create_session().as_default() as session:
var_dict = spt.utils.get_variables_as_dict()
saver = spt.VariableSaver(var_dict, model_name)
#if os.path.exists(model_name):
if False:
print('%s exists' % model_name)
saver.restore()
else:
print('no model here, and start training')
# initialize the network
spt.utils.ensure_variables_initialized()
for [batch_x] in train_flow:
print('Network initialization loss: {:.6g}'.
format(session.run(init_loss, {input_x: batch_x})))
print('')
break
# train the network
with spt.TrainLoop(params,
var_groups=['p_net', 'q_net', 'posterior_flow'],
max_epoch=config.max_epoch,
max_step=config.max_step,
early_stopping=True,
valid_metric_name='valid_loss',
valid_metric_smaller_is_better=True) as loop:
trainer = spt.Trainer(
loop, train_op, [input_x], train_flow,
metrics={'loss': loss}
)
trainer.anneal_after(
learning_rate,
epochs=config.lr_anneal_epoch_freq,
steps=config.lr_anneal_step_freq
)
evaluator = spt.Evaluator(
loop,
metrics={'valid_loss': test_nll},
inputs=[input_x],
data_flow=valid_flow,
time_metric_name='valid_time'
)
trainer.evaluate_after_epochs(evaluator, freq=10)
trainer.log_after_epochs(freq=1)
trainer.run()
saver.save()
# get the answer
print('start testing')
start = time.time()
test_ans = collect_outputs([test_logp], [input_x], test_flow)[0] \
/ config.x_dim
end = time.time()
print("test time: ", end-start)
pd.DataFrame(
{'id': flows_test, 'label': y_test, 'score': test_ans}) \
.to_csv(output_file, index=False)
valid_ans = collect_outputs([test_logp], [input_x], valid_flow)[0] \
/ config.x_dim
pd.DataFrame({'score': valid_ans}).to_csv(valid_file, index=False)
def main():
# parse the arguments
arg_parser = ArgumentParser()
spt.register_config_arguments(config, arg_parser, title='Model options')
spt.register_config_arguments(spt.settings,
arg_parser,
prefix='tfsnippet',
title='TFSnippet options')
arg_parser.parse_args(sys.argv[1:])
# print the config
print_with_title('Configurations', pformat(config.to_dict()), after='\n')
# open the result object and prepare for result directories
results = MLResults(config.result_dir)
results.save_config(config) # save experiment settings for review
results.make_dirs('plotting', exist_ok=True)
results.make_dirs('train_summary', exist_ok=True)
# input placeholders
input_x = tf.placeholder(dtype=tf.int32,
shape=(None, config.x_dim),
name='input_x')
learning_rate = spt.AnnealingVariable('learning_rate', config.initial_lr,
config.lr_anneal_factor)
# derive the loss and lower-bound for training
with tf.name_scope('training'):
train_q_net = q_net(input_x, n_samples=config.train_n_samples)
train_chain = train_q_net.chain(p_net,
latent_axis=0,
observed={'x': input_x})
if config.vi_algorithm == 'reinforce':
baseline = reinforce_baseline_net(input_x)
vae_loss = tf.reduce_mean(
train_chain.vi.training.reinforce(baseline=baseline))
else:
assert (config.vi_algorithm == 'vimco')
vae_loss = tf.reduce_mean(train_chain.vi.training.vimco())
loss = vae_loss + tf.losses.get_regularization_loss()
# derive the nll and logits output for testing
with tf.name_scope('testing'):
test_q_net = q_net(input_x, n_samples=config.test_n_samples)
test_chain = test_q_net.chain(p_net,
latent_axis=0,
observed={'x': input_x})
test_nll = -tf.reduce_mean(test_chain.vi.evaluation.is_loglikelihood())
# derive the classifier via q(y|x)
q_y_given_x = tf.argmax(test_q_net['y'].distribution.logits,
axis=-1,
name='q_y_given_x')
# derive the optimizer
with tf.name_scope('optimizing'):
optimizer = tf.train.AdamOptimizer(learning_rate)
params = tf.trainable_variables()
grads = optimizer.compute_gradients(loss, var_list=params)
with tf.control_dependencies(tf.get_collection(
tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.apply_gradients(grads)
# derive the plotting function
with tf.name_scope('plotting'):
plot_p_net = p_net(
observed={'y': tf.range(config.n_clusters, dtype=tf.int32)},
n_z=10)
x_plots = tf.reshape(
tf.transpose(bernoulli_as_pixel(plot_p_net['x']), (1, 0, 2)),
(-1, 28, 28))
def plot_samples(loop):
with loop.timeit('plot_time'):
images = session.run(x_plots)
save_images_collection(images=images,
filename='plotting/{}.png'.format(
loop.epoch),
grid_size=(config.n_clusters, 10),
results=results)
# derive the final un-supervised classifier
c_classifier = ClusteringClassifier(config.n_clusters, 10)
def train_classifier(loop):
df = bernoulli_flow(x_train,
config.batch_size,
shuffle=False,
skip_incomplete=False)
with loop.timeit('cls_train_time'):
[c_pred] = spt.evaluation.collect_outputs(
outputs=[q_y_given_x],
inputs=[input_x],
data_flow=df,
)
c_classifier.fit(c_pred, y_train)
print(c_classifier.describe())
def evaluate_classifier(loop):
with loop.timeit('cls_test_time'):
[c_pred] = spt.evaluation.collect_outputs(
outputs=[q_y_given_x],
inputs=[input_x],
data_flow=test_flow,
)
y_pred = c_classifier.predict(c_pred)
cls_metrics = {'test_acc': accuracy_score(y_test, y_pred)}
loop.collect_metrics(cls_metrics)
results.update_metrics(cls_metrics)
# prepare for training and testing data
(x_train, y_train), (x_test, y_test) = \
spt.datasets.load_mnist(x_shape=[784])
train_flow = bernoulli_flow(x_train,
config.batch_size,
shuffle=True,
skip_incomplete=True)
test_flow = bernoulli_flow(x_test, config.test_batch_size, sample_now=True)
with spt.utils.create_session().as_default() as session, \
train_flow.threaded(5) as train_flow:
# train the network
with spt.TrainLoop(
params,
var_groups=['p_net', 'q_net', 'gaussian_mixture_prior'],
max_epoch=config.max_epoch,
max_step=config.max_step,
summary_dir=(results.system_path('train_summary')
if config.write_summary else None),
summary_graph=tf.get_default_graph(),
early_stopping=False) as loop:
trainer = spt.Trainer(loop,
train_op, [input_x],
train_flow,
metrics={'loss': loss},
summaries=tf.summary.merge_all(
spt.GraphKeys.AUTO_HISTOGRAM))
trainer.anneal_after(learning_rate,
epochs=config.lr_anneal_epoch_freq,
steps=config.lr_anneal_step_freq)
evaluator = spt.Evaluator(loop,
metrics={'test_nll': test_nll},
inputs=[input_x],
data_flow=test_flow,
time_metric_name='test_time')
evaluator.events.on(
spt.EventKeys.AFTER_EXECUTION,
lambda e: results.update_metrics(evaluator.last_metrics_dict))
trainer.evaluate_after_epochs(evaluator, freq=10)
trainer.evaluate_after_epochs(functools.partial(
plot_samples, loop),
freq=10)
trainer.evaluate_after_epochs(functools.partial(
train_classifier, loop),
freq=10)
trainer.evaluate_after_epochs(functools.partial(
evaluate_classifier, loop),
freq=10)
trainer.log_after_epochs(freq=1)
trainer.run()
# print the final metrics and close the results object
with codecs.open('cluster_classifier.txt', 'wb', 'utf-8') as f:
f.write(c_classifier.describe())
print_with_title('Results', results.format_metrics(), before='\n')
results.close()
def main():
# parse the arguments
arg_parser = ArgumentParser()
spt.register_config_arguments(config, arg_parser, title='Model options')
spt.register_config_arguments(spt.settings, arg_parser, prefix='tfsnippet',
title='TFSnippet options')
arg_parser.parse_args(sys.argv[1:])
# print the config
print_with_title('Configurations', pformat(config.to_dict()), after='\n')
# open the result object and prepare for result directories
results = MLResults(config.result_dir)
results.save_config(config) # save experiment settings for review
results.make_dirs('plotting/sample', exist_ok=True)
results.make_dirs('plotting/z_plot', exist_ok=True)
results.make_dirs('plotting/train.reconstruct', exist_ok=True)
results.make_dirs('plotting/test.reconstruct', exist_ok=True)
results.make_dirs('train_summary', exist_ok=True)
posterior_flow = spt.layers.planar_normalizing_flows(
config.nf_layers, name='posterior_flow')
# input placeholders
input_x = tf.placeholder(
dtype=tf.float32, shape=(None,) + config.x_shape, name='input_x')
warm = tf.placeholder(
dtype=tf.float32, shape=(), name='warm')
learning_rate = spt.AnnealingVariable(
'learning_rate', config.initial_lr, config.lr_anneal_factor)
beta = tf.Variable(initial_value=0.0, dtype=tf.float32, name='beta', trainable=True)
# derive the loss for initializing
with tf.name_scope('initialization'), \
arg_scope([spt.layers.act_norm], initializing=True), \
spt.utils.scoped_set_config(spt.settings, auto_histogram=False):
init_pn_net = p_net(n_z=config.train_n_pz, beta=beta)
init_q_net = q_net(input_x, posterior_flow, n_z=config.train_n_qz)
init_p_net = p_net(observed={'x': input_x, 'z': init_q_net['z']}, n_z=config.train_n_qz, beta=beta)
init_loss = sum(get_all_loss(init_q_net, init_p_net, init_pn_net))
# derive the loss and lower-bound for training
with tf.name_scope('training'), \
arg_scope([batch_norm], training=True):
train_pn_theta = p_net(n_z=config.train_n_pz, beta=beta)
train_pn_omega = p_omega_net(n_z=config.train_n_pz, beta=beta)
train_log_Z = spt.ops.log_mean_exp(-train_pn_theta['z'].log_prob().energy - train_pn_theta['z'].log_prob())
train_q_net = q_net(input_x, posterior_flow, n_z=config.train_n_qz)
train_p_net = p_net(observed={'x': input_x, 'z': train_q_net['z']},
n_z=config.train_n_qz, beta=beta, log_Z=train_log_Z)
VAE_loss, _, VAE_G_loss, VAE_D_real = get_all_loss(train_q_net, train_p_net, train_pn_theta, warm)
_, D_loss, G_loss, D_real = get_all_loss(train_q_net, train_p_net, train_pn_omega, warm)
VAE_loss += tf.losses.get_regularization_loss()
VAE_G_loss += tf.losses.get_regularization_loss()
D_loss += tf.losses.get_regularization_loss()
G_loss += tf.losses.get_regularization_loss()
# derive the nll and logits output for testing
with tf.name_scope('testing'):
test_q_net = q_net(input_x, posterior_flow, n_z=config.test_n_qz)
# test_pd_net = p_net(n_z=config.test_n_pz // 20, mcmc_iterator=20, beta=beta, log_Z=get_log_Z())
test_pn_net = p_net(n_z=config.test_n_pz, mcmc_iterator=0, beta=beta, log_Z=get_log_Z())
test_chain = test_q_net.chain(p_net, observed={'x': input_x}, n_z=config.test_n_qz, latent_axis=0,
beta=beta, log_Z=get_log_Z())
test_recon = tf.reduce_mean(
test_chain.model['x'].log_prob()
)
test_mse = tf.reduce_sum(
(tf.round(test_chain.model['x'].distribution.mean * 128 + 127.5) - tf.round(
test_chain.model['x'] * 128 + 127.5)) ** 2, axis=[-1, -2, -3]) # (sample_dim, batch_dim, x_sample_dim)
test_mse = tf.reduce_min(test_mse, axis=[0])
test_mse = tf.reduce_mean(tf.reduce_mean(tf.reshape(
test_mse, (-1, config.test_x_samples,)
), axis=-1))
test_nll = -tf.reduce_mean(
spt.ops.log_mean_exp(
tf.reshape(
test_chain.vi.evaluation.is_loglikelihood(), (-1, config.test_x_samples,)
), axis=-1)
) + config.x_shape_multiple * np.log(128.0)
test_lb = tf.reduce_mean(test_chain.vi.lower_bound.elbo())
vi = spt.VariationalInference(
log_joint=test_chain.model['x'].log_prob() + test_chain.model['z'].distribution.log_prob(
test_chain.model['z'], group_ndims=1, y=test_chain.model['x']
).log_energy_prob,
latent_log_probs=[test_q_net['z'].log_prob()],
axis=0
)
adv_test_nll = -tf.reduce_mean(
spt.ops.log_mean_exp(
tf.reshape(
vi.evaluation.is_loglikelihood(), (-1, config.test_x_samples,)
), axis=-1)
) + config.x_shape_multiple * np.log(128.0)
adv_test_lb = tf.reduce_mean(vi.lower_bound.elbo())
real_energy = tf.reduce_mean(D_psi(input_origin_x))
reconstruct_energy = tf.reduce_mean(D_psi(test_chain.model['x'].distribution.mean))
pd_energy = tf.reduce_mean(
D_psi(test_pn_net['x'].distribution.mean) * tf.exp(
test_pn_net['z'].log_prob().log_energy_prob - test_pn_net['z'].log_prob()))
pn_energy = tf.reduce_mean(D_psi(test_pn_net['x'].distribution.mean))
log_Z_compute_op = spt.ops.log_mean_exp(
-test_pn_net['z'].log_prob().energy - test_pn_net['z'].log_prob())
kl_adv_and_gaussian = tf.reduce_mean(
test_pn_net['z'].log_prob() - test_pn_net['z'].log_prob().log_energy_prob
)
xi_node = get_var('p_net/xi')
# derive the optimizer
with tf.name_scope('optimizing'):
VAE_params = tf.trainable_variables('q_net') + tf.trainable_variables('G_theta') + tf.trainable_variables(
'beta') + tf.trainable_variables('posterior_flow') + tf.trainable_variables('p_net/xi')
D_params = tf.trainable_variables('D_psi')
VAE_G_params = tf.trainable_variables('G_theta')
G_params = tf.trainable_variables('G_omega')
print("========VAE_params=========")
print(VAE_params)
print("========D_params=========")
print(D_params)
print("========G_params=========")
print(G_params)
with tf.variable_scope('VAE_optimizer'):
VAE_optimizer = tf.train.AdamOptimizer(learning_rate)
VAE_grads = VAE_optimizer.compute_gradients(VAE_loss, VAE_params)
with tf.variable_scope('VAE_G_optimizer'):
VAE_G_optimizer = tf.train.AdamOptimizer(learning_rate, beta1=0.5, beta2=0.999)
VAE_G_grads = VAE_G_optimizer.compute_gradients(VAE_G_loss, VAE_G_params)
with tf.variable_scope('D_optimizer'):
D_optimizer = tf.train.AdamOptimizer(learning_rate, beta1=0.5, beta2=0.999)
D_grads = D_optimizer.compute_gradients(D_loss, D_params)
with tf.variable_scope('G_optimizer'):
G_optimizer = tf.train.AdamOptimizer(learning_rate, beta1=0.5, beta2=0.999)
G_grads = G_optimizer.compute_gradients(G_loss, G_params)
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
VAE_train_op = VAE_optimizer.apply_gradients(VAE_grads)
VAE_G_train_op = VAE_optimizer.apply_gradients(VAE_G_grads)
G_train_op = G_optimizer.apply_gradients(G_grads)
D_train_op = D_optimizer.apply_gradients(D_grads)
# derive the plotting function
with tf.name_scope('plotting'):
sample_n_z = 100
gan_plots = tf.reshape(p_omega_net(n_z=sample_n_z, beta=beta)['x'].distribution.mean,
(-1,) + config.x_shape)
initial_z = q_net(gan_plots, posterior_flow)['z']
gan_plots = 256.0 * gan_plots / 2 + 127.5
initial_z = tf.expand_dims(initial_z, axis=1)
plot_net = p_net(n_z=sample_n_z, mcmc_iterator=20, beta=beta, initial_z=initial_z)
plot_history_e_z = plot_net['z'].history_e_z
plot_history_z = plot_net['z'].history_z
plot_history_pure_e_z = plot_net['z'].history_pure_e_z
plot_history_ratio = plot_net['z'].history_ratio
x_plots = 256.0 * tf.reshape(
plot_net['x'].distribution.mean, (-1,) + config.x_shape) / 2 + 127.5
reconstruct_q_net = q_net(input_x, posterior_flow)
reconstruct_z = reconstruct_q_net['z']
reconstruct_plots = 256.0 * tf.reshape(
p_net(observed={'z': reconstruct_z}, beta=beta)['x'].distribution.mean,
(-1,) + config.x_shape
) / 2 + 127.5
plot_reconstruct_energy = D_psi(reconstruct_plots)
gan_plots = tf.clip_by_value(gan_plots, 0, 255)
x_plots = tf.clip_by_value(x_plots, 0, 255)
reconstruct_plots = tf.clip_by_value(reconstruct_plots, 0, 255)
def plot_samples(loop):
with loop.timeit('plot_time'):
try:
# plot reconstructs
for [x] in reconstruct_train_flow:
x_samples = uniform_sampler.sample(x)
images = np.zeros((300,) + config.x_shape, dtype=np.uint8)
images[::3, ...] = np.round(256.0 * x / 2 + 127.5)
images[1::3, ...] = np.round(256.0 * x_samples / 2 + 127.5)
images[2::3, ...] = np.round(session.run(
reconstruct_plots, feed_dict={input_x: x}))
batch_reconstruct_z = session.run(reconstruct_z, feed_dict={input_x: x})
# print(np.mean(batch_reconstruct_z ** 2, axis=-1))
save_images_collection(
images=images,
filename='plotting/train.reconstruct/{}.png'.format(loop.epoch),
grid_size=(20, 15),
results=results,
)
break
# plot reconstructs
for [x] in reconstruct_test_flow:
x_samples = uniform_sampler.sample(x)
images = np.zeros((300,) + config.x_shape, dtype=np.uint8)
images[::3, ...] = np.round(256.0 * x / 2 + 127.5)
images[1::3, ...] = np.round(256.0 * x_samples / 2 + 127.5)
images[2::3, ...] = np.round(session.run(
reconstruct_plots, feed_dict={input_x: x}))
save_images_collection(
images=images,
filename='plotting/test.reconstruct/{}.png'.format(loop.epoch),
grid_size=(20, 15),
results=results,
)
break
# plot samples
[images, gan_images, batch_history_e_z, batch_history_z, batch_history_pure_e_z,
batch_history_ratio] = session.run(
[x_plots, gan_plots, plot_history_e_z, plot_history_z, plot_history_pure_e_z, plot_history_ratio])
# print(batch_history_e_z)
# print(np.mean(batch_history_z ** 2, axis=-1))
# print(batch_history_pure_e_z)
# print(batch_history_ratio)
save_images_collection(
images=np.round(gan_images),
filename='plotting/sample/gan-{}.png'.format(loop.epoch),
grid_size=(10, 10),
results=results,
)
save_images_collection(
images=np.round(images),
filename='plotting/sample/{}.png'.format(loop.epoch),
grid_size=(10, 10),
results=results,
)
except Exception as e:
print(e)
# prepare for training and testing data
(_x_train, _y_train), (_x_test, _y_test) = \
spt.datasets.load_cifar10(x_shape=config.x_shape)
# train_flow = bernoulli_flow(
# x_train, config.batch_size, shuffle=True, skip_incomplete=True)
x_train = (_x_train - 127.5) / 256.0 * 2
x_test = (_x_test - 127.5) / 256.0 * 2
uniform_sampler = UniformNoiseSampler(-1.0 / 256.0, 1.0 / 256.0, dtype=np.float)
train_flow = spt.DataFlow.arrays([x_train], config.batch_size, shuffle=True, skip_incomplete=True)
# train_flow = train_flow.map(uniform_sampler)
gan_train_flow = spt.DataFlow.arrays(
[np.concatenate([x_train, x_test], axis=0)], config.batch_size, shuffle=True, skip_incomplete=True)
gan_train_flow = gan_train_flow.map(uniform_sampler)
reconstruct_train_flow = spt.DataFlow.arrays(
[x_train], 100, shuffle=True, skip_incomplete=False)
reconstruct_test_flow = spt.DataFlow.arrays(
[x_test], 100, shuffle=True, skip_incomplete=False)
test_flow = spt.DataFlow.arrays(
[np.repeat(x_test, config.test_x_samples, axis=0)], config.test_batch_size)
test_flow = test_flow.map(uniform_sampler)
with spt.utils.create_session().as_default() as session, \
train_flow.threaded(5) as train_flow:
spt.utils.ensure_variables_initialized()
# initialize the network
for [x] in train_flow:
print('Network initialized, first-batch loss is {:.6g}.\n'.
format(session.run(init_loss, feed_dict={input_x: x})))
break
# if config.z_dim == 512:
# restore_checkpoint = '/mnt/mfs/mlstorage-experiments/cwx17/48/19/6f3b6c3ef49ded8ba2d5/checkpoint/checkpoint/checkpoint.dat-390000'
# elif config.z_dim == 1024:
# restore_checkpoint = '/mnt/mfs/mlstorage-experiments/cwx17/cd/19/6f9d69b5d1931e67e2d5/checkpoint/checkpoint/checkpoint.dat-390000'
# elif config.z_dim == 2048:
# restore_checkpoint = '/mnt/mfs/mlstorage-experiments/cwx17/4d/19/6f9d69b5d19398c8c2d5/checkpoint/checkpoint/checkpoint.dat-390000'
# elif config.z_dim == 3072:
# restore_checkpoint = '/mnt/mfs/mlstorage-experiments/cwx17/5d/19/6f9d69b5d1936fb2d2d5/checkpoint/checkpoint/checkpoint.dat-390000'
# else:
restore_checkpoint = None
# train the network
with spt.TrainLoop(tf.trainable_variables(),
var_groups=['q_net', 'p_net', 'posterior_flow', 'G_theta', 'D_psi'],
max_epoch=config.max_epoch,
max_step=config.max_step,
summary_dir=(results.system_path('train_summary')
if config.write_summary else None),
summary_graph=tf.get_default_graph(),
early_stopping=False,
checkpoint_dir=results.system_path('checkpoint'),
checkpoint_epoch_freq=100,
restore_checkpoint=restore_checkpoint
) as loop:
evaluator = spt.Evaluator(
loop,
metrics={'test_nll': test_nll, 'test_lb': test_lb,
'adv_test_nll': adv_test_nll, 'adv_test_lb': adv_test_lb,
'reconstruct_energy': reconstruct_energy,
'real_energy': real_energy,
'pd_energy': pd_energy, 'pn_energy': pn_energy,
'test_recon': test_recon, 'kl_adv_and_gaussian': kl_adv_and_gaussian, 'test_mse': test_mse},
inputs=[input_x],
data_flow=test_flow,
time_metric_name='test_time'
)
loop.print_training_summary()
spt.utils.ensure_variables_initialized()
epoch_iterator = loop.iter_epochs()
n_critical = config.n_critical
# adversarial training
for epoch in epoch_iterator:
step_iterator = MyIterator(train_flow)
while step_iterator.has_next:
if epoch <= config.warm_up_start:
# discriminator training
for step, [x] in loop.iter_steps(limited(step_iterator, n_critical)):
[_, batch_D_loss, batch_D_real] = session.run(
[D_train_op, D_loss, D_real], feed_dict={
input_x: x,
})
loop.collect_metrics(D_loss=batch_D_loss)
loop.collect_metrics(D_real=batch_D_real)
# generator training x
[_, batch_G_loss] = session.run(
[G_train_op, G_loss], feed_dict={
})
loop.collect_metrics(G_loss=batch_G_loss)
else:
# vae training
for step, [x] in loop.iter_steps(limited(step_iterator, n_critical)):
[_, batch_VAE_loss, beta_value, xi_value, batch_train_recon, batch_train_recon_energy,
batch_VAE_D_real, batch_train_kl, batch_train_grad_penalty] = session.run(
[VAE_train_op, VAE_loss, beta, xi_node, train_recon, train_recon_energy, VAE_D_real,
train_kl, train_grad_penalty],
feed_dict={
input_x: x,
warm: 1.0 # min(1.0, 1.0 * epoch / config.warm_up_epoch)
})
loop.collect_metrics(batch_VAE_loss=batch_VAE_loss)
loop.collect_metrics(xi=xi_value)
loop.collect_metrics(beta=beta_value)
loop.collect_metrics(train_recon=batch_train_recon)
loop.collect_metrics(train_recon_energy=batch_train_recon_energy)
loop.collect_metrics(D_real=batch_VAE_D_real)
loop.collect_metrics(train_kl=batch_train_kl)
loop.collect_metrics(train_grad_penalty=batch_train_grad_penalty)
# loop.print_logs()
# generator training x
# [_, batch_VAE_G_loss] = session.run(
# [VAE_G_train_op, VAE_G_loss], feed_dict={
# })
# loop.collect_metrics(VAE_G_loss=batch_VAE_G_loss)
if epoch in config.lr_anneal_epoch_freq:
learning_rate.anneal()
if epoch == config.warm_up_start:
learning_rate.set(config.initial_lr)
if epoch % config.plot_epoch_freq == 0:
plot_samples(loop)
if epoch % config.test_epoch_freq == 0:
log_Z_list = []
for i in range(config.log_Z_times):
log_Z_list.append(session.run(log_Z_compute_op))
from scipy.misc import logsumexp
log_Z = logsumexp(np.asarray(log_Z_list)) - np.log(config.log_Z_times)
get_log_Z().set(log_Z)
print('log_Z_list:{}'.format(log_Z_list))
print('log_Z:{}'.format(log_Z))
with loop.timeit('eval_time'):
evaluator.run()
dataset_img = _x_train
sample_img = []
for i in range((len(x_train)) // 100 + 1):
sample_img.append(session.run(gan_plots))
sample_img = np.concatenate(sample_img, axis=0).astype('uint8')
sample_img = sample_img[:len(dataset_img)]
sample_img = np.asarray(sample_img)
FID = get_fid(sample_img, dataset_img)
# turn to numpy array
IS_mean, IS_std = get_inception_score(sample_img)
loop.collect_metrics(FID=FID)
loop.collect_metrics(IS=IS_mean)
loop.collect_metrics(lr=learning_rate.get())
loop.print_logs()
# print the final metrics and close the results object
print_with_title('Results', results.format_metrics(), before='\n')
results.close()
def main():
# parse the arguments
arg_parser = ArgumentParser()
spt.register_config_arguments(config, arg_parser, title='Model options')
spt.register_config_arguments(spt.settings,
arg_parser,
prefix='tfsnippet',
title='TFSnippet options')
arg_parser.parse_args(sys.argv[1:])
# print the config
print_with_title('Configurations', pformat(config.to_dict()), after='\n')
# open the result object and prepare for result directories
results = MLResults(config.result_dir)
results.save_config(config) # save experiment settings for review
results.make_dirs('plotting/sample', exist_ok=True)
results.make_dirs('plotting/z_plot', exist_ok=True)
results.make_dirs('plotting/train.reconstruct', exist_ok=True)
results.make_dirs('plotting/test.reconstruct', exist_ok=True)
results.make_dirs('train_summary', exist_ok=True)
posterior_flow = spt.layers.planar_normalizing_flows(config.nf_layers,
name='posterior_flow')
# input placeholders
input_x = tf.placeholder(dtype=tf.float32,
shape=(None, ) + config.x_shape,
name='input_x')
input_origin_x = tf.placeholder(dtype=tf.float32,
shape=(None, ) + config.x_shape,
name='input_origin_x')
learning_rate = spt.AnnealingVariable('learning_rate', config.initial_lr,
config.lr_anneal_factor)
beta = tf.Variable(initial_value=0.0,
dtype=tf.float32,
name='beta',
trainable=True)
# derive the nll and logits output for testing
with tf.name_scope('testing'), \
arg_scope([batch_norm], training=True):
test_q_net = q_net(input_x, posterior_flow, n_z=config.test_n_qz)
# test_pd_net = p_net(n_z=config.test_n_pz // 20, mcmc_iterator=20, beta=beta, log_Z=get_log_Z())
test_pn_net = p_net(n_z=config.test_n_pz,
mcmc_iterator=0,
beta=beta,
log_Z=get_log_Z())
test_p_net = p_net(observed={'z': test_q_net['z']},
n_z=config.test_n_qz,
beta=beta,
log_Z=get_log_Z())
pn_abs = tf.abs(
tf.reduce_mean(D_psi(test_pn_net['x']), axis=0) -
D_psi(test_pn_net['x'].distribution.mean))
print(pn_abs)
pn_abs = tf.reduce_mean(pn_abs)
p_abs = tf.abs(
tf.reduce_mean(D_psi(test_p_net['x']), axis=0) -
D_psi(test_p_net['x'].distribution.mean))
p_abs = tf.reduce_mean(p_abs)
xi_node = get_var('p_net/xi')
# prepare for training and testing data
(_x_train,
_y_train), (_x_test,
_y_test) = spt.datasets.load_cifar10(x_shape=config.x_shape)
x_train = (_x_train - 127.5) / 256.0 * 2
x_test = (_x_test - 127.5) / 256.0 * 2
# uniform_sampler = UniformNoiseSampler(-1.0 / 256.0, 1.0 / 256.0, dtype=np.float)
train_flow = spt.DataFlow.arrays([x_train, x_train],
config.test_batch_size)
random_train_flow = spt.DataFlow.arrays([x_train, x_train],
config.test_batch_size,
shuffle=True)
reconstruct_train_flow = spt.DataFlow.arrays([x_train],
100,
shuffle=True,
skip_incomplete=False)
reconstruct_test_flow = spt.DataFlow.arrays([x_test],
100,
shuffle=True,
skip_incomplete=False)
test_flow = spt.DataFlow.arrays([x_test, x_test], config.test_batch_size)
with spt.utils.create_session().as_default() as session, \
train_flow.threaded(5) as train_flow:
spt.utils.ensure_variables_initialized()
# initialize the network
# for [x, origin_x] in train_flow:
# print('Network initialized, first-batch loss is {:.6g}.\n'.
# format(session.run(init_loss, feed_dict={input_x: x, input_origin_x: origin_x})))
# break
# if config.z_dim == 512:
# restore_checkpoint = '/mnt/mfs/mlstorage-experiments/cwx17/48/19/6f3b6c3ef49ded8ba2d5/checkpoint/checkpoint/checkpoint.dat-390000'
# elif config.z_dim == 1024:
# restore_checkpoint = '/mnt/mfs/mlstorage-experiments/cwx17/cd/19/6f9d69b5d1931e67e2d5/checkpoint/checkpoint/checkpoint.dat-390000'
# elif config.z_dim == 2048:
# restore_checkpoint = '/mnt/mfs/mlstorage-experiments/cwx17/4d/19/6f9d69b5d19398c8c2d5/checkpoint/checkpoint/checkpoint.dat-390000'
# elif config.z_dim == 3072:
# restore_checkpoint = '/mnt/mfs/mlstorage-experiments/cwx17/5d/19/6f9d69b5d1936fb2d2d5/checkpoint/checkpoint/checkpoint.dat-390000'
# else:
restore_checkpoint = '/mnt/mfs/mlstorage-experiments/cwx17/3d/0c/d445f4f80a9fee59aed5/checkpoint/checkpoint/checkpoint.dat-312000'
# train the network
with spt.TrainLoop(tf.trainable_variables(),
var_groups=[
'q_net', 'p_net', 'posterior_flow', 'G_theta',
'D_psi', 'G_omega', 'D_kappa'
],
max_epoch=config.max_epoch + 10,
max_step=config.max_step,
summary_dir=(results.system_path('train_summary')
if config.write_summary else None),
summary_graph=tf.get_default_graph(),
early_stopping=False,
checkpoint_dir=results.system_path('checkpoint'),
checkpoint_epoch_freq=100,
restore_checkpoint=restore_checkpoint) as loop:
loop.print_training_summary()
spt.utils.ensure_variables_initialized()
epoch_iterator = loop.iter_epochs()
evaluator = spt.Evaluator(loop,
metrics={
'pn_abs': pn_abs,
'p_abs': p_abs
},
inputs=[input_x, input_origin_x],
data_flow=train_flow,
time_metric_name='test_time')
# adversarial training
for epoch in epoch_iterator:
evaluator.run()
loop.collect_metrics(lr=learning_rate.get())
loop.print_logs()
# print the final metrics and close the results object
print_with_title('Results', results.format_metrics(), before='\n')
results.close()
def main():
# parse the arguments
arg_parser = ArgumentParser()
spt.register_config_arguments(config, arg_parser, title='Model options')
spt.register_config_arguments(spt.settings,
arg_parser,
prefix='tfsnippet',
title='TFSnippet options')
arg_parser.parse_args(sys.argv[1:])
# print the config
print_with_title('Configurations', pformat(config.to_dict()), after='\n')
# open the result object and prepare for result directories
results = MLResults(config.result_dir)
results.save_config(config) # save experiment settings for review
results.make_dirs('train_summary', exist_ok=True)
# input placeholders
input_x = tf.placeholder(dtype=tf.float32,
shape=(None, config.x_dim),
name='input_x')
input_y = tf.placeholder(dtype=tf.int32, shape=[None], name='input_y')
is_training = tf.placeholder(dtype=tf.bool, shape=(), name='is_training')
learning_rate = spt.AnnealingVariable('learning_rate', config.initial_lr,
config.lr_anneal_factor)
multi_gpu = MultiGPU()
# build the model
grads = []
losses = []
y_list = []
acc_list = []
batch_size = spt.utils.get_batch_size(input_x)
params = None
optimizer = tf.train.AdamOptimizer(learning_rate)
for dev, pre_build, [dev_input_x, dev_input_y
] in multi_gpu.data_parallel(batch_size,
[input_x, input_y]):
with tf.device(dev), multi_gpu.maybe_name_scope(dev):
if pre_build:
_ = model(dev_input_x, is_training, channels_last=True)
else:
# derive the loss, output and accuracy
dev_logits = model(dev_input_x,
is_training=is_training,
channels_last=multi_gpu.channels_last(dev))
dev_cls_loss = tf.losses.sparse_softmax_cross_entropy(
dev_input_y, dev_logits)
dev_loss = dev_cls_loss + tf.losses.get_regularization_loss()
dev_y = spt.ops.softmax_classification_output(dev_logits)
dev_acc = spt.ops.classification_accuracy(dev_y, dev_input_y)
losses.append(dev_loss)
y_list.append(dev_y)
acc_list.append(dev_acc)
# derive the optimizer
params = tf.trainable_variables()
grads.append(
optimizer.compute_gradients(dev_loss, var_list=params))
# merge multi-gpu outputs and operations
[loss, acc] = multi_gpu.average([losses, acc_list], batch_size)
[y] = multi_gpu.concat([y_list])
train_op = multi_gpu.apply_grads(grads=multi_gpu.average_grads(grads),
optimizer=optimizer,
control_inputs=tf.get_collection(
tf.GraphKeys.UPDATE_OPS))
# prepare for training and testing data
(x_train, y_train), (x_test, y_test) = \
spt.datasets.load_mnist(x_shape=[784], normalize_x=True)
train_flow = spt.DataFlow.arrays([x_train, y_train],
config.batch_size,
shuffle=True,
skip_incomplete=True)
test_flow = spt.DataFlow.arrays([x_test, y_test], config.test_batch_size)
with spt.utils.create_session().as_default(), \
train_flow.threaded(5) as train_flow:
# train the network
with spt.TrainLoop(params,
max_epoch=config.max_epoch,
max_step=config.max_step,
summary_dir=(results.system_path('train_summary')
if config.write_summary else None),
summary_graph=tf.get_default_graph(),
early_stopping=False) as loop:
trainer = spt.Trainer(loop,
train_op, [input_x, input_y],
train_flow,
feed_dict={is_training: True},
metrics={
'loss': loss,
'acc': acc
},
summaries=tf.summary.merge_all(
spt.GraphKeys.AUTO_HISTOGRAM))
trainer.anneal_after(learning_rate,
epochs=config.lr_anneal_epoch_freq,
steps=config.lr_anneal_step_freq)
evaluator = spt.Evaluator(loop,
metrics={'test_acc': acc},
inputs=[input_x, input_y],
data_flow=test_flow,
feed_dict={is_training: False},
time_metric_name='test_time')
evaluator.events.on(
spt.EventKeys.AFTER_EXECUTION,
lambda e: results.update_metrics(evaluator.last_metrics_dict))
trainer.evaluate_after_epochs(evaluator, freq=5)
trainer.log_after_epochs(freq=1)
trainer.run()
# print the final metrics and close the results object
print_with_title('Results', results.format_metrics(), before='\n')
results.close()