def main(experiment_name, list_experiments=False, gpu_device='/gpu:0'):
"""Create a tensorflow worker to run experiments in your DB."""
if list_experiments:
exps = db.list_experiments()
print '_' * 30
print 'Initialized experiments:'
print '_' * 30
for l in exps:
print l.values()[0]
print '_' * 30
print 'You can add to the DB with: '\
'python prepare_experiments.py --experiment=%s' % \
exps[0].values()[0]
return
if experiment_name is None:
print 'No experiment specified. Pulling one out of the DB.'
experiment_name = db.get_experiment_name()
# Prepare to run the model
config = Config()
condition_label = '%s_%s' % (experiment_name, py_utils.get_dt_stamp())
experiment_label = '%s' % (experiment_name)
log = logger.get(os.path.join(config.log_dir, condition_label))
experiment_dict = experiments.experiments()[experiment_name]()
config = add_to_config(d=experiment_dict, config=config) # Globals
config, exp_params = process_DB_exps(
experiment_name=experiment_name, log=log,
config=config) # Update config w/ DB params
dataset_module = py_utils.import_module(model_dir=config.dataset_info,
dataset=config.dataset)
dataset_module = dataset_module.data_processing() # hardcoded class name
train_data, train_means = get_data_pointers(
dataset=config.dataset,
base_dir=config.tf_records,
cv=dataset_module.folds.keys()[1], # TODO: SEARCH FOR INDEX.
log=log)
val_data, val_means = get_data_pointers(dataset=config.dataset,
base_dir=config.tf_records,
cv=dataset_module.folds.keys()[0],
log=log)
# Initialize output folders
dir_list = {
'checkpoints':
os.path.join(config.checkpoints, condition_label),
'summaries':
os.path.join(config.summaries, condition_label),
'condition_evaluations':
os.path.join(config.condition_evaluations, condition_label),
'experiment_evaluations':
os.path.join( # DEPRECIATED
config.experiment_evaluations, experiment_label),
'visualization':
os.path.join(config.visualizations, condition_label),
'weights':
os.path.join(config.condition_evaluations, condition_label, 'weights')
}
[py_utils.make_dir(v) for v in dir_list.values()]
# Prepare data loaders on the cpu
config.data_augmentations = py_utils.flatten_list(
config.data_augmentations, log)
with tf.device('/cpu:0'):
train_images, train_labels = data_loader.inputs(
dataset=train_data,
batch_size=config.batch_size,
model_input_image_size=dataset_module.model_input_image_size,
tf_dict=dataset_module.tf_dict,
data_augmentations=config.data_augmentations,
num_epochs=config.epochs,
tf_reader_settings=dataset_module.tf_reader,
shuffle=config.shuffle)
val_images, val_labels = data_loader.inputs(
dataset=val_data,
batch_size=config.batch_size,
model_input_image_size=dataset_module.model_input_image_size,
tf_dict=dataset_module.tf_dict,
data_augmentations=config.data_augmentations,
num_epochs=config.epochs,
tf_reader_settings=dataset_module.tf_reader,
shuffle=config.shuffle)
log.info('Created tfrecord dataloader tensors.')
# Load model specification
struct_name = config.model_struct.split(os.path.sep)[-1]
try:
model_dict = py_utils.import_module(
dataset=struct_name,
model_dir=os.path.join('models', 'structs',
experiment_name).replace(os.path.sep, '.'))
except IOError:
print 'Could not find the model structure: %s' % experiment_name
# Inject model_dict with hyperparameters if requested
model_dict.layer_structure = hp_opt_utils.inject_model_with_hps(
layer_structure=model_dict.layer_structure, exp_params=exp_params)
# Prepare model on GPU
with tf.device(gpu_device):
with tf.variable_scope('cnn') as scope:
# Training model
if len(dataset_module.output_size) > 1:
log.warning('Found > 1 dimension for your output size.'
'Converting to a scalar.')
dataset_module.output_size = np.prod(
dataset_module.output_size)
if hasattr(model_dict, 'output_structure'):
# Use specified output layer
output_structure = model_dict.output_structure
else:
output_structure = None
model = model_utils.model_class(
mean=train_means,
training=True,
output_size=dataset_module.output_size)
train_scores, model_summary = model.build(
data=train_images,
layer_structure=model_dict.layer_structure,
output_structure=output_structure,
log=log,
tower_name='cnn')
log.info('Built training model.')
log.debug(json.dumps(model_summary, indent=4), verbose=0)
print_model_architecture(model_summary)
# Prepare the loss function
train_loss, _ = loss_utils.loss_interpreter(
logits=train_scores,
labels=train_labels,
loss_type=config.loss_function,
dataset_module=dataset_module)
# Add weight decay if requested
if len(model.regularizations) > 0:
train_loss = loss_utils.wd_loss(
regularizations=model.regularizations,
loss=train_loss,
wd_penalty=config.regularization_strength)
train_op = loss_utils.optimizer_interpreter(
loss=train_loss,
lr=config.lr,
optimizer=config.optimizer,
constraints=config.optimizer_constraints,
model=model)
log.info('Built training loss function.')
train_accuracy = eval_metrics.metric_interpreter(
metric=dataset_module.score_metric,
pred=train_scores,
labels=train_labels) # training accuracy
if int(train_images.get_shape()[-1]) <= 3:
tf.summary.image('train images', train_images)
tf.summary.scalar('training loss', train_loss)
tf.summary.scalar('training accuracy', train_accuracy)
log.info('Added training summaries.')
# Validation model
scope.reuse_variables()
val_model = model_utils.model_class(
mean=val_means,
training=True,
output_size=dataset_module.output_size)
val_scores, _ = val_model.build( # Ignore summary
data=val_images,
layer_structure=model_dict.layer_structure,
output_structure=output_structure,
log=log,
tower_name='cnn')
log.info('Built validation model.')
val_loss, _ = loss_utils.loss_interpreter(
logits=val_scores,
labels=val_labels,
loss_type=config.loss_function,
dataset_module=dataset_module)
val_accuracy = eval_metrics.metric_interpreter(
metric=dataset_module.score_metric,
pred=val_scores,
labels=val_labels) # training accuracy
if int(train_images.get_shape()[-1]) <= 3:
tf.summary.image('val images', val_images)
tf.summary.scalar('validation loss', val_loss)
tf.summary.scalar('validation accuracy', val_accuracy)
log.info('Added validation summaries.')
# Set up summaries and saver
saver = tf.train.Saver(tf.global_variables())
summary_op = tf.summary.merge_all()
# Initialize the graph
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# Need to initialize both of these if supplying num_epochs to inputs
sess.run(
tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()))
summary_writer = tf.summary.FileWriter(dir_list['summaries'], sess.graph)
# Set up exemplar threading
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# Create dictionaries of important training and validation information
train_dict = {
'train_loss': train_loss,
'train_accuracy': train_accuracy,
'train_images': train_images,
'train_labels': train_labels,
'train_op': train_op,
'train_scores': train_scores
}
val_dict = {
'val_loss': val_loss,
'val_accuracy': val_accuracy,
'val_images': val_images,
'val_labels': val_labels,
'val_scores': val_scores,
}
# Start training loop
np.save(
os.path.join(dir_list['condition_evaluations'],
'training_config_file'), config)
log.info('Starting training')
output_dict = training.training_loop(
config=config,
db=db,
coord=coord,
sess=sess,
summary_op=summary_op,
summary_writer=summary_writer,
saver=saver,
threads=threads,
summary_dir=dir_list['summaries'],
checkpoint_dir=dir_list['checkpoints'],
weight_dir=dir_list['weights'],
train_dict=train_dict,
val_dict=val_dict,
train_model=model,
val_model=val_model,
exp_params=exp_params)
log.info('Finished training.')
model_name = config.model_struct.replace('/', '_')
py_utils.save_npys(data=output_dict,
model_name=model_name,
output_string=dir_list['experiment_evaluations'])
def main(
experiment_name,
list_experiments=False,
load_and_evaluate_ckpt=None,
placeholder_data=None,
grad_images=False,
gpu_device='/gpu:0'):
"""Create a tensorflow worker to run experiments in your DB."""
if list_experiments:
exps = db.list_experiments()
print '_' * 30
print 'Initialized experiments:'
print '_' * 30
for l in exps:
print l.values()[0]
print '_' * 30
if len(exps) == 0:
print 'No experiments found.'
else:
print 'You can add to the DB with: '\
'python prepare_experiments.py --experiment=%s' % \
exps[0].values()[0]
return
if experiment_name is None:
print 'No experiment specified. Pulling one out of the DB.'
experiment_name = db.get_experiment_name()
# Prepare to run the model
config = Config()
condition_label = '%s_%s' % (experiment_name, py_utils.get_dt_stamp())
experiment_label = '%s' % (experiment_name)
log = logger.get(os.path.join(config.log_dir, condition_label))
assert experiment_name is not None, 'Empty experiment name.'
experiment_dict = experiments.experiments()[experiment_name]()
config = add_to_config(d=experiment_dict, config=config) # Globals
config.load_and_evaluate_ckpt = load_and_evaluate_ckpt
if load_and_evaluate_ckpt is not None:
# Remove the train operation and add a ckpt pointer
from ops import evaluation
config, exp_params = process_DB_exps(
experiment_name=experiment_name,
log=log,
config=config) # Update config w/ DB params
dataset_module = py_utils.import_module(
model_dir=config.dataset_info,
dataset=config.dataset)
dataset_module = dataset_module.data_processing() # hardcoded class name
train_key = [k for k in dataset_module.folds.keys() if 'train' in k]
if not len(train_key):
train_key = 'train'
else:
train_key = train_key[0]
train_data, train_means_image, train_means_label = get_data_pointers(
dataset=config.dataset,
base_dir=config.tf_records,
cv=train_key,
log=log)
val_key = [k for k in dataset_module.folds.keys() if 'val' in k]
if not len(val_key):
val_key = 'train'
else:
val_key = val_key[0]
val_data, val_means_image, val_means_label = get_data_pointers(
dataset=config.dataset,
base_dir=config.tf_records,
cv=val_key,
log=log)
# Initialize output folders
dir_list = {
'checkpoints': os.path.join(
config.checkpoints, condition_label),
'summaries': os.path.join(
config.summaries, condition_label),
'condition_evaluations': os.path.join(
config.condition_evaluations, condition_label),
'experiment_evaluations': os.path.join( # DEPRECIATED
config.experiment_evaluations, experiment_label),
'visualization': os.path.join(
config.visualizations, condition_label),
'weights': os.path.join(
config.condition_evaluations, condition_label, 'weights')
}
[py_utils.make_dir(v) for v in dir_list.values()]
# Prepare data loaders on the cpu
if all(isinstance(i, list) for i in config.data_augmentations):
if config.data_augmentations:
config.data_augmentations = py_utils.flatten_list(
config.data_augmentations,
log)
if load_and_evaluate_ckpt is not None:
config.epochs = 1
config.train_shuffle = False
config.val_shuffle = False
with tf.device('/cpu:0'):
if placeholder_data:
placeholder_shape = placeholder_data['train_image_shape']
placeholder_dtype = placeholder_data['train_image_dtype']
original_train_images = tf.placeholder(
dtype=placeholder_dtype,
shape=placeholder_shape,
name='train_images')
placeholder_shape = placeholder_data['train_label_shape']
placeholder_dtype = placeholder_data['train_label_dtype']
original_train_labels = tf.placeholder(
dtype=placeholder_dtype,
shape=placeholder_shape,
name='train_labels')
placeholder_shape = placeholder_data['val_image_shape']
placeholder_dtype = placeholder_data['val_image_dtype']
original_val_images = tf.placeholder(
dtype=placeholder_dtype,
shape=placeholder_shape,
name='val_images')
placeholder_shape = placeholder_data['val_label_shape']
placeholder_dtype = placeholder_data['val_label_dtype']
original_val_labels = tf.placeholder(
dtype=placeholder_dtype,
shape=placeholder_shape,
name='val_labels')
# Apply augmentations
(
train_images,
train_labels
) = data_loader.placeholder_image_augmentations(
images=original_train_images,
model_input_image_size=dataset_module.model_input_image_size,
labels=original_train_labels,
data_augmentations=config.data_augmentations,
batch_size=config.batch_size)
(
val_images,
val_labels
) = data_loader.placeholder_image_augmentations(
images=original_val_images,
model_input_image_size=dataset_module.model_input_image_size,
labels=original_val_labels,
data_augmentations=config.data_augmentations,
batch_size=config.batch_size)
# Store in the placeholder dict
placeholder_data['train_images'] = original_train_images
placeholder_data['train_labels'] = original_train_labels
placeholder_data['val_images'] = original_val_images
placeholder_data['val_labels'] = original_val_labels
else:
train_images, train_labels = data_loader.inputs(
dataset=train_data,
batch_size=config.batch_size,
model_input_image_size=dataset_module.model_input_image_size,
tf_dict=dataset_module.tf_dict,
data_augmentations=config.data_augmentations,
num_epochs=config.epochs,
tf_reader_settings=dataset_module.tf_reader,
shuffle=config.shuffle_train,
resize_output=config.resize_output)
if hasattr(config, 'val_augmentations'):
val_augmentations = config.val_augmentations
else:
val_augmentations = config.data_augmentations
val_images, val_labels = data_loader.inputs(
dataset=val_data,
batch_size=config.batch_size,
model_input_image_size=dataset_module.model_input_image_size,
tf_dict=dataset_module.tf_dict,
data_augmentations=val_augmentations,
num_epochs=config.epochs,
tf_reader_settings=dataset_module.tf_reader,
shuffle=config.shuffle_val,
resize_output=config.resize_output)
log.info('Created tfrecord dataloader tensors.')
# Load model specification
struct_name = config.model_struct.split(os.path.sep)[-1]
try:
model_dict = py_utils.import_module(
dataset=struct_name,
model_dir=os.path.join(
'models',
'structs',
experiment_name).replace(os.path.sep, '.')
)
except IOError:
print 'Could not find the model structure: %s in folder %s' % (
struct_name,
experiment_name)
# Inject model_dict with hyperparameters if requested
model_dict.layer_structure = hp_opt_utils.inject_model_with_hps(
layer_structure=model_dict.layer_structure,
exp_params=exp_params)
# Prepare variables for the models
if len(dataset_module.output_size) == 2:
log.warning(
'Found > 1 dimension for your output size.'
'Converting to a scalar.')
dataset_module.output_size = np.prod(
dataset_module.output_size)
if hasattr(model_dict, 'output_structure'):
# Use specified output layer
output_structure = model_dict.output_structure
else:
output_structure = None
# Correct number of output neurons if needed
if config.dataloader_override and\
'weights' in output_structure[-1].keys():
output_neurons = output_structure[-1]['weights'][0]
size_check = output_neurons != dataset_module.output_size
fc_check = output_structure[-1]['layers'][0] == 'fc'
if size_check and fc_check:
output_structure[-1]['weights'][0] = dataset_module.output_size
log.warning('Adjusted output neurons from %s to %s.' % (
output_neurons,
dataset_module.output_size))
# Prepare model on GPU
if not hasattr(dataset_module, 'input_normalization'):
dataset_module.input_normalization = None
with tf.device(gpu_device):
with tf.variable_scope('cnn') as scope:
# Training model
model = model_utils.model_class(
mean=train_means_image,
training=True,
output_size=dataset_module.output_size,
input_normalization=dataset_module.input_normalization)
train_scores, model_summary, _ = model.build(
data=train_images,
layer_structure=model_dict.layer_structure,
output_structure=output_structure,
log=log,
tower_name='cnn')
if grad_images:
oh_dims = int(train_scores.get_shape()[-1])
target_scores = tf.one_hot(train_labels, oh_dims) * train_scores
train_gradients = tf.gradients(target_scores, train_images)[0]
log.info('Built training model.')
log.debug(
json.dumps(model_summary, indent=4),
verbose=0)
print_model_architecture(model_summary)
# Normalize labels on GPU if needed
if 'normalize_labels' in exp_params.keys():
if exp_params['normalize_labels'] == 'zscore':
train_labels -= train_means_label['mean']
train_labels /= train_means_label['std']
val_labels -= train_means_label['mean']
val_labels /= train_means_label['std']
log.info('Z-scoring labels.')
elif exp_params['normalize_labels'] == 'mean':
train_labels -= train_means_label['mean']
val_labels -= val_means_label['mean']
log.info('Mean-centering labels.')
# Check the shapes of labels and scores
if not isinstance(train_scores, list):
if len(
train_scores.get_shape()) != len(
train_labels.get_shape()):
train_shape = train_scores.get_shape().as_list()
label_shape = train_labels.get_shape().as_list()
val_shape = val_scores.get_shape().as_list()
val_label_shape = val_labels.get_shape().as_list()
if len(
train_shape) == 2 and len(
label_shape) == 1 and train_shape[-1] == 1:
train_labels = tf.expand_dims(train_labels, axis=-1)
val_labels = tf.expand_dims(val_labels, axis=-1)
elif len(
train_shape) == 2 and len(
label_shape) == 1 and train_shape[-1] == 1:
train_scores = tf.expand_dims(train_scores, axis=-1)
val_scores = tf.expand_dims(val_scores, axis=-1)
# Prepare the loss function
train_loss, _ = loss_utils.loss_interpreter(
logits=train_scores, # TODO
labels=train_labels,
loss_type=config.loss_function,
weights=config.loss_weights,
dataset_module=dataset_module)
# Add loss tensorboard tracking
if isinstance(train_loss, list):
for lidx, tl in enumerate(train_loss):
tf.summary.scalar('training_loss_%s' % lidx, tl)
train_loss = tf.add_n(train_loss)
else:
tf.summary.scalar('training_loss', train_loss)
# Add weight decay if requested
if len(model.regularizations) > 0:
train_loss = loss_utils.wd_loss(
regularizations=model.regularizations,
loss=train_loss,
wd_penalty=config.regularization_strength)
assert config.lr is not None, 'No learning rate.' # TODO: Make a QC function
if config.lr > 1:
old_lr = config.lr
config.lr = loss_utils.create_lr_schedule(
train_batch=config.batch_size,
num_training=config.lr)
config.optimizer = 'momentum'
log.info('Forcing momentum classifier.')
else:
old_lr = None
train_op = loss_utils.optimizer_interpreter(
loss=train_loss,
lr=config.lr,
optimizer=config.optimizer,
constraints=config.optimizer_constraints,
model=model)
log.info('Built training loss function.')
# Add a score for the training set
train_accuracy = eval_metrics.metric_interpreter(
metric=dataset_module.score_metric, # TODO: Attach to exp cnfg
pred=train_scores, # TODO
labels=train_labels)
# Add aux scores if requested
train_aux = {}
if hasattr(dataset_module, 'aux_scores'):
for m in dataset_module.aux_scores:
train_aux[m] = eval_metrics.metric_interpreter(
metric=m,
pred=train_scores,
labels=train_labels) # [0] # TODO: Fix for multiloss
# Prepare remaining tensorboard summaries
if config.tensorboard_images:
if len(train_images.get_shape()) == 4:
tf_fun.image_summaries(train_images, tag='Training images')
if (np.asarray(
train_labels.get_shape().as_list()) > 1).sum() > 2:
tf_fun.image_summaries(
train_labels,
tag='Training_targets')
tf_fun.image_summaries(
train_scores,
tag='Training_predictions')
if isinstance(train_accuracy, list):
for tidx, ta in enumerate(train_accuracy):
tf.summary.scalar('training_accuracy_%s' % tidx, ta)
else:
tf.summary.scalar('training_accuracy', train_accuracy)
if config.pr_curve:
if isinstance(train_scores, list):
for pidx, train_score in enumerate(train_scores):
train_label = train_labels[:, pidx]
pr_summary.op(
tag='training_pr_%s' % pidx,
predictions=tf.cast(
tf.argmax(
train_score,
axis=-1),
tf.float32),
labels=tf.cast(train_label, tf.bool),
display_name='training_precision_recall_%s' % pidx)
else:
pr_summary.op(
tag='training_pr',
predictions=tf.cast(
tf.argmax(
train_scores,
axis=-1),
tf.float32),
labels=tf.cast(train_labels, tf.bool),
display_name='training_precision_recall')
log.info('Added training summaries.')
with tf.variable_scope('cnn', tf.AUTO_REUSE) as scope:
# Validation model
scope.reuse_variables()
val_model = model_utils.model_class(
mean=train_means_image, # Normalize with train data
training=False,
output_size=dataset_module.output_size,
input_normalization=dataset_module.input_normalization)
val_scores, _, _ = val_model.build( # Ignore summary
data=val_images,
layer_structure=model_dict.layer_structure,
output_structure=output_structure,
log=log,
tower_name='cnn')
if grad_images:
oh_dims = int(val_scores.get_shape()[-1])
target_scores = tf.one_hot(val_labels, oh_dims) * val_scores
val_gradients = tf.gradients(target_scores, val_images)[0]
log.info('Built validation model.')
# Check the shapes of labels and scores
val_loss, _ = loss_utils.loss_interpreter(
logits=val_scores,
labels=val_labels,
loss_type=config.loss_function,
weights=config.loss_weights,
dataset_module=dataset_module)
# Add loss tensorboard tracking
if isinstance(val_loss, list):
for lidx, tl in enumerate(val_loss):
tf.summary.scalar('validation_loss_%s' % lidx, tl)
val_loss = tf.add_n(val_loss)
else:
tf.summary.scalar('validation_loss', val_loss)
# Add a score for the validation set
val_accuracy = eval_metrics.metric_interpreter(
metric=dataset_module.score_metric, # TODO
pred=val_scores,
labels=val_labels)
# Add aux scores if requested
val_aux = {}
if hasattr(dataset_module, 'aux_scores'):
for m in dataset_module.aux_scores:
val_aux[m] = eval_metrics.metric_interpreter(
metric=m,
pred=val_scores,
labels=val_labels) # [0] # TODO: Fix for multiloss
# Prepare tensorboard summaries
if config.tensorboard_images:
if len(val_images.get_shape()) == 4:
tf_fun.image_summaries(
val_images,
tag='Validation')
if (np.asarray(
val_labels.get_shape().as_list()) > 1).sum() > 2:
tf_fun.image_summaries(
val_labels,
tag='Validation_targets')
tf_fun.image_summaries(
val_scores,
tag='Validation_predictions')
if isinstance(val_accuracy, list):
for vidx, va in enumerate(val_accuracy):
tf.summary.scalar('validation_accuracy_%s' % vidx, va)
else:
tf.summary.scalar('validation_accuracy', val_accuracy)
if config.pr_curve:
if isinstance(val_scores, list):
for pidx, val_score in enumerate(val_scores):
val_label = val_labels[:, pidx]
pr_summary.op(
tag='validation_pr_%s' % pidx,
predictions=tf.cast(
tf.argmax(
val_score,
axis=-1),
tf.float32),
labels=tf.cast(val_label, tf.bool),
display_name='validation_precision_recall_%s' %
pidx)
else:
pr_summary.op(
tag='validation_pr',
predictions=tf.cast(
tf.argmax(
val_scores,
axis=-1),
tf.float32),
labels=tf.cast(val_labels, tf.bool),
display_name='validation_precision_recall')
log.info('Added validation summaries.')
# Set up summaries and saver
if not hasattr(config, 'max_to_keep'):
config.max_to_keep = None
saver = tf.train.Saver(
var_list=tf.global_variables(),
max_to_keep=config.max_to_keep)
summary_op = tf.summary.merge_all()
# Initialize the graph
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# Need to initialize both of these if supplying num_epochs to inputs
sess.run(
tf.group(
tf.global_variables_initializer(),
tf.local_variables_initializer())
)
summary_writer = tf.summary.FileWriter(dir_list['summaries'], sess.graph)
# Set up exemplar threading
if placeholder_data:
coord, threads = None, None
else:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
# Create dictionaries of important training and validation information
train_dict = {
'train_loss': train_loss,
'train_images': train_images,
'train_labels': train_labels,
'train_op': train_op,
'train_scores': train_scores
}
val_dict = {
'val_loss': val_loss,
'val_images': val_images,
'val_labels': val_labels,
'val_scores': val_scores,
}
if grad_images:
train_dict['train_gradients'] = train_gradients
val_dict['val_gradients'] = val_gradients
if isinstance(train_accuracy, list):
for tidx, (ta, va) in enumerate(zip(train_accuracy, val_accuracy)):
train_dict['train_accuracy_%s' % tidx] = ta
val_dict['val_accuracy_%s' % tidx] = va
else:
train_dict['train_accuracy_0'] = train_accuracy
val_dict['val_accuracy_0'] = val_accuracy
if load_and_evaluate_ckpt is not None:
# Remove the train operation and add a ckpt pointer
del train_dict['train_op']
if hasattr(dataset_module, 'aux_score'):
# Attach auxillary scores to tensor dicts
for m in dataset_module.aux_scores:
train_dict['train_aux_%s' % m] = train_aux[m]
val_dict['val_aux_%s' % m] = val_aux[m]
# Start training loop
if old_lr is not None:
config.lr = old_lr
np.save(
os.path.join(
dir_list['condition_evaluations'], 'training_config_file'),
config)
log.info('Starting training')
if load_and_evaluate_ckpt is not None:
return evaluation.evaluation_loop(
config=config,
db=db,
coord=coord,
sess=sess,
summary_op=summary_op,
summary_writer=summary_writer,
saver=saver,
threads=threads,
summary_dir=dir_list['summaries'],
checkpoint_dir=dir_list['checkpoints'],
weight_dir=dir_list['weights'],
train_dict=train_dict,
val_dict=val_dict,
train_model=model,
val_model=val_model,
exp_params=exp_params,
placeholder_data=placeholder_data)
else:
output_dict = training.training_loop(
config=config,
db=db,
coord=coord,
sess=sess,
summary_op=summary_op,
summary_writer=summary_writer,
saver=saver,
threads=threads,
summary_dir=dir_list['summaries'],
checkpoint_dir=dir_list['checkpoints'],
weight_dir=dir_list['weights'],
train_dict=train_dict,
val_dict=val_dict,
train_model=model,
val_model=val_model,
exp_params=exp_params)
log.info('Finished training.')
model_name = config.model_struct.replace('/', '_')
if output_dict is not None:
py_utils.save_npys(
data=output_dict,
model_name=model_name,
output_string=dir_list['experiment_evaluations'])
def build_model(exp_params,
config,
log,
dt_string,
gpu_device,
cpu_device,
use_db=True,
add_config=None,
placeholders=False,
checkpoint=None,
test=False,
map_out=None,
num_batches=None,
tensorboard_images=False):
"""Standard model building routines."""
config = py_utils.add_to_config(d=exp_params, config=config)
if not hasattr(config, 'force_path'):
config.force_path = False
exp_label = '%s_%s_%s' % (exp_params['model'], exp_params['experiment'],
py_utils.get_dt_stamp())
directories = py_utils.prepare_directories(config, exp_label)
dataset_module = py_utils.import_module(pre_path=config.dataset_classes,
module=config.train_dataset)
train_dataset_module = dataset_module.data_processing()
if not config.force_path:
(train_data, _, _) = py_utils.get_data_pointers(
dataset=train_dataset_module.output_name,
base_dir=config.tf_records,
local_dir=config.local_tf_records,
cv='train')
else:
train_data = train_dataset_module.train_path
dataset_module = py_utils.import_module(pre_path=config.dataset_classes,
module=config.val_dataset)
val_dataset_module = dataset_module.data_processing()
if not config.force_path:
val_data, _, _ = py_utils.get_data_pointers(
dataset=val_dataset_module.output_name,
base_dir=config.tf_records,
local_dir=config.local_tf_records,
cv='val')
else:
val_data = train_dataset_module.val_path
# val_means_image, val_means_label = None, None
# Create data tensors
if hasattr(train_dataset_module, 'aux_loss'):
train_aux_loss = train_dataset_module.aux_loss
else:
train_aux_loss = None
with tf.device(cpu_device):
if placeholders and not test:
# Train with placeholders
(pl_train_images, pl_train_labels, pl_val_images, pl_val_labels,
train_images, train_labels, val_images,
val_labels) = get_placeholders(train_dataset=train_dataset_module,
val_dataset=val_dataset_module,
config=config)
train_module_data = train_dataset_module.get_data()
val_module_data = val_dataset_module.get_data()
placeholders = {
'train': {
'images': train_module_data[0]['train'],
'labels': train_module_data[1]['train']
},
'val': {
'images': val_module_data[0]['val'],
'labels': val_module_data[1]['val']
},
}
train_aux, val_aux = None, None
elif placeholders and test:
test_dataset_module = train_dataset_module
# Test with placeholders
(pl_test_images, pl_test_labels, test_images,
test_labels) = get_placeholders_test(
test_dataset=test_dataset_module, config=config)
test_module_data = test_dataset_module.get_data()
placeholders = {
'test': {
'images': test_module_data[0]['test'],
'labels': test_module_data[1]['test']
},
}
train_aux, val_aux = None, None
else:
train_images, train_labels, train_aux = data_loader.inputs(
dataset=train_data,
batch_size=config.train_batch_size,
model_input_image_size=train_dataset_module.
model_input_image_size,
tf_dict=train_dataset_module.tf_dict,
data_augmentations=config.train_augmentations,
num_epochs=config.epochs,
aux=train_aux_loss,
tf_reader_settings=train_dataset_module.tf_reader,
shuffle=config.shuffle_train)
if hasattr(val_dataset_module, 'val_model_input_image_size'):
val_dataset_module.model_input_image_size = val_dataset_module.val_model_input_image_size
val_images, val_labels, val_aux = data_loader.inputs(
dataset=val_data,
batch_size=config.val_batch_size,
model_input_image_size=val_dataset_module.
model_input_image_size,
tf_dict=val_dataset_module.tf_dict,
data_augmentations=config.val_augmentations,
num_epochs=None,
tf_reader_settings=val_dataset_module.tf_reader,
shuffle=config.shuffle_val)
# Build training and val models
model_spec = py_utils.import_module(module=config.model,
pre_path=config.model_classes)
if hasattr(train_dataset_module, 'force_output_size'):
train_dataset_module.output_size = train_dataset_module.force_output_size
if hasattr(val_dataset_module, 'force_output_size'):
val_dataset_module.output_size = val_dataset_module.force_output_size
if hasattr(config, 'loss_function'):
train_loss_function = config.loss_function
val_loss_function = config.loss_function
else:
train_loss_function = config.train_loss_function
val_loss_function = config.val_loss_function
# Route test vs train/val
h_check = [
x for x in tf.trainable_variables()
if 'homunculus' in x.name or 'humonculus' in x.name
]
if not hasattr(config, 'default_restore'):
config.default_restore = False
if test:
assert len(gpu_device) == 1, 'Testing only works with 1 gpu.'
gpu_device = gpu_device[0]
with tf.device(gpu_device):
if not placeholders:
test_images = val_images
test_labels = val_labels
test_dataset_module = val_dataset_module
test_logits, test_vars = model_spec.build_model(
data_tensor=test_images,
reuse=None,
training=False,
output_shape=test_dataset_module.output_size)
if test_logits.dtype is not tf.float32:
test_logits = tf.cast(test_logits, tf.float32)
# Derive loss
if not hasattr(config, 'test_loss_function'):
test_loss_function = val_loss_function
else:
test_loss_function = config.test_loss_function
test_loss = losses.derive_loss(labels=test_labels,
logits=test_logits,
loss_type=test_loss_function)
# Derive score
test_score = losses.derive_score(labels=test_labels,
logits=test_logits,
loss_type=test_loss_function,
score_type=config.score_function)
# Initialize model
(sess, saver, summary_op, summary_writer, coord, threads,
restore_saver) = initialize_tf(config=config,
placeholders=placeholders,
ckpt=checkpoint,
default_restore=config.default_restore,
directories=directories)
if placeholders:
proc_images = test_images
proc_labels = test_labels
test_images = pl_test_images
test_labels = pl_test_labels
_, H, W, _ = test_vars['model_output_y'].shape
jacobian = tf.gradients(test_logits, test_vars['model_output_x'])[
0] # g.batch_jacobian(test_vars['model_output_x'], test_images)
test_dict = {
'test_loss': test_loss,
'test_score': test_score,
'test_images': test_images,
'test_labels': test_labels,
'test_logits': test_logits,
'test_jacobian': jacobian
}
if placeholders:
test_dict['test_proc_images'] = proc_images
test_dict['test_proc_labels'] = proc_labels
if len(h_check):
test_dict['homunculus'] = h_check[0]
if isinstance(test_vars, dict):
for k, v in test_vars.iteritems():
test_dict[k] = v
else:
test_dict['activity'] = test_vars
else:
train_losses, val_losses, tower_grads, norm_updates = [], [], [], []
train_scores, val_scores = [], []
train_image_list, train_label_list = [], []
val_image_list, val_label_list = [], []
train_reuse = None
if not hasattr(config, 'lr_schedule'):
config.lr_schedule = None
if hasattr(config, 'loss_function'):
train_loss_function = config.loss_function
val_loss_function = config.loss_function
else:
train_loss_function = config.train_loss_function
val_loss_function = config.val_loss_function
# Prepare loop
if not placeholders:
train_batch_queue = tf_fun.get_batch_queues(images=train_images,
labels=train_labels,
gpu_device=gpu_device)
val_batch_queue = tf_fun.get_batch_queues(images=val_images,
labels=val_labels,
gpu_device=gpu_device)
config.lr = optimizers.get_lr_schedule(lr=config.lr,
lr_schedule=config.lr_schedule)
opt = optimizers.get_optimizers(optimizer=config.optimizer,
lr=config.lr,
dtype=train_images.dtype)
with tf.device(cpu_device):
global_step = tf.train.get_or_create_global_step()
for i, gpu in enumerate(gpu_device):
# rs = tf.AUTO_REUSE if i > 0 else None
with tf.device(gpu):
with tf.name_scope('tower_%d' % i) as scope:
# Prepare tower data
if placeholders:
# Multi-gpu: will have to split
# train_images per gpu by hand
train_image_batch = train_images
val_image_batch = val_images
train_label_batch = train_labels
val_label_batch = val_labels
else:
(train_image_batch,
train_label_batch) = train_batch_queue.dequeue()
(val_image_batch,
val_label_batch) = val_batch_queue.dequeue()
train_image_list += [train_image_batch]
train_label_list += [train_label_batch]
val_image_list += [val_image_batch]
val_label_list += [val_label_batch]
# Build models
train_logits, train_vars = model_spec.build_model(
data_tensor=train_image_batch,
reuse=train_reuse,
training=True,
output_shape=train_dataset_module.output_size)
num_training_vars = len(tf.trainable_variables())
val_logits, val_vars = model_spec.build_model(
data_tensor=val_image_batch,
reuse=True,
training=False,
output_shape=val_dataset_module.output_size)
num_validation_vars = len(tf.trainable_variables())
assert num_training_vars == num_validation_vars, \
'Found a different # of train and val variables.'
train_reuse = True
# Derive losses
if train_logits.dtype is not tf.float32:
train_logits = tf.cast(train_logits, tf.float32)
if val_logits.dtype is not tf.float32:
val_logits = tf.cast(val_logits, tf.float32)
train_loss = losses.derive_loss(
labels=train_label_batch,
logits=train_logits,
images=train_image_batch,
loss_type=train_loss_function)
val_loss = losses.derive_loss(
labels=val_label_batch,
logits=val_logits,
images=val_image_batch,
loss_type=val_loss_function)
# Derive score
train_score = losses.derive_score(
labels=train_labels,
logits=train_logits,
loss_type=train_loss_function,
score_type=config.score_function)
val_score = losses.derive_score(
labels=val_labels,
logits=val_logits,
loss_type=val_loss_function,
score_type=config.score_function)
# Add aux losses if requested
if hasattr(model_spec, 'weight_decay'):
wd = (model_spec.weight_decay() * tf.add_n([
tf.nn.l2_loss(v)
for v in tf.trainable_variables()
if 'batch_normalization' not in v.name
and 'horizontal' not in v.name
and 'mu' not in v.name and 'beta' not in v.name
and 'intercept' not in v.name
]))
tf.summary.scalar('weight_decay', wd)
train_loss += wd
if hasattr(model_spec, 'bsds_weight_decay'):
wd = (model_spec.bsds_weight_decay()['l2'] *
tf.add_n([
tf.nn.l2_loss(v)
for v in tf.trainable_variables()
if 'horizontal' not in v.name
and 'norm' not in v.name
]))
tf.summary.scalar('weight_decay_readout', wd)
train_loss += wd
wd = (model_spec.bsds_weight_decay()['l1'] *
tf.add_n([
tf.reduce_sum(tf.abs(v))
for v in tf.trainable_variables()
if 'horizontal' in v.name
]))
tf.summary.scalar('weight_decay_horizontal', wd)
train_loss += wd
if hasattr(model_spec, 'orthogonal'):
weights = [
v for v in tf.trainable_variables()
if 'horizontal' in v.name
]
assert len(weights) is not None, \
'No horizontal weights for laplace.'
wd = model_spec.orthogonal() * tf.add_n(
[tf_fun.orthogonal(w) for w in weights])
tf.summary.scalar('weight_decay', wd)
train_loss += wd
if hasattr(model_spec, 'laplace'):
weights = [
v for v in tf.trainable_variables()
if 'horizontal' in v.name
]
assert len(weights) is not None, \
'No horizontal weights for laplace.'
wd = model_spec.laplace() * tf.add_n(
[tf_fun.laplace(w) for w in weights])
tf.summary.scalar('weight_decay', wd)
train_loss += wd
# Derive auxilary losses
if hasattr(config, 'aux_loss'):
aux_loss_type, scale = config.aux_loss.items()[0]
for k, v in train_vars.iteritems():
# if k in train_dataset_module.aux_loss.keys():
# (
# aux_loss_type,
# scale
# ) = train_dataset_module.aux_loss[k]
train_loss += (losses.derive_loss(
labels=train_labels,
logits=v,
loss_type=aux_loss_type) * scale)
# Gather everything
train_losses += [train_loss]
val_losses += [val_loss]
train_scores += [train_score]
val_scores += [val_score]
# Compute and store gradients
with tf.variable_scope(tf.get_variable_scope(),
reuse=tf.AUTO_REUSE):
grads = opt.compute_gradients(train_loss)
optimizers.check_grads(grads)
tower_grads += [grads]
# Gather normalization variables
norm_updates += [
tf.get_collection(tf.GraphKeys.UPDATE_OPS,
scope=scope)
]
# Recompute and optimize gradients
grads = optimizers.average_gradients(tower_grads)
if hasattr(config, 'clip_gradients') and config.clip_gradients:
grads = optimizers.apply_grad_clip(grads, config.clip_gradients)
op_vars = []
if hasattr(config, 'exclusion_lr') and hasattr(config,
'exclusion_scope'):
grads_0 = [
x for x in grads if config.exclusion_scope not in x[1].name
]
grads_1 = [x for x in grads if config.exclusion_scope in x[1].name]
op_vars_0 = optimizers.apply_gradients(opt=opt,
grads=grads_0,
global_step=global_step)
opt_1 = optimizers.get_optimizers(optimizer=config.optimizer,
lr=config.exclusion_lr,
dtype=train_images.dtype)
op_vars_1 = optimizers.apply_gradients(opt=opt_1,
grads=grads_1,
global_step=global_step)
op_vars += [op_vars_0]
op_vars += [op_vars_1]
else:
op_vars += [
optimizers.apply_gradients(opt=opt,
grads=grads,
global_step=global_step)
]
if not hasattr(config, 'variable_moving_average'):
config.variable_moving_average = False
if config.variable_moving_average:
variable_averages = tf.train.ExponentialMovingAverage(
config.variable_moving_average, global_step)
op_vars += [variable_averages.apply(tf.trainable_variables())]
if len(norm_updates):
op_vars += [tf.group(*norm_updates)]
train_op = tf.group(*op_vars)
# Summarize losses and scores
train_loss = tf.reduce_mean(train_losses)
val_loss = tf.reduce_mean(val_losses)
train_score = tf.reduce_mean(train_scores)
val_score = tf.reduce_mean(val_scores)
if len(train_image_list) > 1:
train_image_list = tf.stack(train_image_list, axis=0)
train_label_list = tf.stack(train_label_list, axis=0)
else:
train_image_list = train_image_list[0]
train_label_list = train_label_list[0]
if len(val_image_list) > 1:
val_image_list = tf.stack(val_image_list, axis=0)
val_label_list = tf.stack(val_label_list, axis=0)
else:
val_image_list = val_image_list[0]
val_label_list = val_label_list[0]
tf.summary.scalar('train_loss', train_loss)
tf.summary.scalar('val_loss', val_loss)
if tensorboard_images:
tf.summary.image('train_images', train_images)
tf.summary.image('val_images', val_images)
# Initialize model
(sess, saver, summary_op, summary_writer, coord, threads,
restore_saver) = initialize_tf(config=config,
placeholders=placeholders,
ckpt=checkpoint,
default_restore=config.default_restore,
directories=directories)
# Create dictionaries of important training and validation information
if placeholders:
proc_train_images = train_images
proc_train_labels = train_labels
proc_val_images = val_images
proc_val_labels = val_labels
train_images = pl_train_images
train_labels = pl_train_labels
val_images = pl_val_images
val_labels = pl_val_labels
train_dict = {
'train_loss': train_loss,
'train_score': train_score,
'train_images': train_image_list,
'train_labels': train_label_list,
'train_logits': train_logits,
'train_op': train_op
}
if placeholders:
train_dict['proc_train_images'] = proc_train_images
train_dict['proc_train_labels'] = proc_train_labels
if train_aux is not None:
train_dict['train_aux'] = train_aux
if tf.contrib.framework.is_tensor(config.lr):
train_dict['lr'] = config.lr
else:
train_dict['lr'] = tf.constant(config.lr)
if isinstance(train_vars, dict):
for k, v in train_vars.iteritems():
train_dict[k] = v
else:
train_dict['activity'] = train_vars
if hasattr(config, 'save_gradients') and config.save_gradients:
grad = tf.gradients(train_logits, train_images)[0]
if grad is not None:
train_dict['gradients'] = grad
else:
log.warning('Could not calculate val gradients.')
val_dict = {
'val_loss': val_loss,
'val_score': val_score,
'val_images': val_image_list,
'val_logits': val_logits,
'val_labels': val_label_list,
}
if placeholders:
val_dict['proc_val_images'] = proc_val_images
val_dict['proc_val_labels'] = proc_val_labels
if val_aux is not None:
val_dict['aux'] = val_aux
if isinstance(val_vars, dict):
for k, v in val_vars.iteritems():
val_dict[k] = v
else:
val_dict['activity'] = val_vars
if hasattr(config, 'save_gradients') and config.save_gradients:
grad = tf.gradients(val_logits, val_images)[0]
if grad is not None:
val_dict['gradients'] = grad
else:
log.warning('Could not calculate val gradients.')
if len(h_check):
val_dict['homunculus'] = h_check[0]
# Add optional info to the config
if add_config is not None:
extra_list = add_config.split(',')
for eidx, extra in enumerate(extra_list):
setattr(config, 'extra_%s' % eidx, extra)
# Count parameters
num_params = tf_fun.count_parameters(var_list=tf.trainable_variables())
print 'Model has approximately %s trainable params.' % num_params
if test:
return training.test_loop(log=log,
config=config,
sess=sess,
summary_op=summary_op,
summary_writer=summary_writer,
saver=saver,
restore_saver=restore_saver,
directories=directories,
test_dict=test_dict,
exp_label=exp_label,
num_params=num_params,
checkpoint=checkpoint,
num_batches=num_batches,
save_weights=config.save_weights,
save_checkpoints=config.save_checkpoints,
save_activities=config.save_activities,
save_gradients=config.save_gradients,
map_out=map_out,
placeholders=placeholders)
else:
# Start training loop
training.training_loop(log=log,
config=config,
coord=coord,
sess=sess,
summary_op=summary_op,
summary_writer=summary_writer,
saver=saver,
restore_saver=restore_saver,
threads=threads,
directories=directories,
train_dict=train_dict,
val_dict=val_dict,
exp_label=exp_label,
num_params=num_params,
checkpoint=checkpoint,
use_db=use_db,
save_weights=config.save_weights,
save_checkpoints=config.save_checkpoints,
save_activities=config.save_activities,
save_gradients=config.save_gradients,
placeholders=placeholders)
def model_builder(
params,
config,
model_spec,
gpu_device,
cpu_device,
placeholders=False,
tensorboard_images=False):
"""Standard model building routines."""
config = py_utils.add_to_config(
d=params,
config=config)
exp_label = '%s_%s' % (params['exp_name'], py_utils.get_dt_stamp())
directories = py_utils.prepare_directories(config, exp_label)
dataset_module = py_utils.import_module(
model_dir=config.dataset_info,
dataset=config.dataset)
dataset_module = dataset_module.data_processing() # hardcoded class name
train_key = [k for k in dataset_module.folds.keys() if 'train' in k]
if not len(train_key):
train_key = 'train'
else:
train_key = train_key[0]
(
train_data,
train_means_image,
train_means_label) = py_utils.get_data_pointers(
dataset=config.dataset,
base_dir=config.tf_records,
cv=train_key)
val_key = [k for k in dataset_module.folds.keys() if 'val' in k]
if not len(val_key):
val_key = 'train'
else:
val_key = val_key[0]
if hasattr(config, 'val_dataset'):
val_dataset = config.val_dataset
else:
val_dataset = config.dataset
val_data, val_means_image, val_means_label = py_utils.get_data_pointers(
dataset=val_dataset,
base_dir=config.tf_records,
cv=val_key)
# Create data tensors
with tf.device(cpu_device):
if placeholders:
(
train_images,
train_labels,
val_images,
val_labels) = get_placeholders(dataset_module, config)
placeholders = dataset_module.get_data()
else:
train_images, train_labels = data_loader.inputs(
dataset=train_data,
batch_size=config.batch_size,
model_input_image_size=dataset_module.model_input_image_size,
tf_dict=dataset_module.tf_dict,
data_augmentations=config.data_augmentations,
num_epochs=config.epochs,
tf_reader_settings=dataset_module.tf_reader,
shuffle=config.shuffle_train)
val_images, val_labels = data_loader.inputs(
dataset=val_data,
batch_size=config.batch_size,
model_input_image_size=dataset_module.model_input_image_size,
tf_dict=dataset_module.tf_dict,
data_augmentations=config.val_augmentations,
num_epochs=config.epochs,
tf_reader_settings=dataset_module.tf_reader,
shuffle=config.shuffle_val)
# Build training and val models
with tf.device(gpu_device):
train_logits, train_hgru_act = model_spec(
data_tensor=train_images,
reuse=None,
training=True)
val_logits, val_hgru_act = model_spec(
data_tensor=val_images,
reuse=tf.AUTO_REUSE,
training=False)
# Derive loss
loss_type = None
if hasattr(config, 'loss_type'):
loss_type = config.loss_type
train_loss = losses.derive_loss(
labels=train_labels,
logits=train_logits,
loss_type=loss_type)
val_loss = losses.derive_loss(
labels=val_labels,
logits=val_logits,
loss_type=loss_type)
if hasattr(config, 'metric_type'):
metric_type = config.metric_type
else:
metric_type = 'accuracy'
if metric_type == 'pearson':
train_accuracy = metrics.pearson_score(
labels=train_labels,
pred=train_logits,
REDUCTION=tf.reduce_mean)
val_accuracy = metrics.pearson_score(
labels=val_labels,
pred=val_logits,
REDUCTION=tf.reduce_mean)
else:
train_accuracy = metrics.class_accuracy(
labels=train_labels,
logits=train_logits)
val_accuracy = metrics.class_accuracy(
labels=val_labels,
logits=val_logits)
tf.summary.scalar('train_accuracy', train_accuracy)
tf.summary.scalar('val_accuracy', val_accuracy)
if tensorboard_images:
tf.summary.image('train_images', train_images)
tf.summary.image('val_images', val_images)
# Build optimizer
train_op = optimizers.get_optimizer(
train_loss,
config['lr'],
config['optimizer'])
# Initialize tf variables
saver = tf.train.Saver(
var_list=tf.global_variables())
summary_op = tf.summary.merge_all()
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
sess.run(
tf.group(
tf.global_variables_initializer(),
tf.local_variables_initializer()))
summary_writer = tf.summary.FileWriter(
directories['summaries'],
sess.graph)
if not placeholders:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
else:
coord, threads = None, None
# Create dictionaries of important training and validation information
train_dict = {
'train_loss': train_loss,
'train_images': train_images,
'train_labels': train_labels,
'train_op': train_op,
'train_accuracy': train_accuracy
}
if isinstance(train_hgru_act, dict):
for k, v in train_hgru_act.iteritems():
train_dict[k] = v
else:
train_dict['activity'] = train_hgru_act
val_dict = {
'val_loss': val_loss,
'val_images': val_images,
'val_labels': val_labels,
'val_accuracy': val_accuracy,
}
if isinstance(val_hgru_act, dict):
for k, v in val_hgru_act.iteritems():
val_dict[k] = v
else:
val_dict['activity'] = val_hgru_act
# Count parameters
num_params = np.sum(
[np.prod(x.get_shape().as_list()) for x in tf.trainable_variables()])
print 'Model has approximately %s trainable params.' % num_params
# Create datastructure for saving data
ds = data_structure.data(
batch_size=config.batch_size,
validation_iters=config.validation_iters,
num_validation_evals=config.num_validation_evals,
shuffle_val=config.shuffle_val,
lr=config.lr,
loss_function=config.loss_function,
optimizer=config.optimizer,
model_name=config.model_name,
dataset=config.dataset,
num_params=num_params,
output_directory=config.results)
# Start training loop
training.training_loop(
config=config,
coord=coord,
sess=sess,
summary_op=summary_op,
summary_writer=summary_writer,
saver=saver,
threads=threads,
directories=directories,
train_dict=train_dict,
val_dict=val_dict,
exp_label=exp_label,
data_structure=ds,
placeholders=placeholders)
def extract_vgg_features(cm_type='contextual_vector_vd',
layer_name='pool3',
output_type='sparse_pool',
project_name=None,
model_type='vgg16',
timesteps=5,
dtype=tf.float32):
"""Main extraction and training script."""
assert project_name is not None, 'Need a project name.'
# 1. Get file paths and load config
config = Config()
config.cm_type = cm_type
project_path = config.projects[project_name]
# 2. Assert the model is there and load neural data.
print 'Loading preprocessed data...'
data = np.load(os.path.join(project_path, '%s.npz' % project_name))
neural_data = data['data_matrix']
if config.round_neural_data:
neural_data = np.round(neural_data)
# TODO: across_session_data_matrix is subtracted version
images = data['all_images'].astype(np.float32)
# Remove zeroed columns from neural data
channel_check = np.abs(neural_data).sum(0) > 0
neural_data = neural_data[:, channel_check]
# TODO: create AUX dict with each channel's X/Y
output_aux = {'loss': config.loss_type} # None
rfs = rf_sizes.get_eRFs(model_type)[layer_name]
# 3. Create a output directory if necessary and save a timestamped numpy.
model_description = '%s_%s_%s_%s_%s_%s' % (
cm_type, layer_name, output_type, project_name, model_type, timesteps)
dt_stamp = '%s_%s' % (model_description, str(datetime.now()).replace(
' ', '_').replace(':', '_').replace('-', '_'))
project_dir = os.path.join(config.results, project_name)
out_dir = os.path.join(project_dir, dt_stamp)
checkpoint_dir = os.path.join(out_dir, 'checkpoints')
dirs = [config.results, config.summaries, out_dir]
[py_utils.make_dir(x) for x in dirs]
print '-' * 60
print('Training model:' + out_dir)
print '-' * 60
# 4. Prepare data on CPU
neural_shape = list(neural_data.shape)
num_neurons = neural_shape[-1]
with tf.device('/cpu:0'):
train_images = tf.placeholder(dtype=dtype,
name='train_images',
shape=[config.train_batch_size] +
config.img_shape)
train_neural = tf.placeholder(dtype=dtype,
name='train_neural',
shape=[config.train_batch_size] +
[num_neurons])
val_images = tf.placeholder(dtype=dtype,
name='val_images',
shape=[config.val_batch_size] +
config.img_shape)
val_neural = tf.placeholder(dtype=dtype,
name='val_neural',
shape=[config.val_batch_size] +
[num_neurons])
# 5. Prepare model on GPU
with tf.device('/gpu:0'):
with tf.variable_scope('cnn') as scope:
vgg = vgg16.Vgg16(vgg16_npy_path=config.vgg16_weight_path)
train_mode = tf.get_variable(name='training', initializer=False)
vgg.build(
train_images,
output_shape=1000, # hardcode
train_mode=train_mode,
final_layer=layer_name)
# Select a layer
activities = vgg[layer_name]
# Feature reduce with a 1x1 conv
if config.reduce_features is not None:
vgg, activities, reduce_weights = ff.pool_ff_interpreter(
self=vgg,
it_neuron_op='1x1conv',
act=activities,
it_name='feature_reduce',
out_channels=config.reduce_features,
aux=None)
else:
reduce_weights = None
# Add con-model if requested
if cm_type is not None and cm_type != 'none':
norms = normalizations.normalizations()
activities, cm_weights, _ = norms[cm_type](
x=activities,
r_in=rfs['r_in'],
j_in=rfs['j_in'],
timesteps=timesteps,
lesions=config.lesions,
train=True)
else:
cm_weights = None
# Create output layer for N-recording channels
activities = tf.nn.dropout(activities, 0.5)
vgg, output_activities, output_weights = ff.pool_ff_interpreter(
self=vgg,
it_neuron_op=output_type,
act=activities,
it_name='output',
out_channels=num_neurons,
aux=output_aux)
# Prepare the loss function
loss, _ = loss_utils.loss_interpreter(logits=output_activities,
labels=train_neural,
loss_type=config.loss_type)
# Add contextual model WD
if config.reduce_features is not None and reduce_weights is not None:
loss += loss_utils.add_wd(weights=reduce_weights,
wd_dict=config.wd_types)
# Add contextual model WD
if config.cm_wd_types is not None and cm_weights is not None:
loss += loss_utils.add_wd(weights=cm_weights,
wd_dict=config.cm_wd_types)
# Add WD to output layer
if config.wd_types is not None:
loss += loss_utils.add_wd(weights=output_weights,
wd_dict=config.wd_types)
# Finetune the learning rates
train_op = loss_utils.optimizer_interpreter(
loss=loss, lr=config.lr, optimizer=config.optimizer)
# Calculate metrics
train_accuracy = eval_metrics.metric_interpreter(
metric=config.metric,
pred=output_activities,
labels=train_neural)
# Add summaries for debugging
tf.summary.image('train images', train_images)
tf.summary.image('validation images', val_images)
tf.summary.scalar("loss", loss)
tf.summary.scalar("training accuracy", train_accuracy)
# Setup validation op
scope.reuse_variables()
# Validation graph is the same as training except no batchnorm
val_vgg = vgg16.Vgg16(vgg16_npy_path=config.vgg16_weight_path)
val_vgg.build(val_images,
output_shape=1000,
final_layer=layer_name)
# Select a layer
val_activities = val_vgg[layer_name]
# Add feature reduction if requested
if config.reduce_features is not None:
val_vgg, val_activities, _ = ff.pool_ff_interpreter(
self=val_vgg,
it_neuron_op=config.reduce_type,
act=val_activities,
it_name='feature_reduce',
out_channels=config.reduce_features,
aux=None)
else:
reduce_weights = None
# Add con-model if requested
if cm_type is not None and cm_type != 'none':
val_activities, _, _ = norms[cm_type](x=val_activities,
r_in=rfs['r_in'],
j_in=rfs['j_in'],
timesteps=timesteps,
lesions=config.lesions,
train=False)
# Create output layer for N-recording channels
val_vgg, val_output_activities, _ = ff.pool_ff_interpreter(
self=val_vgg,
it_neuron_op=output_type,
act=val_activities,
it_name='output',
out_channels=num_neurons,
aux=output_aux)
# Prepare the loss function
val_loss, _ = loss_utils.loss_interpreter(
logits=val_output_activities,
labels=val_neural,
loss_type=config.loss_type,
max_spikes=config.max_spikes)
# Calculate metrics
val_accuracy = eval_metrics.metric_interpreter(
metric=config.metric,
pred=val_output_activities,
labels=val_neural)
tf.summary.scalar('validation loss', val_loss)
tf.summary.scalar('validation accuracy', val_accuracy)
# Set up summaries and saver
saver = tf.train.Saver(tf.global_variables())
summary_op = tf.summary.merge_all()
# Initialize the graph
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# Need to initialize both of these if supplying num_epochs to inputs
sess.run(tf.global_variables_initializer())
summary_dir = os.path.join(config.summaries, dt_stamp)
summary_writer = tf.summary.FileWriter(summary_dir, sess.graph)
# Start training loop
train_vars = {
'images': train_images,
'neural_data': train_neural,
'loss': loss,
'score': train_accuracy,
'train_op': train_op
}
if cm_weights is not None:
for k, v in cm_weights.iteritems():
train_vars[k] = v
val_vars = {
'images': val_images,
'neural_data': val_neural,
'loss': val_loss,
'score': val_accuracy,
}
extra_params = {
'cm_type': cm_type,
'layer_name': layer_name,
'output_type': output_type,
'project_name': project_name,
'model_type': model_type,
'lesions': config.lesions,
'timesteps': timesteps
}
np.savez(os.path.join(out_dir, 'training_config_file'),
config=config,
extra_params=extra_params)
train_cv_out, val_cv_out, weights = training.training_loop(
config=config,
neural_data=neural_data,
images=images,
target_size=config.img_shape[:2],
sess=sess,
train_vars=train_vars,
val_vars=val_vars,
summary_op=summary_op,
summary_writer=summary_writer,
checkpoint_dir=checkpoint_dir,
summary_dir=summary_dir,
saver=saver)
np.savez(os.path.join(out_dir, 'data'),
config=config,
extra_params=extra_params,
train_cv_out=train_cv_out,
val_cv_out=val_cv_out,
weight=weights)