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,
config_file=None,
ckpt_file=None,
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))
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
config, exp_params = process_DB_exps(
experiment_name=experiment_name, log=log,
config=config) # Update config w/ DB params
config = np.load(config_file).item()
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_image, train_means_label = 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_image, val_means_label = 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
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)
config.epochs = 1
config.shuffle = False
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_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=['resize_and_crop'],
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 model on GPU
with tf.device(gpu_device):
with tf.variable_scope('cnn') as scope:
# Normalize labels 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']
log.info('Z-scoring labels.')
elif exp_params['normalize_labels'] == 'mean':
train_labels -= train_means_label['mean']
log.info('Mean-centering labels.')
# Training model
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
model = model_utils.model_class(
mean=train_means_image,
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')
eval_graph = tf.Graph()
with eval_graph.as_default():
with eval_graph.gradient_override_map({'selu': 'GradLRP'}):
train_grad_images = tf.gradients(
train_scores[0] * tf.cast(train_labels, tf.float32),
train_images)[0]
log.info('Built training model.')
log.debug(json.dumps(model_summary, indent=4), verbose=0)
print_model_architecture(model_summary)
# 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()
if len(train_shape) == 2 and len(
label_shape) == 1 and train_shape[-1] == 1:
train_labels = tf.expand_dims(train_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)
# 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)
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 len(train_images.get_shape()) == 4:
tf_fun.image_summaries(train_images, tag='Training images')
if len(train_labels.get_shape()) > 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.')
# Validation model
scope.reuse_variables()
val_model = model_utils.model_class(
mean=train_means_image, # Normalize with train data
training=False, # False,
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')
eval_graph = tf.Graph()
with eval_graph.as_default():
with eval_graph.gradient_override_map({'selu': 'GradLRP'}):
val_grad_images = tf.gradients(
val_scores[0] * tf.cast(val_labels, tf.float32),
val_images)[0]
log.info('Built validation model.')
# Check the shapes of labels and scores
if not isinstance(train_scores, list):
if len(val_scores.get_shape()) != len(val_labels.get_shape()):
val_shape = val_scores.get_shape().as_list()
val_label_shape = val_labels.get_shape().as_list()
if len(val_shape) == 2 and len(
val_label_shape) == 1 and val_shape[-1] == 1:
val_labels = tf.expand_dims(val_labels, axis=-1)
if len(val_shape) == 2 and len(
val_label_shape) == 1 and val_shape[-1] == 1:
val_scores = tf.expand_dims(val_scores, axis=-1)
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 len(val_images.get_shape()) == 4:
tf_fun.image_summaries(val_images, tag='Validation')
if len(val_labels.get_shape()) > 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
saver = tf.train.Saver(tf.global_variables())
# 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()))
# 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_images': train_images,
'train_labels': train_labels,
'train_op': train_op,
'train_scores': train_scores,
'train_grad_images': train_grad_images
}
val_dict = {
'val_loss': val_loss,
'val_images': val_images,
'val_labels': val_labels,
'val_scores': val_scores,
'val_grad_images': val_grad_images
}
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
checkpoint_dir = dir_list['checkpoints']
step = 0
train_losses, train_accs, train_aux, timesteps = {}, {}, {}, {}
val_scores, val_aux, val_labels, val_grads = {}, {}, {}, {}
train_images, val_images = {}, {}
train_scores, train_labels = {}, {}
train_aux_check = np.any(['aux_score' in k for k in train_dict.keys()])
val_aux_check = np.any(['aux_score' in k for k in val_dict.keys()])
# Restore model
saver.restore(sess, ckpt_file)
# Start evaluation
try:
while not coord.should_stop():
start_time = time.time()
train_vars = sess.run(train_dict.values())
it_train_dict = {
k: v
for k, v in zip(train_dict.keys(), train_vars)
}
duration = time.time() - start_time
train_losses[step] = it_train_dict['train_loss']
train_accs[step] = it_train_dict['train_accuracy_0']
train_images[step] = it_train_dict['train_images']
train_labels[step] = it_train_dict['train_labels']
train_scores[step] = it_train_dict['train_scores']
timesteps[step] = duration
if train_aux_check:
# Loop through to find aux scores
it_train_aux = {
itk: itv
for itk, itv in it_train_dict.iteritems()
if 'aux_score' in itk
}
train_aux[step] = it_train_aux
assert not np.isnan(it_train_dict['train_loss']).any(
), 'Model diverged with loss = NaN'
if step % config.validation_iters == 0:
it_val_scores, it_val_labels, it_val_aux, it_val_grads, it_val_ims = [], [], [], [], []
for num_vals in range(config.num_validation_evals):
# Validation accuracy as the average of n batches
val_vars = sess.run(val_dict.values())
it_val_dict = {
k: v
for k, v in zip(val_dict.keys(), val_vars)
}
it_val_labels += [it_val_dict['val_labels']]
it_val_scores += [it_val_dict['val_scores']]
it_val_grads += [it_val_dict['val_grad_images']]
it_val_ims += [it_val_dict['val_images']]
if val_aux_check:
iva = {
itk: itv
for itk, itv in it_val_dict.iteritems()
if 'aux_score' in itk
}
it_val_aux += [iva]
val_scores[step] = it_val_scores
val_labels[step] = it_val_labels
val_aux[step] = it_val_aux
val_images[step] = it_val_grads
val_grads[step] = it_val_ims
# End iteration
step += 1
except tf.errors.OutOfRangeError:
print 'Done with evaluation for %d epochs, %d steps.' % (config.epochs,
step)
print 'Saved to: %s' % checkpoint_dir
finally:
coord.request_stop()
coord.join(threads)
sess.close()
import ipdb
ipdb.set_trace()
np.savez(
'val_imgs_grads',
val_images=val_images, # it_val_dict['val_images'],
val_grads=val_grads, # it_val_dict['val_grad_images'],
val_labels=val_labels, # it_val_dict['val_labels'],
val_scores=val_scores) # it_val_dict['val_scores'][0])
def main(
experiment_name,
im_ext='.pdf',
transform_loss=None, # 'log',
colors='Paired',
flip_axis=False,
port_fwd=False,
num_steps=np.inf,
exclude=None,
list_experiments=False,
out_dir='analysis_data'):
"""Plot results of provided experiment name."""
config = Config()
if list_experiments:
db.list_experiments()
return
if port_fwd:
config.db_ssh_forward = True
py_utils.make_dir(out_dir)
# Get experiment data
if ',' in experiment_name:
exps = experiment_name.split(',')
perf = []
for exp in exps:
perf += db.get_performance(experiment_name=exp)
experiment_name = exps[0]
else:
perf = db.get_performance(experiment_name=experiment_name)
if len(perf) == 0:
raise RuntimeError('Could not find any results.')
structure_names = [x['model'].split('/')[-1] for x in perf]
datasets = [x['val_dataset'] for x in perf]
steps = [float(x['step']) for x in perf]
training_loss = [float(x['train_loss']) for x in perf]
validation_loss = [float(x['val_loss']) for x in perf]
training_score = [float(x['train_score']) for x in perf]
validation_score = [float(x['val_score']) for x in perf]
summary_dirs = [x['summary_path'] for x in perf]
ckpts = [x['ckpt_path'] for x in perf]
params = [x['num_params'] for x in perf]
lrs = [x['lr'] for x in perf]
# Pass data into a pandas DF
df = pd.DataFrame(np.vstack(
(structure_names, datasets, steps, params, training_loss,
training_score, validation_loss, validation_score, summary_dirs,
ckpts, lrs)).transpose(),
columns=[
'model names', 'datasets', 'training iteration',
'params', 'training loss', 'training accuracy',
'validation loss', 'validation accuracy',
'summary_dirs', 'checkpoints', 'lrs'
])
df['training loss'] = pd.to_numeric(df['training loss'], errors='coerce')
df['validation accuracy'] = pd.to_numeric(df['validation accuracy'],
errors='coerce')
df['training accuracy'] = pd.to_numeric(df['training accuracy'],
errors='coerce')
df['training iteration'] = pd.to_numeric(df['training iteration'],
errors='coerce')
df['params'] = pd.to_numeric(df['params'], errors='coerce')
df['lrs'] = pd.to_numeric(df['lrs'], errors='coerce')
# Plot TTA
dfs = []
print(len(df))
uni_structure_names = np.unique(structure_names)
max_num_steps = num_steps # (20000 / 32) * num_epochs
# min_num_steps = 1
for m in tqdm(uni_structure_names, total=len(uni_structure_names)):
it_df = df[df['model names'] == m]
it_df = it_df[it_df['training iteration'] < max_num_steps]
# sorted_df = it_df.sort_values('training loss')
# max_vals = sorted_df.groupby(['datasets']).first()
sorted_df = []
different_models = np.unique(it_df['summary_dirs'])
num_models = len(different_models)
for model in different_models:
# Grab each model then sort by training iteration
sel_data = it_df[it_df['summary_dirs'] == model]
sel_data = sel_data.sort_values('training iteration')
# Smooth the sorted validation scores for tta
sel_data['tta'] = ndimage.gaussian_filter1d(
sel_data['validation accuracy'], 3)
sel_data['num_runs'] = num_models
sorted_df += [sel_data]
sorted_df = pd.concat(sorted_df)
dfs += [sorted_df]
# Get max scores and TTAs
dfs = pd.concat(dfs)
scores = dfs.groupby(['lrs', 'datasets', 'model names'],
as_index=False).max() # skipna=True)
losses = dfs.groupby(['lrs', 'datasets', 'model names'],
as_index=False).min() # skipna=True)
ttas = dfs.groupby(['lrs', 'datasets', 'model names'],
as_index=False).mean() # skipna=True)
# Combine into a single DF
print('Sort by val loss, then validate each (make a new dataloader)')
scores['tta'] = ttas['validation accuracy']
scores['validation loss'] = losses['validation loss']
# Save datasets to csv
filename = 'raw_data_%s.csv' % experiment_name
dfs.to_csv(os.path.join(out_dir, filename))
filename = 'scores_%s.csv' % experiment_name
scores.to_csv(os.path.join(out_dir, filename))
# Save an easy-to-parse csv for test datasets and fix for automated processing
trim_ckpts, trim_models = [], []
for idx in range(len(scores)):
ckpt = scores.iloc[idx]['checkpoints']
ckpt = '%s-%s' % (ckpt, ckpt.split('.')[0].split('_')[-1])
model = scores.iloc[idx]['model names']
trim_ckpts += [ckpt]
trim_models += [model]
# trimmed_ckpts = pd.DataFrame(trim_ckpts, columns=['checkpoints'])
# trimmed_models = pd.DataFrame(trim_models, columns=['model'])
trimmed_ckpts = pd.DataFrame(trim_ckpts)
trimmed_models = pd.DataFrame(trim_models)
trimmed_ckpts.to_csv(
os.path.join(out_dir, 'checkpoints_%s.csv' % experiment_name))
trimmed_models.to_csv(
os.path.join(out_dir, 'models_%s.csv' % experiment_name))
# Add indicator variable to group different model types during plotting
scores['model_idx'] = 0
model_groups = ['fgru', 'resnet', 'unet', 'hgru']
for idx, m in enumerate(model_groups):
scores['model_idx'][scores['model names'].str.contains(
m, regex=False)] = idx
keep_groups = np.where(~np.in1d(model_groups, 'hgru'))[0]
scores = scores[scores['model_idx'].isin(keep_groups)]
# Print scores to console
print scores
# Create max accuracy plots and aggregated dataset
num_groups = len(keep_groups)
# agg_df = []
f = plt.figure()
sns.set(context='paper', font='Arial', font_scale=.5)
sns.set_style("white")
sns.despine()
count = 1
for idx in keep_groups:
plt.subplot(1, num_groups, count)
sel_df = scores[scores['model_idx'] == idx]
# sel_df = sel_df.groupby(
# ['datasets', 'model names'], as_index=False).aggregate('max')
# agg_df += [sel_df]
sns.pointplot(data=sel_df,
x='datasets',
y='validation accuracy',
hue='model names')
plt.ylim([0.4, 1.1])
count += 1
plt.savefig(os.path.join(out_dir, 'max_%s.png' % experiment_name), dpi=300)
filename = 'agg_data_%s.csv' % experiment_name
# agg_df = pd.concat(agg_df)
# agg_df.to_csv(os.path.join(out_dir, filename))
plt.close(f)
# Create tta plots
f = plt.figure()
sns.set(context='paper', font='Arial', font_scale=.5)
sns.set_style("white")
sns.despine()
count = 1
for idx in keep_groups:
plt.subplot(1, num_groups, count)
sel_df = scores[scores['model_idx'] == idx]
# sel_df = sel_df.groupby(
# ['datasets', 'model names'], as_index=False).aggregate('mean')
sns.pointplot(data=sel_df, x='datasets', y='tta', hue='model names')
plt.ylim([0.4, 1.1])
count += 1
plt.savefig(os.path.join(out_dir, 'tta_%s.png' % experiment_name), dpi=300)
plt.close(f)
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 main(experiment_name, list_experiments=False):
"""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
return
# Prepare to run the model
config = Config()
condition_label = '%s_%s' % (experiment_name, 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 = 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
# Prepare data loaders on the cpu
with tf.device('/cpu:0'):
# Test issues with data loading? Try placeholders instead.
train_images = tf.placeholder(tf.float32,
name='train_images',
shape=[config.batch_size] +
dataset_module.im_size)
train_labels = tf.placeholder(tf.int64,
name='train_labels',
shape=[config.batch_size])
val_images = tf.placeholder(tf.float32,
name='val_images',
shape=[config.batch_size] +
dataset_module.im_size)
val_labels = tf.placeholder(tf.int64,
name='val_labels',
shape=[config.batch_size])
log.info('Created tfrecord dataloader tensors.')
# Prepare model on GPU
with tf.device('/gpu:0'):
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)
# Click weighting
flat_ims = tf.reshape(
train_images,
[config.batch_size,
np.prod(dataset_module.im_size)])
W = tf.get_variable(
name='W',
initializer=tf.truncated_normal_initializer(stddev=0.1),
shape=[
np.prod(dataset_module.im_size), dataset_module.output_size
])
b = tf.get_variable(
name='b',
initializer=tf.truncated_normal_initializer(stddev=0.1),
shape=[dataset_module.output_size])
output_scores = tf.matmul(flat_ims, W) + b
# Prepare the loss function
train_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=train_labels, logits=output_scores))
train_op = tf.train.GradientDescentOptimizer(
config.lr).minimize(train_loss)
log.info('Built training loss function.')
# 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()))
# Start training loop
step, time_elapsed = 0, 0
train_losses, train_accs, val_losses, val_accs, timesteps = {}, {}, {}, {}, {}
files, labels = dataset_module.get_data()
combined_files = files['train']
combined_labels = labels['train']
batch_size = config.batch_size
num_batches = len(combined_files) // batch_size
log.info('Finished training.')
for image_batch, label_batch, _ in tqdm(image_batcher(
start=0,
num_batches=num_batches,
images=combined_files,
labels=combined_labels,
batch_size=batch_size),
total=num_batches):
feed_dict = {
train_images: image_batch.astype(np.float32),
train_labels: np.asarray(label_batch).astype(int)
}
import ipdb
ipdb.set_trace()
start_time = time.time()
_, loss_value = sess.run([
train_op,
train_loss,
], feed_dict=feed_dict)
files_to_save = {
'training_loss': tr_loss,
'validation_loss': val_loss,
'training_acc': tr_accs,
'validation_acc': val_accs,
'timesteps': timesteps
}
model_name = config.model_struct.replace('/', '_')
py_utils.save_npys(data=files_to_save,
model_name=model_name,
output_string=dir_list['experiment_evaluations'])
# Compare this condition w/ all others.
plotting.plot_data(train_loss=tr_loss,
val_loss=val_loss,
model_name=model_name,
timesteps=timesteps,
config=config,
output=os.path.join(dir_list['condition_evaluations'],
'loss'),
output_ext='.pdf',
data_type='loss')
plotting.plot_data(tr_accs=tr_accs,
val_accs=val_accs,
model_name=model_name,
timesteps=timesteps,
config=config,
output=os.path.join(dir_list['condition_evaluations'],
'acc'),
output_ext='.pdf',
data_type='acc')
log.info('Completed plots.')