def run_training( cfg ):
# set up logging
tf.logging.set_verbosity( tf.logging.INFO )
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = utils.setup_input( cfg, is_training=False, use_filename_queue=True )
RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
RuntimeDeterminedEnviromentVars.populate_registered_variables()
# build model (and losses and train_op)
model = setup_model( inputs, cfg, is_training=False )
# set up metrics to evaluate
names_to_values, names_to_updates = setup_metrics( inputs, model, cfg )
# execute training
start_time = time.time()
utils.print_start_info( cfg, inputs[ 'max_steps' ], is_training=False )
training_runners = { 'sess': tf.Session(), 'coord': tf.train.Coordinator() }
data_prefetch_init_fn = utils.get_data_prefetch_threads_init_fn( inputs, cfg, is_training=False, use_filename_queue=True )
training_runners[ 'threads' ] = data_prefetch_init_fn( training_runners[ 'sess' ], training_runners[ 'coord' ] )
try:
# This just returns the imput as output. It is for testing data
# input only.
for step in xrange( inputs[ 'max_steps' ] ):
input_batch, target_batch, data_idx = training_runners['sess'].run( [
model['input_batch'], model['target_batch'], model[ 'data_idxs' ] ] )
if training_runners['coord'].should_stop():
break
finally:
utils.request_data_loading_end( training_runners )
utils.end_data_loading_and_sess( training_runners )
# else: # Use tf.slim
# train_log_dir = os.path.join( cfg['log_dir'], 'slim-train' )
# # When ready to use a model, use the code below
# train( model[ 'train_op' ],
# train_log_dir,
# get_data_prefetch_threads_init_fn( inputs, cfg ),
# global_step=model[ 'global_step' ],
# number_of_steps=inputs[ 'max_steps' ],
# init_fn=model[ 'init_fn' ],
# save_summaries_secs=300,
# save_interval_secs=600,
# saver=model[ 'saver_op' ] )
end_train_time = time.time() - start_time
print('time to train %d epochs: %.3f hrs' % (cfg['num_epochs'], end_train_time/(60*60)))
print('avg time per epoch: %.3f hrs' % ( (end_train_time/(60*60)) / cfg['num_epochs']) )
def run_extract_representations(args, cfg, save_dir, given_task):
transfer = (cfg['model_type'] == architectures.TransferNet)
if transfer:
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn_transfer
setup_input_fn = utils.setup_input_transfer
else:
setup_input_fn = utils.setup_input
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn
# set up logging
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default() as g:
# create ops and placeholders
tf.logging.set_verbosity(tf.logging.INFO)
inputs = setup_input_fn(cfg,
is_training=False,
use_filename_queue=False)
RuntimeDeterminedEnviromentVars.load_dynamic_variables(inputs, cfg)
RuntimeDeterminedEnviromentVars.populate_registered_variables()
# build model (and losses and train_op)
# set up metrics to evaluate
# names_to_values, names_to_updates = setup_metrics( inputs, model, cfg )
# execute training
start_time = time.time()
utils.print_start_info(cfg, inputs['max_steps'], is_training=False)
# start session and restore model
training_runners = {
'sess': tf.Session(),
'coord': tf.train.Coordinator()
}
try:
if cfg['model_path'] is None:
print('Please specify a checkpoint directory')
return
utils.print_start_info(cfg, inputs['max_steps'], is_training=False)
data_prefetch_init_fn = get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=False, use_filename_queue=False)
prefetch_threads = threading.Thread(
target=data_prefetch_init_fn,
args=(training_runners['sess'], training_runners['coord']))
prefetch_threads.start()
# run one example so that we can calculate some statistics about the representations
filenames = []
representations, data_idx = training_runners['sess'].run(
[inputs['input_batch'], inputs['data_idxs']])
filenames.extend(data_idx)
if type(representations) == list:
representations = representations[0]
representations = representations.reshape(
(-1, np.prod(cfg['input_dim'])))
print('Got first batch representation with size: {0}'.format(
representations.shape))
# run the remaining examples
for step in range(inputs['max_steps'] - 1):
#for step in range( 10 ):
if step % 100 == 0:
print('Step {0} of {1}'.format(step,
inputs['max_steps'] - 1))
# This is just for GAN, for the LEO meeting
encoder_output, data_idx = training_runners['sess'].run(
[inputs['input_batch'], inputs['data_idxs']])
if type(encoder_output) == list:
encoder_output = encoder_output[0]
representations = np.append(representations,
encoder_output.reshape(
(-1,
np.prod(cfg['input_dim']))),
axis=0)
filenames.extend(data_idx)
if training_runners['coord'].should_stop():
break
print(
'The size of representations is %s while we expect it to run for %d steps with batchsize %d'
% (representations.shape, inputs['max_steps'],
cfg['batch_size']))
end_train_time = time.time() - start_time
save_path = os.path.join(
save_dir, '{task}_{split}_representations.pkl'.format(
task='pixels', split=args.data_split))
with open(save_path, 'wb') as f:
pickle.dump(
{
'file_indexes': filenames,
'representations': representations
}, f)
copy_to = None
if args.out_dir:
os.system("sudo cp {fp} {out}/".format(fp=save_path,
out=args.out_dir))
copy_to = args.out_dir
else:
if transfer:
os.system(
"sudo cp {fp} /home/ubuntu/s3/model_log/representations_transfer/"
.format(fp=save_path))
copy_to = '/home/ubuntu/s3/model_log/representations_transfer/'
else:
os.system(
"sudo cp {fp} /home/ubuntu/s3/model_log/representations/"
.format(fp=save_path))
copy_to = "/home/ubuntu/s3/model_log/representations/"
print('saved representations to {0}'.format(save_path))
print('copied representations to {0}'.format(copy_to))
print('time to extract %d epochs: %.3f hrs' %
(cfg['num_epochs'], end_train_time / (60 * 60)))
finally:
utils.request_data_loading_end(training_runners)
utils.end_data_loading_and_sess(training_runners)
def run_extract_losses(args, cfg, save_dir, given_task):
transfer = (cfg['model_type'] == architectures.TransferNet)
if transfer:
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn_transfer
setup_input_fn = utils.setup_input_transfer
else:
setup_input_fn = utils.setup_input
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn
# set up logging
tf.logging.set_verbosity(tf.logging.ERROR)
stats = Statistics()
print_every = int(args.print_every)
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = setup_input_fn(cfg,
is_training=False,
use_filename_queue=False)
#RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
#RuntimeDeterminedEnviromentVars.populate_registered_variables()
max_steps = get_max_steps(inputs['max_steps'], args.data_split)
# pdb.set_trace()
# build model (and losses and train_op)
model = utils.setup_model(inputs, cfg, is_training=False)
loss_names, loss_ops = get_extractable_losses(cfg, model)
if 'l1_loss' in loss_names:
display_loss = 'l1_loss'
elif 'l2_loss' in loss_names:
display_loss = 'l2_loss'
elif 'xentropy' in loss_names:
display_loss = 'xentropy'
elif 'metric_loss' in loss_names:
display_loss = 'metric_loss'
elif 'cycle_loss' in loss_names:
display_loss = 'cycle_loss'
else:
display_loss = 'total_loss'
# set up metrics to evaluate
names_to_values, names_to_updates = setup_metrics(inputs, model, cfg)
# execute training
start_time = time.time()
utils.print_start_info(cfg, max_steps, is_training=False)
# start session and restore model
training_runners = {
'sess': tf.Session(),
'coord': tf.train.Coordinator()
}
try:
if cfg['model_path'] is None:
print('Please specify a checkpoint directory')
return
print('Attention, model_path is ', cfg['model_path'])
model['saver_op'].restore(training_runners['sess'],
cfg['model_path'])
# var = [v for v in tf.global_variables() if 'decoder' in v.name][0]
# print(training_runners[ 'sess' ].run(var))
utils.print_start_info(cfg, max_steps, is_training=False)
data_prefetch_init_fn = get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=False, use_filename_queue=False)
prefetch_threads = threading.Thread(
target=data_prefetch_init_fn,
args=(training_runners['sess'], training_runners['coord']))
prefetch_threads.start()
# run one example so that we can calculate some statistics about the representations
filenames = []
loss_names_to_vals = {name: [] for name in loss_names}
results = training_runners['sess'].run([
inputs['data_idxs'], inputs['target_batch'],
inputs['mask_batch'], *loss_ops
])
#gs = results[1]
data_idx = results[0]
losses = results[3:]
target_input = results[1]
mask_input = results[2]
for i, name in enumerate(loss_names):
loss_names_to_vals[name].append(losses[i])
filenames.extend(data_idx)
print("Step number: {}".format(1), (data_idx))
# print(target_input, target_input.sum())
# return
# training_runners['sess'].run([v for v in tf.global_variables() if "transfer/rep_conv_1/weights" in v.name][0])
# run the remaining examples
start = time.perf_counter()
for step in range(max_steps - 1):
results = training_runners['sess'].run([
inputs['data_idxs'],
# [v for v in tf.global_variables() if "transfer/rep_conv_1/weights/(weights)" in v.name][0],
# model['model'].encoder_endpoints['net1_1_output'],
# model['model'].encoder_endpoints['net1_2_output'],
*loss_ops
])
data_idx = results[0]
# print(data_idx)
losses = results[1:]
# p, t, m = results[1], results[2], results[3]
# losses = results[4:]
# print(p.mean(), t)
for i, name in enumerate(loss_names):
loss_names_to_vals[name].append(losses[i])
filenames.extend(data_idx)
stats.push(loss_names_to_vals[display_loss][-1])
# baseline_loss = get_xentropy_loss(p, t, m)
# tf_loss = loss_names_to_vals[display_loss][-1]
# print('tf {} | ours {}'.format(tf_loss, baseline_loss))
# pdb.set_trace()
if step % print_every == 0 and step > 0:
print(
'Step {0} of {1}: ({5} loss: {2:.3f} || stddev: {3:.3f} :: ({4:.2f} secs/step)'
.format(
step,
max_steps - 1,
stats.mean(),
np.sqrt(stats.variance()),
# stats.variance(),
(time.perf_counter() - start) / print_every,
display_loss))
start = time.perf_counter()
if training_runners['coord'].should_stop():
break
print(
'The size of losses is %s while we expect it to run for %d steps with batchsize %d'
% (len(filenames), inputs['max_steps'], cfg['batch_size']))
end_train_time = time.time() - start_time
if args.out_name:
out_name = args.out_name
else:
out_name = '{task}_{split}_losses.pkl'.format(
task=given_task, split=args.data_split)
save_path = os.path.join(save_dir, out_name)
with open(save_path, 'wb') as f:
loss_names_to_vals['file_indexes'] = filenames
loss_names_to_vals['global_step'] = 0
pickle.dump(loss_names_to_vals, f)
if args.out_dir:
os.makedirs(args.out_dir, exist_ok=True)
os.system("sudo cp {fp} {out}/".format(fp=save_path,
out=args.out_dir))
else:
if transfer:
copy_to = cfg['log_root']
else:
copy_to = os.path.join(cfg['log_root'], given_task)
os.system("sudo mv {fp} {dst}/".format(fp=save_path,
dst=copy_to))
print("sudo mv {fp} {dst}/".format(fp=save_path, dst=copy_to))
# if transfer:
# os.makedirs('/home/ubuntu/s3/model_log/losses_transfer/', exist_ok=True)
# os.system("sudo cp {fp} /home/ubuntu/s3/model_log/losses_transfer/".format(fp=save_path))
# else:
# os.makedirs('/home/ubuntu/s3/model_log/losses/', exist_ok=True)
# os.system("sudo cp {fp} /home/ubuntu/s3/model_log/losses/".format(fp=save_path))
print('saved losses to {0}'.format(save_path))
print('time to extract %d epochs: %.3f hrs' %
(cfg['num_epochs'], end_train_time / (60 * 60)))
finally:
utils.request_data_loading_end(training_runners)
utils.end_data_loading_and_sess(training_runners)
def run_extract_losses( args, cfg, save_dir, given_task ):
transfer = (cfg['model_type'] == architectures.TransferNet)
if transfer:
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn_transfer_imagenet
setup_input_fn = utils.setup_input_transfer_imagenet
else:
setup_input_fn = utils.setup_input
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn
# set up logging
tf.logging.set_verbosity( tf.logging.ERROR )
stats = Statistics()
top5_stats = Statistics()
print_every = int(args.print_every)
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = setup_input_fn( cfg, is_training=False, use_filename_queue=False )
#RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
#RuntimeDeterminedEnviromentVars.populate_registered_variables()
max_steps = get_max_steps(inputs[ 'max_steps' ], args.data_split)
# build model (and losses and train_op)
model = utils.setup_model( inputs, cfg, is_training=False )
# set up metrics to evaluate
names_to_values, names_to_updates = setup_metrics( inputs, model, cfg )
# execute training
start_time = time.time()
utils.print_start_info( cfg, max_steps, is_training=False )
# start session and restore model
training_runners = { 'sess': tf.Session(), 'coord': tf.train.Coordinator() }
try:
if cfg['model_path'] is None:
print('Please specify a checkpoint directory')
return
print('Attention, model_path is ', cfg['model_path'])
model[ 'saver_op' ].restore( training_runners[ 'sess' ], cfg[ 'model_path' ] )
# var = [v for v in tf.global_variables() if 'decoder' in v.name][0]
# print(training_runners[ 'sess' ].run(var))
utils.print_start_info( cfg, max_steps, is_training=False )
data_prefetch_init_fn = get_data_prefetch_threads_init_fn( inputs, cfg,
is_training=False, use_filename_queue=False )
prefetch_threads = threading.Thread(
target=data_prefetch_init_fn,
args=( training_runners[ 'sess' ], training_runners[ 'coord' ] ))
prefetch_threads.start()
# run one example so that we can calculate some statistics about the representations
filenames = []
accuracies = []
if transfer:
accuracy_op = model['model'].decoder.accuracy
final_output = model['model'].decoder.final_output
else:
accuracy_op = model['model'].accuracy
final_output = model['model'].final_output
results = training_runners['sess'].run( [
inputs[ 'data_idxs' ], model['model'].global_step,
accuracy_op ] )
gs = results[1]
data_idx = results[0]
accuracy = results[2]
filenames.extend(data_idx)
accuracies.append(accuracy)
print("Step number: {}".format(gs))
# print(loss_names_to_vals, data_idx)
# return
# run the remaining examples
start = time.perf_counter()
for step in range( max_steps - 1 ):
results = training_runners['sess'].run( [
inputs[ 'data_idxs' ],
final_output,
inputs['target_batch'],
accuracy_op ] )
data_idx = results[0]
accuracy = results[-1]
logits = results[1]
gt = results[2]
sorted_top5 = np.argsort(logits[0])[::-1][:5]
sorted_gt = np.argsort(gt[0])[::-1][0]
top5 = 0.
if sorted_gt in sorted_top5:
top5 = 1.
filenames.extend(data_idx)
accuracies.append(accuracy)
stats.push(accuracy)
top5_stats.push(top5)
if step % print_every == 0 and step > 0:
print( 'Step {0} of {1}: ({5}: {2:.3f} || Top 5: {3:.3f} :: ({4:.2f} secs/step)'.format(
step, max_steps - 1,
stats.mean(),
top5_stats.mean(),
# stats.variance(),
(time.perf_counter() - start) / print_every,
'accuracy'
))
start = time.perf_counter()
if training_runners['coord'].should_stop():
break
os.system("sudo touch /home/ubuntu/s3/imagenet_accuracy/{}_{}_{}.txt".format(
given_task, int(stats.mean() * 1000) / 10., int(top5_stats.mean() * 1000) / 10.))
print('The size of losses is %s while we expect it to run for %d steps with batchsize %d' % (len(filenames), inputs['max_steps'], cfg['batch_size']))
end_train_time = time.time() - start_time
if args.out_name:
out_name = args.out_name
else:
out_name = '{task}_{split}_imagenet_accuracy.pkl'.format(task=given_task, split=args.data_split)
save_path = os.path.join( save_dir, out_name )
val_accuracy = {}
with open( save_path, 'wb' ) as f:
val_accuracy['file_indexes'] = filenames
val_accuracy['global_step'] = gs
val_accuracy['accuracy'] = accuracies
pickle.dump( val_accuracy, f )
if args.out_dir:
os.makedirs(args.out_dir, exist_ok=True)
os.system("sudo cp {fp} {out}/".format(fp=save_path, out=args.out_dir))
else:
if transfer:
copy_to = cfg['log_root']
else:
copy_to = os.path.join(cfg['log_root'], given_task)
os.system("sudo mv {fp} {dst}/".format(fp=save_path, dst=copy_to))
print("sudo mv {fp} {dst}/".format(fp=save_path, dst=copy_to))
# if transfer:
# os.makedirs('/home/ubuntu/s3/model_log/losses_transfer/', exist_ok=True)
# os.system("sudo cp {fp} /home/ubuntu/s3/model_log/losses_transfer/".format(fp=save_path))
# else:
# os.makedirs('/home/ubuntu/s3/model_log/losses/', exist_ok=True)
# os.system("sudo cp {fp} /home/ubuntu/s3/model_log/losses/".format(fp=save_path))
print( 'saved losses to {0}'.format( save_path ))
print('time to extract %d epochs: %.3f hrs' % (cfg['num_epochs'], end_train_time/(60*60)))
finally:
utils.request_data_loading_end( training_runners )
utils.end_data_loading_and_sess( training_runners )
def run_rand_baseline( args, cfg, given_task ):
# set up logging
tf.logging.set_verbosity( tf.logging.INFO )
with tf.Graph().as_default() as g:
# create ops and placeholders
tf.logging.set_verbosity( tf.logging.INFO )
inputs = utils.setup_input( cfg, is_training=False, use_filename_queue=False )
RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
RuntimeDeterminedEnviromentVars.populate_registered_variables()
# build model (and losses and train_op)
model = utils.setup_model( inputs, cfg, is_training=False )
# set up metrics to evaluate
names_to_values, names_to_updates = setup_metrics( inputs, model, cfg )
# execute training
start_time = time.time()
utils.print_start_info( cfg, inputs[ 'max_steps' ], is_training=False )
# start session and restore model
training_runners = { 'sess': tf.Session(), 'coord': tf.train.Coordinator() }
try:
utils.print_start_info( cfg, inputs[ 'max_steps' ], is_training=False )
data_prefetch_init_fn = utils.get_data_prefetch_threads_init_fn( inputs, cfg, is_training=False, use_filename_queue=False )
#training_runners[ 'threads' ] = data_prefetch_init_fn( training_runners[ 'sess' ], training_runners[ 'coord' ] )
prefetch_threads = threading.Thread(
target=data_prefetch_init_fn,
args=( training_runners[ 'sess' ], training_runners[ 'coord' ] ))
prefetch_threads.start()
# run one example so that we can calculate some statistics about the representations
targets = training_runners['sess'].run( inputs[ 'target_batch' ] )
# run the remaining examples
for step in range( inputs[ 'max_steps' ] - 1 ):
#for step in range( 10 ):
if step % 100 == 0:
print( 'Step {0} of {1}'.format( step, inputs[ 'max_steps' ] - 1 ))
target = training_runners['sess'].run( inputs[ 'target_batch' ] )
targets = np.append( targets, target, axis=0)
if training_runners['coord'].should_stop():
break
rand_idx = [random.randint(0, targets.shape[0] - 1) for i in range(targets.shape[0])]
rand_target = [targets[i] for i in rand_idx]
rand_target = np.vstack(rand_target)
counter = 0
sum = 0
for step in range( inputs[ 'max_steps' ] - 1 ):
#for step in range( 10 ):
if step % 100 == 0:
print( 'Step {0} of {1}'.format( step, inputs[ 'max_steps' ] - 1 ))
tar = targets[step*cfg['batch_size']:(step+1)*cfg['batch_size']]
rand = rand_target[step*cfg['batch_size']:(step+1)*cfg['batch_size']]
losses = training_runners['sess'].run( model['model'].losses, feed_dict={
inputs['target_batch']: tar, model['model'].final_output:rand})
sum += losses[0]
counter += 1
if training_runners['coord'].should_stop():
break
print(sum)
print(counter)
print('random_baseline has loss: {loss}'.format(loss=sum/counter))
end_train_time = time.time() - start_time
finally:
utils.request_data_loading_end( training_runners )
utils.end_data_loading_and_sess( training_runners )
def run_training(cfg):
# set up logging
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = utils.setup_input(cfg,
is_training=False,
use_filename_queue=True)
RuntimeDeterminedEnviromentVars.load_dynamic_variables(inputs, cfg)
RuntimeDeterminedEnviromentVars.populate_registered_variables()
# build model (and losses and train_op)
model = setup_model(inputs, cfg, is_training=False)
# set up metrics to evaluate
names_to_values, names_to_updates = setup_metrics(inputs, model, cfg)
# execute training
start_time = time.time()
utils.print_start_info(cfg, inputs['max_steps'], is_training=False)
# start session and restore model
training_runners = {
'sess': tf.Session(),
'coord': tf.train.Coordinator()
}
if cfg['model_path'] is None:
print('Please specify a checkpoint directory')
return
cfg['randomize'] = False
model['saver_op'].restore(training_runners['sess'], cfg['model_path'])
utils.print_start_info(cfg,
inputs['max_steps'],
is_training=is_training)
data_prefetch_init_fn = utils.get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=False, use_filename_queue=True)
training_runners['threads'] = data_prefetch_init_fn(
training_runners['sess'], training_runners['coord'])
representations, input_batch, target_batch, data_idx = training_runners[
'sess'].run([
model['model'].encoder_output, inputs['input_batch'],
inputs['target_batch'], inputs['data_idxs'],
inputs['mask_batch']
])
print('Got first batch representation with size:%s' %
(representations.shape))
for step in xrange(inputs['max_steps'] - 1):
encoder_output, input_batch, target_batch, data_idx = training_runners[
'sess'].run([
model['model'].encoder_output, inputs['input_batch'],
inputs['target_batch'], inputs['data_idxs'],
inputs['mask_batch']
])
representations = np.append(representations,
encoder_output,
axis=0)
if training_runners['coord'].should_stop():
break
print(
'The size of representations is %s while we expect it to run for %d steps with batchsize %d'
% (representations.shape, inputs['max_steps'], cfg['batch_size']))
utils.request_data_loading_end(training_runners)
utils.end_data_loading_and_sess(training_runners)
end_train_time = time.time() - start_time
print('time to train %d epochs: %.3f hrs' %
(cfg['num_epochs'], end_train_time / (60 * 60)))
print('avg time per epoch: %.3f hrs' %
((end_train_time / (60 * 60)) / cfg['num_epochs']))
def run_training(cfg, cfg_dir):
# set up logging
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = utils.setup_input_transfer(cfg, is_training=True)
RuntimeDeterminedEnviromentVars.load_dynamic_variables(inputs, cfg)
RuntimeDeterminedEnviromentVars.populate_registered_variables()
# build model (and losses and train_op)
model = utils.setup_model(inputs, cfg, is_training=True)
# execute training
start_time = time.time()
utils.print_start_info(cfg, inputs['max_steps'], is_training=True)
if cfg['model_type'] == 'empty': # Can't use tf slim because not trainable variables
training_runners = {
'sess': tf.Session(),
'coord': tf.train.Coordinator()
}
data_prefetch_init_fn = utils.get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=True)
training_runners['threads'] = data_prefetch_init_fn(
training_runners['sess'], training_runners['coord'])
try:
# This just returns the imput as output. It is for testing data
# input only.
for step in xrange(inputs['max_steps']):
input_batch, target_batch, data_idx = training_runners[
'sess'].run([
model['input_batch'], model['target_batch'],
model['data_idxs']
])
if training_runners['coord'].should_stop():
break
finally:
utils.request_data_loading_end(training_runners)
utils.end_data_loading_and_sess(training_runners)
else: # Use tf.slim
train_log_dir = os.path.join(cfg['log_dir'], 'slim-train')
permanent_checkpoint_dir = os.path.join(cfg['log_dir'],
'checkpoints')
session_config = tf.ConfigProto()
session_config.gpu_options.allow_growth = True
#max_to_keep = cfg['num_epochs'] * 2
max_to_keep = 10
if 'max_ckpts_to_keep' in cfg:
max_to_keep = cfg['max_ckpts_to_keep']
# When ready to use a model, use the code below
train(
model['train_op'],
train_log_dir,
utils.get_data_prefetch_threads_init_fn_transfer(
inputs, cfg, is_training=True),
train_step_fn=model['train_step_fn'],
train_step_kwargs=model['train_step_kwargs'],
global_step=model['global_step'],
number_of_steps=inputs['max_steps'],
number_of_epochs=cfg['num_epochs'],
init_fn=model['init_fn'],
save_checkpoint_every=max(inputs['max_steps'] // (max_to_keep),
500),
cfg_dir=cfg_dir,
#RuntimeDeterminedEnviromentVars.steps_per_epoch,
permanent_checkpoint_dir=permanent_checkpoint_dir,
save_summaries_secs=cfg['summary_save_every_secs'],
save_interval_secs=cfg['checkpoint_save_every_secs'],
saver=model['saver_op'],
return_accuracy='return_accuracy' in cfg
and cfg['return_accuracy'],
session_config=session_config)
end_train_time = time.time() - start_time
print('time to train %d epochs: %.3f hrs' %
(cfg['num_epochs'], end_train_time / (60 * 60)))
print('avg time per epoch: %.3f hrs' %
((end_train_time / (60 * 60)) / cfg['num_epochs']))
def run_training(cfg, cfg_dir, args):
if args.stat_type == "mean":
statistic = MeanMeter(cfg)
elif args.stat_type == 'median':
statistic = MedianMeter(cfg)
elif args.stat_type == 'marginal':
statistic = DiscreteDistributionMeter(cfg, args.not_one_hot)
elif args.stat_type == 'dense_marginal':
statistic = DenseDiscreteDistributionMeter(cfg)
elif args.stat_type == 'moments':
statistic = MomentsMeter(cfg)
else:
raise NotImplementedError("No average defined for type: {}".format(
args.stat_type))
# set up logging
tf.logging.set_verbosity(tf.logging.ERROR)
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = utils.setup_input(cfg, is_training=False)
RuntimeDeterminedEnviromentVars.load_dynamic_variables(inputs, cfg)
RuntimeDeterminedEnviromentVars.populate_registered_variables()
# execute training
start_time = time.time()
max_steps = get_max_steps(inputs['max_steps'], args.data_split)
utils.print_start_info(cfg, max_steps, is_training=False)
data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=False)
training_runners = {
'sess': tf.Session(),
'coord': tf.train.Coordinator()
}
prefetch_threads = threading.Thread(
target=data_prefetch_threads_init_fn,
args=(training_runners['sess'], training_runners['coord']))
prefetch_threads.start()
target_batch = training_runners['sess'].run(inputs['target_batch'])
# training_runners[ 'threads' ] = data_prefetch_init_fn( training_runners[ 'sess' ], training_runners[ 'coord' ] )
try:
# This just returns the imput as output. It is for testing data
# input only.
start_time = time.time()
batch_time = time.time()
k = int(args.print_every)
for step in range(max_steps):
target_batch, mask_batch = training_runners['sess'].run(
[inputs['target_batch'], inputs['mask_batch']])
target_batch = map_to_img(target_batch.mean(axis=0), cfg)
if len(mask_batch.shape) > 1:
mask_batch = mask_batch.mean(axis=0)
else:
mask_batch = 1
statistic.update(target_batch, mask_batch)
if (step + 1) % k == 0:
print('Step %d/%d: %.2f s/step ' %
(step + 1, max_steps,
(time.time() - batch_time) / k))
batch_time = time.time()
# print(statistic.get())
# break
if training_runners['coord'].should_stop():
break
end_train_time = time.time() - start_time
print('time to train %d epochs: %.3f hrs' %
(cfg['num_epochs'], end_train_time / (60 * 60)))
print('avg time per epoch: %.3f hrs' %
((end_train_time / (60 * 60)) / cfg['num_epochs']))
if args.stat_type == 'moments':
save_moments(statistic, cfg, args)
else:
save_data(statistic, cfg, args)
finally:
utils.request_data_loading_end(training_runners)
utils.end_data_loading_and_sess(training_runners)
def run_extract_losses(args, cfg, save_dir, given_task):
transfer = (cfg['model_type'] == architectures.TransferNet)
if transfer:
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn_transfer
setup_input_fn = utils.setup_input_transfer
if given_task == 'pixels':
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn_transfer_imagenet
setup_input_fn = utils.setup_input_transfer_imagenet
else:
setup_input_fn = utils.setup_input
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn
# set up logging
tf.logging.set_verbosity(tf.logging.ERROR)
stats = Statistics()
print_every = int(args.print_every)
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = setup_input_fn(cfg,
is_training=True,
use_filename_queue=False)
#RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
#RuntimeDeterminedEnviromentVars.populate_registered_variables()
max_steps = get_max_steps(inputs['max_steps'], args.data_split)
# build model (and losses and train_op)
model = utils.setup_model(inputs, cfg, is_training=True)
# set up metrics to evaluate
names_to_values, names_to_updates = setup_metrics(inputs, model, cfg)
train_step_fn = model['train_step_fn']
# execute training
start_time = time.time()
utils.print_start_info(cfg, max_steps, is_training=True)
# start session and restore model
training_runners = {
'sess': tf.Session(),
'coord': tf.train.Coordinator()
}
try:
if cfg['model_path'] is None:
print('Please specify a checkpoint directory')
return
print('Attention, model_path is ', cfg['model_path'])
restore_ckpt = not args.from_scratch
if restore_ckpt:
non_encoder_var = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES)
adams = []
for v in tuple(non_encoder_var):
if 'Adam' in v.name:
non_encoder_var.remove(v)
adams.append(v)
continue
if 'finetune_encoder_imagenet' in cfg:
for x in model['model'].encoder_vars:
if v.name == x.name:
non_encoder_var.remove(v)
if not args.metric_only:
saver_for_transfer = tf.train.Saver(non_encoder_var)
#training_runners['sess'].run(saver_for_transfer)
saver_for_transfer.restore(training_runners['sess'],
cfg['model_path'])
else:
#saver_for_transfer = tf.train.Saver(non_encoder_var)
training_runners['sess'].run(
tf.variables_initializer(non_encoder_var))
training_runners['sess'].run(tf.variables_initializer(adams))
print('Loading Source Encoder:...')
if 'finetune_encoder_imagenet' in cfg:
model['init_fn'](training_runners['sess'])
print('Starting Training:..')
else:
init_op = tf.global_variables_initializer()
training_runners['sess'].run(init_op)
assign_op = model['global_step'].assign(0)
training_runners['sess'].run(assign_op)
# var = [v for v in tf.global_variables() if 'decoder' in v.name][0]
# print(training_runners[ 'sess' ].run(var))
utils.print_start_info(cfg, max_steps, is_training=True)
data_prefetch_init_fn = get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=True, use_filename_queue=False)
prefetch_threads = threading.Thread(
target=data_prefetch_init_fn,
args=(training_runners['sess'], training_runners['coord']))
prefetch_threads.start()
# run one example so that we can calculate some statistics about the representations
start = time.perf_counter()
saver = tf.train.Saver()
save_ckpt_name = 'places'
if args.from_scratch:
save_ckpt_name = 'places_scratch_{}_{}'.format(
args.layers, args.data_used)
if args.train_encoder:
save_ckpt_name = 'places_encoder'
for step in range(max_steps // 2 - 1):
#for step in range(10):
total_loss, should_stop = train_step_fn(
training_runners['sess'],
model['train_op'],
model['global_step'],
train_step_kwargs=model['train_step_kwargs'])
# print(data_idx)
# print(p.mean(), t)
stats.push(total_loss)
if step % print_every == 0 and step > 0:
print(
'Step {0} of {1}: ({5}: {2:.3f} || stddev: {3:.3f} :: ({4:.2f} secs/step)'
.format(
step,
max_steps - 1,
stats.mean(),
np.sqrt(stats.variance()),
# stats.variance(),
(time.perf_counter() - start) / print_every,
'Total_loss'))
start = time.perf_counter()
if step % 3000 == 2999:
saver.save(
training_runners['sess'],
os.path.join(cfg['log_root'], given_task,
'{}_{}'.format(save_ckpt_name, step)))
if training_runners['coord'].should_stop():
break
#print('Heressss')
saver.save(
training_runners['sess'],
os.path.join(cfg['log_root'], given_task, save_ckpt_name))
finally:
utils.request_data_loading_end(training_runners)
utils.end_data_loading_and_sess(training_runners)
def run_extract_losses_5_steps(args, cfg, save_dir, given_task):
transfer = (cfg['model_type'] == architectures.TransferNet)
if transfer:
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn_transfer
setup_input_fn = utils.setup_input_transfer
else:
setup_input_fn = utils.setup_input
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn
# set up logging
tf.logging.set_verbosity(tf.logging.ERROR)
stats = Statistics()
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = setup_input_fn(cfg,
is_training=False,
use_filename_queue=False)
#RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
#RuntimeDeterminedEnviromentVars.populate_registered_variables()
max_steps = get_max_steps(inputs['max_steps'], args.data_split)
# build model (and losses and train_op)
model = utils.setup_model(inputs, cfg, is_training=False)
loss_names, loss_ops = get_extractable_losses(cfg, model)
if 'l1_loss' in loss_names:
display_loss = 'l1_loss'
elif 'l2_loss' in loss_names:
display_loss = 'l2_loss'
elif 'xentropy' in loss_names:
display_loss = 'xentropy'
elif 'metric_loss' in loss_names:
display_loss = 'metric_loss'
elif 'cycle_loss' in loss_names:
display_loss = 'cycle_loss'
else:
display_loss = 'total_loss'
# set up metrics to evaluate
names_to_values, names_to_updates = setup_metrics(inputs, model, cfg)
# execute training
start_time = time.time()
utils.print_start_info(cfg, max_steps, is_training=False)
# start session and restore model
training_runners = {
'sess': tf.Session(),
'coord': tf.train.Coordinator()
}
if cfg['model_path'] is None:
print('Please specify a checkpoint directory')
return
print('Attention, model_path is ', cfg['model_path'])
model['saver_op'].restore(training_runners['sess'], cfg['model_path'])
utils.print_start_info(cfg, max_steps, is_training=False)
data_prefetch_init_fn = get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=False, use_filename_queue=False)
prefetch_threads = threading.Thread(target=data_prefetch_init_fn,
args=(training_runners['sess'],
training_runners['coord']))
prefetch_threads.start()
# run one example so that we can calculate some statistics about the representations
# results = training_runners['sess'].run( [ *loss_ops ] )
# losses = results[0]
x = 0
for step in range(3):
results = training_runners['sess'].run([*loss_ops])
x = x + results[0]
if training_runners['coord'].should_stop():
break
tf.reset_default_graph()
utils.request_data_loading_end(training_runners)
utils.end_data_loading_and_sess(training_runners)
return x / 2.
def run_extract_losses(avg_img, args, cfg, save_dir, given_task):
transfer = (cfg['model_type'] == architectures.TransferNet)
if transfer:
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn_transfer
setup_input_fn = utils.setup_input_transfer
else:
setup_input_fn = utils.setup_input
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn
stats = Statistics()
# set up logging
tf.logging.set_verbosity(tf.logging.ERROR)
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = setup_input_fn(cfg,
is_training=False,
use_filename_queue=False)
RuntimeDeterminedEnviromentVars.load_dynamic_variables(inputs, cfg)
RuntimeDeterminedEnviromentVars.populate_registered_variables()
# build model (and losses and train_op)
# model = utils.setup_model( inputs, cfg, is_training=False )
loss_names = [avg_img_to_loss_type(args.avg_type, given_task)
] # Keep format the same as extract_losses.py
loss_fn = get_loss_op(loss_names[0])
# execute training
start_time = time.time()
max_steps = get_max_steps(inputs['max_steps'], args.data_split)
utils.print_start_info(cfg, max_steps, is_training=False)
# start session and restore model
training_runners = {
'sess': tf.Session(),
'coord': tf.train.Coordinator()
}
try:
utils.print_start_info(cfg, max_steps, is_training=False)
data_prefetch_init_fn = get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=False, use_filename_queue=False)
#training_runners[ 'threads' ] = data_prefetch_init_fn( training_runners[ 'sess' ], training_runners[ 'coord' ] )
prefetch_threads = threading.Thread(
target=data_prefetch_init_fn,
args=(training_runners['sess'], training_runners['coord']))
prefetch_threads.start()
# run one example so that we can calculate some statistics about the representations
filenames = []
loss_names_to_vals = {name: [] for name in loss_names}
start = time.perf_counter()
print_every = int(args.print_every)
# run the remaining examples
for step in range(max_steps):
data_idx, target, mask = training_runners['sess'].run([
inputs['data_idxs'], inputs['target_batch'],
inputs['mask_batch']
])
loss = loss_fn(avg_img, target, mask)
# print(loss)
assert np.isfinite(loss) and loss >= 0.0
loss_names_to_vals[loss_names[0]].append(loss)
filenames.extend(data_idx)
stats.push(loss)
if step % print_every == 0 and step > 0:
print(
'Step {0} of {1}: (Mean {5}: {2:.3f} || stddev: {3:.3f} :: ({4:.2f} secs/step)'
.format(step, max_steps - 1, stats.mean(),
np.sqrt(stats.variance()),
(time.perf_counter() - start) / print_every,
loss_names[0]))
start = time.perf_counter()
if training_runners['coord'].should_stop():
break
print(
'The size of losses is %s while we expect it to run for %d steps with batchsize %d'
% (len(filenames), inputs['max_steps'], cfg['batch_size']))
end_train_time = time.time() - start_time
if args.out_name:
out_name = args.out_name
else:
if args.data_split == "val":
split_name = "train"
if args.data_split == "test":
split_name = "val"
else:
raise ValueError(
"Cannot adequately name output for data split {}".
format(args.data_split))
out_name = '{avg_type}__{task}_{split}_losses.pkl'.format(
task=given_task,
split=split_name,
avg_type="marginal"
if args.avg_type == 'dense_marginal' else args.avg_type)
save_path = os.path.join(save_dir, out_name)
with open(save_path, 'wb') as f:
loss_names_to_vals['file_indexes'] = filenames
loss_names_to_vals['global_step'] = 0
if 'dense_xentropy_loss' in loss_names_to_vals:
loss_names_to_vals['xentropy_loss'] = loss_names_to_vals[
'dense_xentropy_loss']
del loss_names_to_vals['dense_xentropy_loss']
pickle.dump(loss_names_to_vals, f)
if args.out_dir:
os.makedirs(args.out_dir, exist_ok=True)
os.system("sudo mv {fp} {out}/".format(fp=save_path,
out=args.out_dir))
else:
if transfer:
os.makedirs('/home/ubuntu/s3/model_log/losses_transfer/',
exist_ok=True)
os.system(
"sudo mv {fp} /home/ubuntu/s3/model_log/losses_transfer/"
.format(fp=save_path))
else:
os.makedirs('/home/ubuntu/s3/model_log/losses/',
exist_ok=True)
os.system("sudo mv {fp} /home/ubuntu/s3/model_log/losses/".
format(fp=save_path))
print('saved losses to {0}'.format(save_path))
print('time to extract %d epochs: %.3f hrs' %
(cfg['num_epochs'], end_train_time / (60 * 60)))
finally:
utils.request_data_loading_end(training_runners)
utils.end_data_loading_and_sess(training_runners)
def run_extract_representations( args, cfg ):
# set up logging
tf.logging.set_verbosity( tf.logging.INFO )
with tf.Graph().as_default() as g:
cfg['randomize'] = False
cfg['num_epochs'] = 1
# cfg['num_read_threads'] = 5
# cfg['batch_size']=2
#if cfg['model_path'] is None:
# cfg['model_path'] = tf.train.latest_checkpoint( os.path.join( args.cfg_dir, "logs/slim-train/" ) )
cfg['model_path'] = os.path.join( args.cfg_dir, "logs/slim-train/model.ckpt-59690")
# create ops and placeholders
tf.logging.set_verbosity( tf.logging.INFO )
inputs = utils.setup_input( cfg, is_training=False, use_filename_queue=True )
RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
RuntimeDeterminedEnviromentVars.populate_registered_variables()
# build model (and losses and train_op)
model = utils.setup_model( inputs, cfg, is_training=False )
# set up metrics to evaluate
names_to_values, names_to_updates = setup_metrics( inputs, model, cfg )
# execute training
start_time = time.time()
utils.print_start_info( cfg, inputs[ 'max_steps' ], is_training=False )
# start session and restore model
training_runners = { 'sess': tf.Session(), 'coord': tf.train.Coordinator() }
try:
if cfg['model_path'] is None:
print('Please specify a checkpoint directory')
return
model[ 'saver_op' ].restore( training_runners[ 'sess' ], cfg[ 'model_path' ] )
utils.print_start_info( cfg, inputs[ 'max_steps' ], is_training=False )
data_prefetch_init_fn = utils.get_data_prefetch_threads_init_fn( inputs, cfg, is_training=False, use_filename_queue=True )
training_runners[ 'threads' ] = data_prefetch_init_fn( training_runners[ 'sess' ], training_runners[ 'coord' ] )
# run one example so that we can calculate some statistics about the representations
filenames = []
representations, data_idx = training_runners['sess'].run( [
model['model'].encoder_output, inputs[ 'data_idxs' ] ] )
filenames += [ inputs[ 'filepaths_list'][ i ] for i in data_idx ]
print( 'Got first batch representation with size: {0}'.format( representations.shape ) )
# run the remaining examples
for step in xrange( inputs[ 'max_steps' ] - 1 ):
if step % 100 == 0:
print( 'Step {0} of {1}'.format( step, inputs[ 'max_steps' ] - 1 ))
encoder_output, data_idx = training_runners['sess'].run( [
model['model'].encoder_output, inputs[ 'data_idxs' ] ] )
representations = np.append(representations, encoder_output, axis=0)
filenames += [ inputs[ 'filepaths_list'][ i ] for i in data_idx ]
if training_runners['coord'].should_stop():
break
print('The size of representations is %s while we expect it to run for %d steps with batchsize %d' % (representations.shape, inputs['max_steps'], cfg['batch_size']))
end_train_time = time.time() - start_time
save_path = os.path.join( args.cfg_dir, '../representations.pkl' )
with open( save_path, 'wb' ) as f:
pickle.dump( { 'filenames': filenames, 'representations': representations }, f )
print( 'saved representations to {0}'.format( save_path ))
print('time to train %d epochs: %.3f hrs' % (cfg['num_epochs'], end_train_time/(60*60)))
print('avg time per epoch: %.3f hrs' % ( (end_train_time/(60*60)) / cfg['num_epochs']) )
finally:
utils.request_data_loading_end( training_runners )
utils.end_data_loading_and_sess( training_runners )
def run_extract_representations( args, cfg, file_to_process):
setup_input_fn = utils.setup_input
# set up logging
tf.logging.set_verbosity( tf.logging.INFO )
with tf.Graph().as_default() as g:
# create ops and placeholders
tf.logging.set_verbosity( tf.logging.INFO )
inputs = {}
inputs['input_batch'] = tf.placeholder( tf.float32, shape=[1,224,224,3], name='input_placeholder')
inputs['target_batch'] = tf.placeholder( tf.float32, shape=[1,1000], name='target_placeholder' )
inputs['mask_batch'] = tf.placeholder( tf.float32, shape=[1], name='mask_placeholder' )
inputs['data_idxs'] = tf.placeholder( tf.int32, shape=[1], name='data_idx_placeholder')
inputs['num_samples_epoch'] = len(file_to_process)
inputs['max_steps'] = len(file_to_process)
RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
RuntimeDeterminedEnviromentVars.populate_registered_variables()
# build model (and losses and train_op)
model = utils.setup_model( inputs, cfg, is_training=False )
m = model['model']
# execute training
utils.print_start_info( cfg, inputs[ 'max_steps' ], is_training=False )
# start session and restore model
training_runners = { 'sess': tf.Session(), 'coord': tf.train.Coordinator() }
try:
if cfg['model_path'] is None:
print('Please specify a checkpoint directory')
return
to_restore = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
for v in tuple(to_restore):
if 'global_step' in v.name:
to_restore.remove(v)
saver_for_kd = tf.train.Saver(to_restore)
saver_for_kd.restore( training_runners[ 'sess' ], cfg[ 'model_path' ] )
#model[ 'saver_op' ].restore( training_runners[ 'sess' ], cfg[ 'model_path' ] )
for step, filename in enumerate(file_to_process):
start_time = time.time()
if step % 100 == 0:
print( 'Step {0} of {1}'.format( step, inputs[ 'max_steps' ] - 1 ))
m,p,v = filename.decode('UTF-8').split('/')
print(filename)
img_name = '/home/ubuntu/s3/{}/rgb/point_{}_view_{}_domain_rgb.png'.format(m, p, v)
sfm_dir = 's3://taskonomy-unpacked-oregon/{}/softmax_1000'.format(m)
os.system('sudo mkdir -p /home/ubuntu/s3/{}/softmax_1000/'.format(m))
os.system('mkdir -p /home/ubuntu/temp/{}/'.format(m))
npy_name = 'point_{}_view_{}.npy'.format(p, v)
if os.path.isfile('/home/ubuntu/s3/{}/softmax_1000/{}'.format(m, npy_name)):
continue
if not os.path.isfile(img_name):
continue
img = skimage.io.imread(img_name, as_grey=False)
img = resize_rescale_imagenet(img, new_dims=(224,224))
img = np.reshape(img, (1,224,224,3))
feed_dict = {inputs['input_batch'] : img}
predicted = training_runners['sess'].run( model['model'].encoder_output, feed_dict=feed_dict )
# maxs = np.amax(predicted, axis=-1)
# softmax = np.exp(predicted - np.expand_dims(maxs, axis=-1))
# sums = np.sum(softmax, axis=-1)
# softmax = softmax / np.expand_dims(sums, -1)
# print(softmax)
# pdb.set_trace()
local_npy = os.path.join('/home/ubuntu/temp/{}'.format(m), npy_name)
with open(local_npy, 'wb') as fp:
np.save(fp, predicted)
os.system('aws s3 mv {} {}/'.format(local_npy, sfm_dir))
if training_runners['coord'].should_stop():
break
end_train_time = time.time() - start_time
print('time to extract %.3f ' % (end_train_time))
finally:
utils.request_data_loading_end( training_runners )
utils.end_data_loading_and_sess( training_runners )
def run_extract_losses(args, cfg, save_dir, given_task):
transfer = (cfg['model_type'] == architectures.TransferNet)
if transfer:
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn_transfer
setup_input_fn = utils.setup_input_transfer
else:
setup_input_fn = utils.setup_input
get_data_prefetch_threads_init_fn = utils.get_data_prefetch_threads_init_fn
# set up logging
tf.logging.set_verbosity(tf.logging.ERROR)
stats = Statistics()
print_every = int(args.print_every)
with tf.Graph().as_default() as g:
# create ops and placeholders
inputs = setup_input_fn(cfg,
is_training=False,
use_filename_queue=False)
#RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
#RuntimeDeterminedEnviromentVars.populate_registered_variables()
max_steps = get_max_steps(inputs['max_steps'], args.data_split)
# pdb.set_trace()
# build model (and losses and train_op)
model = utils.setup_model(inputs, cfg, is_training=False)
# set up metrics to evaluate
names_to_values, names_to_updates = setup_metrics(inputs, model, cfg)
# execute training
start_time = time.time()
utils.print_start_info(cfg, max_steps, is_training=False)
# start session and restore model
training_runners = {
'sess': tf.Session(),
'coord': tf.train.Coordinator()
}
try:
if cfg['model_path'] is None:
print('Please specify a checkpoint directory')
return
print('Attention, model_path is ', cfg['model_path'])
if given_task == 'rgb2depth_sota':
non_encoder_var = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES)
for v in tuple(non_encoder_var):
if 'global_step' in v.name:
non_encoder_var.remove(v)
if 'Adam' in v.name:
non_encoder_var.remove(v)
saver_sota = tf.train.Saver(non_encoder_var)
saver_sota.restore(training_runners['sess'], cfg['model_path'])
else:
model['saver_op'].restore(training_runners['sess'],
cfg['model_path'])
# var = [v for v in tf.global_variables() if 'decoder' in v.name][0]
# print(training_runners[ 'sess' ].run(var))
utils.print_start_info(cfg, max_steps, is_training=False)
data_prefetch_init_fn = get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=False, use_filename_queue=False)
prefetch_threads = threading.Thread(
target=data_prefetch_init_fn,
args=(training_runners['sess'], training_runners['coord']))
prefetch_threads.start()
m = model['model']
# run one example so that we can calculate some statistics about the representations
filenames = []
losses = []
results = training_runners['sess'].run([
inputs['data_idxs'], m.target_images, m.masks, m.decoder_output
])
data_idx = results[0]
target = results[1]
mask = results[2]
predicted = results[3]
import pdb
#pdb.set_trace()
normalized_loss = get_norm_loss(target, mask, predicted)
filenames.extend(data_idx)
losses.append(normalized_loss)
print("Step number: {}".format(0), (data_idx))
# print(target_input, target_input.sum())
# return
# training_runners['sess'].run([v for v in tf.global_variables() if "transfer/rep_conv_1/weights" in v.name][0])
# run the remaining examples
start = time.perf_counter()
max_steps = 3000
for step in range(max_steps - 1):
results = training_runners['sess'].run([
inputs['data_idxs'], m.target_images, m.masks,
m.decoder_output
])
data_idx = results[0]
target = results[1]
mask = results[2]
predicted = results[3]
# p, t, m = results[1], results[2], results[3]
# losses = results[4:]
normalized_loss = get_norm_loss(target, mask, predicted)
# print(p.mean(), t)
filenames.extend(data_idx)
losses.append(normalized_loss)
stats.push(normalized_loss)
if step % print_every == 0 and step > 0:
print(
'Step {0} of {1}: ({5} loss: {2:.3f} || stddev: {3:.3f} :: ({4:.2f} secs/step)'
.format(
step,
max_steps - 1,
stats.mean(),
np.sqrt(stats.variance()),
# stats.variance(),
(time.perf_counter() - start) / print_every,
'norm_depth'))
start = time.perf_counter()
if training_runners['coord'].should_stop():
break
print(
'The size of losses is %s while we expect it to run for %d steps with batchsize %d'
% (len(filenames), inputs['max_steps'], cfg['batch_size']))
end_train_time = time.time() - start_time
if args.out_name:
out_name = args.out_name
else:
out_name = '{task}_{split}_losses_normed.pkl'.format(
task=given_task, split=args.data_split)
save_path = os.path.join(save_dir, out_name)
loss_names_to_vals = {}
with open(save_path, 'wb') as f:
loss_names_to_vals['file_indexes'] = filenames
loss_names_to_vals['global_step'] = 0
loss_names_to_vals['l1_loss'] = losses
pickle.dump(loss_names_to_vals, f)
if args.out_dir:
os.makedirs(args.out_dir, exist_ok=True)
os.system("sudo cp {fp} {out}/".format(fp=save_path,
out=args.out_dir))
else:
if transfer:
copy_to = cfg['log_root']
else:
copy_to = os.path.join(cfg['log_root'], given_task)
os.system("sudo mv {fp} {dst}/".format(fp=save_path,
dst=copy_to))
print("sudo mv {fp} {dst}/".format(fp=save_path, dst=copy_to))
print('saved losses to {0}'.format(save_path))
print('time to extract %d epochs: %.3f hrs' %
(cfg['num_epochs'], end_train_time / (60 * 60)))
finally:
utils.request_data_loading_end(training_runners)
utils.end_data_loading_and_sess(training_runners)