def test_single_dialog_existing_file_one_update(user_input,
prepare_config_file,
monkeypatch):
"""This test checks that key update in the config works:
* User input:
* 'n' -> no update
* 'y' -> update
* '' -> update
* If update happens - key is really updated
"""
config, config_text = prepare_config_file
new_value = "new_value"
test_input = [
"Y", # overwrite file
user_input, # whether the value should be updated
new_value, # new value for the parameter
"Y", # save the file
]
setup_input(monkeypatch, test_input)
assert config_dialog(config, ["update_node"])
if user_input in ["Y", ""]:
assert (parse(config.read_text())["update_node"] == new_value
), "Existing value was not updated"
else:
assert config.read_text() == config_text
def test_single_dialog_new_config(tmp_path, monkeypatch, save_agree):
"""Tests single run of config_dialog by populating a blank file"""
path: Path = tmp_path / "config.toml"
# File will be created as a result of the test
file_created = save_agree or save_agree is None
if save_agree:
save_agree = "Y"
elif save_agree is None:
save_agree = ""
else:
save_agree = "N"
# The monkey patched test cli_input. Should be the same length as the list of params fed to config_dialog later
# Note: click provides validations for confirmation prompts, no need to test for garbage cli_input
test_input = [
"author name", "https://confluence.local", "page title", save_agree
]
setup_input(monkeypatch, test_input)
# In this scenario file should be created when and only when the function returned true
assert (config_dialog(
Path(path),
["author", "auth.url", "pages.page1.page_title"]) == path.exists())
if file_created:
assert (path.read_text() == """author = "author name"
[auth]
url = "https://confluence.local"
[pages]
[pages.page1]
page_title = "page title"
""")
def test_incremental_config_dialog(prepare_config_file, monkeypatch, capsys):
config, config_text = prepare_config_file
new_value = "new_value"
setup_input(monkeypatch, [new_value, "Y"])
assert config_dialog(config, attributes=["new_path"], incremental=True)
captured = capsys.readouterr()
assert ("Current content" not in captured.out
), "incremental should suppress printing file altogether"
with pytest.raises(Exception):
config_dialog(config, attributes=["update_node"], incremental=True)
def test_sensitive_parameter_file_mode(tmp_path, monkeypatch, capsys):
"""Checks that if there is a parameter - the wizard will create file with 600 permissions and alert the user"""
monkeypatch.setattr("getpass.getpass", lambda x: "password")
setup_input(monkeypatch, ("password", "Y"))
config_path: Path = tmp_path / "config.toml"
assert config_dialog(
config_path,
attributes=[DialogParameter("hidden param", hide_input=True)])
captured = capsys.readouterr()
assert "sensitive parameter was passed" in captured.out
assert config_path.stat().st_mode == 33152
def test_dialog_converts_filename_to_path(tmp_path, monkeypatch):
"""Makes sure the dialog accepts both Path and strings for config file"""
path_as_path: Path = tmp_path / "config_path.toml"
path_as_string: str = str(tmp_path / "config_string.toml")
# In this scenario file should be created when and only when the function returned true
for tested_type in [path_as_path, path_as_string]:
# Taken from previous test
test_input = [
"author name", "https://confluence.local", "page title", "Y"
]
setup_input(monkeypatch, test_input)
assert config_dialog(tested_type,
["author", "auth.url", "pages.page1.page_title"])
assert Path(tested_type).exists()
def test_prompt_function_confirm_function(monkeypatch, filter_mode):
s = StateConfig()
if filter_mode is not None:
s.filter_mode = filter_mode
else:
assert not s.filter_mode
setup_input(monkeypatch, cli_input=("Y", "Y"))
if s.filter_mode:
with pytest.raises(Exception):
s.prompt_function("Prompt?")
with pytest.raises(Exception):
s.confirm_function("Prompt?")
else:
_ = s.prompt_function("Prompt?")
assert _ == "Y"
_ = s.confirm_function("Prompt?")
assert _
def test_single_dialog_existing_file_base(mode, user_agrees_to_overwrite_file,
prepare_config_file, monkeypatch):
config, config_text = prepare_config_file
new_value = "new_value"
if user_agrees_to_overwrite_file:
user_input = ["Y", new_value, "Y"]
else:
user_input = ["N"]
setup_input(monkeypatch, user_input)
if user_agrees_to_overwrite_file:
node_path = "new_node"
assert config_dialog(config, [node_path])
assert (parse(config.read_text())[node_path] == new_value
), "Value was not set to the new one"
else:
assert config_dialog(config, ["update"]) is None
assert config.read_text() == config_text
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 test_single_dialog_existing_file_multiple_updates(user_updates_values,
prepare_config_file,
monkeypatch):
"""In a scenario where there are more than 1 keys to update - make sure that permutations of user input are
handled correctly"""
new_value = "new_value"
update_attrs = ["update_node", "parent.parent_update_node"]
answer_1, answer_2 = user_updates_values
config, config_text = prepare_config_file
test_input = ["Y"]
for user_answer in [answer_1, answer_2]:
if user_answer:
test_input += ["Y", new_value]
else:
test_input += ["N"]
test_input += ["Y"]
setup_input(monkeypatch, test_input)
assert config_dialog(config, update_attrs)
for user_answer, attr in zip([answer_1, answer_2], update_attrs):
# The value should be updated <=> user said yes
assert user_answer == (get_attribute_by_path(
attribute_path=attr,
config=parse(config.read_text())) == new_value)
def run_to_task(task_to):
import general_utils
from general_utils import RuntimeDeterminedEnviromentVars
import models.architectures as architectures
from data.load_ops import resize_rescale_image
import utils
from data.task_data_loading import load_and_specify_preprocessors_for_representation_extraction
import lib.data.load_ops as load_ops
import pdb
global synset
synset_1000 = [" ".join(i.split(" ")[1:]) for i in synset]
select = np.asarray([ 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 1.,
1., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0.,
0., 0., 0., 0., 1., 0., 1., 0., 0., 0., 0., 0., 1.,
1., 0., 0., 0., 1., 0., 0., 0., 0., 1., 0., 1., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 0.,
0., 0., 1., 0., 1., 0., 0., 0., 0., 1., 0., 1., 0.,
0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 1., 0., 0., 1., 0., 1., 0., 0., 1.,
0., 1., 0., 1., 0., 1., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
1., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 1., 1., 0., 1., 0., 0.,
1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0.,
1., 0., 0., 0., 0., 0., 0., 1., 0., 1., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 1., 1., 0., 0., 1., 0., 1.,
0., 1., 0., 0., 0., 0., 1., 0., 1., 0., 0., 0., 0.,
0., 0., 0., 0., 1., 0., 0., 0., 0., 1., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 1., 1., 1., 0., 0., 1.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1.,
0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 1.,
0., 0., 0., 0., 0., 1., 1., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 1., 0., 0., 0., 1., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 1., 0., 0., 1., 0., 0., 0., 0.,
0., 1., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 1., 0.])
with open('/home/ubuntu/task-taxonomy-331b/lib/data/places_class_names.txt', 'r') as fp:
synset_places = [x.rstrip()[4:-1] for x,y in zip(fp.readlines(), select) if y == 1.]
tf.logging.set_verbosity(tf.logging.ERROR)
args = parser.parse_args()
if args.task is not 'NONE':
args.idx = list_of_tasks.index(args.task)
for idx, task in enumerate(list_of_tasks):
if idx != args.idx and args.idx != -1:
continue
if task == 'class_places':
synset = synset_places
elif task == 'class_1000':
synset = synset_1000
print("Doing {task}".format(task=task))
general_utils = importlib.reload(general_utils)
tf.reset_default_graph()
training_runners = { 'sess': tf.InteractiveSession(), 'coord': tf.train.Coordinator() }
# task = '{f}__{t}__{hs}'.format(f=task_from, t=task_to, hs=args.hs)
CONFIG_DIR = '/home/ubuntu/task-taxonomy-331b/experiments/final/{TASK}'.format(TASK=task)
############## Load Configs ##############
cfg = utils.load_config( CONFIG_DIR, nopause=True )
RuntimeDeterminedEnviromentVars.register_dict( cfg )
split_file = os.path.join('/home/ubuntu/task-taxonomy-331b/assets/aws_data/', 'video2_info.pkl')
cfg['train_filenames'] = split_file
cfg['val_filenames'] = split_file
cfg['test_filenames'] = split_file
cfg['num_epochs'] = 2
cfg['randomize'] = False
root_dir = cfg['root_dir']
cfg['num_read_threads'] = 1
print(cfg['log_root'])
cfg['model_path'] = os.path.join(
cfg['log_root'],
task,
'model.permanent-ckpt'
)
print( cfg['model_path'])
if cfg['model_path'] is None:
continue
cfg['dataset_dir'] = '/home/ubuntu'
cfg['preprocess_fn'] = load_and_specify_preprocessors_for_representation_extraction
############## Set Up Inputs ##############
# tf.logging.set_verbosity( tf.logging.INFO )
inputs = utils.setup_input( cfg, is_training=ON_TEST_SET, use_filename_queue=False ) # is_training determines whether to use train/validaiton
RuntimeDeterminedEnviromentVars.load_dynamic_variables( inputs, cfg )
RuntimeDeterminedEnviromentVars.populate_registered_variables()
start_time = time.time()
# utils.print_start_info( cfg, inputs[ 'max_steps' ], is_training=False )
############## Set Up Model ##############
model = utils.setup_model( inputs, cfg, is_training=IN_TRAIN_MODE )
m = model[ 'model' ]
model[ 'saver_op' ].restore( training_runners[ 'sess' ], cfg[ 'model_path' ] )
############## Start dataloading workers ##############
data_prefetch_init_fn = utils.get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=ON_TEST_SET, use_filename_queue=False )
prefetch_threads = threading.Thread(
target=data_prefetch_init_fn,
args=( training_runners[ 'sess' ], training_runners[ 'coord' ] ))
prefetch_threads.start()
list_of_fname = np.load('/home/ubuntu/task-taxonomy-331b/assets/aws_data/video2_fname.npy')
import errno
try:
os.mkdir('/home/ubuntu/{}'.format(task))
os.mkdir('/home/ubuntu/{}/vid1'.format(task))
os.mkdir('/home/ubuntu/{}/vid2'.format(task))
os.mkdir('/home/ubuntu/{}/vid3'.format(task))
os.mkdir('/home/ubuntu/{}/vid4'.format(task))
except OSError as e:
if e.errno != errno.EEXIST:
raise
curr_comp = np.zeros((3,64))
curr_fit_img = np.zeros((256,256,3))
embeddings = []
############## Run First Batch ##############
for step_num in range(inputs['max_steps'] - 1):
#for step_num in range(1):
#if step_num > 0 and step_num % 20 == 0:
print(step_num)
if not hasattr(m, 'masks'):
(
input_batch, target_batch,
data_idx,
predicted, loss,
) = training_runners['sess'].run( [
m.input_images, m.targets,
model[ 'data_idxs' ],
m.decoder_output, m.total_loss] )
mask_batch = 1.
else:
(
input_batch, target_batch, mask_batch,
data_idx,
predicted, loss,
) = training_runners['sess'].run( [
m.input_images, m.targets, m.masks,
model[ 'data_idxs' ],
m.decoder_output, m.total_loss] )
if task == 'segment2d' or task == 'segment25d':
from sklearn.decomposition import PCA
x = np.zeros((32,256,256,3), dtype='float')
k_embed = 8
# for i in range(predicted.shape[0]):
# embedding_flattened = np.squeeze(predicted[i]).reshape((-1,64))
# pca = PCA(n_components=3)
# pca.fit(embedding_flattened)
# min_order = None
# min_dist = float('inf')
# for order in itertools.permutations([0,1,2]):
# reordered = pca.components_[list(order), :]
# dist = np.linalg.norm(curr_comp-reordered)
# if dist < min_dist:
# min_order = list(order)
# min_dist = dist
# print(min_order)
# pca.components_ = pca.components_[min_order, :]
# curr_comp = pca.components_
# lower_dim = pca.transform(embedding_flattened).reshape((256,256,-1))
# lower_dim = (lower_dim - lower_dim.min()) / (lower_dim.max() - lower_dim.min())
# x[i] = lower_dim
for i in range(predicted.shape[0]):
embedding_flattened = np.squeeze(predicted[i]).reshape((-1,64))
embeddings.append(embedding_flattened)
if len(embeddings) > k_embed:
embeddings.pop(0)
pca = PCA(n_components=3)
pca.fit(np.vstack(embeddings))
min_order = None
min_dist = float('inf')
copy_of_comp = np.copy(pca.components_)
for order in itertools.permutations([0,1,2]):
#reordered = pca.components_[list(order), :]
#dist = np.linalg.norm(curr_comp-reordered)
pca.components_ = copy_of_comp[order, :]
lower_dim = pca.transform(embedding_flattened).reshape((256,256,-1))
lower_dim = (lower_dim - lower_dim.min()) / (lower_dim.max() - lower_dim.min())
dist = np.linalg.norm(lower_dim - curr_fit_img)
if dist < min_dist:
min_order = order
min_dist = dist
pca.components_ = copy_of_comp[min_order, :]
lower_dim = pca.transform(embedding_flattened).reshape((256,256,-1))
lower_dim = (lower_dim - lower_dim.min()) / (lower_dim.max() - lower_dim.min())
curr_fit_img = np.copy(lower_dim)
x[i] = lower_dim
predicted = x
if task == 'curvature':
std = [31.922, 21.658]
mean = [123.572, 120.1]
predicted = (predicted * std) + mean
predicted[:,0,0,:] = 0.
predicted[:,1,0,:] = 1.
predicted = np.squeeze(np.clip(predicted.astype(int) / 255., 0., 1. )[:,:,:,0])
just_rescale = ['autoencoder', 'denoise', 'edge2d',
'edge3d', 'keypoint2d', 'keypoint3d',
'reshade', 'rgb2sfnorm']
if task in just_rescale:
predicted = (predicted + 1.) / 2.
predicted = np.clip(predicted, 0., 1.)
predicted[:,0,0,:] = 0.
predicted[:,1,0,:] = 1.
just_clip = ['rgb2depth', 'rgb2mist']
if task in just_clip:
predicted[:,0,0,:] = 0.
predicted[:,1,0,:] = 1.
if task == 'segmentsemantic_rb':
label = np.argmax(predicted, axis=-1)
COLORS = ('white','red', 'blue', 'yellow', 'magenta',
'green', 'indigo', 'darkorange', 'cyan', 'pink',
'yellowgreen', 'black', 'darkgreen', 'brown', 'gray',
'purple', 'darkviolet')
rgb = (input_batch + 1.) / 2.
preds = [color.label2rgb(np.squeeze(x), np.squeeze(y), colors=COLORS, kind='overlay')[np.newaxis,:,:,:] for x,y in zip(label, rgb)]
predicted = np.vstack(preds)
if task in ['class_1000', 'class_places']:
for file_idx, predict_output in zip(data_idx, predicted):
to_store_name = list_of_fname[file_idx].decode('utf-8').replace('video', task)
to_store_name = os.path.join('/home/ubuntu', to_store_name)
sorted_pred = np.argsort(predict_output)[::-1]
top_5_pred = [synset[sorted_pred[i]] for i in range(5)]
to_print_pred = "Top 5 prediction: \n {}\n {}\n {}\n {} \n {}".format(*top_5_pred)
img = Image.new('RGBA', (400, 200), (255, 255, 255))
d = ImageDraw.Draw(img)
fnt = ImageFont.truetype('/usr/share/fonts/truetype/dejavu/DejaVuSerifCondensed.ttf', 25)
d.text((20, 5), to_print_pred, fill=(255, 0, 0), font=fnt)
img.save(to_store_name, 'PNG')
else:
for file_idx, predict_output in zip(data_idx, predicted):
to_store_name = list_of_fname[file_idx].decode('utf-8').replace('video', task)
to_store_name = os.path.join('/home/ubuntu', to_store_name)
scipy.misc.toimage(np.squeeze(predict_output), cmin=0.0, cmax=1.0).save(to_store_name)
subprocess.call('tar -czvf /home/ubuntu/{t}.tar.gz /home/ubuntu/{t}'.format(t=task), shell=True)
subprocess.call('aws s3 cp /home/ubuntu/{t}.tar.gz s3://task-preprocessing-512-oregon/video2/'.format(t=task), shell=True)
subprocess.call('ffmpeg -r 29.97 -f image2 -s 256x256 -i /home/ubuntu/{t}/vid2/020%04d.png -vcodec libx264 -crf 15 -pix_fmt yuv420p {t}_2.mp4'.format(t=task), shell=True)
subprocess.call('aws s3 cp {t}_2.mp4 s3://task-preprocessing-512-oregon/video2/'.format(t=task), shell=True)
############## Clean Up ##############
training_runners[ 'coord' ].request_stop()
training_runners[ 'coord' ].join()
# if os.path.isfile(pickle_dir):
# with open(pickle_dir, 'rb') as fp:
# all_outputs = pickle.load(fp)
############## Store to dict ##############
print("Done: {}".format(task))
# os.system("sudo cp {d} /home/ubuntu/s3/model_log".format(d=pickle_dir))
############## Reset graph and paths ##############
tf.reset_default_graph()
training_runners['sess'].close()
return
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, 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_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_val_test(cfg):
# set up logging
tf.logging.set_verbosity(tf.logging.INFO)
tf.reset_default_graph()
training_runners = {
'sess': tf.InteractiveSession(),
'coord': tf.train.Coordinator()
}
# create ops and placeholders
inputs = utils.setup_input(cfg, is_training=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)
# full_path = tf.train.latest_checkpoint(checkpoint_dir)
# step = full_path.split('-')[-1]
# model_path = os.path.join('/home/ubuntu/s3/model_log', cfg['task_name'], 'model.permanent-ckpt')
model_path = os.path.join('/home/ubuntu/s3/model_log_final',
cfg['task_name'], 'model.permanent-ckpt')
model['saver_op'].restore(training_runners['sess'], model_path)
m = model['model']
# execute training
start_time = time.time()
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)
prefetch_threads = threading.Thread(target=data_prefetch_init_fn,
args=(training_runners['sess'],
training_runners['coord']))
prefetch_threads.start()
print("Dataloading workers dispatched....")
return_accuracy = 'return_accuracy' in cfg and cfg['return_accuracy'],
losses_mean = AverageMeter()
accuracy_mean = AverageMeter()
for step in range(inputs['max_steps']):
#print(step)
if return_accuracy:
(data_idx, loss, accuracy) = training_runners['sess'].run(
[model['data_idxs'], m.losses[0], m.accuracy])
losses_mean.update(loss)
accuracy_mean.update(accuracy)
if step % 100 == 0:
print(
'Step: {step} with Current Losses mean: {loss}; with accuracy: {accur}'
.format(step=step,
loss=losses_mean.avg,
accur=accuracy_mean.avg))
else:
(data_idx, loss) = training_runners['sess'].run(
[model['data_idxs'], m.losses[0]])
losses_mean.update(loss)
if step % 100 == 0:
print('Step: {step} with Current Losses mean: {loss}'.format(
step=step, loss=losses_mean.avg))
if return_accuracy:
print('Final Losses mean: {loss}; with accuracy: {accur}'.format(
loss=losses_mean.avg, accur=accuracy_mean.avg))
else:
print('Final Losses mean: {loss}'.format(loss=losses_mean.avg))
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(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
# When ready to use a model, use the code below
train(
model['train_op'],
train_log_dir,
utils.get_data_prefetch_threads_init_fn(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=inputs['max_steps'] //
(cfg['num_epochs'] * 2),
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_to_task(task_to):
import general_utils
from general_utils import RuntimeDeterminedEnviromentVars
import models.architectures as architectures
from data.load_ops import resize_rescale_image
import utils
from data.task_data_loading import load_and_specify_preprocessors_for_representation_extraction
import lib.data.load_ops as load_ops
tf.logging.set_verbosity(tf.logging.ERROR)
all_outputs = {}
pickle_dir = 'viz_output_single_task.pkl'
import os
if os.path.isfile(pickle_dir):
with open(pickle_dir, 'rb') as fp:
all_outputs = pickle.load(fp)
for task in list_of_tasks:
if task in all_outputs:
print("{} already exists....\n\n\n".format(task))
continue
print("Doing {task}".format(task=task))
general_utils = importlib.reload(general_utils)
tf.reset_default_graph()
training_runners = {
'sess': tf.InteractiveSession(),
'coord': tf.train.Coordinator()
}
# task = '{f}__{t}__{hs}'.format(f=task_from, t=task_to, hs=args.hs)
CONFIG_DIR = '/home/ubuntu/task-taxonomy-331b/experiments/final/{TASK}'.format(
TASK=task)
############## Load Configs ##############
cfg = utils.load_config(CONFIG_DIR, nopause=True)
RuntimeDeterminedEnviromentVars.register_dict(cfg)
split_file = cfg['test_filenames'] if ON_TEST_SET else cfg[
'val_filenames']
cfg['train_filenames'] = split_file
cfg['val_filenames'] = split_file
cfg['test_filenames'] = split_file
cfg['num_epochs'] = 1
cfg['randomize'] = False
root_dir = cfg['root_dir']
cfg['num_read_threads'] = 1
print(cfg['log_root'])
if task == 'jigsaw':
continue
cfg['model_path'] = os.path.join(cfg['log_root'], task,
'model.permanent-ckpt')
print(cfg['model_path'])
if cfg['model_path'] is None:
continue
############## Set Up Inputs ##############
# tf.logging.set_verbosity( tf.logging.INFO )
inputs = utils.setup_input(
cfg, is_training=ON_TEST_SET, use_filename_queue=False
) # is_training determines whether to use train/validaiton
RuntimeDeterminedEnviromentVars.load_dynamic_variables(inputs, cfg)
RuntimeDeterminedEnviromentVars.populate_registered_variables()
start_time = time.time()
# utils.print_start_info( cfg, inputs[ 'max_steps' ], is_training=False )
############## Set Up Model ##############
model = utils.setup_model(inputs, cfg, is_training=IN_TRAIN_MODE)
m = model['model']
model['saver_op'].restore(training_runners['sess'], cfg['model_path'])
############## Start dataloading workers ##############
data_prefetch_init_fn = utils.get_data_prefetch_threads_init_fn(
inputs, cfg, is_training=ON_TEST_SET, use_filename_queue=False)
prefetch_threads = threading.Thread(target=data_prefetch_init_fn,
args=(training_runners['sess'],
training_runners['coord']))
prefetch_threads.start()
############## Run First Batch ##############
if not hasattr(m, 'masks'):
(
input_batch,
target_batch,
data_idx,
predicted,
loss,
) = training_runners['sess'].run([
m.input_images, m.targets, model['data_idxs'],
m.decoder_output, m.total_loss
])
mask_batch = 1.
else:
(
input_batch,
target_batch,
mask_batch,
data_idx,
predicted,
loss,
) = training_runners['sess'].run([
m.input_images, m.targets, m.masks, model['data_idxs'],
m.decoder_output, m.total_loss
])
if task == 'segment2d' or task == 'segment25d':
from sklearn.decomposition import PCA
x = np.zeros((32, 256, 256, 3), dtype='float')
for i in range(predicted.shape[0]):
embedding_flattened = np.squeeze(predicted[i]).reshape(
(-1, 64))
pca = PCA(n_components=3)
pca.fit(embedding_flattened)
lower_dim = pca.transform(embedding_flattened).reshape(
(256, 256, -1))
lower_dim = (lower_dim - lower_dim.min()) / (lower_dim.max() -
lower_dim.min())
x[i] = lower_dim
predicted = x
############## Clean Up ##############
training_runners['coord'].request_stop()
training_runners['coord'].join()
# if os.path.isfile(pickle_dir):
# with open(pickle_dir, 'rb') as fp:
# all_outputs = pickle.load(fp)
############## Store to dict ##############
to_store = {
'input': input_batch,
'target': target_batch,
'mask': mask_batch,
'data_idx': data_idx,
'output': predicted
}
all_outputs[task] = to_store
print("Done: {}".format(task))
# os.system("sudo cp {d} /home/ubuntu/s3/model_log".format(d=pickle_dir))
############## Reset graph and paths ##############
tf.reset_default_graph()
training_runners['sess'].close()
try:
del sys.modules['config']
except:
pass
sys.path = remove_dups(sys.path)
print("FINISHED: {}\n\n\n\n\n\n".format(task))
pickle_dir = 'viz_output_single_task.pkl'
with open(pickle_dir, 'wb') as fp:
pickle.dump(all_outputs, fp)
try:
subprocess.call(
"aws s3 cp {} s3://task-preprocessing-512-oregon/visualizations/"
.format(pickle_dir),
shell=True)
except:
subprocess.call(
"sudo cp {} /home/ubuntu/s3/visualizations/".format(
pickle_dir),
shell=True)
return
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 test_page_add_dialog(prepare_config_file, monkeypatch):
config, config_text = prepare_config_file
setup_input(monkeypatch, ["title", "testfile", "", "SPC", "Y"])
assert page_add_dialog(config)
assert (get_attribute_by_path("pages.page1.page_title",
parse(config.read_text())) == "title")
