def validate(val_dirs: ValidationDirs, images_iterator: ImagesIterator,
flags: OutputFlags):
"""
Saves in val_dirs.log_dir/val/dataset_name/measures.csv:
- `img_name,bpp,psnr,ms-ssim forall img_name`
"""
print(_VALIDATION_INFO_STR)
validated_checkpoints = val_dirs.get_validated_checkpoints(
) # :: [10000, 18000, ..., 256000], ie, [int]
all_ckpts = Saver.all_ckpts_with_iterations(val_dirs.ckpt_dir)
if len(all_ckpts) == 0:
print('No checkpoints found in {}'.format(val_dirs.ckpt_dir))
return
# if ckpt_step is -1, then all_ckpt[:-1:flags.ckpt_step] === [] because of how strides work
ckpt_to_check = all_ckpts[:-1:flags.ckpt_step] + [
all_ckpts[-1]
] # every ckpt_step-th checkpoint plus the last one
if flags.ckpt_step == -1:
assert len(ckpt_to_check) == 1
print('Validating {}/{} checkpoints (--ckpt_step {})...'.format(
len(ckpt_to_check), len(all_ckpts), flags.ckpt_step))
missing_checkpoints = [(ckpt_itr, ckpt_path)
for ckpt_itr, ckpt_path in ckpt_to_check
if ckpt_itr not in validated_checkpoints]
if len(missing_checkpoints) == 0:
print('All checkpoints validated, stopping...')
return
# ---
# create networks
autoencoder_config_path, probclass_config_path = logdir_helpers.config_paths_from_log_dir(
val_dirs.log_dir,
base_dirs=[constants.CONFIG_BASE_AE, constants.CONFIG_BASE_PC])
ae_config, ae_config_rel_path = config_parser.parse(
autoencoder_config_path)
pc_config, pc_config_rel_path = config_parser.parse(probclass_config_path)
ae_cls = autoencoder.get_network_cls(ae_config)
pc_cls = probclass.get_network_cls(pc_config)
# Instantiate autoencoder and probability classifier
ae = ae_cls(ae_config)
pc = pc_cls(pc_config, num_centers=ae_config.num_centers)
x_val_ph = tf.placeholder(tf.uint8, (3, None, None), name='x_val_ph')
x_val_uint8 = tf.expand_dims(x_val_ph, 0, name='batch')
x_val = tf.to_float(x_val_uint8, name='x_val')
enc_out_val = ae.encode(x_val, is_training=False)
x_out_val = ae.decode(enc_out_val.qhard, is_training=False)
bc_val = pc.bitcost(enc_out_val.qbar,
enc_out_val.symbols,
is_training=False,
pad_value=pc.auto_pad_value(ae))
bpp_val = bits.bitcost_to_bpp(bc_val, x_val)
x_out_val_uint8 = tf.cast(x_out_val, tf.uint8, name='x_out_val_uint8')
# Using numpy implementation due to dynamic shapes
msssim_val = ms_ssim_np.tf_msssim_np(x_val_uint8,
x_out_val_uint8,
data_format='NCHW')
psnr_val = psnr_np(x_val_uint8, x_out_val_uint8)
restorer = Saver(val_dirs.ckpt_dir,
var_list=Saver.get_var_list_of_ckpt_dir(
val_dirs.ckpt_dir))
# create fetch_dict
fetch_dict = {
'bpp': bpp_val,
'ms-ssim': msssim_val,
'psnr': psnr_val,
}
if flags.real_bpp:
fetch_dict['sym'] = enc_out_val.symbols # NCHW
if flags.save_ours:
fetch_dict['img_out'] = x_out_val_uint8
# ---
fw = tf.summary.FileWriter(val_dirs.out_dir, graph=tf.get_default_graph())
def full_summary_tag(summary_name):
return '/'.join(['val', images_iterator.dataset_name, summary_name])
# Distance
try:
codec_distance_ms_ssim = CodecDistance(images_iterator.dataset_name,
codec='bpg',
metric='ms-ssim')
codec_distance_psnr = CodecDistance(images_iterator.dataset_name,
codec='bpg',
metric='psnr')
except CodecDistanceReadException as e: # no codec distance values stored for the current setup
print('*** Distance to BPG not available for {}:\n{}'.format(
images_iterator.dataset_name, str(e)))
codec_distance_ms_ssim = None
codec_distance_psnr = None
# Note that for each checkpoint, the structure of the network will be the same. Thus the pad depending image
# loading can be cached.
# create session
with tf_helpers.create_session() as sess:
if flags.real_bpp:
pred = probclass.PredictionNetwork(pc, pc_config,
ae.get_centers_variable(), sess)
checker = probclass.ProbclassNetworkTesting(pc, ae, sess)
bpp_fetcher = bpp_helpers.BppFetcher(pred, checker)
fetcher = sess.make_callable(fetch_dict, feed_list=[x_val_ph])
last_ckpt_itr = missing_checkpoints[-1][0]
for ckpt_itr, ckpt_path in missing_checkpoints:
if not ckpt_still_exists(ckpt_path):
# May happen if job is still training
print('Checkpoint disappeared: {}'.format(ckpt_path))
continue
print(_CKPT_ITR_INFO_STR.format(ckpt_itr))
restorer.restore_ckpt(sess, ckpt_path)
values_aggregator = ValuesAggregator('bpp', 'ms-ssim', 'psnr')
# truncates the previous measures.csv file! This way, only the last valid checkpoint is saved.
measures_writer = MeasuresWriter(val_dirs.out_dir)
# ----------------------------------------
# iterate over images
# images are padded to work with current auto encoder
for img_i, (img_name, img_content) in enumerate(
images_iterator.iter_imgs(
pad=ae.get_subsampling_factor())):
otp = fetcher(img_content)
measures_writer.append(img_name, otp)
if flags.real_bpp:
# Calculate
bpp_real, bpp_theory = bpp_fetcher.get_bpp(
otp['sym'],
bpp_helpers.num_pixels_in_image(img_content))
# Logging
bpp_loss = otp['bpp']
diff_percent_tr = (bpp_theory / bpp_real) * 100
diff_percent_lt = (bpp_loss / bpp_theory) * 100
print('BPP: Real {:.5f}\n'
' Theoretical: {:.5f} [{:5.1f}% of real]\n'
' Loss: {:.5f} [{:5.1f}% of real]'.format(
bpp_real, bpp_theory, diff_percent_tr, bpp_loss,
diff_percent_lt))
assert abs(
bpp_theory - bpp_loss
) < 1e-3, 'Expected bpp_theory to match loss! Got {} and {}'.format(
bpp_theory, bpp_loss)
if flags.save_ours and ckpt_itr == last_ckpt_itr:
save_img(img_name, otp['img_out'], val_dirs)
values_aggregator.update(otp)
print('{: 10d} {img_name} | Mean: {avgs}'.format(
img_i,
img_name=img_name,
avgs=values_aggregator.averages_str()),
end=('\r' if not flags.real_bpp else '\n'),
flush=True)
measures_writer.close()
print() # add newline
avgs = values_aggregator.averages()
avg_bpp, avg_ms_ssim, avg_psnr = avgs['bpp'], avgs[
'ms-ssim'], avgs['psnr']
tf_helpers.log_values(
fw, [(full_summary_tag('avg_bpp'), avg_bpp),
(full_summary_tag('avg_ms_ssim'), avg_ms_ssim),
(full_summary_tag('avg_psnr'), avg_psnr)],
iteration=ckpt_itr)
if codec_distance_ms_ssim and codec_distance_psnr:
try:
d_ms_ssim = codec_distance_ms_ssim.distance(
avg_bpp, avg_ms_ssim)
d_pnsr = codec_distance_psnr.distance(avg_bpp, avg_psnr)
print('Distance to BPG: {:.3f} ms-ssim // {:.3f} psnr'.
format(d_ms_ssim, d_pnsr))
tf_helpers.log_values(
fw,
[(full_summary_tag('distance_BPG_MS-SSIM'), d_ms_ssim),
(full_summary_tag('distance_BPG_PSNR'), d_pnsr)],
iteration=ckpt_itr)
except ValueError as e: # out of range errors from distance calls
print(e)
val_dirs.add_validated_checkpoint(ckpt_itr)
print('Validation completed {}'.format(val_dirs))
def train(autoencoder_config_path, probclass_config_path,
restore_manager: RestoreManager, log_dir_root, datasets: Datasets,
train_flags: TrainFlags, ckpt_interval_hours: float,
description: str):
ae_config, ae_config_rel_path = config_parser.parse(
autoencoder_config_path)
pc_config, pc_config_rel_path = config_parser.parse(probclass_config_path)
print_configs(('ae_config', ae_config), ('pc_config', pc_config))
continue_in_ckpt_dir = restore_manager and restore_manager.continue_in_ckpt_dir
if continue_in_ckpt_dir:
logdir = restore_manager.log_dir
else:
logdir = logdir_helpers.create_unique_log_dir(
[ae_config_rel_path, pc_config_rel_path],
log_dir_root,
restore_dir=restore_manager.ckpt_dir if restore_manager else None)
print(_LOG_DIR_FORMAT.format(logdir))
if description:
_write_to_sheets(logdir_helpers.log_date_from_log_dir(logdir),
ae_config_rel_path,
pc_config_rel_path,
description,
git_ref=_get_git_ref(),
log_dir_root=log_dir_root,
is_continue=continue_in_ckpt_dir)
ae_cls = autoencoder.get_network_cls(ae_config)
pc_cls = probclass.get_network_cls(pc_config)
# Instantiate autoencoder and probability classifier
ae = ae_cls(ae_config)
pc = pc_cls(pc_config, num_centers=ae_config.num_centers)
# train ---
ip_train = inputpipeline.InputPipeline(
inputpipeline.get_dataset(datasets.train),
ae_config.crop_size,
batch_size=ae_config.batch_size,
shuffle=False,
num_preprocess_threads=NUM_PREPROCESS_THREADS,
num_crops_per_img=NUM_CROPS_PER_IMG)
x_train = ip_train.get_batch()
enc_out_train = ae.encode(
x_train, is_training=True) # qbar is masked by the heatmap
x_out_train = ae.decode(enc_out_train.qbar, is_training=True)
# stop_gradient is beneficial for training. it prevents multiple gradients flowing into the heatmap.
pc_in = tf.stop_gradient(enc_out_train.qbar)
bc_train = pc.bitcost(pc_in,
enc_out_train.symbols,
is_training=True,
pad_value=pc.auto_pad_value(ae))
bpp_train = bits.bitcost_to_bpp(bc_train, x_train)
d_train = Distortions(ae_config, x_train, x_out_train, is_training=True)
# summing over channel dimension gives 2D heatmap
heatmap2D = (tf.reduce_sum(enc_out_train.heatmap, 1)
if enc_out_train.heatmap is not None else None)
# loss ---
total_loss, H_real, pc_comps, ae_comps = get_loss(ae_config, ae, pc,
d_train.d_loss_scaled,
bc_train,
enc_out_train.heatmap)
train_op = get_train_op(ae_config, pc_config, ip_train, pc.variables(),
total_loss)
# test ---
with tf.name_scope('test'):
ip_test = inputpipeline.InputPipeline(
inputpipeline.get_dataset(datasets.test),
ae_config.crop_size,
batch_size=ae_config.batch_size,
num_preprocess_threads=NUM_PREPROCESS_THREADS,
num_crops_per_img=1,
big_queues=False,
shuffle=False)
x_test = ip_test.get_batch()
enc_out_test = ae.encode(x_test, is_training=False)
x_out_test = ae.decode(enc_out_test.qhard, is_training=False)
bc_test = pc.bitcost(enc_out_test.qhard,
enc_out_test.symbols,
is_training=False,
pad_value=pc.auto_pad_value(ae))
bpp_test = bits.bitcost_to_bpp(bc_test, x_test)
d_test = Distortions(ae_config, x_test, x_out_test, is_training=False)
try: # Try to get codec distnace for current dataset
codec_distance_ms_ssim = CodecDistance(datasets.codec_distance,
codec='bpg',
metric='ms-ssim')
get_distance = functools_ext.catcher(ValueError,
handler=functools_ext.const(
np.nan),
f=codec_distance_ms_ssim.distance)
get_distance = functools_ext.compose(np.float32,
get_distance) # cast to float32
d_BPG_test = tf.py_func(get_distance, [bpp_test, d_test.ms_ssim],
tf.float32,
stateful=False,
name='d_BPG')
d_BPG_test.set_shape(())
except CodecDistanceReadException as e:
print('Cannot compute CodecDistance: {}'.format(e))
d_BPG_test = tf.constant(np.nan, shape=(), name='ConstNaN')
# ---
train_logger = Logger()
test_logger = Logger()
distortion_name = ae_config.distortion_to_minimize
train_logger.add_summaries(d_train.summaries_with_prefix('train'))
# Visualize components of losses
train_logger.add_summaries([
tf.summary.scalar('train/PC_loss/{}'.format(name), comp)
for name, comp in pc_comps
])
train_logger.add_summaries([
tf.summary.scalar('train/AE_loss/{}'.format(name), comp)
for name, comp in ae_comps
])
train_logger.add_summaries([tf.summary.scalar('train/bpp', bpp_train)])
train_logger.add_console_tensor('loss={:.3f}', total_loss)
train_logger.add_console_tensor('ms_ssim={:.3f}', d_train.ms_ssim)
train_logger.add_console_tensor('bpp={:.3f}', bpp_train)
train_logger.add_console_tensor('H_real={:.3f}', H_real)
test_logger.add_summaries(d_test.summaries_with_prefix('test'))
test_logger.add_summaries([
tf.summary.scalar('test/bpp', bpp_test),
tf.summary.scalar('test/distance_BPG_MS-SSIM', d_BPG_test),
tf.summary.image('test/x_in',
prep_for_image_summary(x_test, n=3, name='x_in')),
tf.summary.image('test/x_out',
prep_for_image_summary(x_out_test, n=3, name='x_out'))
])
if heatmap2D is not None:
test_logger.add_summaries([
tf.summary.image(
'test/hm',
prep_for_grayscale_image_summary(heatmap2D,
n=3,
autoscale=True,
name='hm'))
])
test_logger.add_console_tensor('ms_ssim={:.3f}', d_test.ms_ssim)
test_logger.add_console_tensor('bpp={:.3f}', bpp_test)
test_logger.add_summaries([
tf.summary.histogram('centers', ae.get_centers_variable()),
tf.summary.histogram(
'test/qbar', enc_out_test.qbar[:ae_config.batch_size // 2, ...])
])
test_logger.add_console_tensor('d_BPG={:.6f}', d_BPG_test)
test_logger.add_console_tensor(Logger.Numpy1DFormatter('centers={}'),
ae.get_centers_variable())
print('Starting session and queues...')
with tf_helpers.start_queues_in_sess(
init_vars=restore_manager is None) as (sess, coord):
train_logger.finalize_with_sess(sess)
test_logger.finalize_with_sess(sess)
if restore_manager:
restore_manager.restore(sess)
saver = Saver(Saver.ckpt_dir_for_log_dir(logdir),
max_to_keep=1,
keep_checkpoint_every_n_hours=ckpt_interval_hours)
train_loop(ae_config,
sess,
coord,
train_op,
train_logger,
test_logger,
train_flags,
logdir,
saver,
is_restored=restore_manager is not None)