def main(unused_argv):
os.environ['CUDA_VISIBLE_DEVICES'] = FLAGS.gpu_to_use
dconf = data_config.get_config(FLAGS.data_name)
dataset = input_fn_amoabanet(False, False, FLAGS.data_dir, FLAGS.batch_size,
dataset_name=FLAGS.data_name,
preprocessing_type=FLAGS.preprocessing_type)
iterator = dataset.make_one_shot_iterator()
images, labels = iterator.get_next()
module = hub.Module("https://tfhub.dev/google/imagenet/amoebanet_a_n18_f448/classification/1")
logits = module(images) # Logits with shape [batch_size, num_classes].
pred = tf.nn.softmax(logits)
top1 = tf.argmax(logits, axis=1)
np_preds = np.zeros(dconf.num_images['validation'], dtype=np.int64)
np_labels = np.zeros(dconf.num_images['validation'], dtype=np.int64)
np_i = 0
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
n_loop = dconf.num_images['validation'] // FLAGS.batch_size
for _ in tqdm(range(n_loop + 1)):
try:
_pred, _top1, _labels = sess.run([pred, top1, labels])
np_preds[np_i:np_i + _pred.shape[0]] = _top1
np_labels[np_i:np_i + _labels.shape[0]] = _labels
np_i += _pred.shape[0]
except tf.errors.OutOfRangeError:
break
print('Accuracy:')
print(np.count_nonzero(np_preds == np_labels) / dconf.num_images['validation'])
def model_fn_bfe(features, labels, mode, params):
assert int(params['resnet_size']) >= 50
dataset = data_config.get_config(params['dataset_name'])
learning_rate_fn = learning_rate_with_decay(
learning_rate_decay_type=params['learning_rate_decay_type'],
batch_size=params['batch_size'],
batch_denom=params['batch_size'],
num_images=dataset.num_images['train'],
num_epochs_per_decay=params['num_epochs_per_decay'],
learning_rate_decay_factor=params['learning_rate_decay_factor'],
end_learning_rate=params['end_learning_rate'],
piecewise_lr_boundary_epochs=params['piecewise_lr_boundary_epochs'],
piecewise_lr_decay_rates=params['piecewise_lr_decay_rates'],
base_lr=params['base_learning_rate'],
train_epochs=params['train_epochs'],
warmup=params['lr_warmup'])
if params['use_dropblock']:
starter_kp = params['dropblock_kp'][0]
end_kp = params['dropblock_kp'][1]
batches_per_epoch = dataset.num_images['train'] / params['batch_size']
decay_steps = int(params['train_epochs'] * batches_per_epoch)
keep_prob_fn = keep_prob_decay(starter_kp, end_kp, decay_steps)
else:
keep_prob_fn = None
return run_loop_bfe.resnet_ir_model_fn(
features=features,
labels=labels,
mode=mode,
model_class=Model,
resnet_size=params['resnet_size'],
weight_decay=params['weight_decay'],
learning_rate_fn=learning_rate_fn,
momentum=params['momentum'],
data_format=params['data_format'],
dim_features=params['dim_features'],
num_classes=dataset.num_classes,
resnet_version=params['resnet_version'],
loss_scale=params['loss_scale'],
loss_filter_fn=None,
dtype=params['dtype'],
fine_tune=params['fine_tune'],
zero_gamma=params['zero_gamma'],
use_resnet_d=params['use_resnet_d'],
no_downsample=params['no_downsample'],
weight_softmax_loss=params['weight_softmax_loss'],
weight_ranking_loss=params['weight_ranking_loss'],
label_smoothing=params['label_smoothing'],
use_bfe=params['use_bfe'],
use_global_branch=params['use_global_branch'],
ranking_loss_type=params['ranking_loss_type'],
off_bfe_ranking=params['off_bfe_ranking'],
margin=params['margin'],
use_summary_image=params['use_summary_image'],
use_se_block=params['use_se_block'],
rollback_period=params['rollback_period'],
rollback_lr_multiplier=params['rollback_lr_multiplier'],
keep_prob_fn=keep_prob_fn)
def input_fn_amoabanet(is_training,
use_random_crop,
data_dir,
batch_size,
num_epochs=1,
num_gpus=None,
dtype=tf.float32,
with_drawing_bbox=False,
autoaugment_type=None,
dataset_name=None,
drop_remainder=False,
preprocessing_type='imagenet',
return_logits=False,
dct_method="",
train_regex='train*',
val_regex='validation*'):
filenames = data_util.get_filenames(is_training, data_dir,
train_regex=train_regex,
val_regex=val_regex)
dataset = data_config.get_config(dataset_name)
return input_fn(is_training, filenames, use_random_crop, batch_size,
dataset.num_train_files, dataset.num_images['train'],
dataset.shuffle_buffer, dataset.num_channels, num_epochs,
num_gpus, dtype,
autoaugment_type=autoaugment_type,
with_drawing_bbox=with_drawing_bbox,
drop_remainder=drop_remainder,
preprocessing_type=preprocessing_type,
return_logits=return_logits,
dct_method=dct_method)
def input_fn_ir_eval(is_training,
data_dir,
batch_size,
num_epochs=1,
num_gpus=0,
dtype=tf.float32,
preprocessing_type='imagenet',
dataset_name=None,
dct_method="",
val_regex='validation-*'):
filenames = data_util.get_filenames(is_training,
data_dir,
val_regex=val_regex)
assert len(filenames) > 0
dataset_config = data_config.get_config(dataset_name)
return input_fn(is_training,
filenames,
False,
batch_size,
dataset_config.num_train_files,
dataset_config.num_images['validation'],
dataset_config.shuffle_buffer,
dataset_config.num_channels,
num_epochs,
num_gpus,
dtype,
preprocessing_type=preprocessing_type,
dct_method=dct_method)
def model_fn_cls(features, labels, mode, params):
if int(params['resnet_size']) < 50:
assert not params['use_dropblock']
assert not params['use_se_block']
assert not params['use_resnet_d']
dataset = data_config.get_config(params['dataset_name'])
learning_rate_fn = learning_rate_with_decay(
learning_rate_decay_type=params['learning_rate_decay_type'],
batch_size=params['batch_size'],
batch_denom=params['batch_size'],
num_images=dataset.num_images['train'],
num_epochs_per_decay=params['num_epochs_per_decay'],
learning_rate_decay_factor=params['learning_rate_decay_factor'],
end_learning_rate=params['end_learning_rate'],
piecewise_lr_boundary_epochs=params['piecewise_lr_boundary_epochs'],
piecewise_lr_decay_rates=params['piecewise_lr_decay_rates'],
base_lr=params['base_learning_rate'],
train_epochs=params['train_epochs'],
warmup=params['lr_warmup'])
if params['use_dropblock']:
starter_kp = params['dropblock_kp'][0]
end_kp = params['dropblock_kp'][1]
batches_per_epoch = dataset.num_images['train'] / params['batch_size']
decay_steps = int(params['train_epochs'] * batches_per_epoch)
keep_prob_fn = keep_prob_decay(starter_kp, end_kp, decay_steps)
else:
keep_prob_fn = None
return run_loop_classification.resnet_model_fn(
features=features,
labels=labels,
num_classes=dataset.num_classes,
mode=mode,
model_class=Model,
resnet_size=params['resnet_size'],
weight_decay=params['weight_decay'],
learning_rate_fn=learning_rate_fn,
momentum=params['momentum'],
zero_gamma=params['zero_gamma'],
use_resnet_d=params['use_resnet_d'],
label_smoothing=params['label_smoothing'],
data_format=params['data_format'],
resnet_version=params['resnet_version'],
loss_scale=params['loss_scale'],
loss_filter_fn=None,
dtype=params['dtype'],
fine_tune=params['fine_tune'],
use_se_block=params['use_se_block'],
display_raw_images_with_bbox=params['display_raw_images_with_bbox'],
use_ranking_loss=params['use_ranking_loss'],
mixup_type=params['mixup_type'],
rollback_period=params['rollback_period'],
rollback_lr_multiplier=params['rollback_lr_multiplier'],
keep_prob_fn=keep_prob_fn)
def input_fn_cls(is_training,
use_random_crop,
data_dir,
batch_size,
num_epochs=1,
num_gpus=None,
dtype=tf.float32,
with_drawing_bbox=False,
autoaugment_type=None,
dataset_name=None,
drop_remainder=False,
dct_method=""):
"""Input function which provides batches for train or eval.
Args:
is_training: A boolean denoting whether the input is for training.
use_random_crop: Whether to randomly crop a training image.
data_dir: The directory containing the input data.
batch_size: The number of samples per batch.
num_epochs: The number of epochs to repeat the dataset.
num_gpus: The number of gpus used for training.
dtype: Data type to use for images/features
autoaugment_type: Auto augmentation type. 'imagenet', 'svhn', 'cifar', 'good'
with_drawing_bbox: If True, return the dataset including raw image tensor with bbox.
dct_method: An optional `string`. Defaults to `""`.
string specifying a hint about the algorithm used for
decompression. Defaults to "" which maps to a system-specific
default. Currently valid values are ["INTEGER_FAST",
"INTEGER_ACCURATE"]. The hint may be ignored (e.g., the internal
jpeg library changes to a version that does not have that specific
option.)
Returns:
A dataset that can be used for iteration.
"""
filenames = data_util.get_filenames(is_training, data_dir)
dataset = data_config.get_config(dataset_name)
return input_fn(is_training,
filenames,
use_random_crop,
batch_size,
dataset.num_train_files,
dataset.num_images['train'],
dataset.shuffle_buffer,
dataset.num_channels,
num_epochs,
num_gpus,
dtype,
autoaugment_type=autoaugment_type,
with_drawing_bbox=with_drawing_bbox,
drop_remainder=drop_remainder,
dct_method=dct_method)
def input_fn_cls(is_training, use_random_crop, num_epochs, flags_obj):
if flags_obj.mixup_type == 1 and is_training:
batch_size = flags_obj.batch_size * 2
num_epochs = num_epochs * 2
else:
batch_size = flags_obj.batch_size
batch_size = distribution_utils.per_device_batch_size(
batch_size, flags_core.get_num_gpus(flags_obj))
filenames_sup = data_util.get_filenames(is_training,
flags_obj.data_dir,
train_regex=flags_obj.train_regex,
val_regex=flags_obj.val_regex)
tf.logging.info('The # of Supervised tfrecords: {}'.format(
len(filenames_sup)))
dataset_meta = data_config.get_config(flags_obj.dataset_name)
datasets = []
dataset_sup = input_fn(is_training,
filenames_sup,
use_random_crop,
batch_size,
dataset_meta.num_train_files,
dataset_meta.num_images['train'],
dataset_meta.shuffle_buffer,
dataset_meta.num_channels,
num_epochs,
flags_core.get_num_gpus(flags_obj),
flags_core.get_tf_dtype(flags_obj),
autoaugment_type=flags_obj.autoaugment_type,
with_drawing_bbox=flags_obj.with_drawing_bbox,
drop_remainder=False,
preprocessing_type=flags_obj.preprocessing_type,
return_logits=flags_obj.kd_temp > 0,
dct_method=flags_obj.dct_method,
parse_record_fn=data_util.parse_record_sup)
datasets.append(dataset_sup)
def flatten_input(*features):
images_dict = {}
for feature in features:
for key in feature:
if key == 'label':
label = feature[key]
else:
images_dict[key] = feature[key]
return images_dict, label
dataset = tf.data.Dataset.zip(tuple(datasets))
dataset = dataset.map(flatten_input)
tf.logging.info('dataset = dataset.map(flatten_input)')
tf.logging.info(dataset)
return dataset
def input_fn_bfe_eval(is_training,
data_dir,
batch_size,
num_epochs=1,
num_gpus=0,
dtype=tf.float32,
preprocessing_type='imagenet',
dataset_name=None,
dct_method=""):
"""Input function which provides batches for train or eval.
Args:
is_training: A boolean denoting whether the input is for training.
data_dir: The directory containing the input data.
batch_size: The number of samples per batch.
num_epochs: The number of epochs to repeat the dataset.
num_gpus: The number of gpus used for training.
dtype: Data type to use for images/features
preprocessing_type: TODO
dataset_name: TODO
dct_method: An optional `string`. Defaults to `""`.
string specifying a hint about the algorithm used for
decompression. Defaults to "" which maps to a system-specific
default. Currently valid values are ["INTEGER_FAST",
"INTEGER_ACCURATE"]. The hint may be ignored (e.g., the internal
jpeg library changes to a version that does not have that specific
option.)
Returns:
A dataset that can be used for iteration.
"""
filenames = data_util.get_filenames(is_training,
data_dir,
val_regex='validation-label*')
dataset_config = data_config.get_config(dataset_name)
return input_fn(is_training,
filenames,
False,
batch_size,
dataset_config.num_train_files,
dataset_config.num_images['validation'],
dataset_config.shuffle_buffer,
dataset_config.num_channels,
num_epochs,
num_gpus,
dtype,
preprocessing_type=preprocessing_type,
dct_method=dct_method)
def model_fn_cls(features, labels, mode, params):
if int(params['resnet_size']) < 50:
assert not params['use_dropblock']
assert not params['use_se_block']
assert not params['use_sk_block']
assert not params['use_resnet_d']
dataset = data_config.get_config(params['dataset_name'])
learning_rate_fn = learning_rate_with_decay(
learning_rate_decay_type=params['learning_rate_decay_type'],
batch_size=params['batch_size'],
batch_denom=params['batch_size'],
num_images=dataset.num_images['train'],
num_epochs_per_decay=params['num_epochs_per_decay'],
learning_rate_decay_factor=params['learning_rate_decay_factor'],
end_learning_rate=params['end_learning_rate'],
piecewise_lr_boundary_epochs=params['piecewise_lr_boundary_epochs'],
piecewise_lr_decay_rates=params['piecewise_lr_decay_rates'],
base_lr=params['base_learning_rate'],
train_epochs=params['train_epochs'],
warmup_epochs=params['lr_warmup_epochs'])
if params['use_dropblock']:
starter_kp = params['dropblock_kp'][0]
end_kp = params['dropblock_kp'][1]
batches_per_epoch = dataset.num_images['train'] / params['batch_size']
decay_steps = int(params['train_epochs'] * batches_per_epoch)
keep_prob_fn = keep_prob_decay(starter_kp, end_kp, decay_steps)
else:
keep_prob_fn = None
return run_loop_classification.resnet_model_fn(
features=features,
labels=labels,
num_classes=dataset.num_classes,
mode=mode,
model_class=Model,
learning_rate_fn=learning_rate_fn,
keep_prob_fn=keep_prob_fn,
loss_filter_fn=None,
p=params)
def run(flags_obj):
"""Run ResNet ImageNet training and eval loop.
Args:
flags_obj: An object containing parsed flag values.
"""
if not flags_obj.eval_only:
config_utils.dump_hparam()
dataset = data_config.get_config(flags_obj.dataset_name)
run_loop_classification.resnet_main(flags_obj,
model_fns.model_fn_cls,
input_fns.input_fn_cls,
dataset.dataset_name,
shape=[
dataset.default_image_size,
dataset.default_image_size,
dataset.num_channels
],
num_images=dataset.num_images,
zeroshot_eval=flags_obj.zeroshot_eval)
def resnet_main(flags_obj,
model_function,
input_function,
dataset_name,
shape=None,
num_images=None,
zeroshot_eval=False):
model_helpers.apply_clean(flags.FLAGS)
# Using the Winograd non-fused algorithms provides a small performance boost.
os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
# Create session config based on values of inter_op_parallelism_threads and
# intra_op_parallelism_threads. Note that we default to having
# allow_soft_placement = True, which is required for multi-GPU and not
# harmful for other modes.
session_config = config_utils.get_session_config(flags_obj)
run_config = config_utils.get_run_config(flags_obj, flags_core,
session_config,
num_images['train'])
tf.logging.info("ERR1!!!!")
def gen_estimator(period=None):
resnet_size = int(flags_obj.resnet_size)
data_format = flags_obj.data_format
batch_size = flags_obj.batch_size
resnet_version = int(flags_obj.resnet_version)
loss_scale = flags_core.get_loss_scale(flags_obj)
dtype_tf = flags_core.get_tf_dtype(flags_obj)
num_epochs_per_decay = flags_obj.num_epochs_per_decay
learning_rate_decay_factor = flags_obj.learning_rate_decay_factor
end_learning_rate = flags_obj.end_learning_rate
learning_rate_decay_type = flags_obj.learning_rate_decay_type
weight_decay = flags_obj.weight_decay
zero_gamma = flags_obj.zero_gamma
lr_warmup_epochs = flags_obj.lr_warmup_epochs
base_learning_rate = flags_obj.base_learning_rate
use_resnet_d = flags_obj.use_resnet_d
use_dropblock = flags_obj.use_dropblock
dropblock_kp = [float(be) for be in flags_obj.dropblock_kp]
label_smoothing = flags_obj.label_smoothing
momentum = flags_obj.momentum
bn_momentum = flags_obj.bn_momentum
train_epochs = flags_obj.train_epochs
piecewise_lr_boundary_epochs = [
int(be) for be in flags_obj.piecewise_lr_boundary_epochs
]
piecewise_lr_decay_rates = [
float(dr) for dr in flags_obj.piecewise_lr_decay_rates
]
use_ranking_loss = flags_obj.use_ranking_loss
use_se_block = flags_obj.use_se_block
use_sk_block = flags_obj.use_sk_block
mixup_type = flags_obj.mixup_type
dataset_name = flags_obj.dataset_name
kd_temp = flags_obj.kd_temp
no_downsample = flags_obj.no_downsample
anti_alias_filter_size = flags_obj.anti_alias_filter_size
anti_alias_type = flags_obj.anti_alias_type
cls_loss_type = flags_obj.cls_loss_type
logit_type = flags_obj.logit_type
embedding_size = flags_obj.embedding_size
pool_type = flags_obj.pool_type
arc_s = flags_obj.arc_s
arc_m = flags_obj.arc_m
bl_alpha = flags_obj.bl_alpha
bl_beta = flags_obj.bl_beta
exp = None
if install_hyperdash and flags_obj.use_hyperdash:
exp = Experiment(flags_obj.model_dir.split("/")[-1])
resnet_size = exp.param("resnet_size", int(flags_obj.resnet_size))
batch_size = exp.param("batch_size", flags_obj.batch_size)
exp.param("dtype", flags_obj.dtype)
learning_rate_decay_type = exp.param(
"learning_rate_decay_type", flags_obj.learning_rate_decay_type)
weight_decay = exp.param("weight_decay", flags_obj.weight_decay)
zero_gamma = exp.param("zero_gamma", flags_obj.zero_gamma)
lr_warmup_epochs = exp.param("lr_warmup_epochs",
flags_obj.lr_warmup_epochs)
base_learning_rate = exp.param("base_learning_rate",
flags_obj.base_learning_rate)
use_dropblock = exp.param("use_dropblock", flags_obj.use_dropblock)
dropblock_kp = exp.param(
"dropblock_kp", [float(be) for be in flags_obj.dropblock_kp])
piecewise_lr_boundary_epochs = exp.param(
"piecewise_lr_boundary_epochs",
[int(be) for be in flags_obj.piecewise_lr_boundary_epochs])
piecewise_lr_decay_rates = exp.param(
"piecewise_lr_decay_rates",
[float(dr) for dr in flags_obj.piecewise_lr_decay_rates])
mixup_type = exp.param("mixup_type", flags_obj.mixup_type)
dataset_name = exp.param("dataset_name", flags_obj.dataset_name)
exp.param("autoaugment_type", flags_obj.autoaugment_type)
classifier = tf.estimator.Estimator(
model_fn=model_function,
model_dir=flags_obj.model_dir,
config=run_config,
params={
'resnet_size': resnet_size,
'data_format': data_format,
'batch_size': batch_size,
'resnet_version': resnet_version,
'loss_scale': loss_scale,
'dtype': dtype_tf,
'num_epochs_per_decay': num_epochs_per_decay,
'learning_rate_decay_factor': learning_rate_decay_factor,
'end_learning_rate': end_learning_rate,
'learning_rate_decay_type': learning_rate_decay_type,
'weight_decay': weight_decay,
'zero_gamma': zero_gamma,
'lr_warmup_epochs': lr_warmup_epochs,
'base_learning_rate': base_learning_rate,
'use_resnet_d': use_resnet_d,
'use_dropblock': use_dropblock,
'dropblock_kp': dropblock_kp,
'label_smoothing': label_smoothing,
'momentum': momentum,
'bn_momentum': bn_momentum,
'embedding_size': embedding_size,
'train_epochs': train_epochs,
'piecewise_lr_boundary_epochs': piecewise_lr_boundary_epochs,
'piecewise_lr_decay_rates': piecewise_lr_decay_rates,
'with_drawing_bbox': flags_obj.with_drawing_bbox,
'use_ranking_loss': use_ranking_loss,
'use_se_block': use_se_block,
'use_sk_block': use_sk_block,
'mixup_type': mixup_type,
'kd_temp': kd_temp,
'no_downsample': no_downsample,
'dataset_name': dataset_name,
'anti_alias_filter_size': anti_alias_filter_size,
'anti_alias_type': anti_alias_type,
'cls_loss_type': cls_loss_type,
'logit_type': logit_type,
'arc_s': arc_s,
'arc_m': arc_m,
'pool_type': pool_type,
'bl_alpha': bl_alpha,
'bl_beta': bl_beta,
'train_steps': total_train_steps,
})
return classifier, exp
run_params = {
'batch_size': flags_obj.batch_size,
'dtype': flags_core.get_tf_dtype(flags_obj),
'resnet_size': flags_obj.resnet_size,
'resnet_version': flags_obj.resnet_version,
'synthetic_data': flags_obj.use_synthetic_data,
'train_epochs': flags_obj.train_epochs,
}
if flags_obj.use_synthetic_data:
dataset_name = dataset_name + '-synthetic'
benchmark_logger = logger.get_benchmark_logger()
benchmark_logger.log_run_info('resnet',
dataset_name,
run_params,
test_id=flags_obj.benchmark_test_id)
train_hooks = hooks_helper.get_train_hooks(flags_obj.hooks,
model_dir=flags_obj.model_dir,
batch_size=flags_obj.batch_size)
def input_fn_train(num_epochs):
return input_function(is_training=True,
use_random_crop=flags_obj.training_random_crop,
num_epochs=num_epochs,
flags_obj=flags_obj)
def input_fn_eval():
return input_function(is_training=False,
use_random_crop=False,
num_epochs=1,
flags_obj=flags_obj)
ckpt_keeper = checkpoint_utils.CheckpointKeeper(
save_dir=flags_obj.model_dir,
num_to_keep=flags_obj.num_best_ckpt_to_keep,
keep_epoch=flags_obj.keep_ckpt_every_eval,
maximize=True)
if zeroshot_eval:
dataset = data_config.get_config(dataset_name)
model = model_fns_predict.Model(
int(flags_obj.resnet_size),
flags_obj.data_format,
resnet_version=int(flags_obj.resnet_version),
num_classes=dataset.num_classes,
zero_gamma=flags_obj.zero_gamma,
use_se_block=flags_obj.use_se_block,
use_sk_block=flags_obj.use_sk_block,
no_downsample=flags_obj.no_downsample,
anti_alias_filter_size=flags_obj.anti_alias_filter_size,
anti_alias_type=flags_obj.anti_alias_type,
bn_momentum=flags_obj.bn_momentum,
embedding_size=flags_obj.embedding_size,
pool_type=flags_obj.pool_type,
bl_alpha=flags_obj.bl_alpha,
bl_beta=flags_obj.bl_beta,
dtype=flags_core.get_tf_dtype(flags_obj),
loss_type=flags_obj.cls_loss_type)
def train_and_evaluate(hooks):
tf.logging.info('Starting cycle: %d/%d', cycle_index, int(n_loops))
if num_train_epochs:
classifier.train(input_fn=lambda: input_fn_train(num_train_epochs),
hooks=hooks,
steps=flags_obj.max_train_steps)
tf.logging.info('Starting to evaluate.')
if zeroshot_eval:
tf.reset_default_graph()
eval_results = recall_metric.recall_at_k(
flags_obj,
flags_core,
input_fns.input_fn_ir_eval,
model,
num_images['validation'],
eval_similarity=flags_obj.eval_similarity,
return_embedding=True)
else:
eval_results = classifier.predict(input_fn=input_fn_eval)
return eval_results
total_train_steps = flags_obj.train_epochs * int(
num_images['train'] / flags_obj.batch_size)
if flags_obj.eval_only or not flags_obj.train_epochs:
schedule, n_loops = [0], 1
elif flags_obj.export_only:
schedule, n_loops = [], 0
else:
n_loops = math.ceil(flags_obj.train_epochs /
flags_obj.epochs_between_evals)
schedule = [
flags_obj.epochs_between_evals for _ in range(int(n_loops))
]
schedule[-1] = flags_obj.train_epochs - sum(
schedule[:-1]) # over counting.
schedule = config_utils.get_epoch_schedule(flags_obj, schedule,
num_images)
tf.logging.info('epoch schedule:')
tf.logging.info(schedule)
classifier, exp = gen_estimator()
if flags_obj.pretrained_model_checkpoint_path:
warm_start_hook = WarmStartHook(
flags_obj.pretrained_model_checkpoint_path)
train_hooks.append(warm_start_hook)
for cycle_index, num_train_epochs in enumerate(schedule):
tf.logging.info("ERR123!!!!")
eval_results = train_and_evaluate(train_hooks)
return eval_results
if zeroshot_eval:
metric = eval_results['recall_at_1']
else:
metric = eval_results['accuracy']
tf.logging.info("ERR1234!!!!")
ckpt_keeper.save(metric, flags_obj.model_dir)
if exp:
exp.metric("accuracy", metric)
benchmark_logger.log_evaluation_result(eval_results)
tf.logging.info("ERR12345!!!!")
if model_helpers.past_stop_threshold(flags_obj.stop_threshold, metric):
break
if model_helpers.past_stop_threshold(total_train_steps,
eval_results['global_step']):
break
if exp:
exp.end()
if flags_obj.export_dir is not None:
export_utils.export_pb(flags_core, flags_obj, shape, classifier)
def input_fn_npair_train(is_training,
use_random_crop,
data_dir,
batch_size,
train_epochs,
num_gpus=0,
dtype=tf.float32,
with_drawing_bbox=False,
autoaugment_type=None,
preprocessing_type='imagenet',
dct_method="",
dataset_name=None):
"""Input function which provides batches for train or eval.
See https://oss.navercorp.com/VisualSearch/food-fighters/pull/49#issue-566301.
Args:
is_training: A boolean denoting whether the input is for training.
use_random_crop: Whether to randomly crop a training image.
data_dir: The directory containing the input data.
batch_size: The number of samples per batch.
train_epochs: TODO
num_gpus: The number of gpus used for training.
dtype: Data type to use for images/features
dct_method: An optional `string`. Defaults to `""`.
string specifying a hint about the algorithm used for
decompression. Defaults to "" which maps to a system-specific
default. Currently valid values are ["INTEGER_FAST",
"INTEGER_ACCURATE"]. The hint may be ignored (e.g., the internal
jpeg library changes to a version that does not have that specific
option.)
autoaugment_type: Auto augmentation type. 'imagenet', 'svhn', 'cifar', `good`
with_drawing_bbox: If True, return the dataset including raw image tensor with bbox.
Returns:
A dataset that can be used for iteration.
"""
dconf = data_config.get_config(dataset_name)
all_choices_ds = []
for filename in data_util.get_filenames(True,
data_dir,
train_regex='train-label*'):
dataset = tf.data.Dataset.from_tensors(filename)
# The cycle_length isn't necessary now because there's only one tfrecord file.
# Use this feature when you want to increase the number of file shards.
dataset = dataset.apply(
tf.contrib.data.parallel_interleave(tf.data.TFRecordDataset,
cycle_length=1))
# shuffling records by class. A larger shuffle buffer's size results in better randomness,
# but smaller size reduce startup time and use less memory.
dataset = dataset.shuffle(buffer_size=100, seed=0)
dataset = dataset.repeat()
all_choices_ds.append(dataset)
# A Repeat number must be mutliples of two which means anchor and positive.
max_train_steps = train_epochs * int(
dconf.num_images['train'] / batch_size)
choice_dataset = tf.data.Dataset.range(len(all_choices_ds)).repeat(
max_train_steps * 2)
dataset = tf.contrib.data.choose_from_datasets(all_choices_ds,
choice_dataset)
return data_util.process_record_dataset_ir(
dataset=dataset,
is_training=is_training,
batch_size=batch_size,
parse_record_fn=data_util.parse_record,
num_classes=dconf.num_classes,
num_channels=dconf.num_channels,
num_gpus=num_gpus,
use_random_crop=use_random_crop,
dtype=dtype,
with_drawing_bbox=with_drawing_bbox,
autoaugment_type=autoaugment_type,
num_instances=2,
preprocessing_type=preprocessing_type,
is_aggregated=False,
dct_method=dct_method)
def export_test(bin_export_path, flags_obj, ir_eval):
ds = tf.data.Dataset.list_files(flags_obj.data_dir + '/' +
flags_obj.val_regex)
ds = ds.interleave(tf.data.TFRecordDataset, cycle_length=10)
def parse_tfr(example_proto):
feature_def = {
'image/class/label':
tf.FixedLenFeature([], dtype=tf.int64, default_value=-1),
'image/encoded':
tf.FixedLenFeature([], dtype=tf.string, default_value='')
}
features = tf.io.parse_single_example(serialized=example_proto,
features=feature_def)
return features['image/encoded'], features['image/class/label']
ds = ds.map(parse_tfr)
ds = ds.batch(flags_obj.val_batch_size)
iterator = ds.make_one_shot_iterator()
images, labels = iterator.get_next()
dconf = data_config.get_config(flags_obj.dataset_name)
num_val_images = dconf.num_images['validation']
if flags_obj.zeroshot_eval or ir_eval:
feature_dim = flags_obj.embedding_size if flags_obj.embedding_size > 0 else flags_obj.num_features
np_features = np.zeros((num_val_images, feature_dim), dtype=np.float32)
np_labels = np.zeros(num_val_images, dtype=np.int64)
np_i = 0
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
tf.saved_model.load(sess=sess,
export_dir=bin_export_path,
tags={"serve"})
for _ in tqdm(
range(int(num_val_images / flags_obj.val_batch_size) + 1)):
try:
np_image, np_label = sess.run([images, labels])
np_predict = sess.run(
'embedding_tensor:0',
feed_dict={'input_tensor:0': np_image})
np_features[np_i:np_i +
np_predict.shape[0], :] = np_predict
np_labels[np_i:np_i + np_label.shape[0]] = np_label
np_i += np_predict.shape[0]
except tf.errors.OutOfRangeError:
break
assert np_i == num_val_images
from sklearn.preprocessing import normalize
x = normalize(np_features)
np_sim = x.dot(x.T)
np.fill_diagonal(np_sim, -10) # removing similarity for query.
num_correct = 0
for i in range(num_val_images):
cur_label = np_labels[i]
rank1_label = np_labels[np.argmax(np_sim[i, :])]
if rank1_label == cur_label:
num_correct += 1
recall_at_1 = num_correct / num_val_images
metric = recall_at_1
else:
np_i = 0
correct_cnt = 0
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
tf.saved_model.load(sess=sess,
export_dir=bin_export_path,
tags={"serve"})
for _ in tqdm(
range(int(num_val_images / flags_obj.val_batch_size) + 1)):
try:
np_image, np_label = sess.run([images, labels])
np_predict = sess.run(
'ArgMax:0', feed_dict={'input_tensor:0': np_image})
np_i += np_predict.shape[0]
correct_cnt += np.sum(np_predict == np_label)
except tf.errors.OutOfRangeError:
break
assert np_i == num_val_images
metric = correct_cnt / np_i
return metric
def model_fn_npair(features, labels, mode, params):
if int(params['resnet_size']) < 50:
assert not params['use_dropblock']
assert not params['use_resnet_d']
dataset = data_config.get_config(params['dataset_name'])
learning_rate_fn = learning_rate_with_decay(
learning_rate_decay_type=params['learning_rate_decay_type'],
batch_size=params['global_batch_size'],
batch_denom=params['global_batch_size'],
num_images=dataset.num_images['train'],
num_epochs_per_decay=params['num_epochs_per_decay'],
learning_rate_decay_factor=params['learning_rate_decay_factor'],
end_learning_rate=params['end_learning_rate'],
piecewise_lr_boundary_epochs=params['piecewise_lr_boundary_epochs'],
piecewise_lr_decay_rates=params['piecewise_lr_decay_rates'],
base_lr=params['base_learning_rate'],
train_epochs=params['train_epochs'],
warmup=params['lr_warmup'])
if params['use_dropblock']:
starter_kp = params['dropblock_kp'][0]
end_kp = params['dropblock_kp'][1]
batches_per_epoch = dataset.num_images['train'] / params['batch_size']
decay_steps = int(params['train_epochs'] * batches_per_epoch)
keep_prob_fn = keep_prob_decay(starter_kp, end_kp, decay_steps)
else:
keep_prob_fn = None
steps_per_epoch = int(dataset.num_images['train'] /
params['global_batch_size'])
return run_loop_npair.resnet_model_fn(
features=features,
labels=labels,
mode=mode,
model_class=Model,
num_classes=dataset.num_classes,
resnet_size=params['resnet_size'],
weight_decay=params['weight_decay'],
learning_rate_fn=learning_rate_fn,
momentum=params['momentum'],
data_format=params['data_format'],
dim_features=params['dim_features'],
resnet_version=params['resnet_version'],
loss_scale=params['loss_scale'],
loss_filter_fn=None,
dtype=params['dtype'],
fine_tune=params['fine_tune'],
zero_gamma=params['zero_gamma'],
use_resnet_d=params['use_resnet_d'],
rollback_period=params['rollback_period'],
rollback_lr_multiplier=params['rollback_lr_multiplier'],
use_se_block=params['use_se_block'],
HDML_type=params['HDML_type'],
batch_size=params['batch_size'],
steps_per_epoch=steps_per_epoch,
optimizer_name=params['optimizer'],
loss_l2_reg=params['loss_l2_reg'],
Softmax_factor=params['Softmax_factor'],
lr_gen=params['lr_gen'],
s_lr=params['s_lr'],
alpha=params['alpha'],
beta=params['beta'],
_lambda=params['_lambda'],
keep_prob_fn=keep_prob_fn)
