def __init__(
self,
data_path,
save_path,
eval_paths=[],
basic_model=-2,
model=None,
compile=True,
batch_size=128,
lr_base=0.001,
lr_decay=0.05,
lr_min=0,
eval_freq=1,
random_status=0,
custom_objects={},
):
self.model, self.basic_model = None, None
if isinstance(model, str):
if model.endswith(".h5") and os.path.exists(model) and isinstance(
basic_model, int):
print(">>>> Load model from h5 file: %s..." % model)
custom_objects.update({
"NormDense": NormDense,
"margin_softmax": losses.margin_softmax,
"arcface_loss": losses.arcface_loss,
"ArcfaceLoss": losses.ArcfaceLoss,
"CenterLoss": losses.CenterLoss,
})
with keras.utils.custom_object_scope(custom_objects):
self.model = keras.models.load_model(
model, compile=compile, custom_objects=custom_objects)
self.basic_model = keras.models.Model(
self.model.inputs[0],
self.model.layers[basic_model].output)
self.model.summary()
elif isinstance(model, keras.models.Model):
self.model = model
self.basic_model = keras.models.Model(
self.model.inputs[0], self.model.layers[basic_model].output)
elif isinstance(basic_model, str):
if basic_model.endswith(".h5") and os.path.exists(basic_model):
custom_objects.update({
"batch_hard_triplet_loss":
losses.batch_hard_triplet_loss,
"batch_all_triplet_loss":
losses.batch_all_triplet_loss,
"BatchHardTripletLoss":
losses.BatchHardTripletLoss,
"BatchAllTripletLoss":
losses.BatchAllTripletLoss,
})
print(">>>> Load basic_model from h5 file: %s..." %
basic_model)
with keras.utils.custom_object_scope(custom_objects):
self.basic_model = keras.models.load_model(
basic_model,
compile=compile,
custom_objects=custom_objects)
elif isinstance(basic_model, keras.models.Model):
self.basic_model = basic_model
if self.basic_model == None:
print(
"Initialize model by:\n"
"| basicmodel | model |\n"
"| ---------------------------------------- | --------------- |\n"
"| model structure | None |\n"
"| basic model .h5 file | None |\n"
"| model layer index for basic model output | model .h5 file |\n"
"| model layer index for basic model output | model structure |\n"
)
return
self.softmax, self.arcface, self.triplet = "softmax", "arcface", "triplet"
self.batch_size = batch_size
if tf.distribute.has_strategy():
strategy = tf.distribute.get_strategy()
self.batch_size = batch_size * strategy.num_replicas_in_sync
print(">>>> num_replicas_in_sync: %d, batch_size: %d" %
(strategy.num_replicas_in_sync, self.batch_size))
self.data_path, self.random_status = data_path, random_status
self.train_ds, self.steps_per_epoch, self.classes = None, 0, 0
self.is_triplet_dataset = False
self.default_optimizer = "adam"
self.metrics = ["accuracy"]
my_evals = [
evals.eval_callback(self.basic_model,
ii,
batch_size=batch_size,
eval_freq=eval_freq) for ii in eval_paths
]
if len(my_evals) != 0:
my_evals[-1].save_model = os.path.splitext(save_path)[0]
basic_callbacks = myCallbacks.basic_callbacks(checkpoint=save_path,
evals=my_evals,
lr=lr_base,
lr_decay=lr_decay,
lr_min=lr_min)
self.my_evals = my_evals
self.basic_callbacks = basic_callbacks
self.my_hist = self.basic_callbacks[-2]
def __init__(
self,
data_path,
save_path,
eval_paths=[],
basic_model=None,
model=None,
compile=True,
output_wd_multiply=1, # works ONLY for SGDW
custom_objects={},
batch_size=128,
lr_base=0.001,
lr_decay=0.05, # lr_decay < 1 for exponential, or it's cosine decay_steps
lr_decay_steps=0, # lr_decay_steps < 1 for update lr on epoch, or update on every [NUM] batches, or list for ConstantDecayScheduler
lr_min=0,
eval_freq=1,
random_status=0,
dataset_cache=False,
):
custom_objects.update({
"NormDense":
NormDense,
"L2_decay_wdm":
L2_decay_wdm,
"margin_softmax":
losses.margin_softmax,
"MarginSoftmax":
losses.MarginSoftmax,
"arcface_loss":
losses.arcface_loss,
"ArcfaceLossT4":
losses.ArcfaceLossT4,
"ArcfaceLoss":
losses.ArcfaceLoss,
"CenterLoss":
losses.CenterLoss,
"batch_hard_triplet_loss":
losses.batch_hard_triplet_loss,
"batch_all_triplet_loss":
losses.batch_all_triplet_loss,
"BatchHardTripletLoss":
losses.BatchHardTripletLoss,
"BatchAllTripletLoss":
losses.BatchAllTripletLoss,
})
self.model, self.basic_model, self.save_path = None, None, save_path
if isinstance(model, str):
if model.endswith(".h5") and os.path.exists(model):
print(">>>> Load model from h5 file: %s..." % model)
with keras.utils.custom_object_scope(custom_objects):
self.model = keras.models.load_model(
model, compile=compile, custom_objects=custom_objects)
embedding_layer = basic_model if basic_model is not None else self.__search_embedding_layer__(
self.model)
self.basic_model = keras.models.Model(
self.model.inputs[0],
self.model.layers[embedding_layer].output)
# self.model.summary()
elif isinstance(model, keras.models.Model):
self.model = model
embedding_layer = basic_model if basic_model is not None else self.__search_embedding_layer__(
self.model)
self.basic_model = keras.models.Model(
self.model.inputs[0],
self.model.layers[embedding_layer].output)
elif isinstance(basic_model, str):
if basic_model.endswith(".h5") and os.path.exists(basic_model):
print(">>>> Load basic_model from h5 file: %s..." %
basic_model)
with keras.utils.custom_object_scope(custom_objects):
self.basic_model = keras.models.load_model(
basic_model,
compile=compile,
custom_objects=custom_objects)
elif isinstance(basic_model, keras.models.Model):
self.basic_model = basic_model
if self.basic_model == None:
print(
"Initialize model by:\n"
"| basic_model | model |\n"
"| --------------------------------------------------------------- | --------------- |\n"
"| model structure | None |\n"
"| basic model .h5 file | None |\n"
"| None for 'embedding' layer or layer index of basic model output | model .h5 file |\n"
"| None for 'embedding' layer or layer index of basic model output | model structure |\n"
)
return
self.softmax, self.arcface, self.triplet, self.center = "softmax", "arcface", "triplet", "center"
self.batch_size = batch_size
if tf.distribute.has_strategy():
strategy = tf.distribute.get_strategy()
self.batch_size = batch_size * strategy.num_replicas_in_sync
print(">>>> num_replicas_in_sync: %d, batch_size: %d" %
(strategy.num_replicas_in_sync, self.batch_size))
self.data_path, self.random_status = data_path, random_status
self.train_ds, self.steps_per_epoch, self.classes = None, None, 0
self.is_triplet_dataset = False
self.default_optimizer = "adam"
self.metrics = ["accuracy"]
my_evals = [
evals.eval_callback(self.basic_model,
ii,
batch_size=batch_size,
eval_freq=eval_freq) for ii in eval_paths
]
if len(my_evals) != 0:
my_evals[-1].save_model = os.path.splitext(save_path)[0]
basic_callbacks = myCallbacks.basic_callbacks(
checkpoint=save_path,
evals=my_evals,
lr=lr_base,
lr_decay=lr_decay,
lr_min=lr_min,
lr_decay_steps=lr_decay_steps)
self.my_evals = my_evals
self.basic_callbacks = basic_callbacks
self.my_hist = [
ii for ii in self.basic_callbacks
if isinstance(ii, myCallbacks.My_history)
][0]
self.custom_callbacks = []
self.output_wd_multiply = output_wd_multiply
self.dataset_cache = dataset_cache
self.is_distiller = False
def __init__(
self,
data_path,
save_path,
eval_paths=[],
basic_model=None,
model=None,
compile=True,
output_weight_decay=0, # L2 regularizer for output layer, 0 for None, >=1 for value in basic_model, (0, 1) for specific value
custom_objects={},
batch_size=128,
lr_base=0.001,
lr_decay=0.05, # for cosine it's m_mul, or it's decay_rate for exponential or constant
lr_decay_steps=0, # <=1 for Exponential, (1, 500) for Cosine decay on epoch, >= 500 for Cosine decay on batch, list for Constant
lr_min=0,
eval_freq=1,
random_status=0,
image_per_class=0, # For triplet, image_per_class will be `4` if it's `< 4`
teacher_model_interf=None, # Teacher model to generate embedding data, used for distilling training.
):
from inspect import getmembers, isfunction, isclass
custom_objects.update(
dict([
ii for ii in getmembers(losses)
if isfunction(ii[1]) or isclass(ii[1])
]))
custom_objects.update({"NormDense": models.NormDense})
self.model, self.basic_model, self.save_path, self.default_type = None, None, save_path, None
if isinstance(model, str):
if model.endswith(".h5") and os.path.exists(model):
print(">>>> Load model from h5 file: %s..." % model)
with keras.utils.custom_object_scope(custom_objects):
self.model = keras.models.load_model(
model, compile=compile, custom_objects=custom_objects)
embedding_layer = basic_model if basic_model is not None else self.__search_embedding_layer__(
self.model)
self.basic_model = keras.models.Model(
self.model.inputs[0],
self.model.layers[embedding_layer].output)
# self.model.summary()
elif isinstance(model, keras.models.Model):
self.model = model
embedding_layer = basic_model if basic_model is not None else self.__search_embedding_layer__(
self.model)
self.basic_model = keras.models.Model(
self.model.inputs[0],
self.model.layers[embedding_layer].output)
self.default_type = "MODEL"
print(
">>>> Specified model structure, output layer will keep from changing"
)
elif isinstance(basic_model, str):
if basic_model.endswith(".h5") and os.path.exists(basic_model):
print(">>>> Load basic_model from h5 file: %s..." %
basic_model)
with keras.utils.custom_object_scope(custom_objects):
self.basic_model = keras.models.load_model(
basic_model,
compile=compile,
custom_objects=custom_objects)
elif isinstance(basic_model, keras.models.Model):
self.basic_model = basic_model
if self.basic_model == None:
print(
"Initialize model by:\n"
"| basic_model | model |\n"
"| --------------------------------------------------------------- | --------------- |\n"
"| model structure | None |\n"
"| basic model .h5 file | None |\n"
"| None for 'embedding' layer or layer index of basic model output | model .h5 file |\n"
"| None for 'embedding' layer or layer index of basic model output | model structure |\n"
)
return
self.softmax, self.arcface, self.triplet, self.center, self.distill = "softmax", "arcface", "triplet", "center", "distill"
if output_weight_decay >= 1:
l2_weight_decay = 0
for ii in self.basic_model.layers:
if hasattr(ii, "kernel_regularizer") and isinstance(
ii.kernel_regularizer, keras.regularizers.L2):
l2_weight_decay = ii.kernel_regularizer.l2
break
print(">>>> L2 regularizer value from basic_model:",
l2_weight_decay)
output_weight_decay *= l2_weight_decay * 2
self.output_weight_decay = output_weight_decay
self.batch_size = batch_size
if tf.distribute.has_strategy():
strategy = tf.distribute.get_strategy()
self.batch_size = batch_size * strategy.num_replicas_in_sync
print(">>>> num_replicas_in_sync: %d, batch_size: %d" %
(strategy.num_replicas_in_sync, self.batch_size))
self.data_options = tf.data.Options()
self.data_options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.DATA
my_evals = [
evals.eval_callback(self.basic_model,
ii,
batch_size=batch_size,
eval_freq=eval_freq) for ii in eval_paths
]
if len(my_evals) != 0:
my_evals[-1].save_model = os.path.splitext(save_path)[0]
self.my_history, self.model_checkpoint, self.lr_scheduler, self.gently_stop = myCallbacks.basic_callbacks(
checkpoint=save_path,
evals=my_evals,
lr=lr_base,
lr_decay=lr_decay,
lr_min=lr_min,
lr_decay_steps=lr_decay_steps)
self.my_evals, self.custom_callbacks = my_evals, []
self.metrics = ["accuracy"]
self.default_optimizer = "adam"
self.data_path, self.random_status, self.image_per_class, self.teacher_model_interf = data_path, random_status, image_per_class, teacher_model_interf
self.train_ds, self.steps_per_epoch, self.classes, self.is_triplet_dataset, self.is_distill_ds = None, None, 0, False, False
self.distill_emb_map_layer = None
