def uniary_convert():
with tf.device('gpu:0'):
model = None
input_size = [416, 416, 3]
# config = exp_cfg.YoloTask(
# model=exp_cfg.Yolo(_input_size = [416, 416, 3],
# base='v4tiny',
# min_level=4,
# norm_activation = exp_cfg.common.NormActivation(activation="leaky"),
# _boxes = ['(10, 14)', '(23, 27)', '(37, 58)', '(81, 82)', '(135, 169)', '(344, 319)'],
# #_boxes = ['(20, 28)', '(46, 54)', '(74, 116)', '(81, 82)', '(135, 169)', '(344, 319)'],
# #_boxes = ["(10, 13)", "(16, 30)", "(33, 23)","(30, 61)", "(62, 45)", "(59, 119)","(116, 90)", "(156, 198)", "(373, 326)"],
# #_boxes = ['(12, 16)', '(19, 36)', '(40, 28)', '(36, 75)','(76, 55)', '(72, 146)', '(142, 110)', '(192, 243)','(459, 401)'],
# filter = exp_cfg.YoloLossLayer(use_nms=False)
# ))
config = exp_cfg.YoloTask(model=exp_cfg.Yolo(
_input_size=input_size,
base='v3',
min_level=3,
norm_activation=exp_cfg.common.NormActivation(activation='leaky'),
_boxes=[
'(10, 13)', '(16, 30)', '(33, 23)', '(30, 61)', '(62, 45)',
'(59, 119)', '(116, 90)', '(156, 198)', '(373, 326)'
],
filter=exp_cfg.YoloLossLayer(use_nms=False)))
task = YoloTask(config)
model = task.build_model()
task.initialize(model)
#model.build((1, 416, 416, 3))
model(tf.ones((1, 416, 416, 3), dtype=tf.float32), training=False)
image = url_to_image(
'https://raw.githubusercontent.com/zhreshold/mxnet-ssd/master/data/demo/dog.jpg'
)
image = cv2.resize(image, (416, 416))
image = tf.expand_dims(image, axis=0)
func = conversion(model)
model.summary()
# Convert the model
converter = tf.lite.TFLiteConverter.from_concrete_functions(
[func.get_concrete_function(image)])
converter.optimizations = [tf.lite.Optimize.DEFAULT]
#converter.representative_dataset = get_rep_data()
try:
tflite_model = converter.convert()
except BaseException:
print('here')
# st.print_exc()
import sys
sys.exit()
with open('detect.tflite', 'wb') as f:
f.write(tflite_model)
def get_model(model, input_size=[416, 416, 3], training=True):
with tf.device("gpu:0"):
if model == "v4tiny":
config = exp_cfg.YoloTask(model=exp_cfg.Yolo(
_input_size=input_size,
base="v4tiny",
min_level=4,
norm_activation=exp_cfg.common.NormActivation(
activation="leaky"),
_boxes=[
"(10, 14)", "(23, 27)", "(37, 58)", "(81, 82)",
"(135, 169)", "(344, 319)"
],
#_boxes = ['(20, 28)', '(46, 54)', '(74, 116)', '(81, 82)', '(135, 169)', '(344, 319)'],
#_boxes = ["(10, 13)", "(16, 30)", "(33, 23)","(30, 61)", "(62, 45)", "(59, 119)","(116, 90)", "(156, 198)", "(373, 326)"],
#_boxes = ['(12, 16)', '(19, 36)', '(40, 28)', '(36, 75)','(76, 55)', '(72, 146)', '(142, 110)', '(192, 243)','(459, 401)'],
filter=exp_cfg.YoloLossLayer(use_nms=False)))
name = f"detect-{input_size[0]}.tflite"
else:
config = exp_cfg.YoloTask(model=exp_cfg.Yolo(
_input_size=input_size,
base="v3",
min_level=3,
norm_activation=exp_cfg.common.NormActivation(
activation="leaky"),
_boxes=[
"(10, 13)", "(16, 30)", "(33, 23)", "(30, 61)", "(62, 45)",
"(59, 119)", "(116, 90)", "(156, 198)", "(373, 326)"
],
filter=exp_cfg.YoloLossLayer(use_nms=False)))
name = f"detect-large-{input_size[0]}.tflite"
task = YoloTask(config)
model = task.build_model()
task.initialize(model)
model(tf.ones((1, *input_size), dtype=tf.float32), training=False)
return model, name
def test_yolo_input_task():
with tf.device('/CPU:0'):
config = yolocfg.YoloTask(model=yolocfg.Yolo(
base='v4',
min_level=3,
norm_activation=yolocfg.common.NormActivation(activation='mish'),
#norm_activation = yolocfg.common.NormActivation(activation="leaky"),
#_boxes = ['(10, 14)', '(23, 27)', '(37, 58)', '(81, 82)', '(135, 169)', '(344, 319)'],
#_boxes = ["(10, 13)", "(16, 30)", "(33, 23)","(30, 61)", "(62, 45)", "(59, 119)","(116, 90)", "(156, 198)", "(373, 326)"],
_boxes=[
'(12, 16)', '(19, 36)', '(40, 28)', '(36, 75)', '(76, 55)',
'(72, 146)', '(142, 110)', '(192, 243)', '(459, 401)'
],
filter=yolocfg.YoloLossLayer(use_nms=False)))
task = yolo.YoloTask(config)
# loading both causes issues, but oen at a time is not issue, why?
train_data = task.build_inputs(config.train_data)
test_data = task.build_inputs(config.validation_data)
return train_data, test_data
mixed_precision.set_policy('mixed_float16')
# mixed_precision.set_policy("float32")
config = exp_cfg.YoloTask(
model=exp_cfg.Yolo(
base='v4',
min_level=3,
norm_activation=exp_cfg.common.NormActivation(activation='mish'),
#norm_activation = exp_cfg.common.NormActivation(activation="leaky"),
#_boxes = ['(10, 14)', '(23, 27)', '(37, 58)', '(81, 82)', '(135, 169)', '(344, 319)'],
#_boxes = ["(10, 13)", "(16, 30)", "(33, 23)","(30, 61)", "(62, 45)", "(59, 119)","(116, 90)", "(156, 198)", "(373, 326)"],
_boxes=[
'(12, 16)', '(19, 36)', '(40, 28)', '(36, 75)', '(76, 55)',
'(72, 146)', '(142, 110)', '(192, 243)', '(459, 401)'
],
filter=exp_cfg.YoloLossLayer(use_nms=False)))
# config = exp_cfg.YoloTask(model=exp_cfg.Yolo(base='v3',
# min_level=3,
# #norm_activation = exp_cfg.common.NormActivation(activation="mish"),
# norm_activation = exp_cfg.common.NormActivation(activation="leaky"),
# #_boxes = ['(10, 14)', '(23, 27)', '(37, 58)', '(81, 82)', '(135, 169)', '(344, 319)'],
# #_boxes = ["(10, 13)", "(16, 30)", "(33, 23)","(30, 61)", "(62, 45)", "(59, 119)","(116, 90)", "(156, 198)", "(373, 326)"],
# _boxes = ['(12, 16)', '(19, 36)', '(40, 28)', '(36, 75)','(76, 55)', '(72, 146)', '(142, 110)', '(192, 243)','(459, 401)'],
# filter = exp_cfg.YoloLossLayer(use_nms=False)
# ))
task = YoloTask(config)
model = task.build_model()
task.initialize(model)
# model(tf.ones((1, 416, 416, 3), dtype = tf.float32))
from yolo.utils.training import schedule
from yolo.utils.training import history_full_callback as hfc
from yolo.tasks.yolo import YoloTask
config = exp_cfg.YoloTask(
model=exp_cfg.Yolo(
base="v4",
min_level=3,
norm_activation=exp_cfg.common.NormActivation(activation="mish"),
#norm_activation = exp_cfg.common.NormActivation(activation="leaky"),
#_boxes = ['(10, 14)', '(23, 27)', '(37, 58)', '(81, 82)', '(135, 169)', '(344, 319)'],
_boxes=[
"(12, 16)", "(19, 36)", "(40, 28)", "(36, 75)", "(76, 55)",
"(72, 146)", "(142, 110)", "(192, 243)", "(459, 401)"
],
filter=exp_cfg.YoloLossLayer(use_nms=True)),
darknet_load_decoder=False)
task = YoloTask(config)
trainer = Trainer(task)
train, test = trainer.build_datasets()
size = tf.data.experimental.cardinality(train).numpy()
tsize = tf.data.experimental.cardinality(test).numpy()
EPOCHS = 1
step_changes = [int(size * EPOCHS * 0.8), int(size * EPOCHS * 0.9)]
schedule = schedule.Schedule(1000,
0.001,
total_steps=size * EPOCHS,
type="STEPWISE",
step_changes=step_changes,