network.build(input_shape=(None, IMAGE_HEIGHT, IMAGE_WIDTH, CHANNELS))
network.summary()
if __name__ == '__main__':
print("Train start\n")
# GPU settings
gpus = tf.config.list_physical_devices("GPU")
if gpus:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
print("Generate the Data Set\n")
dataset = TFDataset()
train_data, train_count = dataset.generate_datatset()
ssd = SSD()
print_model_summary(network=ssd)
if load_weights_before_training:
ssd.load_weights(filepath=save_model_dir+"epoch-{}".format(load_weights_from_epoch))
print("Successfully load weights!")
else:
load_weights_from_epoch = -1
# loss
loss = SSDLoss()
# optimizer
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(initial_learning_rate=1e-3,
decay_steps=20000,
decay_rate=0.96)
def main():
dataset = TFDataset()
train_data, train_count = dataset.generate_datatset()
model = SSD()
print_model_summary(model)
if load_weights_from_epoch >= 0:
model.load_weights(filepath=save_model_dir +
"epoch-{}".format(load_weights_from_epoch))
print("成功从epoch-{}加载模型权重！".format(load_weights_from_epoch))
loss_fn = MultiBoxLoss(num_classes=NUM_CLASSES,
overlap_thresh=0.5,
neg_pos=3)
# optimizer
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
initial_learning_rate=1e-3, decay_steps=20000, decay_rate=0.96)
optimizer = tf.optimizers.Adam(learning_rate=lr_schedule)
loss_metric = tf.metrics.Mean()
cls_loss_metric = tf.metrics.Mean()
reg_loss_metric = tf.metrics.Mean()
for epoch in range(load_weights_from_epoch + 1, EPOCHS):
start_time = time.time()
for step, batch_data in enumerate(train_data):
images, labels = ReadDataset().read(batch_data)
with tf.GradientTape() as tape:
predictions = model(images, training=True)
loss_l, loss_c = loss_fn(y_true=labels, y_pred=predictions)
total_loss = loss_l + loss_c
gradients = tape.gradient(total_loss, model.trainable_variables)
optimizer.apply_gradients(
grads_and_vars=zip(gradients, model.trainable_variables))
loss_metric.update_state(values=total_loss)
cls_loss_metric.update_state(values=loss_c)
reg_loss_metric.update_state(values=loss_l)
time_per_step = (time.time() - start_time) / (step + 1)
print(
"Epoch: {}/{}, step: {}/{}, speed: {:.2f}s/step, loss: {:.10f}, "
"cls loss: {:.10f}, reg loss: {:.10f}".format(
epoch, EPOCHS, step,
tf.math.ceil(train_count / BATCH_SIZE), time_per_step,
loss_metric.result(), cls_loss_metric.result(),
reg_loss_metric.result()))
loss_metric.reset_states()
cls_loss_metric.reset_states()
reg_loss_metric.reset_states()
if epoch % save_frequency == 0:
model.save_weights(filepath=save_model_dir +
"epoch-{}".format(epoch),
save_format="tf")
if test_images_during_training:
visualize_training_results(pictures=test_images_dir_list,
model=model,
epoch=epoch)
model.save_weights(filepath=save_model_dir + "epoch-{}".format(EPOCHS),
save_format="tf")
if is_object_exist:
image_with_boxes = draw_boxes_on_image(cv2.imread(picture_dir), boxes, scores, classes)
else:
print("No objects were detected.")
image_with_boxes = cv2.imread(picture_dir)
return image_with_boxes
if __name__ == '__main__':
# GPU settings
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
# 텐서플로가 첫 번째 GPU만 사용하도록 제한
try:
print('start with GPU 4')
tf.config.experimental.set_visible_devices(gpus[4], 'GPU')
tf.config.experimental.set_memory_growth(gpus[4], True)
except RuntimeError as e:
# 프로그램 시작시에 접근 가능한 장치가 설정되어야만 합니다
print(e)
ssd_model = SSD()
print_model_summary(network=ssd_model, shape=(None, 224, 224, 3))
ssd_model.load_weights(filepath=save_model_dir+"mobilenet_v2_new-epoch-49.h5")
image = test_single_picture(picture_dir=test_picture_dir, model=ssd_model)
cv2.namedWindow("detect result", flags=cv2.WINDOW_NORMAL)
cv2.imshow("detect result", image)
cv2.waitKey(0)
if __name__ == '__main__':
# GPU settings
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
# 텐서플로가 첫 번째 GPU만 사용하도록 제한
try:
print('start with GPU 4')
tf.config.experimental.set_visible_devices(gpus[4], 'GPU')
tf.config.experimental.set_memory_growth(gpus[4], True)
except RuntimeError as e:
# 프로그램 시작시에 접근 가능한 장치가 설정되어야만 합니다
print(e)
ssd = SSD()
dataset = TFDataset()
train_data, train_count = dataset.generate_datatset()
print_model_summary(network=ssd, shape=(None, 224, 224, 3))
if load_weights_before_training:
ssd.load_weights(filepath=save_model_dir +
"mobilenet_v2_new-epoch-{}.h5".format(49))
print("Successfully load weights!")
else:
load_weights_from_epoch = -1
# loss
loss = SSDLoss()
# optimizer
def test_single_picture(picture_dir, model):
image_tensor = preprocess_image(picture_dir)
image_tensor = tf.expand_dims(image_tensor, axis=0)
procedure = InferenceProcedure(model=model)
is_object_exist, boxes, scores, classes = procedure.get_final_boxes(image=image_tensor)
if is_object_exist:
image_with_boxes = draw_boxes_on_image(cv2.imread(picture_dir), boxes, scores, classes)
else:
print("No objects were detected.")
image_with_boxes = cv2.imread(picture_dir)
return image_with_boxes
if __name__ == '__main__':
# GPU settings
gpus = tf.config.list_physical_devices("GPU")
if gpus:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
ssd_model = SSD()
ssd_model.load_weights(filepath=save_model_dir+"saved_model")
image = test_single_picture(picture_dir=test_picture_dir, model=ssd_model)
cv2.namedWindow("detect result", flags=cv2.WINDOW_NORMAL)
cv2.imshow("detect result", image)
cv2.waitKey(0)
tf.config.experimental.set_memory_growth(gpus[4], True)
except RuntimeError as e:
# 프로그램 시작시에 접근 가능한 장치가 설정되어야만 합니다
print(e)
predicted_dir_path = '../mAP/predicted'
ground_truth_dir_path = '../mAP/ground-truth'
if os.path.exists(predicted_dir_path): shutil.rmtree(predicted_dir_path)
if os.path.exists(ground_truth_dir_path): shutil.rmtree(ground_truth_dir_path)
if os.path.exists("./dataset/detection/"): shutil.rmtree("./dataset/detection/")
os.mkdir(predicted_dir_path)
os.mkdir(ground_truth_dir_path)
os.mkdir("./dataset/detection/")
ssd_model = SSD()
print_model_summary(network=ssd_model, shape=(None, 224, 224, 3))
ssd_model.load_weights(filepath=save_model_dir + "vgg16-epoch-49.h5")
dataset = TFDataset()
test_data, train_count = dataset.generate_datatset()
# image = test_single_picture(picture_dir=test_picture_dir, model=ssd_model)
def test_step(batch_images, batch_labels, name):
print('=> ground truth of %s:' % name[0])
ground_truth_path = os.path.join(ground_truth_dir_path, str(name[0]) + '.txt')
with open(ground_truth_path, 'w') as f:
i = 0
while batch_labels[0][i][4] != -1:
class_name = str(find_class_name(batch_labels[0][i][4]))
boxes = resize_box(boxes, h, w)
classes = tf.cast(results[:, -1], dtype=tf.int32).numpy()
# print("检测结果：boxes: {}, classes: {}, scores: {}".format(boxes, classes, scores))
image_with_boxes = draw_boxes_on_image(image_array, boxes, scores, classes)
return image_with_boxes
if __name__ == '__main__':
gpus = tf.config.list_physical_devices('GPU')
if gpus:
try:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
logical_gpus = tf.config.list_logical_devices('GPU')
print(len(gpus), "Physical GPUs,", len(logical_gpus),
"Logical GPUs")
except RuntimeError as e:
print(e)
ssd_model = SSD()
# 始终加载最新的权重文件
last_epoch = os.listdir(save_model_dir)[-2].split(".")[0]
ssd_model.load_weights(filepath=save_model_dir + last_epoch)
image = test_single_picture(picture_dir=test_picture_dir, model=ssd_model)
cv2.namedWindow("detect result", flags=cv2.WINDOW_NORMAL)
cv2.imshow("detect result", image)
cv2.waitKey(0)
def test_single_picture(picture_dir, model):
image_tensor = preprocess_image(picture_dir)
image_tensor = tf.expand_dims(image_tensor, axis=0)
procedure = InferenceProcedure(model=model)
is_object_exist, boxes, scores, classes = procedure.get_final_boxes(
image=image_tensor)
if is_object_exist:
image_with_boxes = draw_boxes_on_image(cv2.imread(picture_dir), boxes,
scores, classes)
else:
print("No objects were detected.")
image_with_boxes = cv2.imread(picture_dir)
return image_with_boxes
if __name__ == '__main__':
# GPU settings
gpus = tf.config.experimental.list_physical_devices("GPU")
if gpus:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
ssd_model = SSD()
ssd_model.load_weights(filepath=save_model_dir + 'epoch-0')
image = test_single_picture(picture_dir=test_picture_path, model=ssd_model)
cv2.imwrite('dataset/result.png', image)
# cv2.namedWindow("detect result", flags=cv2.WINDOW_NORMAL)
# cv2.imshow("detect result", image)
# cv2.waitKey(0)