def validation(self):
valid_data = parse_annotation(YoloParams.VALIDATION_ANN_PATH,
YoloParams.VALIDATION_IMG_PATH)
generator = YoloDataGenerator(valid_data, shuffle=True)
yolo_eval = YoloEvaluate(generator=generator, model=self.inf_model)
AP = yolo_eval.comp_map()
_AP_items = [[class_label, ap] for class_label, ap in AP.items()]
AP_items = sorted(_AP_items, key=lambda x: x[1], reverse=True)
for class_label, ap in AP_items:
print("AP( %s ): %.3f" % (class_label, ap))
# Store AP results as csv
#df_ap = pd.DataFrame.from_dict(AP, orient='index')
#df_ap.loc['mAP'] = df_ap.mean()
#df_ap.to_csv('validation_maP.csv', header=False)
print('-------------------------------')
print("mAP: %.3f" % (np.mean(list(AP.values()))))
return AP
def gen_anchors(fname):
imgs = parse_annotation(YoloParams.TRAIN_ANN_PATH,
YoloParams.TRAIN_IMG_PATH)
data_wh = []
for img in imgs:
data_wh += exrtract_wh(img)
clustering = KMeans(NUM_CENTROIDS)
centroids, _ = clustering.fit(np.array(data_wh))
anchors = list(centroids.flatten())
anchors_text = "".join(["%.5f, "%a \
if i < len(anchors)-1 else "%.5f"%a for i,a in enumerate(anchors)])
fname = fname if fname != 'custom_' else 'custom_anchors.txt'
with open(fname, 'w') as f:
f.write("%s" % anchors_text)
print("\nAnchors: \n")
print(anchors_text)
print("\n\tSored at: %s\n" % (fname))
return anchors
def training(self):
train_data = parse_annotation(YoloParams.TRAIN_ANN_PATH,
YoloParams.TRAIN_IMG_PATH)
valid_data = parse_annotation(YoloParams.VALIDATION_ANN_PATH,
YoloParams.VALIDATION_IMG_PATH)
train_gen = YoloDataGenerator(train_data, shuffle=True, augment=True)
valid_gen = YoloDataGenerator(valid_data, shuffle=True)
early_stop = EarlyStopping(monitor='val_loss',
min_delta=0.001,
patience=3,
mode='min',
verbose=1)
log_path = setup_logging()
checkpoint_path = os.path.join(log_path, self.trained_model_name)
checkpoint = ModelCheckpoint(
checkpoint_path,
#monitor='val_loss',
monitor='loss',
verbose=1,
save_best_only=True,
mode='min',
period=1)
#tb_path = os.path.join(log_path, )
tensorboard = YoloTensorBoard(log_dir=log_path,
histogram_freq=0,
write_graph=True,
write_images=False)
"""
optimizer = SGD(
lr=YoloParams.L_RATE,
momentum=0.9,
decay=0.0005
)
"""
optimizer = Adam(lr=YoloParams.L_RATE,
beta_1=0.9,
beta_2=0.999,
decay=0.0)
map_cbck = Callback_MAP(generator=valid_gen,
model=self.inf_model,
tensorboard=tensorboard)
yolo_recall.__name__ = 'recall'
metrics = [
self.yolo_loss.l_coord, self.yolo_loss.l_obj,
self.yolo_loss.l_class, yolo_recall
]
self.model.compile(loss=self.yolo_loss,
optimizer=optimizer,
metrics=metrics)
self.model.fit_generator(
generator=train_gen,
steps_per_epoch=len(train_gen),
verbose=YoloParams.TRAIN_VERBOSE,
validation_data=valid_gen,
validation_steps=len(valid_gen),
callbacks=[early_stop, checkpoint, tensorboard, map_cbck],
#callbacks=[tensorboard, map_cbck],
epochs=YoloParams.NUM_EPOCHS,
max_queue_size=20)