keras-YOLOv3-mobilenet
keras-YOLOv3-mobilenet copied to clipboard
Adamdad
no test result
as i trained the model on widerface dataset for face detection, the loss 875/875 [==============================] - 1327s 2s/step - loss: 13.1126 - val_loss: 13.7300 loos not decrease , and test no bbox output result.
do you know the reason
one drawback of this implementation is that the loss can not reach 0 due to l2 regularization. You should check your threshold and the data format.
my dataset format is /home/data/JPEGImages/34--Baseball_34_Baseball_Baseball_34_465.jpg 227,250,256,281,0 928,938,75,87,0 in one line each image and threshold is default as score_threshold=.6, iou_threshold=.5 also i just set one classes and label is 0
you should try a smaller score_threshold(0.1 or 0.01 for example) to see if there is any result
i tried 0.1,as some has result, but seems not correct, so i retrained based on the final model, but loss keep on >13 , so how can i reduce the loss
input_shape = (416,416) batch_size = 16 learning rate as use your default not changed lr=1e-3 (freeze) lr=1e-4(unfreeze)
do you train when some layers are freezing？You should unfreeze all the layers when training.
also what about the memory usage with your GPU , as i tested with 980TI > 5GB
You should train without frozen layers, so the change the code to 'False' in the frozen part
` if False: model.compile(optimizer=Adam(lr=1e-3), loss={ # use custom yolo_loss Lambda layer. 'yolo_loss': lambda y_true, y_pred: y_pred}) batch_size = 16 print('Train on {} samples, val on {} samples, with batch size {}.'.format(num_train, num_val, batch_size)) model.fit_generator(data_generator_wrapper(t_lines, batch_size, input_shape, anchors, num_classes), steps_per_epoch=max(1, num_train//batch_size), validation_data=data_generator_wrapper(v_lines, batch_size, input_shape, anchors, num_classes), validation_steps=max(1, num_val//batch_size), epochs=30, initial_epoch=0, callbacks=[logging, checkpoint]) model.save_weights(log_dir + 'trained_weights_stage_1.h5')
The GPU size is ok, because the batch size when testing is set to 1. So just 2GB of memory is enough.
first time i trained with both frozen and unfrozen, second time ,i retrain with set the frozen as false like you said above based on the first step result model
but loss still not descrease
as the dataset i trained on yolo v3 dosn't has such issue , and loss is ~ 1.8
I am not pretty sure about the reason about the bad performance, but keras implementation has such fault that the loss cannot be as low as darknet implementation. A higher loss does not mean bad performance naturally, because the keras implementation use L2 regularization for so the loss the we see is naturally bigger than the real one.
as i just want a good performance ,but now seems not , thanks for your replay, i will check it more
as i just want a good performance ,but now seems not , thanks for your replay, i will check it more
hi kkang,have you solved the problem yet? I used the same WiderFace set as you and my train and val loss decrease to ~7 in 5th epoch. But my test result is very bad too. If you have any information about the problem, please contact me.
I am not pretty sure about the reason about the bad performance, but keras implementation has such fault that the loss cannot be as low as darknet implementation. A higher loss does not mean bad performance naturally, because the keras implementation use L2 regularization for so the loss the we see is naturally bigger than the real one.
hi, Adamad.I have the same problem as kkang and perhaps because we used the same WiderFace dataset. My train and val loss decrease to ~7 in 5th freeze epoch. But my test result is very bad too. Most score below 0.1 and the bbox is not correct too. My train data is organized by the following format: D:/WIDER_train/images/0--Parade/0_Parade_marchingband_1_849.jpg 449,330,122,149,0 150,262,1276,74,1 and my parameters are: input_shape = (416,416) batch_size = 16 lr=1e-3 (freeze) lr=1e-4(unfreeze)