search

hizhangp / yolo_tensorflow

Tensorflow implementation of YOLO, including training and test phase.
MIT License
800 stars 442 forks source link

Strange phenomenon during my experiment #34

Open feixuetuba opened 7 years ago

feixuetuba commented 7 years ago

@hizhangp I have use the pascal_voc2007 dataset to train the network, and try to use the trained network to do some prediction. I found that whatever image feed to the network, It always generate the same output(both bonding-box and confidence in different image are the same), through the total_loss is quite low. So I review the code, and try to understand what have happened during the training progress, and There are two place is really hard for me to get it through, they both in "loss_layer" First of all, “boxes = tf.tile(boxes, [1, 1, 1, self.boxes_per_cell, 1]) / self.image_size”， why the "boxes" needed to be divided by "image_size" , when the predict_boxes_train only divided by cell_size Secondly, why the predict_box should be plused with "offset" when we calculate the predict_boxes_train,?can we remove it?

ddsrt-sean commented 6 years ago

Hello， Sir I used pascal_voc2007 dataset to train the network with default config. I tried it many times but the Loss always is keeping around 11。the default ckpt(YOLO_small.ckpt) train the network, the loss is 4。 Can you tell me the loss you got with default config, and you can help me to reduce loss ？ Thanks You!

BlankAix commented 6 years ago

@feixuetuba Have you figured out the problem? I hope you can give me some advice about this.

lei9wang commented 6 years ago

@ddsrt-sean , how about your test result with your trained weights? I also trained the network with default config, the loss is about 9. but there is not effect when I test it with my trained weights. I don't know why, is it loss too big?

guker commented 5 years ago

Hello， Sir I used pascal_voc2007 dataset to train the network with default config. I tried it many times but the Loss always is keeping around 11。the default ckpt(YOLO_small.ckpt) train the network, the loss is 4。 Can you tell me the loss you got with default config, and you can help me to reduce loss ？ Thanks You!

I have same problem，and you must pretrain，not train from scrath，otherwise，loss is about 10.0 and does not decrease

guker commented 5 years ago

@ddsrt-sean , how about your test result with your trained weights? I also trained the network with default config, the loss is about 9. but there is not effect when I test it with my trained weights. I don't know why, is it loss too big?

check https://github.com/hizhangp/yolo_tensorflow/issues/18#issue-229843655