How to reproduce the model deeplabv3_xception.h5?

[zhimengf]

Hi, I'm trying to train deeplabv3 with xception backbone on voc + SBD dataset. You provided the voc pretrained model deeplabv3_xception.h5. But If I want to reproduce your training result, I should not use it as pretrained model, right? So I comment out the line in train.py to not loading pretrained weights. But after 100 epochs, my model accuracy is poor, compared with your model. Did I miss something, do I need something like an ImageNet pretrained model or COCO pretrained model? Thanks!

[zhimengf]

@bubbliiiing

[bubbliiiing]

It will be difficult to reproduce the results of xception, and it should require a large memory of the graphics card for training.I remember that the effect of xception as the backbone in the paper should be about 80%. My previous pre training weight comes from https://github.com/bonlime/keras-deeplab-v3-plus. Its feature extraction effect is good.
If you want to start with backbone, you may need to know this https://github.com/keras-team/keras-applications/blob/master/keras_applications/xception.py

[zhimengf]

@bubbliiiing do you mean you used cityscapes pretrained model: https://github.com/bonlime/keras-deeplab-v3-plus/releases/download/1.2/deeplabv3_xception_tf_dim_ordering_tf_kernels_cityscapes.h5 to finetune on pascal voc? Let me try it. I'm not going to reach SOTA, but just to reproduce what you did to generate the voc model myself.

[zhimengf]

thanks! I can get reasonable accuracy with cityscapes pretrained model.

===>background:	mIou-92.38; mPA-95.9
===>aeroplane:	mIou-85.0; mPA-88.47
===>bicycle:	mIou-40.77; mPA-91.49
===>bird:	mIou-82.09; mPA-94.45
===>boat:	mIou-58.61; mPA-86.91
===>bottle:	mIou-71.73; mPA-86.56
===>bus:	mIou-89.62; mPA-91.48
===>car:	mIou-87.12; mPA-93.32
===>cat:	mIou-89.15; mPA-94.4
===>chair:	mIou-34.24; mPA-60.78
===>cow:	mIou-60.49; mPA-64.41
===>diningtable:	mIou-51.83; mPA-59.9
===>dog:	mIou-87.77; mPA-93.22
===>horse:	mIou-81.91; mPA-88.78
===>motorbike:	mIou-81.64; mPA-91.21
===>person:	mIou-82.8; mPA-88.29
===>pottedplant:	mIou-41.24; mPA-78.89
===>sheep:	mIou-84.88; mPA-95.24
===>sofa:	mIou-43.24; mPA-56.91
===>train:	mIou-77.98; mPA-86.84
===>tvmonitor:	mIou-66.68; mPA-81.11
===> mIoU: 71.01; mPA: 84.22

[zhimengf]

closing now.

[bubbliiiing]

3Q