Change Backbone to Swin Transformer about transreid

[sssqfaa]

Hello, I modified the code according to the method you said, but it still failed to reach the result you showed. I suspect it may be related to the learning rate. Could you please tell me the learning rate and batchsize you have used？

[Soar-Sir]

Hello, I modified the code according to the method you said, but it still failed to reach the result you showed. I suspect it may be related to the learning rate. Could you please tell me the learning rate and batchsize you have used？

I also encountered the same situation, the accuracy of the swin model is not as high as expected.

[michuanhaohao]

Hi.

Swin transformer performs well in person ReID. There are some results:

Model	Market	Duke
Vit-base (Imagenet21k pre-training)	86.8/94.4	78.9/89.3
Swin-base (Imagenet21k pre-training)	89.3/95.4	81.2/90.4

[Soar-Sir]

Thanks! I want to know what codes you have changed？After I changed Backbone to swin-transformer, at the same time, I changed some codes in make_model.py, but the improvement effect is not obvious. Maybe my parameters are not set properly? If possible, can you share your files? It will be very helpful to me. Thanks again.

[michuanhaohao]

I only changed the window size for some special resolutions. You can try 224*224 resolution at first., which needs no change. The pre-trained models are pre-trained on ImageNet21K.

def pre_settings(img_size, drop_rate, attn_drop_rate, drop_path_rate):
    if img_size in ([256,128],[256,256],[384,128]):
        window_size = 8
    elif img_size in ([224,224],[224,112]):
        window_size =7
    elif img_size in ([384,384],[384,192]):
        window_size = 12
    elif img_size in ([192,192],):
        window_size = 6
    else:
        print('Window size dose not match!')
    print('Window size is set to %d'%window_size)
    print('using drop_out rate is : {}'.format(drop_rate))
    print('using attn_drop_out rate is : {}'.format(attn_drop_rate))
    print('using drop_path rate is : {}'.format(drop_path_rate))
    return window_size

def swin_base_patch4_window7_224(img_size=224,drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.1,camera_num=0, view_num=0, **kwargs):
    window_size = pre_settings(img_size, drop_rate, attn_drop_rate, drop_path_rate)
    model = SwinTransformer(img_size = img_size, patch_size=4, window_size=window_size, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32), drop_path_rate=drop_path_rate, drop_rate=drop_rate, attn_drop_rate=attn_drop_rate, **kwargs)
    return model

def swin_small_patch4_window7_224(img_size=224,drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.1,camera_num=0, view_num=0, **kwargs):
    window_size = pre_settings(img_size, drop_rate, attn_drop_rate, drop_path_rate)
    model = SwinTransformer(img_size = img_size, patch_size=4, window_size=window_size, embed_dim=96, depths=(2, 2, 18, 2), num_heads=(3, 6, 12, 24), drop_path_rate=drop_path_rate, drop_rate=drop_rate, attn_drop_rate=attn_drop_rate, **kwargs)
    return model

[CarrieYpi]

I only changed the window size for some special resolutions. You can try 224*224 resolution at first., which needs no change. The pre-trained models are pre-trained on ImageNet21K.

def pre_settings(img_size, drop_rate, attn_drop_rate, drop_path_rate):
    if img_size in ([256,128],[256,256],[384,128]):
        window_size = 8
    elif img_size in ([224,224],[224,112]):
        window_size =7
    elif img_size in ([384,384],[384,192]):
        window_size = 12
    elif img_size in ([192,192],):
        window_size = 6
    else:
        print('Window size dose not match!')
    print('Window size is set to %d'%window_size)
    print('using drop_out rate is : {}'.format(drop_rate))
    print('using attn_drop_out rate is : {}'.format(attn_drop_rate))
    print('using drop_path rate is : {}'.format(drop_path_rate))
    return window_size

def swin_base_patch4_window7_224(img_size=224,drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.1,camera_num=0, view_num=0, **kwargs):
    window_size = pre_settings(img_size, drop_rate, attn_drop_rate, drop_path_rate)
    model = SwinTransformer(img_size = img_size, patch_size=4, window_size=window_size, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32), drop_path_rate=drop_path_rate, drop_rate=drop_rate, attn_drop_rate=attn_drop_rate, **kwargs)
    return model

def swin_small_patch4_window7_224(img_size=224,drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.1,camera_num=0, view_num=0, **kwargs):
    window_size = pre_settings(img_size, drop_rate, attn_drop_rate, drop_path_rate)
    model = SwinTransformer(img_size = img_size, patch_size=4, window_size=window_size, embed_dim=96, depths=(2, 2, 18, 2), num_heads=(3, 6, 12, 24), drop_path_rate=drop_path_rate, drop_rate=drop_rate, attn_drop_rate=attn_drop_rate, **kwargs)
    return model

我也在更换为swin-transformer时,遇见了问题，代码中我是加载了预训练模型但似乎没有加载上，效果非常差，请问您能够分享更改为swin transformer完整的backbone代码以及make_model中的代码吗，非常感谢！！