Caffe: How to finetune with the VGG_coco_SSD_500x500_iter_200000.caffemodel ？

[yanxp]

I miss the problem Cannot copy param 0 weights from layer 'conv4_3_norm_mbox_loc'; shape mismatch. Source param shape is 16 512 3 3 (73728); target param shape is 12 512 3 3 (55296). To learn this layer's parameters from scratch rather than copying from a saved net, rename the layer.?

[chienyiwang]

You can rename the mbox layer in network by providing the postfix parameter (already implemented in the function)

[yanxp]

@chienyiwang which function ? thank you !

[chienyiwang]

For example, in "ssd_pascal.py" script, the mbox_layers creation:

mbox_layers = CreateMultiBoxHead(net, data_layer='data', from_layers=mbox_source_layers,
use_batchnorm=use_batchnorm, min_sizes=min_sizes, max_sizes=max_sizes,
aspect_ratios=aspect_ratios, normalizations=normalizations,
num_classes=num_classes, share_location=share_location, flip=flip, clip=clip,
prior_variance=prior_variance, kernel_size=3, pad=1)

you can pass in "conf_postfix" and "loc_postfix" in the end to change the name of those layers:

mbox_layers = CreateMultiBoxHead(net, data_layer='data', from_layers=mbox_source_layers,
use_batchnorm=use_batchnorm, min_sizes=min_sizes, max_sizes=max_sizes,
aspect_ratios=aspect_ratios, normalizations=normalizations,
num_classes=num_classes, share_location=share_location, flip=flip, clip=clip,
prior_variance=prior_variance, kernel_size=3, pad=1, conf_postfix='_my_layer', loc_postfix='_my_layer')

[yanxp]

Thank you very much !

[zgplvyou]

@chienyiwang I got the same error with him. I modified "video_ssd_pascal.py" script as your solution. But the output is strange, there are so many bounding boxes (labed in 'train' ' person' ‘shape’ ‘cat’ ...VOC_20_class)in the video.I trained the model with own data just for detecting person , so it should have appeared only two classes(person and background) in the output. I didn't modify the train.prototxt and test.prototxt before training the model.

[Anhaoxu]

@yanxp @chienyiwang
I use this method,but too many bbox appeared, and they are wrong,any suggestions?

[rain2008204]

@zgplvyou I have the same problems, so, how did you solve this? thank you very much.

[unanan]

https://stackoverflow.com/questions/39811936/using-bvlc-googlenet-as-pretrained-model-in-digits-errors/39837047#39837047
This SO page may help?

When you rename the layer, Caffe doesn't try to copy the weights for that layer and you get randomly initialized weights