Yet-another-efficientdet-pytorch: can not load model weights

[wanghuajia]

Thanks for sharing your code. When I loaded the code weights, I found that the dimensions were wrong, but I strictly followed your code to load.

`
def get. net():

 nun_classes = 7
 anchors_ratios = '[(1.0， 1.0)，(1.4, 0.7)，(0.7, 1.4)]'
 anchors_scales = '[2 ** 0, 2 ** (1.0 1 3.0), 2 ** (2.0 / 3.0)] '
 compound_coef = 2
 my_model = EfficientDetBackbone(nun_ classes = nun_ classes, compound_coef = 
                       compound_coef, ratios = eval(anchors_ ratios), scales = eval(anchor))
weights_path = ' ./efficientdet-d2 . pth'
my_model.load_state_dict(torch. load(weights_ path), strict = False)

return my_model

if __ name__== "__ main__" :
model = get_net()
`

RuntineError: Error(s) in loading state_ dict for EfficientDetBackbone:
size mismatch for classifier .header . pointwise_ _conv . conv . weight :
copying a paran with shape torch.size([810， 112，1，1]) from
checkpoint, the shape in current moJdeil is torch. size([63, 112, 1, 1]).
size mismatch for classifier .header . pointwise_ conv. conv. bias:
copying a param with shape torch. Size([810]) from checkpoint, the shape
in current model is torch. Size([63]).

[zylo117]

Please provide more info

[wanghuajia]

Please provide more info

OK, as shown above

[zylo117]

There is no method or function named 'get_net' in this repo.

[prfans]

You can try this.

state_dict = torch.load(weights_path)
state_dict.pop('classifier.header.pointwise_conv.conv.weight')
state_dict.pop('classifier.header.pointwise_conv.conv.bias')
model.load_state_dict(state_dict, strict=False)

[plusczh]

check num_classes or obj_list
810 = 9*90 = num_anchors * num_classes

[xtzd]

I guess you use your own dataset but you wanted to use pretrained model. The point is num_classes is not the same. According to your issues, your dataset seems to have 7 classes, but coco pretrained model have 80 classes. Maybe you can try to use only the backbone.

[zylo117]

It doesn't matter, the training program will skip the classifier's header if num_classes don't match while loading weights.

[BenBerCao]

I used my own data to train(two classes) and got similar problem:

Traceback (most recent call last):
File "coco_eval.py", line 156, in
model.load_state_dict(torch.load(weights_path, map_location=torch.device('cpu')))
File "C:UsersASUSAnaconda3envspytorch-gpulibsite-packagestorchnnmodulesmodule.py", line 830, in load_state_dict
self.__class__.__name__, "nt".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for EfficientDetBackbone:

It confused me,how to solve this?

[zylo117]

pls provide more info

[BenBerCao]

Thank you for your reply

Traceback (most recent call last):
File "coco_eval.py", line 156, in
model.load_state_dict(torch.load(weights_path, map_location=torch.device('cpu')))
File "C:UsersASUSAnaconda3envspytorch-gpulibsite-packagestorchnnmodulesmodule.py", line 830, in load_state_dict
self.__class__.__name__, "nt".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for EfficientDetBackbone:
Missing key(s) in state_dict: "bifpn.4.p6_w1", "bifpn.4.p5_w1", "bifpn.4.p4_w1", "bifpn.4.p3_w1", "bifpn.4.p4_w2", "bifpn.4.p5_w2", "bifpn.4.p6_w2", "bifpn.4.p7_w2", "bifpn.4.conv6_up.depthwise_conv.conv.weight", "bifpn.4.conv6_up.pointwise_conv.conv.weight", "bifpn.4.conv6_up.pointwise_conv.conv.bias", "bifpn.4.conv6_up.bn.weight", "bifpn.4.conv6_up.bn.bias", "bifpn.4.conv6_up.bn.running_mean", "bifpn.4.conv6_up.bn.running_var", "bifpn.4.conv5_up.depthwise_conv.conv.weight", "bifpn.4.conv5_up.pointwise_conv.conv.weight", "bifpn.4.conv5_up.pointwise_conv.conv.bias", "bifpn.4.conv5_up.bn.weight", "bifpn.4.conv5_up.bn.bias", "bifpn.4.conv5_up.bn.running_mean", "bifpn.4.conv5_up.bn.running_var", "bifpn.4.conv4_up.depthwise_conv.conv.weight", "bifpn.4.conv4_up.pointwise_conv.conv.weight", "bifpn.4.conv4_up.pointwise_conv.conv.bias", "bifpn.4.conv4_up.bn.weight", "bifpn.4.conv4_up.bn.bias", "bifpn.4.conv4_up.bn.running_mean", "bifpn.4.conv4_up.bn.running_var", "bifpn.4.conv3_up.depthwise_conv.conv.weight", "bifpn.4.conv3_up.pointwise_conv.conv.weight", "bifpn.4.conv3_up.pointwise_conv.conv.bias", "bifpn.4.conv3_up.bn.weight", "bifpn.4.conv3_up.bn.bias", "bifpn.4.conv3_up.bn.running_mean", "bifpn.4.conv3_up.bn.running_var", "bifpn.4.conv4_down.depthwise_conv.conv.weight", "bifpn.4.conv4_down.pointwise_conv.conv.weight", "bifpn.4.conv4_down.pointwise_conv.conv.bias", "bifpn.4.conv4_down.bn.weight", "bifpn.4.conv4_down.bn.bias", "bifpn.4.conv4_down.bn.running_mean", "bifpn.4.conv4_down.bn.running_var", "bifpn.4.conv5_down.depthwise_conv.conv.weight", "bifpn.4.conv5_down.pointwise_conv.conv.weight", "bifpn.4.conv5_down.pointwise_conv.conv.bias",

the rest is too long
this happened when I run coco_eval.py(used the model I trained)

[qtw1998]

pls provide more info
@BenBerCao I met the same problems @wanghuajia

[BenBerCao]

pls provide more info
@BenBerCao I met the same problems @wanghuajia

I tried auther's EfficientDet Training On A Custom Dataset(https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch/blob/master/tutorial/train_shape.ipynb) , I trained it and evalled successfully, However,when I evalled my own model ,I still have this problem, if you have any solutions,please reply to me,thanks a lot.

[ManojKesani]

pls provide more info
@BenBerCao I met the same problems @wanghuajia

I tried auther's EfficientDet Training On A Custom Dataset(https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch/blob/master/tutorial/train_shape.ipynb) , I trained it and evalled successfully, However,when I evalled my own model ,I still have this problem, if you have any solutions,please reply to me,thanks a lot.

i met the same problem as well,
After training on a custom dataset with 2 objects . I used the efficientdet_test_video.py script to test the models. D0 seems to work fine but D1,D2,.. gives this error.

[williamscott701]

any fix?

[zylo117]

you need to specify the network architecture before inference, loading a d1 weights into a d0 network is not possible

[ManojKesani]

Go through your imports , some where in them you might be still using d0 rather than d1,d2....
Change it as required and it might solve this.

[williamscott701]

Go through your imports , some where in them you might be still using d0 rather than d1,d2....
Change it as required and it might solve this.

Found it, Thank you ;)