Darknet: Can not load pretrained YOLOv3-spp weights

[YangZhang4065]

Hello!
I saw that the spatial pyramid pooling was recently introduced into YOLOV3 and a new pretrained model was uploaded. However when I downloaded weight from https://pjreddie.com/media/files/yolov3-spp.weights and ran
./darknet detect cfg/yolov3-spp.cfg yolov3-spp.weights data/dog.jpg
I got the following error message

   84 Layer before convolutional layer must output image.: File exists
darknet: ./src/utils.c:256: error: Assertion `0' failed.
Aborted (core dumped)

I have no issue running default yolov3 model with the above code.

[crasmuss]

Looking at src/parse.c, this happens because the h and w don't match up for the 4 different layers (82 80 78 77) feeding into the route layer (83) preceding the convolutional layer. In other instances of a route layer having multiple arguments, h and w are the same for the different sources while only c is allowed to differ (and is summed for the output).

Could there have been a change to parsing that wasn't committed along with this new config file?

[nikkymen]

Confirm, the same issue

[pjreddie]

Sorry, i messed with maxpool layers a bit, should be fixed now from e209b3bbbf52d7cae56b0b8e320ab1bd31a60bf8

[nikkymen]

@pjreddie

Now it`s works, thank you!

[AlexeyAB]

@pjreddie
Does it brake [maxpool]-layer with stride=1 in the yolov3-tiny.cfg?

• It was pad_mode = SAME (in terms of Caffe/Tensorflow) in the Darknet https://github.com/pjreddie/darknet/blob/49ba88d9f73cf80ed657823a80fefb4b929414a5/src/maxpool_layer.c#L30-L31

• But currently with new fix https://github.com/pjreddie/darknet/blob/e209b3bbbf52d7cae56b0b8e320ab1bd31a60bf8/src/maxpool_layer.c#L30-L31

• the pad_mode = VALID (in terms of Caffe/Tensorflow) should be: https://www.tensorflow.org/api_guides/python/nn#Convolution

 l.out_w = (w - size + 1)/stride; 
 l.out_h = (h - size + 1)/stride; 

https://stackoverflow.com/a/37675359/1558037

[pjreddie]

well hm....

[pjreddie]

I have a fix but i'd have to commit everything i've been working on and i'm not sure if it breaks stuff, let me check

[pjreddie]

It is distinctly possible that b13f67bfdd87434e141af532cdb5dc1b8369aa3b fixes this

[AlexeyAB]

@pjreddie Thanks

But why is there int w_offset = -pad/2; in the maxpool_layer_kernels.cu

• https://github.com/pjreddie/darknet/commit/b13f67bfdd87434e141af532cdb5dc1b8369aa3b#diff-9b5ceb8892a4a6227260f670a226d503R27
• https://github.com/pjreddie/darknet/commit/b13f67bfdd87434e141af532cdb5dc1b8369aa3b#diff-9b5ceb8892a4a6227260f670a226d503R69

But there is int w_offset = -l.pad/l.stride; in the maxpool_layer.c

• https://github.com/pjreddie/darknet/commit/b13f67bfdd87434e141af532cdb5dc1b8369aa3b#diff-9f8eac4bd4bbc86cabfe69de7a0a7a60R82

[trn84]

Are you going to update the cfg and weights or are they unaffected?

[smaazs71]

ERROR :
cifar_small
1
First section must be [net] or [network]: No error
assertion "0" failed: file "./src/utils.c", line 256, function: ### ### error
Aborted (core dumped)

still exist while training cifar

[shouryasimha]

Hello,
I am trying to train the model using yolov3 with image resolution 200200 and I'm getting no results when I used the yolov3 .cfg after changing the number of classes. If I change the width and height in the config file from 608 * 608 to 200200 respectively, I'm getting an error
87 Layer before convolutional layer must output image.: File exists
darknet: ./src/utils.c:256: error: Assertion `0' failed.

[AlexeyAB]

@shouryasimha width and height should be multiple of 32 due to 5 subsampling layers. pow(2,5)=32

[shouryasimha]

I have tried with a multiple of 32 , i gave the value as 224 for width and
height but my model is failing to recognize any object. getting my IOU = -nan
i edited the config file as follows:
number of class = 6
steps =12000
filters =33
width = 224
height = 244
and my image width and height are 200 * 200 respectively.

Kindly help me solve this issue.

On Fri, Mar 20, 2020 at 9:43 PM Alexey [e-mail] wrote:

@shouryasimha https://github.com/shouryasimha width and height should
be multiple of 32 due to 5 subsampling layers. pow(2,5)=32

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/pjreddie/darknet/issues/1018#issuecomment-601975429,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/ALKV2F2HZ7Y67JLXL6FAW2LRIQLV3ANCNFSM4FNOWBZQ
.