yolov3-spp.cfg Support

[happog]

Could you add support for yolov3-spp model? https://github.com/pjreddie/darknet/blob/master/cfg/yolov3-spp.cfg

[glenn-jocher]

@happog I added yolov3-spp.cfg in commit b93839dea7bf814ed0374fbe01dccd5f4c251c28. You just need to specify it in train.py before you start training now. I'm unfamiliar with the SPP variant, what does it do?

https://github.com/ultralytics/yolov3/blob/b93839dea7bf814ed0374fbe01dccd5f4c251c28/train.py#L12

[happog]

@glenn-jocher Thank you for answering. SPP means spatial pyramid pooling. As the spp cfg needs the maxpool (stride=1) layer, so it should not work correctly if this layer is not implemented in the model.py file.

[glenn-jocher]

@happog models.py correctly implements any of the following layers in any *.cfg file. If the SPP file only uses these existing types, then everything should be fine. If the SPP is introducing a new type of layer then there could be a problem (if it is not recognized it may be skipped). When I start training with SPP I see 3 new layers added, layer count goes from 221 to 224. How many layers are you expecting the SPP cfg to produce?

• convolutional
• batch_norm
• upsample
• route
• shortcut
• yolo

[nirbenz]

@happog @glenn-jocher

Re-opening this. I gave this a shot, trying to simply add support for a maxpool layer as it should be (in models.py):

        elif module_def['type'] == 'maxpool':
            stride = int(module_def['stride'])
            kernel_size = int(module_def['size'])
            pad = (kernel_size - 1) / 2
            maxpool_layer = nn.MaxPool2d(stride=stride, kernel_size=kernel_size, padding=pad)
            modules.add_module('maxpool_%d' % i, maxpool_layer)

And this indeed works, however network performance is actually significantly worse so I'm obviously doing something wrong.

[glenn-jocher]

I added support for yolov3-tiny using the maxpool layer (and implementing a fix to the routing layer, and modifying the stride calculation in models.py yolo forward) in https://github.com/ultralytics/yolov3/issues/51, so if yolov3-spp was suffering from issues before perhaps these are now resolved as well. I'll try to run inference with yolov3-spp and see the result now:

python3 detect.py --cfg cfg/yolov3.cfg --weights weights/yolov3.weights

python3 detect.py --cfg cfg/yolov3-spp.cfg --weights weights/yolov3-spp.weights

Ok, results look comparable to yolov3. Well SPP certainly seems to be working, though this limited anecdotal evidence makes it hard to conclude whether it represents an improvement over standard yolov3, or if there are pytorch implementation issues. To resolve implementation issues I must compare against darknet directly. I'll try that now. Hmm, ok still inconclusive.

./darknet detect cfg/yolov3.cfg yolov3.weights data/...

./darknet detect cfg/yolov3-spp.cfg yolov3-spp.weights data/...

Strangely enough tie used to get 1.00 also using the repo originally.

[LukeAI]

What was the conclusion here? Does Yolov3-SPP work now with reasonable performance and precision?

[glenn-jocher]

@LukeAI works perfectly. Yolov3-SPP is the default model for all operations now in train.py, detect.py and test.py. It acheives .607 mAP as you can see: https://github.com/ultralytics/yolov3#map.

[LukeAI]

Ah yes, thankyou! Why do you think this repo. achieves slightly higher accuracy?

[sanazss]

should we use yolove-1cls or yolov3-spp-1cls for single class and image size of 512*512? does it make difference? all default size in the code is 416 which is in yolov3-1cl

[glenn-jocher]

@sanazss yolov3-spp-1cls.cfg will give you better results for training single class models.