Darknet: How to train my data with densnet201?

[ChuckGithub]

[wangqiqi]

the same question, how to use the densnet201 weights and how to train

[AlexeyAB]

To use densenet201:

• Update code from this repo
• download: https://pjreddie.com/media/files/densenet201.weights
• and run: build\darknet\x64\classifier_densenet201.cmd
  or use this command: darknet.exe classifier predict cfg/imagenet1k.data cfg/densenet201.cfg densenet201.weights
• enter image filename: dog.jpg

[AlexeyAB]

@ChuckGithub @wangqiqi

So, if you want to train Densenet201 for object detection as Yolo (not for classify) then:

1. Download pre-trained weights densenet201.300 https://drive.google.com/open?id=0BwRgzHpNbsWBOFpORVB4UUtfT0U

2. Download cfg-file densenet201_yolo.cfg (+2 conv-layers): https://drive.google.com/open?id=0BwRgzHpNbsWBMXJLX3hHZllFdzQ
   Or larger densenet201_yolo2.cfg (+3 conv-layers): https://drive.google.com/open?id=0BwRgzHpNbsWBeTlpajNWc21jZ0k

3. Make changes in densenet201_yolo.cfg the same as described here for yolo-voc.2.0.cfg: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

   • i.e. change filters=125 in the last convolutional layer: https://github.com/AlexeyAB/darknet/blob/c48f8c157361e8d2edbf9af00a587ceb3be37059/build/darknet/x64/densenet201_yolo.cfg#L1944

   • change classes=20: https://github.com/AlexeyAB/darknet/blob/c48f8c157361e8d2edbf9af00a587ceb3be37059/build/darknet/x64/densenet201_yolo.cfg#L1951

4. Prepare your images dataset and launch training: darknet.exe detector train data/obj.data densenet201_yolo.cfg densenet201.300

5. When training iterations is equal to ~2000*number_of_objects or more, then stop training: https://github.com/AlexeyAB/darknet#when-should-i-stop-training

   And run detection: darknet.exe detector test data/obj.data densenet201_yolo.cfg backup/densenet201_yolo_300.weights -thresh 0.24

---

For classification task the network Densenet201 has higher precision than Darknet_448 (on which Yolo v2 is based), so probably you can achive more precision using densenet201_yolo.cfg instead of yolo-voc.cfg/yolo-voc.2.0.cfg: https://pjreddie.com/darknet/imagenet/

Probably Densenet201 should be a good solution for small and big objects - this network concatinate unchanged features from many scales using [route]-layers, so Densenet201 at the end has information about ~100 different scales.

But densenet201_yolo.cfg ~2x slower than yolo-voc.cfg.
(if you want more speed, try to decrease width=416 and height=416 in cfg to lower value multiple of 32)

---

• https://arxiv.org/abs/1707.06990

  On the ImageNet ILSVRC classification dataset, this large DenseNet obtains a state-of-the-art single-crop top-1 error of 20.26%.

• https://arxiv.org/abs/1608.06993

Whereas traditional convolutional networks with L layers have L connections - one between each layer and its subsequent layer - our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less memory and computation to achieve high performance. Code and models are available at this https URL - CVPR 2017 (Best Paper Award): https://github.com/liuzhuang13/DenseNet

• https://arxiv.org/abs/1611.09326

The idea of DenseNets is based on the observation that if each layer is directly connected to every other layer in a feed-forward fashion then the network will be more accurate and easier to train.
In this paper, we extend DenseNets to deal with the problem of semantic segmentation. We achieve state-of-the-art results on urban scene benchmark datasets such as CamVid and Gatech, without any further post-processing module nor pretraining. Moreover, due to smart construction of the model, our approach has much less parameters than currently published best entries for these datasets.
Code to reproduce the experiments is available here : https://github.com/SimJeg/FC-DenseNet/blob/master/train.py

[wangqiqi]

Thank you, @AlexeyAB , I will try DenseNet.

I have another question, have you ever heard or used mask image to speed the inference? The mask image maybe a white-black image, the black pixels needn't do conv ?

[wangqiqi]

hello, @AlexeyAB, the densenet201_yolo_300.weights seems to be not good as the image below, because the iter is only 300, right?

[AlexeyAB]

@wangqiqi Hi,

Yes, you should train about ~2000 * number_of_classes iterations.

Also read: https://github.com/AlexeyAB/darknet/issues/196#issuecomment-329914554

You can use one of these cfg-files:

1. cfg-file densenet201_yolo.cfg (+2 conv-layers): https://drive.google.com/open?id=0BwRgzHpNbsWBMXJLX3hHZllFdzQ
2. Or larger densenet201_yolo2.cfg (+3 conv-layers): https://drive.google.com/open?id=0BwRgzHpNbsWBeTlpajNWc21jZ0k

---

I have another question, have you ever heard or used mask image to speed the inference? The mask image maybe a white-black image, the black pixels needn't do conv ?

Yes, I heared it. Darknet doesn't support it, but maybe someday I'll add this.

[AlexeyAB]

I updated links to densenet201_yolo.cfg above.

Now for DenseNet201:

1. pre-trained weights densenet201.300 https://drive.google.com/open?id=0BwRgzHpNbsWBOFpORVB4UUtfT0U
2. cfg file - use on of:
   
   
   
   • cfg-file densenet201_yolo.cfg (+2 conv-layers): https://drive.google.com/open?id=0BwRgzHpNbsWBMXJLX3hHZllFdzQ
   
   
   • larger densenet201_yolo2.cfg (+3 conv-layers): https://drive.google.com/open?id=0BwRgzHpNbsWBeTlpajNWc21jZ0k
   
   

---

For ResNet50:

1. pre-trained weights resnet50.65 https://drive.google.com/open?id=0BwRgzHpNbsWBemRnNTRJWjNIc28
2. cfg file - use on of:
   
   
   
   • cfg-file resnet50_yolo.cfg (+2 conv-layers): https://drive.google.com/open?id=0BwRgzHpNbsWBc2MtQXcxV29hTFU
   
   

---

Comparison of the field of view for various networks:
receptive_field.xlsx

| Network name | receptive field of a single activation of the last layer, pixels |
|---|---|
| yolo-voc.cfg | 566 x 566 |
| tiny-yolo-voc.cfg | 318 x 318 |
| yolo9000.cfg | 545 x 545 |
| densenet201.cfg | 1934 x 1934 |
| densenet201-yolo.cfg | ~1934 x 1934 |
| densenet201-yolo2.cfg | ~1934 x 1934 |
| resnet50-yolo.cfg | 76 x 76 |
| resnet50.cfg | 72 x 72 |
| resnet101.cfg | 140 x 140 |
| resnet152.cfg | 208 x 208 |

• resnet50-yolo.cfg can be used only for small objects.
• densenet201-yolo.cfg can be used for small and big objects.
  Also DenseNet201 can be used for very big and very small objects using high resolution detector, for train and detect set in cfg-file: width=832 height=832, or width=1080 height=1080 or width=1920 height=1920 - if you have enough GPU-RAM.

[phongnhhn92]

@AlexeyAB I have a question, how did you calculate the receptive field of a single activation of the last layer in the table above ?

[iraadit]

@AlexeyAB , I have the same question than @phongnhhn92

[wait1988]

How much gpu memeory dose densenet201-yolo need during trainning for input size 416? @AlexeyAB
I change the bese network to densenet201 for yolov3,but I can't train it for limited gpu memory.

[JamesChenChina]

@AlexeyAB
On Windows I compile build\darknet\yolo_cpp_dll.sln to generate C++ DLL-file yolo_cpp_dll.dll.
Now I want to use darknet classifier predict function rather than detector.
Is there an API that I can use?
It seems that https://github.com/AlexeyAB/darknet/blob/master/include/yolo_v2_class.hpp and https://github.com/AlexeyAB/darknet/blob/master/src/yolo_console_dll.cpp only support detector.

Thanks.