lddl/yolo-go

Port of Darknet YOLO for Go ecosystem

WIP. PRs are welcome

Port of Darknet YOLO, but via Gorgonia Both YOLOv3 and tiny-YOLOv3 are implemented.

Usage

Navigate to example/yolo-v3 folder and run main.go.

Available flags

go run main.go -h

  -cfg string
        Path to net configuration file (default "../../test_network_data/yolov3-tiny.cfg")
  -image string
        Path to image file for 'detector' mode (default "../../test_network_data/dog_416x416.jpg")
  -mode string
        Choose the mode: detector/training (default "detector")
  -train string
        Path to folder with labeled data (default "../../test_yolo_op_data")
  -weights string
        Path to weights file (default "../../test_network_data/yolov3-tiny.weights")

For testing tiny-yolov3:

go run main.go --mode detector --cfg ../../test_network_data/yolov3-tiny.cfg --weights ../../test_network_data/yolov3-tiny.weights --image ../../test_network_data/dog_416x416.jpg

For testing yolov3:

go run main.go --mode detector --cfg ../../test_network_data/yolov3.cfg --weights ../../test_network_data/yolov3.weights --image ../../test_network_data/dog_416x416.jpg

For training WIP. PRs are welcome:

go run main.go --mode training --cfg ../../test_network_data/yolov3-tiny.cfg --weights ../../test_network_data/yolov3-tiny.weights --image ../../test_network_data/dog_416x416.jpg --train ../../test_yolo_op_data

Weights and configuration

Weights can be downloaded via curl-scripts download_weights_yolo_v3.sh and download_weights_yolo_tiny_v3.sh. Configuration files: yolov3-tiny.cfg and yolov3.cfg

Network Architecture

Tiny-YOLOv3 Architecture is:

Convolution layer: Filters->16 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Maxpooling layer: Size->2 Stride->2
Convolution layer: Filters->32 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Maxpooling layer: Size->2 Stride->2
Convolution layer: Filters->64 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Maxpooling layer: Size->2 Stride->2
Convolution layer: Filters->128 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Maxpooling layer: Size->2 Stride->2
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Maxpooling layer: Size->2 Stride->2
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Maxpooling layer: Size->2 Stride->1
Convolution layer: Filters->1024 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->255 Padding->0 Kernel->1x1 Stride->1 Activation->linear Batch->0 Bias->true
YOLO layer: Mask->3 Anchors->[81, 82]   |       Mask->4 Anchors->[135, 169]     |       Mask->5 Anchors->[344, 319]
Route layer: Start->13
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Upsample layer: Scale->2
Route layer: Start->19 End->8
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->255 Padding->0 Kernel->1x1 Stride->1 Activation->linear Batch->0 Bias->true
YOLO layer: Mask->0 Anchors->[10, 14]   |       Mask->1 Anchors->[23, 27]       |       Mask->2 Anchors->[37, 58] 

YOLOv3 Architecture is:

Convolution layer: Filters->32 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->64 Padding->1 Kernel->3x3 Stride->2 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->32 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->64 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->1 | index to->3
Convolution layer: Filters->128 Padding->1 Kernel->3x3 Stride->2 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->64 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->128 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->5 | index to->7
Convolution layer: Filters->64 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->128 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->8 | index to->10
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->2 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->12 | index to->14
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->15 | index to->17
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->18 | index to->20
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->21 | index to->23
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->24 | index to->26
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->27 | index to->29
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->30 | index to->32
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->33 | index to->35
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->2 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->37 | index to->39
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->40 | index to->42
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->43 | index to->45
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->46 | index to->48
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->49 | index to->51
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->52 | index to->54
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->55 | index to->57
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->58 | index to->60
Convolution layer: Filters->1024 Padding->1 Kernel->3x3 Stride->2 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->1024 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->62 | index to->64
Convolution layer: Filters->512 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->1024 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->65 | index to->67
Convolution layer: Filters->512 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->1024 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->68 | index to->70
Convolution layer: Filters->512 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->1024 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Shortcut layer: index from->71 | index to->73
Convolution layer: Filters->512 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->1024 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->1024 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->1024 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->255 Padding->0 Kernel->1x1 Stride->1 Activation->linear Batch->0 Bias->true
YOLO layer: Mask->6 Anchors->[116, 90]  |       Mask->7 Anchors->[156, 198]     |       Mask->8 Anchors->[373, 326]
Route layer: Start->79
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Upsample layer: Scale->2
Route layer: Start->85 End->61
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->512 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->255 Padding->0 Kernel->1x1 Stride->1 Activation->linear Batch->0 Bias->true
YOLO layer: Mask->3 Anchors->[30, 61]   |       Mask->4 Anchors->[62, 45]       |       Mask->5 Anchors->[59, 119]
Route layer: Start->91
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Upsample layer: Scale->2
Route layer: Start->97 End->36
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->128 Padding->0 Kernel->1x1 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->256 Padding->1 Kernel->3x3 Stride->1 Activation->leaky Batch->1 Bias->false
Convolution layer: Filters->255 Padding->0 Kernel->1x1 Stride->1 Activation->linear Batch->0 Bias->true
YOLO layer: Mask->0 Anchors->[10, 13]   |       Mask->1 Anchors->[16, 30]       |       Mask->2 Anchors->[33, 23]

yolo-go's Issues

[FEATURE REQUEST] Loss function

Is your feature request related to a problem? Please describe.
Current trainer - https://github.com/LdDl/yolo-go/blob/master/yolo_trainer.go#L11
Current loss function for YOLO bounding boxes - https://github.com/LdDl/yolo-go/blob/master/yolo_trainer.go#L100
I have doubts that it is right code (or even approach). Need to implement it step-by-step and compare with classic C-based implementation (may be take inspirations by some tensorflow/pytorch implementation)

Describe the solution you'd like and provide pseudocode examples if you can
According to source code of YOLO:
IOU's - https://github.com/AlexeyAB/darknet/blob/master/src/yolo_layer.c#L166:

float delta_yolo_box(box truth, float *x, float *biases, int n, int index, int i, int j, int lw, int lh, int w, int h, float *delta, float scale, int stride)
{
    box pred = get_yolo_box(x, biases, n, index, i, j, lw, lh, w, h, stride);
    float iou = box_iou(pred, truth);

    float tx = (truth.x*lw - i);
    float ty = (truth.y*lh - j);
    float tw = log(truth.w*w / biases[2*n]);
    float th = log(truth.h*h / biases[2*n + 1]);

    delta[index + 0*stride] = scale * (tx - x[index + 0*stride]);
    delta[index + 1*stride] = scale * (ty - x[index + 1*stride]);
    delta[index + 2*stride] = scale * (tw - x[index + 2*stride]);
    delta[index + 3*stride] = scale * (th - x[index + 3*stride]);
    return iou;
}

classes - https://github.com/AlexeyAB/darknet/blob/master/src/yolo_layer.c#L276:

void delta_yolo_class(float *output, float *delta, int index, int class, int classes, int stride, float *avg_cat)
{
    int n;
    if (delta[index]){
        delta[index + stride*class] = 1 - output[index + stride*class];
        if(avg_cat) *avg_cat += output[index + stride*class];
        return;
    }
    for(n = 0; n < classes; ++n){
        delta[index + stride*n] = ((n == class)?1 : 0) - output[index + stride*n];
        if(n == class && avg_cat) *avg_cat += output[index + stride*n];
    }
}

Describe alternatives you've considered and provide pseudocode examples if you can
nope

Additional context
nope