search

hhk7734 / tensorflow-yolov4

YOLOv4 Implemented in Tensorflow 2.
MIT License
136 stars 75 forks source link

modify or remove head and predict #47

Closed hhk7734 closed 3 years ago

hhk7734 commented 3 years ago

PAN

... -> PAN -> Dim(batch, y, x, #anchors * (xywh + conf + #classes)) ~#size

Head

raw -> sigmoid -> raw, sig
raw, sig -> split -> Dim(batch, y, x, xywh + conf + #classes) * #anchors
... -> split -> Dim(batch, y, x, #anchors, xy) 
                Dim(batch, y, x, #anchors, wh)
                Dim(batch, y, x, #anchors, conf + #classes)

Dim(batch, y, x, #anchors, xy)  -> scale -> add grid
Dim(batch, y, x, #anchors, wh) -> exp -> mul anchor

... -> concat -> Dim(batch, y, x, xywh + conf + #classes)
... -> concat -> Dim(batch, y, x, #anchors * (xywh + conf + #classes))

Predict

... -> reshape -> Dim(1 , y * x * #anchors, (xywh + conf + #classes)) ~#size
... -> concat(#size) -> Dim(1 , y * x * #anchors ~ #size, (xywh + conf + #classes)) 
... -> find max prob(#classes) -> conf * prob -> -> threshold class_prob
                               -> Dim(1 , threshold, (xywh + class_id + class_prob))
... -> NMS -> result

It can do the head after finding max prob. I will change the order in a direction that can reduce operations or increase speed.

hhk7734 commented 3 years ago

Ref: https://coral.ai/docs/edgetpu/models-intro/#model-requirements

LOGISTIC: tf.keras.activations.sigmoid SPLIT: tf.split(x, num, axis) :x:SPLIT_V: tf.split(x, [num, ...], axis)

hhk7734 commented 3 years ago

https://github.com/AlexeyAB/darknet/wiki/CFG-Parameters-in-the-%5Bnet%5D-section https://github.com/AlexeyAB/darknet/wiki/CFG-Parameters-in-the-different-layers

hhk7734 commented 3 years ago

https://wiki.loliot.net/docs/lang/python/libraries/yolov4/model/python-yolov4-model-backbone/

hhk7734 commented 3 years ago

Do not include fixed type constants to the model. #54

hhk7734 commented 3 years ago

Check: https://github.com/hhk7734/tensorflow-yolov4/blob/4f75970d02b60960257f9f7baa091e0e6874a890/py_src/yolov4/common/predict.py#L27

hhk7734 commented 3 years ago 
 GPU isn't used 
 Used AVX 
 Used FMA & AVX2 
 OpenCV isn't used - data augmentation will be slow 
mini_batch = 1, batch = 8, time_steps = 1, train = 0 
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1    608 x 608 x   3 ->  608 x 608 x  32 0.639 BF
   1 conv     64       3 x 3/ 2    608 x 608 x  32 ->  304 x 304 x  64 3.407 BF
   2 conv     64       1 x 1/ 1    304 x 304 x  64 ->  304 x 304 x  64 0.757 BF
   3 route  1                                  ->  304 x 304 x  64 
   4 conv     64       1 x 1/ 1    304 x 304 x  64 ->  304 x 304 x  64 0.757 BF
   5 conv     32       1 x 1/ 1    304 x 304 x  64 ->  304 x 304 x  32 0.379 BF
   6 conv     64       3 x 3/ 1    304 x 304 x  32 ->  304 x 304 x  64 3.407 BF
   7 Shortcut Layer: 4,  wt = 0, wn = 0, outputs: 304 x 304 x  64 0.006 BF
   8 conv     64       1 x 1/ 1    304 x 304 x  64 ->  304 x 304 x  64 0.757 BF
   9 route  8 2                                ->  304 x 304 x 128 
  10 conv     64       1 x 1/ 1    304 x 304 x 128 ->  304 x 304 x  64 1.514 BF
  11 conv    128       3 x 3/ 2    304 x 304 x  64 ->  152 x 152 x 128 3.407 BF
  12 conv     64       1 x 1/ 1    152 x 152 x 128 ->  152 x 152 x  64 0.379 BF
  13 route  11                                 ->  152 x 152 x 128 
  14 conv     64       1 x 1/ 1    152 x 152 x 128 ->  152 x 152 x  64 0.379 BF
  15 conv     64       1 x 1/ 1    152 x 152 x  64 ->  152 x 152 x  64 0.189 BF
  16 conv     64       3 x 3/ 1    152 x 152 x  64 ->  152 x 152 x  64 1.703 BF
  17 Shortcut Layer: 14,  wt = 0, wn = 0, outputs: 152 x 152 x  64 0.001 BF
  18 conv     64       1 x 1/ 1    152 x 152 x  64 ->  152 x 152 x  64 0.189 BF
  19 conv     64       3 x 3/ 1    152 x 152 x  64 ->  152 x 152 x  64 1.703 BF
  20 Shortcut Layer: 17,  wt = 0, wn = 0, outputs: 152 x 152 x  64 0.001 BF
  21 conv     64       1 x 1/ 1    152 x 152 x  64 ->  152 x 152 x  64 0.189 BF
  22 route  21 12                              ->  152 x 152 x 128 
  23 conv    128       1 x 1/ 1    152 x 152 x 128 ->  152 x 152 x 128 0.757 BF
  24 conv    256       3 x 3/ 2    152 x 152 x 128 ->   76 x  76 x 256 3.407 BF
  25 conv    128       1 x 1/ 1     76 x  76 x 256 ->   76 x  76 x 128 0.379 BF
  26 route  24                                 ->   76 x  76 x 256 
  27 conv    128       1 x 1/ 1     76 x  76 x 256 ->   76 x  76 x 128 0.379 BF
  28 conv    128       1 x 1/ 1     76 x  76 x 128 ->   76 x  76 x 128 0.189 BF
  29 conv    128       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 128 1.703 BF
  30 Shortcut Layer: 27,  wt = 0, wn = 0, outputs:  76 x  76 x 128 0.001 BF
  31 conv    128       1 x 1/ 1     76 x  76 x 128 ->   76 x  76 x 128 0.189 BF
  32 conv    128       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 128 1.703 BF
  33 Shortcut Layer: 30,  wt = 0, wn = 0, outputs:  76 x  76 x 128 0.001 BF
  34 conv    128       1 x 1/ 1     76 x  76 x 128 ->   76 x  76 x 128 0.189 BF
  35 conv    128       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 128 1.703 BF
  36 Shortcut Layer: 33,  wt = 0, wn = 0, outputs:  76 x  76 x 128 0.001 BF
  37 conv    128       1 x 1/ 1     76 x  76 x 128 ->   76 x  76 x 128 0.189 BF
  38 conv    128       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 128 1.703 BF
  39 Shortcut Layer: 36,  wt = 0, wn = 0, outputs:  76 x  76 x 128 0.001 BF
  40 conv    128       1 x 1/ 1     76 x  76 x 128 ->   76 x  76 x 128 0.189 BF
  41 conv    128       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 128 1.703 BF
  42 Shortcut Layer: 39,  wt = 0, wn = 0, outputs:  76 x  76 x 128 0.001 BF
  43 conv    128       1 x 1/ 1     76 x  76 x 128 ->   76 x  76 x 128 0.189 BF
  44 conv    128       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 128 1.703 BF
  45 Shortcut Layer: 42,  wt = 0, wn = 0, outputs:  76 x  76 x 128 0.001 BF
  46 conv    128       1 x 1/ 1     76 x  76 x 128 ->   76 x  76 x 128 0.189 BF
  47 conv    128       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 128 1.703 BF
  48 Shortcut Layer: 45,  wt = 0, wn = 0, outputs:  76 x  76 x 128 0.001 BF
  49 conv    128       1 x 1/ 1     76 x  76 x 128 ->   76 x  76 x 128 0.189 BF
  50 conv    128       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 128 1.703 BF
  51 Shortcut Layer: 48,  wt = 0, wn = 0, outputs:  76 x  76 x 128 0.001 BF
  52 conv    128       1 x 1/ 1     76 x  76 x 128 ->   76 x  76 x 128 0.189 BF
  53 route  52 25                              ->   76 x  76 x 256 
  54 conv    256       1 x 1/ 1     76 x  76 x 256 ->   76 x  76 x 256 0.757 BF
  55 conv    512       3 x 3/ 2     76 x  76 x 256 ->   38 x  38 x 512 3.407 BF
  56 conv    256       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 256 0.379 BF
  57 route  55                                 ->   38 x  38 x 512 
  58 conv    256       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 256 0.379 BF
  59 conv    256       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 256 0.189 BF
  60 conv    256       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 256 1.703 BF
  61 Shortcut Layer: 58,  wt = 0, wn = 0, outputs:  38 x  38 x 256 0.000 BF
  62 conv    256       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 256 0.189 BF
  63 conv    256       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 256 1.703 BF
  64 Shortcut Layer: 61,  wt = 0, wn = 0, outputs:  38 x  38 x 256 0.000 BF
  65 conv    256       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 256 0.189 BF
  66 conv    256       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 256 1.703 BF
  67 Shortcut Layer: 64,  wt = 0, wn = 0, outputs:  38 x  38 x 256 0.000 BF
  68 conv    256       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 256 0.189 BF
  69 conv    256       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 256 1.703 BF
  70 Shortcut Layer: 67,  wt = 0, wn = 0, outputs:  38 x  38 x 256 0.000 BF
  71 conv    256       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 256 0.189 BF
  72 conv    256       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 256 1.703 BF
  73 Shortcut Layer: 70,  wt = 0, wn = 0, outputs:  38 x  38 x 256 0.000 BF
  74 conv    256       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 256 0.189 BF
  75 conv    256       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 256 1.703 BF
  76 Shortcut Layer: 73,  wt = 0, wn = 0, outputs:  38 x  38 x 256 0.000 BF
  77 conv    256       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 256 0.189 BF
  78 conv    256       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 256 1.703 BF
  79 Shortcut Layer: 76,  wt = 0, wn = 0, outputs:  38 x  38 x 256 0.000 BF
  80 conv    256       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 256 0.189 BF
  81 conv    256       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 256 1.703 BF
  82 Shortcut Layer: 79,  wt = 0, wn = 0, outputs:  38 x  38 x 256 0.000 BF
  83 conv    256       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 256 0.189 BF
  84 route  83 56                              ->   38 x  38 x 512 
  85 conv    512       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 512 0.757 BF
  86 conv   1024       3 x 3/ 2     38 x  38 x 512 ->   19 x  19 x1024 3.407 BF
  87 conv    512       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x 512 0.379 BF
  88 route  86                                 ->   19 x  19 x1024 
  89 conv    512       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x 512 0.379 BF
  90 conv    512       1 x 1/ 1     19 x  19 x 512 ->   19 x  19 x 512 0.189 BF
  91 conv    512       3 x 3/ 1     19 x  19 x 512 ->   19 x  19 x 512 1.703 BF
  92 Shortcut Layer: 89,  wt = 0, wn = 0, outputs:  19 x  19 x 512 0.000 BF
  93 conv    512       1 x 1/ 1     19 x  19 x 512 ->   19 x  19 x 512 0.189 BF
  94 conv    512       3 x 3/ 1     19 x  19 x 512 ->   19 x  19 x 512 1.703 BF
  95 Shortcut Layer: 92,  wt = 0, wn = 0, outputs:  19 x  19 x 512 0.000 BF
  96 conv    512       1 x 1/ 1     19 x  19 x 512 ->   19 x  19 x 512 0.189 BF
  97 conv    512       3 x 3/ 1     19 x  19 x 512 ->   19 x  19 x 512 1.703 BF
  98 Shortcut Layer: 95,  wt = 0, wn = 0, outputs:  19 x  19 x 512 0.000 BF
  99 conv    512       1 x 1/ 1     19 x  19 x 512 ->   19 x  19 x 512 0.189 BF
 100 conv    512       3 x 3/ 1     19 x  19 x 512 ->   19 x  19 x 512 1.703 BF
 101 Shortcut Layer: 98,  wt = 0, wn = 0, outputs:  19 x  19 x 512 0.000 BF
 102 conv    512       1 x 1/ 1     19 x  19 x 512 ->   19 x  19 x 512 0.189 BF
 103 route  102 87                             ->   19 x  19 x1024 
 104 conv   1024       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x1024 0.757 BF
 105 conv    512       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x 512 0.379 BF
 106 conv   1024       3 x 3/ 1     19 x  19 x 512 ->   19 x  19 x1024 3.407 BF
 107 conv    512       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x 512 0.379 BF
 108 max                5x 5/ 1     19 x  19 x 512 ->   19 x  19 x 512 0.005 BF
 109 route  107                                    ->   19 x  19 x 512 
 110 max                9x 9/ 1     19 x  19 x 512 ->   19 x  19 x 512 0.015 BF
 111 route  107                                    ->   19 x  19 x 512 
 112 max               13x13/ 1     19 x  19 x 512 ->   19 x  19 x 512 0.031 BF
 113 route  112 110 108 107                        ->   19 x  19 x2048 
 114 conv    512       1 x 1/ 1     19 x  19 x2048 ->   19 x  19 x 512 0.757 BF
 115 conv   1024       3 x 3/ 1     19 x  19 x 512 ->   19 x  19 x1024 3.407 BF
 116 conv    512       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x 512 0.379 BF
 117 conv    256       1 x 1/ 1     19 x  19 x 512 ->   19 x  19 x 256 0.095 BF
 118 upsample                 2x    19 x  19 x 256 ->   38 x  38 x 256
 119 route  85                                 ->   38 x  38 x 512 
 120 conv    256       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 256 0.379 BF
 121 route  120 118                                ->   38 x  38 x 512 
 122 conv    256       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 256 0.379 BF
 123 conv    512       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 512 3.407 BF
 124 conv    256       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 256 0.379 BF
 125 conv    512       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 512 3.407 BF
 126 conv    256       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 256 0.379 BF
 127 conv    128       1 x 1/ 1     38 x  38 x 256 ->   38 x  38 x 128 0.095 BF
 128 upsample                 2x    38 x  38 x 128 ->   76 x  76 x 128
 129 route  54                                 ->   76 x  76 x 256 
 130 conv    128       1 x 1/ 1     76 x  76 x 256 ->   76 x  76 x 128 0.379 BF
 131 route  130 128                                ->   76 x  76 x 256 
 132 conv    128       1 x 1/ 1     76 x  76 x 256 ->   76 x  76 x 128 0.379 BF
 133 conv    256       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 256 3.407 BF
 134 conv    128       1 x 1/ 1     76 x  76 x 256 ->   76 x  76 x 128 0.379 BF
 135 conv    256       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 256 3.407 BF
 136 conv    128       1 x 1/ 1     76 x  76 x 256 ->   76 x  76 x 128 0.379 BF
 137 conv    256       3 x 3/ 1     76 x  76 x 128 ->   76 x  76 x 256 3.407 BF
 138 conv    255       1 x 1/ 1     76 x  76 x 256 ->   76 x  76 x 255 0.754 BF
 139 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, obj_norm: 1.00, cls_norm: 1.00, delta_norm: 1.00, scale_x_y: 1.20
nms_kind: greedynms (1), beta = 0.600000 
 140 route  136                                    ->   76 x  76 x 128 
 141 conv    256       3 x 3/ 2     76 x  76 x 128 ->   38 x  38 x 256 0.852 BF
 142 route  141 126                                ->   38 x  38 x 512 
 143 conv    256       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 256 0.379 BF
 144 conv    512       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 512 3.407 BF
 145 conv    256       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 256 0.379 BF
 146 conv    512       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 512 3.407 BF
 147 conv    256       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 256 0.379 BF
 148 conv    512       3 x 3/ 1     38 x  38 x 256 ->   38 x  38 x 512 3.407 BF
 149 conv    255       1 x 1/ 1     38 x  38 x 512 ->   38 x  38 x 255 0.377 BF
 150 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, obj_norm: 1.00, cls_norm: 1.00, delta_norm: 1.00, scale_x_y: 1.10
nms_kind: greedynms (1), beta = 0.600000 
 151 route  147                                    ->   38 x  38 x 256 
 152 conv    512       3 x 3/ 2     38 x  38 x 256 ->   19 x  19 x 512 0.852 BF
 153 route  152 116                                ->   19 x  19 x1024 
 154 conv    512       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x 512 0.379 BF
 155 conv   1024       3 x 3/ 1     19 x  19 x 512 ->   19 x  19 x1024 3.407 BF
 156 conv    512       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x 512 0.379 BF
 157 conv   1024       3 x 3/ 1     19 x  19 x 512 ->   19 x  19 x1024 3.407 BF
 158 conv    512       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x 512 0.379 BF
 159 conv   1024       3 x 3/ 1     19 x  19 x 512 ->   19 x  19 x1024 3.407 BF
 160 conv    255       1 x 1/ 1     19 x  19 x1024 ->   19 x  19 x 255 0.189 BF
 161 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, obj_norm: 1.00, cls_norm: 1.00, delta_norm: 1.00, scale_x_y: 1.05
nms_kind: greedynms (1), beta = 0.600000 
Total BFLOPS 128.459 
avg_outputs = 1068395
hhk7734 commented 3 years ago

v2 input_shape = (384, 512, 3) compile with -a, FPS: 11 ~ 13

v3-beta input_shape = (384, 512, 3) compile with -a, FPS: 11 ~ 13

without head input_shape = (384, 512, 3) compile with -a, FPS: 9 ~ 10 more accuracy! :)

v3-release input_shape = (384, 512, 3) compile with -a, FPS: 11 ~ 13

hhk7734 commented 3 years ago

https://github.com/AlexeyAB/darknet/issues/2139

hhk7734 commented 3 years ago

https://analysisbugs.tistory.com/103

hhk7734 commented 3 years ago

211afddf827d558fd7cfcef2cee21beda1c34ce4