记录一下yolo v5从零训练COCO数据集的情况
关于coco2017数据集
coco2017 80个类别
训练集118287 验证集 5000 测试集40670 一共163957
训练集中有117266被标注(每张图片有多个不同种类的目标) 验证集中有4952张被标注
关于混合精度训练
yolov5默认开启混合精度训练：
# Forward
with torch.cuda.amp.autocast(amp):
pred = model(imgs) # forward
loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size
if RANK != -1:
loss *= WORLD_SIZE # gradient averaged between devices in DDP mode
if opt.quad:
loss *= 4.
官方说的是在单卡V100显卡上大概训练s模型需要2天
我这里是单卡3090，如果是基于预训练模型训练，用时不到4天半，55个小时，但是从零训练应该不能基于预训练模型，如果weights不传参，训练就很慢了，大概测算需要多半个月。
基于预训练模型训练完后的日志，忘了保存，只有一部分：
fork 5000 215 0.646 0.349 0.431 0.248
knife 5000 325 0.516 0.206 0.223 0.111
spoon 5000 253 0.54 0.19 0.229 0.117
bowl 5000 623 0.616 0.475 0.516 0.356
banana 5000 370 0.486 0.332 0.343 0.175
apple 5000 236 0.433 0.275 0.219 0.138
sandwich 5000 177 0.61 0.458 0.487 0.314
orange 5000 285 0.492 0.373 0.349 0.255
broccoli 5000 312 0.485 0.394 0.377 0.184
carrot 5000 365 0.381 0.359 0.309 0.184
hot dog 5000 125 0.62 0.472 0.477 0.31
pizza 5000 284 0.712 0.634 0.668 0.455
donut 5000 328 0.547 0.491 0.5 0.372
cake 5000 310 0.618 0.439 0.508 0.31
chair 5000 1771 0.596 0.404 0.451 0.257
couch 5000 261 0.707 0.489 0.596 0.398
potted plant 5000 342 0.564 0.415 0.435 0.235
bed 5000 163 0.709 0.509 0.591 0.368
dining table 5000 695 0.592 0.357 0.394 0.241
toilet 5000 179 0.805 0.76 0.829 0.607
tv 5000 288 0.758 0.664 0.735 0.521
laptop 5000 231 0.765 0.649 0.706 0.538
mouse 5000 106 0.804 0.708 0.762 0.543
remote 5000 283 0.547 0.406 0.433 0.215
keyboard 5000 153 0.652 0.575 0.658 0.423
cell phone 5000 262 0.595 0.454 0.488 0.296
microwave 5000 55 0.631 0.636 0.712 0.494
oven 5000 143 0.626 0.448 0.535 0.304
toaster 5000 9 0.793 0.333 0.515 0.355
sink 5000 225 0.656 0.498 0.556 0.338
refrigerator 5000 126 0.759 0.603 0.69 0.489
book 5000 1129 0.455 0.181 0.219 0.0885
clock 5000 267 0.777 0.704 0.735 0.471
vase 5000 274 0.566 0.496 0.482 0.311
scissors 5000 36 0.566 0.222 0.246 0.18
teddy bear 5000 190 0.721 0.526 0.611 0.373
hair drier 5000 11 1 0 0.0206 0.00303
toothbrush 5000 57 0.529 0.298 0.344 0.19
Evaluating pycocotools mAP... saving runs/train/exp/_predictions.json...
loading annotations into memory...
Done (t=0.53s)
creating index...
index created!
Loading and preparing results...
DONE (t=3.21s)
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=41.97s).
Accumulating evaluation results...
DONE (t=6.99s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.372
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.566
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.402
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.209
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.422
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.477
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.309
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.514
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.570
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.374
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.631
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.716
Results saved to runs/train/exp
训练好的模型和结果在 从零训练yolov5在COCO数据集上的模型和结果-深度学习文档类资源-CSDN下载