YOLOv4-pytorch icon indicating copy to clipboard operation
YOLOv4-pytorch copied to clipboard

Mobilenet-YOLOv4 mAP 0.85怎么也到不了啊,训练只有0.57,参数也没改啥,代码是不是有问题?

Open cqray1990 opened this issue 3 years ago • 18 comments

MODEL_TYPE = { "TYPE": "Mobilenet-YOLOv4" } # YOLO type:YOLOv4, Mobilenet-YOLOv4 or Mobilenetv3-YOLOv4

CONV_TYPE = {"TYPE": "DO_CONV"} # conv type:DO_CONV or GENERAL

ATTENTION = {"TYPE": "NONE"} # attention type:SEnet、CBAM or NONE

train

TRAIN = { "DATA_TYPE": "VOC", # DATA_TYPE: VOC ,COCO or Customer "TRAIN_IMG_SIZE": 416, "AUGMENT": True, "BATCH_SIZE": 1, "MULTI_SCALE_TRAIN": True, "IOU_THRESHOLD_LOSS": 0.5, "YOLO_EPOCHS": 50, "Mobilenet_YOLO_EPOCHS": 120, "NUMBER_WORKERS": 0, "MOMENTUM": 0.9, "WEIGHT_DECAY": 0.0005, "LR_INIT": 1e-4, "LR_END": 1e-6, "WARMUP_EPOCHS": 2, # or None "showatt": False }

val

VAL = { "TEST_IMG_SIZE": 416, "BATCH_SIZE": 1, "NUMBER_WORKERS": 0, "CONF_THRESH": 0.005, "NMS_THRESH": 0.45, "MULTI_SCALE_VAL": False, "FLIP_VAL": False, "Visual": False, "showatt": False }

cqray1990 avatar Apr 28 '21 04:04 cqray1990

0.85的不是mobilenet,是YOLOv4,需要训练200轮?(某一个issue里面提到了),代码中设置的是50轮

li-ju-bazhong avatar May 12 '21 10:05 li-ju-bazhong

我测试的mobileyolo可以达到79.62, CSP那个源代码可以到72.31。50轮,也还行叭

li-ju-bazhong avatar May 12 '21 10:05 li-ju-bazhong

@li-ju-bazhong 我训练的都不只200

cqray1990 avatar May 13 '21 06:05 cqray1990

可能还有一些其他的trick?, 也可能是随机初始化得到的最优结果?

li-ju-bazhong avatar May 13 '21 07:05 li-ju-bazhong

@li-ju-bazhong 请问确定0.851是yolov4的吗,那为什么作者的readme里面写的是mobilenetv2=0.851?

哎,就是哈,我之前看还是yolov4的,修改过?

li-ju-bazhong avatar May 19 '21 05:05 li-ju-bazhong

@Anleeno-Xu 我按照代码测试下来,就是0.78的样子,之前看到有个问题说作者是跑了200epoch,我只是跑了120,结果是0.78

li-ju-bazhong avatar May 19 '21 06:05 li-ju-bazhong

有算力就跑一下呗,估计3-4天就能出结果了吧

li-ju-bazhong avatar May 19 '21 06:05 li-ju-bazhong

或者有其他的trick哈哈哈哈哈, I don‘t know

li-ju-bazhong avatar May 19 '21 06:05 li-ju-bazhong

你 eval_voc的路径没对,斜杠和反斜杠,你修改成统一的,才能正确寻址

li-ju-bazhong avatar May 19 '21 07:05 li-ju-bazhong

他有的地方有问题。。。。

li-ju-bazhong avatar May 19 '21 07:05 li-ju-bazhong

你用那个 get_map能够得到更好的结果

li-ju-bazhong avatar May 19 '21 09:05 li-ju-bazhong

@li-ju-bazhong 你算力是多少,我用的2080ti ,batchsize=6训练了600epoch,mAP才61

cqray1990 avatar May 28 '21 10:05 cqray1990

我3090,batch_size=6,mobilenet训练了120,mAP=0.78,CSPdarknet,50个epoch,0.74的mAP

li-ju-bazhong avatar Jun 01 '21 06:06 li-ju-bazhong

@li-ju-bazhong 你确定啥也没改?

cqray1990 avatar Jun 10 '21 02:06 cqray1990

@li-ju-bazhong 有加预训练模型?

cqray1990 avatar Jun 10 '21 02:06 cqray1990

@li-ju-bazhong 有加预训练模型?

肯定加载了预训练模型,作为初始参数。我用的mobilenetv3应该是,然后做了一次,他官方给的不是0.84么。我看我这个训练出来时0.74,觉得还行叭。

li-ju-bazhong avatar Jun 10 '21 07:06 li-ju-bazhong

话说,用官方给的best_voc来做val,为啥我换台机器就全是0了,help

li-ju-bazhong avatar Jun 10 '21 07:06 li-ju-bazhong

我3090,batch_size=6,mobilenet训练了120,mAP=0.78,CSPdarknet,50个epoch,0.74的mAP

老兄解决了吗

xaioffff avatar Jan 06 '22 03:01 xaioffff