cl886699

> 这个loss要降到0.几，你的图片很少？ 我有八张gpu，这个loss应该是reduce_sum的结果，图片数量确实很少，平均一张图有十来个目标，训练集一百张，验证集20张。 贴图的loss是第二阶段的loss，第一阶段的loss差不多降到了6左右，6除以8应该也就是不到1了。 我先检测了一遍输入数据，解码回原图没有问题，这是我解码部分的代码 ` from utils.utils import BBoxUtility from utils.anchors import get_anchors priors = get_anchors(640) bbox_util = BBoxUtility(1, priors) tf_record_path = 'D:/datasets/bjod/' train_datasets, val_data = ZipDataset(tf_record_path, 1, 96,...

> 增大训练集再说，我做过归一化了兄弟 我之前用fpn 跑过该数据集，验证集map在0.4左右，也用过tf自带的object detection aip中的efficientdet跑过，训练集的map能达到0.7，但是验证集只有0.1

> @flyinggx OK [email protected] thanks a lot

in my case，I put the annotation *.txt files in HR,LR and Bic， but it should be put in HR/x4，after fixed this，I meet RuntimeError: CUDA out of memory. by the way,...

> @cl886699 From where did you get those annotation files? Even I don't know how to generate the validation set. mybe copy_folder_name_for_valid_image (line of 92 in scripts_GAN_HR-LR.py) this function can...

> @cl886699 But I don't have the data for `/home/jakaria/Super_Resolution/Filter_Enhance_Detect/saved_ESRGAN/val_images/*/`. What should it be? I haven't done any training. I'm trying to reproduce the results using the pretrained weights. val_images...

我改了一个多gup训练的版本，但只支持固定大小的输入

> Hi @cl886699 , I am afraid I cannot help if I don't see the code. You should be getting MCC ~ 98% after 60 epochs, as per the manuscript...

> I can take a look at the code if you want to post it here, if there is any obvious error that I can spot quickly, it may help....

that's the result of overviting one image, just trained 1000 steps  that's the result of training 50 epochs on LEVIRCD  it sames falling into local optimality, all segmentations...