zhengpeng7 / gconet_plus Goto Github PK

Why does an epoch stop early and go straight to the next epoch？
（中文：为什么一个epoch提前停止并直接进入了下一个epoch）

I didn't notice this situation when I was training before. Later, it happened when I changed to multi-GPU training. I thought it was a problem with the multi-GPU training I changed, but I re-downloaded the code for single-GPU training and still had this problem:
Each epoch ends early and a new epoch is performed
Is this correct? Or that it was
（中文：之前我训练时并没有注意到有没有出现这个状况，后来我改成多GPU训练时出现了，我以为是改的多GPU训练有问题，可是我重新下了代码单GPU训练还是有这个问题：
每个epoch提前结束，进行新的epoch
这是正确的吗？还是说原本就这样）

想知道RefUnet(nn.Module)模块的作用是什么，在论文中有对应的部分吗？

如能回答，万分感谢！

祝好~

select_results.py这个文件是用来和别的算法结果比较的吗？之所以报错是不是因为没有别的算法的结果？还是说有其他的地方要调整

您好！请问可以提供类激活图可视化代码吗？

您好，在运行sort_results.py和select_resulets.py时各有一个错误，第二个错误可能是因为第一个引起的，还没有看，对于第一个错误，我看到是来源于130行这句代码

其中best_ckpt的值为'gconet_-ep315:'，在130行代码中它取了一个空值，错误原因应该是在这

请问trail这个变量原本应该是取什么值的？原本以为是取数字，但是后面的ep变量取的是数字315

Hello, thanks for nice project and you had done solid work.

I have some about the inference time measurement.
I also write script to check the inference time.
while my result is 10 ms per image(100 fps).
It is only inference time, the time of data loading and image resizing are not included.

I am wondering how to make 250 fps.
My hardware are 2080ti GPU and i7 cpu.
Thanks in advance!

I want to make simple inference for some images I have and the results are bad

from models.GCoNet import GCoNet
import torch
from pathlib import Path
from torchvision.io import read_image
from torchvision.utils import make_grid, save_image
from torchvision import transforms

dataset_path = Path("/path/to/data")

device = torch.device("cuda")
model = GCoNet()
model = model.to(device)
gconet_dict = torch.load("/path/to/weights")
# gconet_dict = torch.load("/shared_volume/rashad/salient_app/ultimate.pth")
trans = transforms.Compose([
            transforms.Resize((320, 320)),
            # transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ])

model.to(device)
model.load_state_dict(gconet_dict)

model.eval()
images = [read_image(str(img_path))[None, ...] for img_path in dataset_path.glob("11/*")]
# print(images)
torch_images = torch.concat(images).float()
torch_images = trans(torch_images)
print(torch_images.shape)
with torch.no_grad():
    result = model(torch_images.cuda())[-1]
    result = torch.concat([torch.nn.functional.interpolate(res.unsqueeze(0), size=(res.shape[1], res.shape[2]), mode='bilinear', align_corners=True) for res in result]).sigmoid()
result = (result > result.mean()).float() * 255
print(result.shape)
print(result.max(), result.min())
save_image(make_grid(result), "salient_example.png")

And this is the result,

@HzFu @ZhengPeng7

您好，我从您google drive下载的权重包中用 （ultimate_duts_cocoseg (The best one).pth）评估，但和论文中还略微有点差距，且这个权重和（ultimate_duts_cocoseg.pth）应该是同一个，是上传错误了吗？

所提供的代码和训练命令训练时是默认使用了DUTS class和COCO-SEG吗？我看到在参数上只写了DUTS_class

CUDA_VISIBLE_DEVICES=0 python train.py --trainset DUTS_class --size ${size} --ckpt_dir ckpt/${method} --epochs ${epochs} --val_dir tmp4val_${method}

但是我在训练脚本里看到的是对两个训练集都遍历了

还有他的训练方式是训练一个batch的DUTS，然后一个bach的COCO-SEG，再更新梯度，是这样吗？