-
-
Notifications
You must be signed in to change notification settings - Fork 15.9k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We鈥檒l occasionally send you account related emails.
Already on GitHub? Sign in to your account
W&B Bug Report: Information missing from 'Overview' Panel #2741
Comments
@AyushExel this second bug appeared at the same time as #2740 and may be related. Can you please look into this? I'm interested in launching a new activation study but I can't start this currently as the trainings won't be logged correctly in W&B. |
@glenn-jocher I think this isn't related to integration. Probably someone introduced a front-end bug which doesn't display all the info. This will be resolved. I'll raise a ticket. Looking into the other issue now. That too, I think isn't related to the integration. Looking. |
@glenn-jocher I cannot reproduce this. I just ran this run from the official yolo colab -> https://wandb.ai/cayush/yoloV5/runs/3fgnniz5?workspace=user-cayush |
@AyushExel that's strange. I restarted my computer and retrained again just now with the same problem. Checked on Chrome and Safari, and on my mobile device but no images are showing. Two new training runs are https://wandb.ai/glenn-jocher/YOLOv5 I see your results are showing up well though, https://wandb.ai/cayush/yoloV5/runs/3fgnniz5?workspace=user-cayush looks good. |
@glenn-jocher That's weird. I think this isn't a problem originating from the repo as the same colab is working fine for me and no one else has reported this. I'll get someone to look into this issue today.
I'm looking into this and I'll reach out later today as soon as I know more. |
@glenn-jocher Let me know as soon as you try these out. It will definitely help in debugging. Also, can you move the description of Issue #2740 here as well, so that it's easier to tend to both issues in one place |
@AyushExel I deleted YOLOv5 project yesterday when it first happened, but I'll try again today, and also try to log to --project bug. Oh, this is new. Now some images appear and others not. Overview panel appears better. I checked my storage to see if I was full, but it says 46G/100G so I should be fine there. |
@glenn-jocher The above run loads perfectly for me. And all the images are present in the files tab. |
@AyushExel I tried again just now at same link and now all images appear correctly (!). Tried a new training to default project though and same missing images: https://wandb.ai/glenn-jocher/YOLOv5/runs/17tt5p8w EDIT: I'll try to delete both bug and YOLOv5 projects and rerun |
@glenn-jocher Awesome, this means that there's some edge case with that project in particular. This is helpful. Thanks. |
@AyushExel ok I've deleted both and rerun one training in both projects:
Result is that bug project works well, YOLOv5 does not. Hmm yes problem seems mainly in default project! |
@glenn-jocher yes that is helpful. I'll get back to you as soon as I have more info about this |
@AyushExel ok thanks! This should be enough to get our activations study launched today at least :) |
@glenn-jocher I have no idea what's activation study is though or maybe I forgot if we discussed it 馃槀 |
@AyushExel it's just an idea I had to test the effect of different activations on training, i.e.: class Conv(nn.Module):
# Standard convolution
def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
super(Conv, self).__init__()
self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)
self.bn = nn.BatchNorm2d(c2)
# self.act = nn.Identity() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
# self.act = nn.Tanh() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
# self.act = nn.Sigmoid() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
# self.act = nn.ReLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
# self.act = nn.LeakyReLU(0.1) if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
# self.act = nn.Hardswish() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
# self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
self.act = Mish() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
def forward(self, x):
return self.act(self.bn(self.conv(x)))
def fuseforward(self, x):
return self.act(self.conv(x)) |
Ohh awesome. btw, fun fact, Diganta Mishra, who wrote the Mish activation function paper, works on my team :) |
@AyushExel wow interesting! I've had a few conversations with him before when he first published his function. Is he part of W&B now? We haven't officially adopted Mish() due to the increased resource requirements and the relative similarity with SiLU() results in our earlier experiments. Let's see what this activation study shows, should be ready in a few days if I can start it today. Was first delayed by this issue and then today by my own bug I fixed in #2750, took me a while to isolate the source. |
@AyushExel activation study is live here! Looks like we have 3 close candidates at the top, and 1 clear loser so far. Many epochs to go still though. |
@glenn-jocher awesome :) |
Today I had the same problem as @glenn-jocher had above. During the first training sessions this morning, the images were displayed. See picture 1 Unfortunately not tonight. See picture 2. These are both specially created projects. |
@cyberFoxi It can happen because the images were not uploaded by the time you checked. Can you re-run and wait for thr run to sync completely and see if the problem still exists? Also, have you tried changing the project name using |
@AyushExel Thank you for your very quick reply. Now the pictures are there. Then there was probably a problem with the project Thank you. |
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. |
I had this issue and I debugged it to the name I gave the keys of my log dictionary. ie: wandb.log({"/val/image": wandb.Image(my_image)}) doesn't work, but wandb.log({"val_image": wandb.Image(my_image)}) does. |
@andimarafioti are you facing this issue in the current yolov5 repository? |
Hi @AyushExel , No, I found this issue while trying to solve the bug I was having with B&W, I actually thought this issue was from some wandb repository. It's probably not related to the issue you had here or your repository. |
@andimarafioti okay thanks. I think this is to do with how windows paths are handled. You can't use '/' in the names. The same works for linux. Are you on windows? |
I'm seeing the records on windows, but the code is running on linux. |
Also, using '/' in the names works for the rest of the things I'm logging, just not for media. |
@andimarafioti thanks for confirming. |
馃悰 Bug
Default W&B logging is not recording the same information to the 'Overview' screen as before, in particular 'Command', which is required to reproduce, as well as System Hardware, Python Version, and others.
To Reproduce (REQUIRED)
In Colab Notebook: https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb
View bug at W&B console: https://wandb.ai/glenn-jocher/YOLOv5
Earlier Overview (GOOD):
https://wandb.ai/glenn-jocher/batch_size/runs/1hogmvhh/overview?workspace=user-glenn-jocher
Current Overview (BAD):
https://wandb.ai/glenn-jocher/YOLOv5/runs/23u3uq0x/overview?workspace=user-glenn-jocher
The text was updated successfully, but these errors were encountered: