Closed ma3252788 closed 1 year ago
I changed several servers, all of which are like this
I changed several servers, all of which are like this
This is the logging info about loading pretrained weights
, which will print the matched keys
of your pretrained weights and models, I think it's a normal, all the unmatched keys will be logged after printing this with green and purple colors
I changed several servers, all of which are like this
This is the logging info about
loading pretrained weights
, which will print thematched keys
of your pretrained weights and models, I think it's a normal, all the unmatched keys will be logged after printing this with green and purple colors
I used the officially provided pre-trained models, as well as the COCO2017 dataset. There are very, very many of these empty lines, and since there are no automatic line breaks, I need to scroll with the mouse for a long time to see the beginning output content.
I changed several servers, all of which are like this
This is the logging info about
loading pretrained weights
, which will print thematched keys
of your pretrained weights and models, I think it's a normal, all the unmatched keys will be logged after printing this with green and purple colorsI used the officially provided pre-trained models, as well as the COCO2017 dataset. There are very, very many of these empty lines, and since there are no automatic line breaks, I need to scroll with the mouse for a long time to see the beginning output content.
we will take a look into it and try to improve the logging function
I changed several servers, all of which are like this
This is the logging info about
loading pretrained weights
, which will print thematched keys
of your pretrained weights and models, I think it's a normal, all the unmatched keys will be logged after printing this with green and purple colorsI used the officially provided pre-trained models, as well as the COCO2017 dataset. There are very, very many of these empty lines, and since there are no automatic line breaks, I need to scroll with the mouse for a long time to see the beginning output content.
we will take a look into it and try to improve the logging function
I just looked at log.txt and I noticed that in the output of checkpoint, there are a lot of spaces in each line and each line has 14501 characters, which I think is the main reason for showing a lot of empty lines in the terminal.
But this problem does not exist when outputting the network structure, I think it can be outputted according to the network structure.
I changed several servers, all of which are like this
This is the logging info about
loading pretrained weights
, which will print thematched keys
of your pretrained weights and models, I think it's a normal, all the unmatched keys will be logged after printing this with green and purple colorsI used the officially provided pre-trained models, as well as the COCO2017 dataset. There are very, very many of these empty lines, and since there are no automatic line breaks, I need to scroll with the mouse for a long time to see the beginning output content.
we will take a look into it and try to improve the logging function
I just looked at log.txt and I noticed that in the output of checkpoint, there are a lot of spaces in each line and each line has 14501 characters, which I think is the main reason for showing a lot of empty lines in the terminal.
But this problem does not exist when outputting the network structure, I think it can be outputted according to the network structure.
Yes, in detectron2, If the names of two nodes are similar, they will be merged into one line in order to simplify the output. This may lead to too much text in a single line, making it unable to be completely displayed. It's a good way to just print the nodes' name and not merge them, but which may need to hack the DetectionCheckpointer
I had the same problem. When loading weights with unmatched keys, the output log is filled with A LOT of blank spaces among those keys. Is there any way to solve that problem? @rentainhe
The unmatched keys in the logs are too difficult to read.
The unmatched keys in the logs are too difficult to read.
We will take a look into this logging info recently, the blanks means the matched keys actually, unmatched keys will be printed in other ways~
The unmatched keys in the logs are too difficult to read.
We will take a look into this logging info recently, the blanks means the matched keys actually, unmatched keys will be printed in other ways~
Thanks,hope this issue could be addressed soon.
We have already fixed this issue in this PR: https://github.com/IDEA-Research/detrex/pull/203, which will be merged into main recently
There's no blank any more in this case (DINO-R50
).
@ma3252788 @YTEP-ZHI
We simply modified the detectron2.checkpoint.DetectionCheckpointer
, but we did not set this checkpoint as the default one because there will be much more output in this case sometimes. After this PR is merged, you can try to update the tools/train_net.py
by replacing the old one with the detrex version:
...
# from detectron2.checkpoint import DetectionCheckpointer
from detrex.checkpoint import DetectionCheckpointer
...
@rentainhe Thanks for your hard work.
I'm closing this issue, feel free to reopen it if needed~ @ma3252788 @YTEP-ZHI
When I use detrex for training, I got a lot of blanks: