Dear community members and @John6666 ,
As seen inmy profile and continuation from this question, I have successfully trained a model with two labels from a model with hundreds of labels called segformer-b0-finetuned-ade-512-512. I got a pretty much a satisfied visually output I cannot send rn (will add on update or reply because the said laptop is freezing-occupied in larger batch training) but with a little bit incomplete final conclusion stats? Is this often happen when I trained on small batch-small epoch or is there a mistake in my program?
The stats are such:
edit 2: The final product is surprisingly⌠very similar to 10 random image. Is this what they called overfitting?
I mean⌠not that similar, it shouldnât be that roundy either, I havenât test it side by side with the image or make it colored like most segmentation product. This one is done with
epochs = 0.1
lr = 0.00006
batch_size = 8
Thatâs all the details I could give, feel free to ask for more information, thanks in advance for the community help
1 Like
I think you probably donât have enough epochs. Iâve heard that itâs okay to train up to around 20 epochs. It seems that overfitting can occur if you train with the same data too much, but anyway, I think youâre probably short on epochs right now. I think it would be a good idea to train the model youâve trained using the same procedure again.
I was gonna put big epoch but doing 1 is quite restraining and colab doesnât seem to be a proper solution. Is there a holy triangle of epoch, batch_size, that wonât sacrifice result quality and computer ability?
1 Like
If the model is small, I think it is possible to train it in the free CPU space of HF or in the virtual environment without GPU in Colab. It will take time, butâŚ
In your case, the model is a little special, so it is difficult to use, but AutoTrainAdvanced, etc. actually work even in CPU space. This uses HF Trainer internally.
Also, with other company services, you can use a little GPU for free with Lightning.ai.
For some reason Lightning.ai doesnât have services in my area, is there actually a lot services like this other than google collab? I have ran my model with big epoch and still running 24 hours after⌠looks like going with cpu doesnât look good while itâs working.
Iâll try that AutoTrainAdvanced perhaps, even though it doesnât have segmentation presetâŚ
1 Like
I think github also offers a free virtual machine service. However, it doesnât have a GPU. In HF, you can create as many weak VMs without GPUs as you like for free, so it will take time but it wonât cost you any money.
By the way, the program below is a sample program I made for a different project, which adds a simple GUI to HF Trainer.
Itâs just like doing manually what AutoTrainAdvanced does automatically, but this is more convenient if there are parts you want to do manually. In your case, it might be faster to modify this.
You are truly a godsend, let me try that quickly
1 Like
âError occured: âImageâ object has no attribute ânamesââ is this from your code or
Ohh is it from my dataset format
1 Like
The key relationships and structure of the data set are hard-coded, so think of this as a normal script that is model-dependent, with a GUI added to make it work in HF space rather than an app. Concept model?
I donât understand, so should I convert the images to other format? Where should I read the hardcoded code so I can suit my dataset to it?
1 Like
I think the collate function is just causing problems, so I think it would be better to rewrite it or stop using it.
# data_collator=collate_fn,
For better or worse, the GUI âdoes nothingâ.
Well, you can think of it as a CLI in effect. Thatâs what it was made for. (dummy GUI)
