No foreground is detected when testing about keras-frcnn HOT 6 CLOSED

Hi @Arham-Aalam,
I finally figure out I got this exception because I changed default settings in keras_frcnn/config.py. When I reset the default settings and train after 100 epochs, I got output finally. Before that, only background is outputed.
You can try to train more epochs and see if everything is okay. I have trained almost 1000 epochs and I plan to share the checkpoints in the future.

from keras-frcnn.

Hi @vra
are you getting this exception

Could not load pretrained model weights. Weights can be found in the keras application folder https://github.com/fchollet/keras/tree/master/keras/applications

same problem happening with me. I think we need to start with pretrained models using

--input_weight_path
but I'm still unable to solve this problem. is there any solution you found?
Thanks

from keras-frcnn.

okay @vra,
Do you know how to train it using pretrained models like resnet50, I'm stuck in large training time.
could you help?
Thanks for your response.

from keras-frcnn.

Yes It's very slow to train the model. I use 8 2080Ti GPUs and I need almost 500 seconds to train one epoch! When I review the code, I found maybe the batchsize is too small, because the code here seems shows that only one sample is yielded at each time. Maybe you can try to improve the code near here to speed up your training. I searched before but I didn't figure a good solution yet. Some discussiosn at here and here may be useful for you.

from keras-frcnn.

thanks, it started predicting when I train it with more ephocs, but one more thing I need to know that How can it be easily useful for i3 CPU in production. I test it but it is too slow and taking something about 30 sec. to predict a single image.
is there any way to do fast prediction?
thanks

from keras-frcnn.

I think there are some solution to your situation:

1. Use GPUs. You can buy a vps with GPUs on Google Cloud, AWS etc
2. Use CPUs, but with smaller input size, e.g., change 512x512 to 256x256
3. Use CPUs, but with less parameter settings, e.g., you can change to #num_ious to a smaller value thus it will be faster
4. Use CPUs, but use ConvNet Prune techniques to make a network smaller.

Solution 3 need more understanding of details of code in this project, Solution 4 need more knowledge about network pruning.

Since discussion going on here is no relations with original issue, so I will close this issue and any further disscussion should be opened in a new place.

from keras-frcnn.