Very slow Training process about keras-frcnn HOT 12 OPEN

I'm trying to use the pre-trained weights file to train it . Maybe it's useful.
https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels.h5

from keras-frcnn.

I'm trying to use the pre-trained weights file to train it . Maybe it's useful.
https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels.h5

How did it go? Is it faster?

from keras-frcnn.

I'm trying to use the pre-trained weights file to train it . Maybe it's useful.
https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels.h5

How did it go? Is it faster?

nope. Someone said that he used 8RTX2080TI to train it. But it also taked 500seconds per round

from keras-frcnn.

Agree, It took me 1 day to complete one epoch, I have only 1500 images.
All images are 150x150 and black/white

FYI I trained the model on google colab.

from keras-frcnn.

Is there a way to increase the speed or an updated rep. Where we can use GPU ?

from keras-frcnn.

I'm trying to use the pre-trained weights file to train it . Maybe it's useful.
https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels.h5

How did it go? Is it faster?

nope. Someone said that he used 8RTX2080TI to train it. But it also taked 500seconds per round

@SmartInvoker How was he able to train it on GPU ?

from keras-frcnn.

Hi,

Can anyone suggest how to increase training speed ? I am having GPU but unable to utilise with required keras version.

For 10 epochs only it took almost 2.5 hours on my PC.

Thanks.

@ketul93 Hi, Did you find a solution ?

from keras-frcnn.

Hi,
Can anyone suggest how to increase training speed ? I am having GPU but unable to utilise with required keras version.
For 10 epochs only it took almost 2.5 hours on my PC.
Thanks.

@ketul93 Hi, Did you find a solution ?

Till now not....

from keras-frcnn.

Agree, It took me 1 day to complete one epoch, I have only 1500 images.
All images are 150x150 and black/white

FYI I trained the model on google colab.

Agree, It took me 1 day to complete one epoch, I have only 1500 images.
All images are 150x150 and black/white

FYI I trained the model on google colab.

how did you train on google colab
what changes did you make to the code and shell commands?

from keras-frcnn.

Agree, It took me 1 day to complete one epoch, I have only 1500 images.
All images are 150x150 and black/white
FYI I trained the model on google colab.

Agree, It took me 1 day to complete one epoch, I have only 1500 images.
All images are 150x150 and black/white
FYI I trained the model on google colab.

how did you train on google colab
what changes did you make to the code and shell commands?

Agree, It took me 1 day to complete one epoch, I have only 1500 images.
All images are 150x150 and black/white

FYI I trained the model on google colab.

If you are using google colab you need to manual implement the GPU provided by colab

To do that, click on 'Runtime' on the top toolbar, then 'change runtime type', then set the hardware acceleration to GPU

from keras-frcnn.

You can reduce epoch_length = 50 in train_frcnn.py, I don't know what this parameter is doing.

from keras-frcnn.

You can reduce epoch_length = 50 in train_frcnn.py, I don't know what this parameter is doing.

The model doesn't learn well from data on very low epoch length/epochs.

Btw, you can try running it on Theano backend, I've noticed a great performance increase without much resource usage.

from keras-frcnn.