Transfer exisiting simple DNN Python code (Tensorflow) to the OneFlow framework
Budget: €30 – €250 EUR
Hello!
I am doing some personal researches in artificial intelligence for fun and for learning. To do so I create (more or less) simple deep learning networks to find solutions for self-made problems. Often with computer generated data as input. I am using Tensorflow/Keras as my framework.
What I want is to transform my simple DL model of my current project code (flatten layers, categorical_crossentropy, relu, softmax) to use the OneFlow framework for improving the training speed for my project drastically.
Also it is very obvious that in my project the data processing is a bottleneck. The GPU has an utilization of just 30% while the CPU is not the limiting factor. My graphic card is a Geforce 2070. Letting the training process run on a Geforce 3090 the GPU utilization has just 17%. I think that here are also optimization options in preparing the input data, reducing overhead, how to optimal use the VRAM of a Geforce card and especially how to improve the processes between CPU and GPU regarding all options of CUDA and cudnn. May be that OneFlow automatical manage some of that steps by itsself. I don't know.
I want to use the vast.ai service for training so a multi GPU support is also required.
If you have any questions, don't hesitate to ask me!
Marc
I am doing some personal researches in artificial intelligence for fun and for learning. To do so I create (more or less) simple deep learning networks to find solutions for self-made problems. Often with computer generated data as input. I am using Tensorflow/Keras as my framework.
What I want is to transform my simple DL model of my current project code (flatten layers, categorical_crossentropy, relu, softmax) to use the OneFlow framework for improving the training speed for my project drastically.
Also it is very obvious that in my project the data processing is a bottleneck. The GPU has an utilization of just 30% while the CPU is not the limiting factor. My graphic card is a Geforce 2070. Letting the training process run on a Geforce 3090 the GPU utilization has just 17%. I think that here are also optimization options in preparing the input data, reducing overhead, how to optimal use the VRAM of a Geforce card and especially how to improve the processes between CPU and GPU regarding all options of CUDA and cudnn. May be that OneFlow automatical manage some of that steps by itsself. I don't know.
I want to use the vast.ai service for training so a multi GPU support is also required.
If you have any questions, don't hesitate to ask me!
Marc