cuda CNN Model Training on CPU Instead of GPU on my laptop.

[Vatsal_09]

Problem Statement:
While attempting to train a Convolutional Neural Network (CNN) model using TensorFlow/Keras, the model runs on the CPU instead of the GPU, causing high CPU usage and prolonged training time. This issue persists despite having a compatible GPU available.

Details:

1. Environment Setup:

   • Framework: TensorFlow
   • GPU Hardware: Geforce RTX 4050 (mobile)
   • CPU Hardware: Intel i5-13420H
   • Operating System: Windows 11 
   • Python Version: Tried on both ( 3.10.5 and 3.12 )
   • Additional Libraries: Installed and set up CUDA, cuDNN

2. Observed Behavior:

   • Upon running model.fit(x=x_train, y=y_train, batch_size=128, epochs=100, validation_data=(x_test, y_test)), the system utilizes the CPU at 100%, leading to slower training times.
   • Despite having a compatible GPU, the training process does not leverage GPU resources, as confirmed by the absence of GPU activity in monitoring tools like task manager and nvidia-smi.

3. Expected Behavior:

   • The model training should utilize the GPU to accelerate computations and reduce CPU usage.
   • With GPU-enabled training, the training process should be faster, and the CPU load should remain minimal.

4. Attempted Solutions:

   • Reinstalled TensorFlow with GPU support (tensorflow-gpu).
   • Checked CUDA and cuDNN installation to ensure compatibility with the TensorFlow version.
   • Created a different environment variable using (anaconda) for py3.10.5 and did the same steps again.
   • Tried installing WSL (assuming in case if it wasn't compitible with windows and needed linux insted)

5. Technical Constraints:

   • Ensuring TensorFlow, CUDA, and cuDNN versions are compatible.
   • Checking for any conflicting CPU-only dependencies that may override GPU settings.

trainmodel.ipynb

[Tridefender]

Did you install torch on gpu, they use a seperate command. Or did you neglect the device select so it defaulted to CPU0?

[Vatsal_09]

Yeah i did install torch, also tried with different set of commands. 

And I could not find any such settings such as device select.

Can you be more specific where should i check.

[Tridefender]

import torch
for i in range(torch.cuda.device_count()):
   print(torch.cuda.get_device_properties(i).name)

Punch this line in and run it, see what pops up. Basically it tells you devices that torch can choose from.

 

Fix the indentations yourself tho, copied it using a phone.

 

If it throws out an exception then perhaps your torch ain't working 

[Vatsal_09]

Appreciated @Tridefender i ran it but didn't print anything, So, i ran 

import torch

print(torch.__version__)

print(torch.cuda.is_available())

and the output is:

2.2.1+cpu
1
False

So, i get it something is wrong but don't know the solution   

[Tridefender]

Just now, Vatsal_09 said:

Appreciated @Tridefender i ran it but didn't print anything, So, i ran 

import torch

print(torch.__version__)

print(torch.cuda.is_available())

and the output is:

2.2.1+cpu
1
False

So, i get it something is wrong but don't know the solution   

Are you sure the environment has cuda?? Because it says FALSE

[Vatsal_09]

I dont know man re-installed twice. even edited system variables and added 3 new paths (x64,bin,lib)

[Tridefender]

What does nvcc -v gives you? Because clearly the Jupyter environment has no cuda support.