GPU (Tesla T4) not being utilised for mixed precision

[nish16111]

I have been assigned the Tesla T4 Gpu by google colab, but when I run the fit method I have been getting ETA of 90 hours for a single epoch.

I have turned on the mixed precision and have the Tesla T4 gpu allotted to me by google colab, then why am I getting the ETA(time for completion) of such a gigantic magnitude?

Can someone please help??

[SaketMunda]

Make sure before fitting the model, when you ask for listing the physical devices,
tf.config.list_physical_devices()

it should output, a GPU as well.

[PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU'),
PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

T4 GPU might have been allotted to you but it isn't necessary that your code is utilising GPU.

Also reviewing your screenshot looks like you're trying FoodVision101 app, and if you're using EfficientNetB0 model. I would suggest rather use tensorflow==2.7 instead tensorflow==2.4.1.

[nish16111]

Thanks!!

[mrdbourke]

Hi @nish16111,

Looks like there's a few issues with mixed precision training depending on the version of TensorFlow you're using.

As of now, 2.11.x seems to work.

You can see more here: #508

[nish16111]

Thanks!!