Tensorflow-directml doesnt work with AMD Radeon RX 6700 XT about tensorflow-directml HOT 6 CLOSED

This was caused by a bug in one of our kernels and will be fixed in the next release. Thanks for reporting this!

from tensorflow-directml.

Do you have more detailed steps to set up your repro? The script you attached imports some dependencies (e.g. object_detection) and you also passed in some arguments like the .pb file and .pbtxt file. We'll need these same files to run your script.

Have you tried running any other TF models/scripts?

Thanks!

from tensorflow-directml.

The script uses the method "create_category_index_from_labelmap" from the Tensorflow Object Detection API. I am using the MobilenetV2-model from https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf1_detection_zoo.md and the Labelmap
mscoco_label_map_pbtxt.txt (uploaded as txt because otherwise not supported).
I also tried to run a training script using InceptionV2 and got the same error.
Same scripts on the same PC worked fine using tensorflow-CPU-version.
What would be the recommended way to run pretrained SSD-MobilenetV2 inference with USB-Webcam using Tensorflow-directml?

from tensorflow-directml.

Thanks. I'm able to repro this now. The issue is not specific to your graphics card.

I can't run this using tensorflow-gpu with CUDA either, but it could just be a coincidence. This is with a discrete NVIDIA GPU that has 8GB of memory:

2021-05-24 11:35:28.248310: I tensorflow/stream_executor/cuda/cuda_driver.cc:831] failed to allocate 2.91G (3123724288 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY: out of memory

I can only run this with tensorflow-cpu. I haven't tried debugging yet but I'll get back to you when I have the time to look into this a bit more.

Until we've taken a deeper look I can't recommend how to run this specific model with tensorflow-directml. If you're simply interested in a pretrained object detection model with webcam input you can try the YOLOv3 sample here.

from tensorflow-directml.

Thank you for the answer.
I also tried the Yolov3 sample and the inference worked fine.
But when I tried to run the training script from the sample, I always got a driver crash, even with a batch size of 1.

from tensorflow-directml.

The latest tensorflow-directml package on pypi should resolve this issue.

from tensorflow-directml.