CODEWITHSUNDEEP
tensorflowobject-detectionobject-detection-api
NKK
NKK

Reputation: 13

Tensorflow Object Detection API TF Nightly

Is it possible to use the Tensorflow Object Detection API with the current tf-nightly build and how do I replace tensorflow with tf-nightly? I have a RTX 3080 which needs CUDA 11 and that is only supported in tf-nightly as of now.

Upvotes: 0

Views: 317

Answers (1)

Craidos
Craidos

Reputation: 11

I got it working with my RTX 3070 on Windows with TF2.

  1. Create a python 3.8 conda environment and install tf-nightly-gpu via pip

pip install tf-nightly-gpu==2.5.0.dev20210109

  1. Install cuda 11.0 and cuDNN 8.0.2

  2. Install cuda 11.1

  3. Replace ptxas.exe in the v11.0 bin directory with the v11.1 version (the 11.0 version was causing errors for me)

  4. Make sure your path/cuda path point to cuda 11.0 (not 11.1)

  5. install object detection api from tensorflow

  6. Add this to your model_main_tf2.py

os.environ['TF_CPP_MIN_LOG_LEVEL']='2'

from absl import flags
import tensorflow.compat.v2 as tf


physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0, "Not enough GPU hardware devices available"
config = tf.config.experimental.set_memory_growth(physical_devices[0], True)

from object_detection import model_lib_v2

With this unfortunately i only can train and not evaluate the model.

Upvotes: 1

Related Questions