Skip to content

Latest commit

 

History

History
98 lines (74 loc) · 4.21 KB

README.md

File metadata and controls

98 lines (74 loc) · 4.21 KB
Raw

Bonsai - MoveNet

How to install

Bonsai.TensorFlow.MoveNet can be downloaded through the Bonsai package manager. However, in order to use it for either CPU or GPU inference, you need to pair it with a compiled native TensorFlow binary. You can find precompiled binaries for Windows 64-bit at https://www.tensorflow.org/install/lang_c.

To use GPU TensorFlow (highly recommended for live inference), you also need to install the CUDA Toolkit and the cuDNN libraries. The current package was developed and tested with CUDA v11.3 and cuDNN 8.2. Additionally, make sure you have a CUDA compatible GPU with the latest NVIDIA drivers.

After downloading the native TensorFlow binary and cuDNN, you can follow these steps to get the required native files into the Extensions folder of your local Bonsai install:

  1. The easiest way to find your Bonsai install folder is to right-click on the Bonsai shortcut > Properties. The path to the folder will be shown in the "Start in" textbox;
  2. Copy tensorflow.dll file from either the CPU or GPU tensorflow release to the Extensions folder;
  3. If you are using TensorFlow GPU, make sure to add the cuda/bin folder of your cuDNN download to the PATH environment variable, or copy all DLL files to the Extensions folder.

How to use

The package already includes a pre-trained MoveNet .pb file that is loaded when using the PredictMoveNet operator. In order to run inference on an incoming image, simply connect a node that provides an IplImage type to PredictMoveNet. E.g:

WorkflowExample

The output of the model (Pose) can be indexed using one of the following 17 human bodypart keypoints:

- nose
- left_eye
- right_eye
- left_ear
- right_ear
- left_shoulder
- right_shoulder
- left_elbow
- right_elbow
- left_wrist
- right_wrist
- left_hip
- right_hip
- left_knee
- right_knee
- left_ankle
- right_ankle

How to download pre-exported networks:

Network files are available from TensorFlow Hub:

All downloaded network .pb files should be placed inside src/Externals/Networks in order to build the project successfully.

How to export the .pb files

The pre-trained models are provided as .pb files. If you want to export your own .pb files, you can use the following code:

import tensorflow as tf
import tensorflow_hub as hub
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2

## Pick one of the networks
networkurl = r"https://tfhub.dev/google/movenet/singlepose/lightning/4"
#networkurl = r"https://tfhub.dev/google/movenet/singlepose/thunder/4"
#networkurl = r"https://tfhub.dev/google/movenet/multipose/lightning/1"

## Download the network
module = hub.load(networkurl)
networkName = networkurl.split("/")[-3] + "_" + networkurl.split("/")[-2]

## Freeze the network
model = module.signatures['serving_default']
full_model = tf.function(lambda x: model(x))
full_model = full_model.get_concrete_function(
    tf.TensorSpec(model.inputs[0].shape, model.inputs[0].dtype))
frozen_func = convert_variables_to_constants_v2(full_model)
frozen_func.graph.as_graph_def()

## Print the network layers and output/input specs
layers = [op.name for op in frozen_func.graph.get_operations()]
print("-" * 50)
print("Frozen model layers: ")
for layer in layers:
    print(layer)

print("-" * 50)
print("Frozen model inputs: ")
print(frozen_func.inputs)
print("Frozen model outputs: ")
print(frozen_func.outputs)

# Save frozen graph from frozen ConcreteFunction to hard drive
tf.io.write_graph(graph_or_graph_def=frozen_func.graph,
                logdir="./frozen_models",
                name=f"frozen_graph{networkName}.pb",
                as_text=False)