fpgarun.paper01.zenodo

Same as FPGARun.Paper01. (This release is issued for Zenodo).

FPGA Run - Results

The results for the fpgarun.paper01.

The project has been built on AWS R5.2xLarge (11 hours) and run on AWS F1.2xLarge.

• The runs with OCL profiling and sdaccel.timeline enabled:
  • ModelNet40, batch size 2:1:20, 22:2:30, 35, and 40
  • ShapeNet2, batch size 2:1:20, 22:2:30, 35, and 40
• The runs with OCL profiling and sdaccel.timeline disabled:
  • ModelNet40, batch size 2:1:20, 22, and 24
  • ShapeNet2, batch size 2:1:20, 22, and 24
• AWS F1 Average and Maximum Vcc(Int) Power Measurements:
  Done under repetitive feed-forward launches with BatchSize=40.

data

Sample data to run the project with. It contains trained weights and raw samples (with labels) for the two datasets discussed in our paper ( ModelNet40 and ShapeNet V2 Core).

The conversion from mesh to point cloud for ShapeNet V2 Core is done using our opensource utility and its python script.

The learning model is trained using Tensorflow 1.x on an Nvidia GTX1070. To accelerate the training procedure of the model on ShapeNet V2 Core, the weights of the trained model for ModelNet40 have been used as the initial values.

Please unzip the archive into data/.