andrearosasco/hyperpcr

This repository contains the code relative to the paper "Towards Confidence-guided Shape Completion for Robotic Applications". This code contain a python package that allows you to do inference with our model. If you want to look at the training and evaluation code check out https://github.com/andrearosasco/confidence_shape_completion or https://github.com/andrearosasco/pcr/tree/mc_dropout mc_dropout branch for a messier version.

Installation

To install the code run pip install hyperpcr If you are on Linux this should automatically install all the dependencies. On windows you might need to manually install pytorch.

Usage

The following script download a checkpoint and a partial point cloud. Then it process it and generates its reconstruction. If you wanna visualize the results you have to additionally install Open3D python library.

import numpy as np
import torch
try:
    from open3d.cpu.pybind.geometry import PointCloud
    from open3d.cpu.pybind.utility import Vector3dVector
    from open3d.visualization import draw
    visualize = True
except ImportError:
    visualize = False

from pcr.model import PCRNetwork as Model
from pcr.utils import Normalize, Denormalize
from pcr.default_config import Config
from pcr.misc import download_checkpoint, download_asset


def main():

    ckpt_path = download_checkpoint(f'grasping.ckpt')
    asset_path = download_asset(f'partial_bleach_317.npy')

    model = Model(config=Config.Model)
    model.load_state_dict(torch.load(ckpt_path)['state_dict'])
    model.cuda()
    model.eval()

    partial = np.load(asset_path)

    partial, ctx = Normalize(Config.Processing)(partial)

    partial = torch.tensor(partial, dtype=torch.float32).cuda().unsqueeze(0)
    complete, probabilities = model(partial)

    complete = complete.squeeze(0).cpu().numpy()
    partial = partial.squeeze(0).cpu().numpy()

    complete = Denormalize(Config.Processing)(complete, ctx)
    partial = Denormalize(Config.Processing)(partial, ctx)

    if visualize:
        draw([
              PointCloud(points=Vector3dVector(partial)).paint_uniform_color([0, 0, 1]),
              PointCloud(points=Vector3dVector(complete)).paint_uniform_color([0, 1, 1]),
              ])

    print(complete)

If you want to change the Decoder behaviour you can change the configuration values of pcr.default_config.Config.

hyperpcr's Issues

Question about how to train this net

Hello!Can you tell me how to train this network? For example, which file to run or what environment to configure?

Question about training

Hello, thank you for your great work. I have a question about training. I found that training this network is difficult to converge, meaning that the loss decreases to around 0.6 and then remains largely unchanged. I am using the ShapeNetV2 dataset and training the network through pcr/ours/train.py. The network is configured using ours/configs/local_config.py, and the dataloader is generated through ours/datasets/ShapeNetPOVRemoval.py. I look forward to your response.