Skip to content

sunyong2016/Deep_Multimodal_Embedding

Repository files navigation

Deep_Multimodal_Embedding

Reproduction of 'Deep Multimodal Embedding: Manipulating Novel Objects with Point-clouds, Language and Trajectories'

(http://robobarista.cs.cornell.edu/assets/papers/sung_icra2017_dme.pdf)

Requirements:

Use python2

  1. rospy
  2. keras
  3. tensorflow

To run:

Prepare Robobarista dataset

(See: http://robobarista.cs.cornell.edu/dataset)

Modify load_data.py in line 12: data_path = 'path/to/Robobarista/robobarista_dataset/dataset'.

Script discription:

Autoencoder pretraining:

pretraining_pointcloud.py: pretrain pointcloud to h2 layer
pretraining_language.py: pretrain language to h2 layer
pretraining_trajectory.py: pretrain trajectory to h2 layer

Joint training:

joint_training_p_l.py: joint pointcloud, language
joint_training_traj.py: joint trajectory
joint_training_p_l_tau.py: joint (pointcloud, language), trajectory

Simply use 'python xxx.py' to train and test.

Others:

traj_distance_matrix.npy: numpy array that stores 1225x1225 trajectory distances.
load_data.py: data parsing and preprocessing, no need to run this script.
Processed_data: processed data is in this directory
tools: some useful tools for trajectory data
Weights: all weights are in this directory
History: training histories

Note:

Currently the result of trajectory joint training is not good. Now I am trying to modify the similarity definition and the loss function.

For more details, please turn to the authors of http://robobarista.cs.cornell.edu/assets/papers/sung_icra2017_dme.pdf

About

Reproduction of 'Deep Multimodal Embedding: Manipulating Novel Objects with Point-clouds, Language and Trajectories'

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages