update supplement

2d_reaching/README.md

@@ -6,7 +6,7 @@ We used a bi-dimensional space sequential reaching dataset. Starting from point
 
 ## Label Table
 
-|         | Degree of factors | Ground truth factor     |
+|         | Degree of factors | Ground-truth factor     |
 | ------- | ----------------- | ----------------------- |
 | factor1 | 2                 | Goal position           |
 | factor2 | 2                 | Curved inward / outward |

2d_wavy_reaching/README.md

@@ -8,7 +8,7 @@ This sequential datasets generated from more complex dynamic factors. Five diffe
 
 ## Label Table
 
-|         | Degree of factors | Ground truth factors               |
+|         | Degree of factors | Ground-truth factors               |
 | ------- | ----------------- | ---------------------------------- |
 | factor1 | 5                 | Goal position                      |
 | factor2 | 4                 | 1st trajectory shape               |

README.md

@@ -9,9 +9,9 @@ These dataset contain multiple dynamic factors.
 ## Contents
 
 - [Dataset](#dataset)
-- [Note of Information Bottleneck Principle forSequential Data](./note/Information Bottleneck_Principle_for_Sequential_Data.pdf)
+- [Note of Information Bottleneck Principle for Sequential Data](./note/Information_Bottleneck_Principle_for_Sequential_Data.pdf)
 - [addtional experiments (Gripper)](./note/gripper_result.md)
-- [Experiments parameter's detail](./note/model_details.md)
+- [Experiment Details](./note/model_details.md)
 
 ##  <a name="dataset">Dataset</a>
 

gripper/README.md

@@ -1,18 +1,17 @@
 # Gripper Dataset.
 
 ![](./recon.gif)
-
-We implemented the end-effector only rather than the entire robot arm since controlling the robot arm during the picking task is easily computable by calculating the inverse kinematics and inverse dynamics. Gripper is a 12 dimensional data set: [joint x position, joint y position, finger1 joint position(angle), finger2 joint position(angle), box1 x position, box1 y position, box2 x position, box2 y position, ball1 x position, ball1 y position, ball2 x position, ball2 y position]. Eight factors are represented in this dataset: 1) color of ball to pick up, 2) initial location of red ball, 3) initial location of blue ball, 4) initial location of blue basket, 5) initial location of red basket, 6) plan for using end effector to move to ball to pick it up [first, moving horizontally to the x-location of ball and then descending horizontally to the y-location of ball, like the movement of the doll drawing machine (perpendicular motion); second, moving straight to the location of he ball to pick it up (oblique motion)], 7) plan for using end effector to move to point above basket after picking up ball (perpendicular or oblique motion), 8) plan for placing ball in basket (by dropping ball or descending to basket and gently placing ball in basket). Among the four initial positions, the two balls and two baskets are placed randomly. The movement of the robot is hard-coded on the basis of a goal-position-based script. To reduce collision detection errors during the simulation, we used a large physical model (end effector size ~1 m), which does not affect the overall validity. The length of the data-point sequence was 400.
+We implemented only the end-effector rather than the entire robot arm since controlling the robot arm during a picking task is easily computable by calculating the inverse kinematics and inverse dynamics. Gripper is a 12-dimensional dataset: [joint x position, joint y position, finger1 joint position (angle), finger2 joint position (angle), box1 x position, box1 y position, box2 x position, box2 y position, ball1 x position, ball1 y position, ball2 x position, ball2 y position]. Eight factors are represented in this dataset: 1) color of ball to pick up, 2) initial location of red ball, 3) initial location of blue ball, 4) initial location of blue basket, 5) initial location of red basket, 6) plan for using end-effector to move to ball to pick it up [first, moving horizontally to the x-location of ball then descending horizontally to the y-location of ball, like the movement of the doll drawing machine (perpendicular motion); second, moving straight to the location of ball to pick it up (oblique motion)], 7) plan for using end-effector to move to point above basket after picking up ball (perpendicular or oblique motion), 8) plan for placing ball into basket (by dropping ball or descending to basket and gently placing ball into basket). The two balls and two baskets are placed randomly among the four initial positions. The movement of the robot is hard-coded on the basis of a goal-position-based script. To reduce collision-detection errors during the simulation, we used a large physical model (end-effector size ~1 m), which did not affect the overall validity. The length of the data-point sequence was 400.
 
 ## Label Table
 
-| Degree of factors | Ground truth factors               |
+| Degree of factors | Ground-truth factors               |
 | ----------------- | ---------------------------------- |
 | 4                 | Red ball initial position          |
 | 4                 | Blue ball initial position         |
 | 4                 | Red basket position                |
 | 4                 | Blue basket position               |
 | 2                 | Plan to reach target ball          |
 | 2                 | Plan to reach corresponding basket |
-| 2                 | Plan to place ball in basket       |
+| 2                 | Plan to place ball into basket     |
 

note/gripper_result.md

@@ -11,4 +11,6 @@ FAVAE learned the disentangled factors of the Gripper dataset. Examples visualiz
 **(c)  Targeted ball (red or blue)**
 ![](../gripper/1_8_-0.8.gif)
 **(d)  Drop the ball or placing it into the basket gently**
-![](../gripper/1_3_-1.20.gif)
+![](../gripper/1_3_-1.20.gif)
+
+Visualization of learned disentangled representations for Gripper dataset. Traversed latent variable shows changing position of blue ball and blue basket: (b) (9th z in 1st ladder), changing target ball from red one to blue one and moving accordingly:(c) (1st z in 2nd ladder), changing plan from dropping ball to placing it gently: (d) (4th z in 2nd ladder).

note/model_details.md

@@ -1,10 +1,8 @@
-## Architecture details
-
-
+# Experiment Details
 
 ## FAVAE (proposal model)
 
-We describe the hyper parameters used for the results in 2D Reaching dataset, 2D Wavy Reaching dataset and Sprites dataset. We search the parameters beta and best C is shown in dictionary bellow.  In Sprites dataset, we add the pre-encoder and post-decoder for encoding and decoding of images at each time step. The pre-encoding and post-decoding architecture is same as "encode_frames" and "decode_frames" in the code implementing Disentangled-Sequential-Autoencoder [This code is available](https://github.com/yatindandi/Disentangled-Sequential-Autoencoder) with conv_dim=48. We search the parameters from beta = 0 to beta = 100 and the best parameter is beta = 20 with C = [20, 10, 5].
+We describe the hyper parameters used for the results in 2D Reaching dataset, 2D Wavy Reaching dataset and Sprites dataset. We search the parameters beta and best C is shown in dictionary bellow.  In Sprites dataset, we add the pre-encoder and post-decoder for encoding and decoding of images at each time step. The pre-encoding and post-decoding architectures are the same as "encode_frames" and "decode_frames" in the code implementing the Disentangled-Sequential-Autoencoder [This code is available](https://github.com/yatindandi/Disentangled-Sequential-Autoencoder) with conv_dim=48. We search the parameters from beta = 0 to beta = 100 and the best parameter is beta = 20 with C = [20, 10, 5].
 
 ```python
 import numpy as np
@@ -31,4 +29,4 @@ c_dict = {
 ## FHVAE (baseline model)
 
 For the FHVAE experiments, we used the code from the implementation at [here](https://github.com/wnhsu/FactorizedHierarchicalVAE).
-We used the recurrent setting with LSTM encoders and decoder with unit size=256 and batch size=80. Dimensions of latent variable z1 and z2 are 7 for each. We used Adam optimizer with learning rate=0.001. We applied goal position factors (2 classes for 2D Reaching and 5 classes for 2D wavy Reaching) as label inputs. We varied alpha from 1.0 to 30.0, and the best alpha was alpha=1.0 for both 2D Reaching and 2D wavy reaching.
+We used the recurrent setting with LSTM encoders and a decoder with unit size=256 and batch size=80. The dimensions of latent variables z1 and z2 wrer 7 for each. We used Adam optimizer with learning rate=0.001. We applied goal-position factors (2 classes for 2D Reaching and 5 classes for 2D Wavy Reaching) as label inputs. We varied alpha from 1.0 to 30.0, and the best alpha was alpha=1.0 for both 2D Reaching and 2D Wavy Reaching.

sprites/README.md

@@ -1,19 +1,19 @@
 # Sprites Dataset.
 
-Sprits dataset is video data of video game "sprites". It used was used in [**Universal-LPC-spritesheet**](https://github.com/jrconway3/Universal-LPC-spritesheet) for confirming the extraction of disentangled representation between static factors and dynamic factors. The dataset consists of sequences with T = 8 frames of dimension 3 x 64 x 64. 
+Sprites dataset contains video data of the video game "sprites". It used was used in [**Universal-LPC-spritesheet**](https://github.com/jrconway3/Universal-LPC-spritesheet) for confirming the extraction of disentangled representation between static factors and dynamic factors. The dataset consists of sequences with T = 8 frames of dimension 3 x 64 x 64. 
 
 ![](./sprites.gif)
 
 ## Label Table
 
-| Degree of factors | Ground truth factors                         |
-| ----------------- | -------------------------------------------- |
-| 2                 | Body color (light or dark2)                  |
-| 2                 | Shirts color (brown or teal)                 |
-| 2                 | Hair color (green or pink)                   |
-| 2                 | Pants color (red or teal)                    |
-| 3                 | Direction of character (left, foward, right) |
-| 3                 | Motion (Spellcast, Walk, Slash )             |
+| Degree of factors | Ground-truth factors                          |
+| ----------------- | --------------------------------------------- |
+| 2                 | Skin color (light or dark2)                   |
+| 2                 | Shirts color (brown or teal)                  |
+| 2                 | Hair color (green or pink)                    |
+| 2                 | Pants color (red or teal)                     |
+| 3                 | Direction of character (left, forward, right) |
+| 3                 | Motion (Spellcast, Walk, Slash )              |
 
 ## You can use a download script.