data.md

Dataset

This file is a description about the folder structure of data folder

Generate and load data

These data is generated by the function create_dataset() or create_dataset_gt() in pack.py, for more detial please check these code. The process of loading data is done by the class PACKDataset in pack.py.

Folder structure:

• rand_2d/: Random data on 2D case
• rand_3d/: Random data on 2D case
• gt_2d/: PPSG data on 2D case
• gt_3d/: PPSG data on 3D case

Each dataset will generate under these main folder, for example:

• rand_2d/
  • pack-train-10-12800-7-1-5/
    • blocks.txt : The 2D / 3D size of blocks
    • pos.txt : The position of blocks in intial container
    • dep_move.txt : The top block states of each block
    • dep_small.txt : The left / forward access block states of each block
    • dep_large.txt : The right / backward access block states of each block
    • container.txt : The target container id of each block (only use for multi-container case)

Under these folders, we have many kinds for dataset, pack-train-10-1280-7-1-5 means the training dataset with 1280 data, each data has 10 blocks, the initial container width is 7, the block size is from 1~5. pack-valid-6-100-5-1-4 means the testing dataset with 100 data, each data has 6 blocks, the initial container width is 5, the block size is from 1~4.

Data structure:

Let number of blocks is n, data size is D, inital container size if w

• blocks.txt

  In 2D case, the number of input block states is (n*2), (n*6) for 3D case. Each line of data is n blocks size unber a rotation state, the total number of lines of the file is (D*2) for 2D, (D*6) for 3D

• pos.txt

  Each line of data is the position of each block in initial container, the total number of lines of the file is (D)

• dep_move.txt

  We generate a 2D top block/precedence array with size of (n)*(n), then flatten into a 1D array (n*n). Each line of data is the top block/precedence data, the total number of lines of the file is (D)

• dep_small.txt

  Since we have 2 rotation states for 2D, 6 rotation states for 3D, we should generate 2D left/forward block/precedence array with size of (n)*(n), then flatten into a 1D array (n*n) for each rotation state. Each line of data is the left/forward block/precedence data for a rotation state, the total number of lines of the file is (D*2) for 2D, (D*6) for 3D

• dep_large.txt

  Since we have 2 rotation states for 2D, 6 rotation states for 3D, we should generate 2D right/backward block/precedence array with size of (n)*(n), then flatten into a 1D array (n*n) for each rotation state. Each line of data is the right/backward block/precedence data for a rotation state, the total number of lines of the file is (D*2) for 2D, (D*6) for 3D

• container.txt

  Each line of data is a targe container id of each block, which is generate randomly, only use for multi-container case

Example

Take the data shows in the figure as example, we want to generate 2 groups of data of 3 blocks in 2D case

• block.txt

  • For data 1, we have 3 blocks with size: [3,2],[1,1],[1,2]
  • For data 2, we have 3 blocks with size: [1,2],[2,3],[1,2]
  • The block.txt will be:

  3 2 1 1 1 2  #  data 1 original state
  2 3 1 1 2 1  #  data 1 rotated by Z axis
  1 2 2 3 1 2  #  data 2 original state
  2 1 3 2 2 1  #  data 2 rotated by Z axis
  

• pos.txt

  • For data 1, we have 3 positons of each block: [0,0],[3,0],[3,1]
  • For data 2, we have 3 positons of each block: [0,0],[1,0],[3,0]
  • The pos.txt will be:

  0 0 3 0 3 1  #  data 1 block positions
  0 0 1 0 3 0  #  data 2 block positions
  

• dep_move.txt

  • For data 1, we can generate the 2D top block array, denote the array as B, B[i][j] means block j is blocked by block i:

    0 0 0
    0 0 0
    0 1 0
    

  • For data 2, we can generate the 2D top block array, denote the array as B, B[i][j] means block j is blocked by block i:

    0 0 0
    0 0 0
    0 0 0
    

  • The dep_move.txt will be:

  0 0 0 0 0 0 0 1 0  #  data 1 top block array (after flattened)
  0 0 0 0 0 0 0 0 0  #  data 2 top block array (after flattened)
  

• dep_small.txt

  • For data 1, we can generate the 2D left/forward block array, denote the array as B, B[i][j] means block j is blocked by block i:

    1 1 1
    0 0 0
    0 0 0
    

  • For data 2, we can generate the 2D left/forward block array, denote the array as B, B[i][j] means block j is blocked by block i:

    1 1 0
    0 0 1
    0 0 0
    

  • The dep_small.txt will be:

  0 0 0 0 0 0 0 0 0  #  data 1 block array (after flattened) of rotation state 1, since rotation state 1 (original state) doesn't need to grasp from left size of block, we don't need to consider about its left block state
  1 1 1 0 0 0 0 0 0  #  data 1 left block array (after flattened) of rotation state 2
  0 0 0 0 0 0 0 0 0  #  data 2 block array (after flattened) of rotation state 1, since rotation state 1 (original state) doesn't need to grasp from left size of block, we don't need to consider about its left block state
  1 1 0 0 0 1 0 0 0  #  data 2 left block array (after flattened) of rotation state 2
  

• dep_large.txt

  • For data 1, we can generate the 2D right/backward block array, denote the array as B, B[i][j] means block j is blocked by block i:

    0 0 0
    1 1 0
    1 0 1
    

  • For data 2, we can generate the 2D right/backward block array, denote the array as B, B[i][j] means block j is blocked by block i:

    0 0 0
    1 0 0
    0 1 1
    

  • The dep_large.txt will be:

  0 0 0 0 0 0 0 0 0  #  data 1 block array (after flattened) of rotation state 1, since rotation state 1 (original state) doesn't need to grasp from right size of block, we don't need to consider about its right block state
  0 0 0 1 1 0 1 0 1  #  data 1 right block array (after flattened) of rotation state 2
  0 0 0 0 0 0 0 0 0  #  data 2 block array (after flattened) of rotation state 1, since rotation state 1 (original state) doesn't need to grasp from right size of block, we don't need to consider about its right block state
  0 0 0 1 0 0 0 1 1  #  data 2 right block array (after flattened) of rotation state 2