Feature conv3d

[NewBornRustacean]

what does this PR do

• implementation of conv3d
• rank6 permutation macro added
• related issues: 3D Convolution #28 v0.3 #29 [Question] implementing Conv3d #50

test log

• dimension mismatch is resolved(24. 05. 03)
• test fails at this time of writing(24. 05. 02)
• I got index out of bounds error when the graph execute() called
  • in the test case, len(inp1)==720, because example input data has 5 channels, 16x9x5=720 volume.
  • rust says, the function op.rs::get_index() fails where data[ind.exec_single_var_stack(index, stack)]

   fn get_index(
     data: &[f32],
     (ind, val): &(BigExpression, BigExpression),
     stack: &mut Vec<i64>,
     index: usize,
   ) -> f32 {
     if val.exec_single_var_stack(index, stack) != 0 {
       data[ind.exec_single_var_stack(index, stack)]
     } else {
       0.0
     }
  }

need help (__)

I have no idea where and how to start debug.. I'd really appreciate any ideas you could suggest. @jafioti @TheSeamau5

println!("{:?}", cx)

Graph:
  Tensors: {}
  Dynamic Map: {}
  Graph:
    Type: Directed
    Node Count: 17
    Edge Count: 14
    Edges:
      - (Node 1, Node 4)
      - (Node 4, Node 5)
      - (Node 5, Node 6)
      - (Node 6, Node 7)
      - (Node 7, Node 8)
      - (Node 8, Node 9)
      - (Node 9, Node 10)
      - (Node 10, Node 11)
      - (Node 11, Node 12)
      - (Node 12, Node 13)
      - (Node 13, Node 14)
      - (Node 3, Node 15)
      - (Node 14, Node 15)
      - (Node 15, Node 16)
    Node Weights:
      - Node 0: Tensor Load
      - Node 1: Tensor Load
      - Node 2: Tensor Load
      - Node 3: Weight Load
      - Node 4: Contiguous
      - Node 5: Contiguous
      - Node 6: Contiguous
      - Node 7: Contiguous
      - Node 8: Contiguous
      - Node 9: Contiguous
      - Node 10: Contiguous
      - Node 11: Contiguous
      - Node 12: Contiguous
      - Node 13: Contiguous
      - Node 14: Contiguous
      - Node 15: Multiplication
      - Node 16: Sum Reduction (axis: 2)
    Edge Weights:
      - Edge 0: Data (input_order: 0, output_order: 0, shape: [5, 16, 9, 5, 2])
      - Edge 1: Data (input_order: 0, output_order: 0, shape: [5, 16, 9, 10])
      - Edge 2: Data (input_order: 0, output_order: 0, shape: [5, 16, 9, 2, 7])
      - Edge 3: Data (input_order: 0, output_order: 0, shape: [5, 16, 9, 2, 2, 8])
      - Edge 4: Data (input_order: 0, output_order: 0, shape: [5, 9, 2, 2, 128])
      - Edge 5: Data (input_order: 0, output_order: 0, shape: [5, 9, 2, 2, 8, 18])
      - Edge 6: Data (input_order: 0, output_order: 0, shape: [5, 9, 2, 2, 8, 2])
      - Edge 7: Data (input_order: 0, output_order: 0, shape: [5, 2, 2, 64, 32])
      - Edge 8: Data (input_order: 0, output_order: 0, shape: [5, 2, 2, 2048])
      - Edge 9: Data (input_order: 0, output_order: 0, shape: [5, 2, 2, 32, 66])
      - Edge 10: Data (input_order: 0, output_order: 0, shape: [5, 2, 2, 32, 2])
      - Edge 11: Data (input_order: 0, output_order: 0, shape: [2, 40, 64])
      - Edge 12: Data (input_order: 1, output_order: 0, shape: [40, 64, 2])
      - Edge 13: Data (input_order: 0, output_order: 0, shape: [2, 64, 40])
  Free Node: NodeIndex(4294967295)
  Free Edge: EdgeIndex(4294967295)
  Nodes to Keep:
    - NodeIndex(16)
    - NodeIndex(2)
  Nodes to Retrieve:
    - NodeIndex(16): Shape ([2, 8, 4, 2])
    - NodeIndex(2): Shape ([2, 8, 4, 2])
  Linearized Graph: None
  Consumers Map: None

[jafioti]

Amazing job! This is great work for your first contribution. I'll take a look at debugging later today or tomorrow. Should have it merged by the weekend

[NewBornRustacean]

thanks for your comment! I'm still debugging..
btw, "print cx" is awesome(also display() is). I'm really impressed static graph strategy could be like this.

[NewBornRustacean]

Now, the dim. mismatch resolved!

[jafioti]

Amazing work, merged!