如何在TensorFlow中按列切割2D张量并转换为3D张量？

Hey there! Let's break down how to transform your [10,12] 2D tensor into the [3,10,4] 3D tensor you need, and even get the flattened version matching your example.

Step 1: Understand the Split Logic #

Your original tensor has 12 columns, and you want to split it vertically (by columns) into 3 chunks of 4 columns each:

• First chunk: Columns 0-3 (before the 3rd column, 0-indexed)
• Second chunk: Columns 4-7 (after 3rd column, before 7th)
• Third chunk: Columns 8-11 (after 7th column)

Each chunk will be a [10,4] sub-matrix, which we'll then stack into a 3D tensor with batch_size=3 as the first dimension.

Step 2: Implementation with NumPy #

Since you provided a NumPy array example, here's how to do it with NumPy:

import numpy as np

# Define your original 2D tensor K
K = np.array([[1, 9, 5, 9, 9, 2, 0, 9, 1, 9, 0, 6], 
              [0, 4, 8, 4, 3, 3, 8, 8, 7, 0, 3, 8], 
              [7, 7, 1, 8, 4, 7, 0, 4, 9, 0, 6, 4], 
              [2, 4, 6, 3, 3, 7, 8, 5, 0, 8, 5, 4], 
              [7, 4, 1, 3, 3, 9, 2, 5, 2, 3, 5, 7], 
              [2, 7, 1, 6, 5, 0, 0, 3, 1, 9, 9, 6], 
              [6, 7, 8, 8, 7, 0, 8, 6, 8, 9, 8, 3], 
              [6, 1, 7, 4, 9, 2, 0, 8, 2, 7, 8, 4], 
              [4, 1, 7, 6, 9, 4, 1, 5, 9, 7, 1, 3], 
              [5, 7, 3, 6, 6, 7, 9, 1, 9, 6, 0, 3]])

# Split into three [10,4] sub-matrices
chunk1 = K[:, :4]   # Columns 0-3
chunk2 = K[:, 4:8]  # Columns 4-7
chunk3 = K[:, 8:]   # Columns 8-11

# Stack the chunks along axis=0 to get [3,10,4] 3D tensor
K_new = np.stack([chunk1, chunk2, chunk3], axis=0)

# Optional: Flatten to get the 1D array matching your example
K_new_flat = K_new.flatten()
print(K_new_flat)

Step 3: Verify the Results #

• K_new.shape will return (3, 10, 4), which matches your requirement of batch_size=3 and sequence_length=4.
• The flattened K_new_flat will exactly match the 1D array you provided, since NumPy's flatten() uses row-major (C-style) order by default.

Alternative: PyTorch Implementation #

If you're working with PyTorch tensors instead, the logic is almost identical:

import torch

# Define your tensor as a PyTorch tensor
K = torch.tensor([[1, 9, 5, 9, 9, 2, 0, 9, 1, 9, 0, 6], 
                  [0, 4, 8, 4, 3, 3, 8, 8, 7, 0, 3, 8], 
                  [7, 7, 1, 8, 4, 7, 0, 4, 9, 0, 6, 4], 
                  [2, 4, 6, 3, 3, 7, 8, 5, 0, 8, 5, 4], 
                  [7, 4, 1, 3, 3, 9, 2, 5, 2, 3, 5, 7], 
                  [2, 7, 1, 6, 5, 0, 0, 3, 1, 9, 9, 6], 
                  [6, 7, 8, 8, 7, 0, 8, 6, 8, 9, 8, 3], 
                  [6, 1, 7, 4, 9, 2, 0, 8, 2, 7, 8, 4], 
                  [4, 1, 7, 6, 9, 4, 1, 5, 9, 7, 1, 3], 
                  [5, 7, 3, 6, 6, 7, 9, 1, 9, 6, 0, 3]])

# Split into chunks
chunk1 = K[:, :4]
chunk2 = K[:, 4:8]
chunk3 = K[:, 8:]

# Stack into 3D tensor
K_new = torch.stack([chunk1, chunk2, chunk3], dim=0)

# Flatten to 1D
K_new_flat = K_new.flatten()
print(K_new_flat)

内容的提问来源于stack exchange，提问作者Sarath R Nair