63 lines
1.8 KiB
Python
63 lines
1.8 KiB
Python
from typing import Any, Dict
|
|
import torch
|
|
import ast
|
|
|
|
|
|
class PtView:
|
|
"""
|
|
Reshapes a PyTorch tensor into a specified shape using `torch.view()`.
|
|
The total number of elements must remain unchanged.
|
|
|
|
For example, if a tensor has shape (2, 3, 4), and `new_shape="[6, 4]"`,
|
|
the output will have shape (6, 4).
|
|
|
|
Use -1 to automatically infer a dimension, e.g.:
|
|
```
|
|
[2, -1]
|
|
```
|
|
category: PyTorch wrapper - Transform
|
|
"""
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(cls) -> Dict[str, Any]:
|
|
"""
|
|
Defines the input types for the function.
|
|
|
|
Returns:
|
|
Dict[str, Any]: A dictionary specifying required input types.
|
|
"""
|
|
return {
|
|
"required": {
|
|
"tens": ("TENSOR", {}),
|
|
"new_shape": ("STRING", {"multiline": True})
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES: tuple = ("TENSOR",)
|
|
FUNCTION: str = "f"
|
|
CATEGORY: str = "Data Analysis"
|
|
|
|
def f(self, tens: torch.Tensor, new_shape: str) -> tuple:
|
|
"""
|
|
Reshapes the input tensor based on the specified shape.
|
|
|
|
Args:
|
|
tens (torch.Tensor): The input tensor.
|
|
new_shape (str): A string representation of the desired shape,
|
|
e.g., "[2, -1]" or "(6, 4)".
|
|
|
|
Returns:
|
|
tuple: A tuple containing the reshaped PyTorch tensor.
|
|
"""
|
|
# Convert string to a tuple of integers
|
|
shape_order = tuple(ast.literal_eval(new_shape))
|
|
|
|
# Ensure the parsed shape is valid
|
|
if not isinstance(shape_order, tuple) or not all(isinstance(i, int) for i in shape_order):
|
|
raise ValueError(f"Invalid format for new_shape: {new_shape}. Must be a tuple or list of integers.")
|
|
|
|
# Apply view operation
|
|
tens = tens.view(*shape_order)
|
|
|
|
return (tens,)
|