In PyTorch, prefer creating tensors directly using torch.rand, torch.tensor, or other Torch methods instead of converting from NumPy arrays. Avoid using torch.tensor(np.random.rand(…​)) or similar patterns when the same result can be achieved directly with PyTorch.

Non Compliant Code Example

import torch
import numpy as np

def non_compliant_random_rand():
    tensor = torch.tensor(np.random.rand(1000, 1000))

Compliant Solution

import torch

def compliant_random_rand():
    tensor = torch.rand([1000, 1000])

Relevance Analysis

Experiments were conducted to compare the performance and environmental impact of two tensor creation methods in PyTorch:

  • Using NumPy for random data generation followed by conversion to PyTorch tensor

  • Direct creation using native PyTorch tensor functions (torch.rand, torch.tensor, etc.)

Configuration

  • Processor: Intel® Xeon® CPU 3.80GHz

  • RAM: 64GB

  • GPU: NVIDIA Quadro RTX 6000

  • CO₂ Emissions Measurement: CodeCarbon

  • Framework: PyTorch

  • Dataset: MNIST

  • Model: Simple 2-layer fully connected network

Context

Two workflows were benchmarked: - NumPy-based: Data created using NumPy and converted to PyTorch - Torch-based: Data created natively using PyTorch tensor operations

Metrics assessed: - Training execution time - CO₂ emissions - Final model accuracy

Impact Analysis

image

  • Execution Time: Torch-based method reduced total training time by more than 50%

  • Carbon Emissions: Torch-based method lowered emissions by approximatively 50%

  • Accuracy: Both approaches yielded comparable model accuracy

Conclusion

Using native PyTorch methods to create tensors: