PyTorch 张量与自动微分：形状、设备和梯度

import torch

values = torch.tensor([[1.0, 2.0], [3.0, 4.0]])
print(values.shape)
print(values.dtype)
print(values.device)
print(values.requires_grad)

import numpy as np

array = np.array([1.0, 2.0, 3.0], dtype="float32")
shared = torch.from_numpy(array)
copied = torch.tensor(array)
array[0] = 99.0
print(shared, copied)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tensor = torch.ones(3, 2, device=device)
print(device, tensor.device)

import torch

torch.manual_seed(42)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

features = torch.tensor(
    [[1.0, 2.0], [2.0, 0.5], [3.0, 1.5]],
    device=device,
)
weight = torch.tensor([0.3, -0.2], device=device, requires_grad=True)
bias = torch.tensor(0.1, device=device, requires_grad=True)
target = torch.tensor([1.0, 0.5, 1.4], device=device)

prediction = features @ weight + bias
loss = ((prediction - target) ** 2).mean()
loss.backward()

print("device:", device)
print("prediction shape:", prediction.shape)
print("loss:", round(loss.item(), 6))
print("weight grad:", weight.grad.detach().cpu().numpy().round(4))
print("bias grad:", round(bias.grad.item(), 4))

weight.grad.zero_()
bias.grad.zero_()
print("cleared:", weight.grad, bias.grad)

with torch.inference_mode():
    inference_prediction = features @ weight + bias
print(inference_prediction.requires_grad)