Torch Memo

Preliminaries

• .__dir__(): Getting all the functions in the module (Tool function)

TensorBoard

from torch.utils.tensorboard import SummaryWriter
import numpy as np
from PIL import Image

writer = SummaryWriter("logs")

def compute(x):
    return (torch.sin(torch.tensor(x)) + torch.randn(1) * torch.sin(torch.tensor(x)))


# writer.add_scalar()
for i in range(100):
    writer.add_scalar("y=x", compute(i).item(), i)


# writer.add_image()
img = Image.open("003.jpg")
img_array = np.array(img)

# dataformats is HWC
writer.add_image("test", img_array, 1, dataformats='HWC')

Transforms

Use Transforms to make Data Augmentation and design your own class of data!

Use GPU in torch

# Memo: Using GPU in Pytorch

import torch

import argparse

# Using parser
parser = argparse.ArgumentParser(description="Selecting device for Pytorch")
parser.add_argument(
    "--device",
    type=str,
    choices=["cuda", "cpu"],
    default="cuda" if torch.cuda.is_available() else "cpu",
    help= "Device to use, cuda or cpu"
)
args = parser.parse_args()

# Usage
device = args.device
print(f"Device selectd, using {device}")

Tensor Operations

torch中最关键的数据结构就是张量，因此对张量做操作是数据预处理的关键操作之一。