Machine Learning Quickref
Matplotlib
Recommend using OO-style:
from matplotlib import pyplot as plt
def draw_image(data: dict):
"""
data = {"data1": [x0, y00, y01], "data2": [x1, y10, y11],
"param": {"structure", "lr", "activation"}
}
"""
fig, axs = plt.subplots(nrows=1, ncols=3, layout="tight", figsize=[9.6, 4.8])
axs[0].plot(data["data1"][0], data["data1"][1], label="target")
axs[0].plot(data["data1"][0], data["data1"][2], label="model")
axs[0].set_xlabel("x")
axs[0].set_ylabel("y = f(x)")
axs[0].set_title("Fitting")
axs[0].legend()
...
fig.suptitle(f'structure: {data["param"]["structure"]}\
lr: {data["param"]["lr"]}\
activation: {data["param"]["activation"]}')
fig.savefig('./figure/figure.png')
plt.show()
When operating on images, image input and ndarray input are both supported:
from matplotlib.images import imread
#...
axs[0].imshow(imread('a.png'))
import numpy as np
from PIL import Image
#...
axs[1].imshow(np.asarray(Image.open('a.png')))
plt.show()
PyTorch
Learn the Basics — PyTorch Tutorials 2.0.0+cu117 documentation
Device
device = "cuda" if torch.cude.is_available() else "cpu"
print(f"Device: {device}")
tensor = tensor.to(device)
model = model.to(device)
Tensors
torch.Tensor - PyTorch 2.0 documentation
Create tensors
shape = (2, 3)
tensorr = torch.rand(shape)
tensor0 = torch.zeros(shape)
tensor1 = torch.ones(shape)
tensorn = torch.from_numpy(np.array([...]))
Attributes
tensor.shape
tensor.dtype
tensor.device
tensor.require_grad
Methods
torch.Tensor.detach(): returns a tensor sharing same storage with original tensor but never requiring grad (so that can be convert to normal datatypes if on cpu)
torch.Tensor.item(): works when only one element in tensor
tensor = Tensor([12.0])
print(tensort.item(), type(tensor.item())
# 12.0 <class 'float'>
Add “_” suffix to operations means inplace operations. Example: tensor.add_().
Datasets & DataLoaders
primitives:
torch.utils.data.DataLoader: wraps an iterable around Dataset
torch.utils.data.Dataset: samples and labels
Load FashionMNIST
import torch
from torch.utils.data import Dataset
from torchvision import datasets
from torchvision.transforms import ToTensor
import matplotlib.pyplot as plt
training_data = datasets.FashionMNIST(
root="data", # path to data
train=True, # specify training/test dataset
download=True, # if not available at root, whether download from internet
transform=ToTensor()
)
test_data = datasets.FashionMNIST(
root="data",
train=False,
download=True,
transform=ToTensor()
)
We can index Datasets manually like a list: training_data[index].
We use matplotlib to visualize some samples in our training data.
labels_map = {
0: "T-Shirt",
1: "Trouser",
2: "Pullover",
3: "Dress",
4: "Coat",
5: "Sandal",
6: "Shirt",
7: "Sneaker",
8: "Bag",
9: "Ankle Boot",
}
sample_idx = torch.randint(len(training_data), size=(1,)).item()
img, label = training_data[sample_idx]
plt.title(labels_map[label])
Wrap my data to dataset
PyTorch has 2 kinds of dataset: map and iterate. We focus on the first kind.
All datasets should subclass torch.utils.data.Dataset.
All datasets should overwrite __getitem()__, and optionally overwrite __len__().
My data should have: a file containing the map each data’s filename to its label; a data directory.
A sample dataset:
from torch.utils.data import Dataset
import pandas as pd
from pathlib import Path
class MyDataset(Dataset):
def __init__(self, annotation_file="./dataset/label.csv",\
data_dir="./dataset/data/", transform=None, target_transform=None):
self.labels = pd.read_csv(annotation_file)
self.data_dir = data_dir
def __getitem__(self, index):
path = Path(self.dir) + Path(self.labels.iloc[index, 0])
item = read_item(path) # read_item() TODO
label = self.labels.iloc[index, 1]
if self.transform:
item = self.transform(item)
if self.target_transform:
label = self.target_transform(label)
return item, label
def __len__(self):
return len(self.labels)