Skip to content

PyTorch

Init

import torch
from torch import nn, optim
from torch.utils.data import Dataset, DataLoader

import torchvision
from torchvision import transforms
from torchvision.datasets import ImageFolder
import timm

import matplotlib.pyplot as plt ## For data viz
import pandas as pd
import numpy as np

import sys
from tqdm.notebook import tqdm

print('System Version:', sys.version)
print('PyTorch version', torch.__version__)
print('Torchvision version', torchvision.__version__)
print('Numpy version', np.__version__)
print('Pandas version', pd.__version__)

device = (
"mps" if getattr(torch, "has_mps", False)
else
"cuda" if torch.cuda().is_available()
else
"cpu"
)

Tensors

torch.mean(image_data, axis=0) ## column-wise mean

luminance_approx = torch.mean(image_array, axis=-1) ## color_channel-wise mean

values, indices = torch.max(data, axis=-1)
  • int8 is an integer type, it can be used for any operation which needs integers
  • qint8 is a quantized tensor type which represents a compressed floating point tensor, it has an underlying int8 data layer, a scale, a zero_point and a qscheme

Conversion To Tensor

# ❌ creates a copy
tensor = torch.tensor(array)

# ✅ avoids copying
tensor = torch.as_tensor(array)
tensor = torch.from_numpy(array)
# however, changing array will also affect tensor

Conversion From Tensor

# ❌
tensor.cpu()
tensor.item()
tensor.numpy()

# ✅
tensor.detach()

Creating Tensors

# ❌
tensor.rand(2, 2).cuda()

# ✅
tensor.rand(2, 2, device=torch.device("cuda:0"))

API

Model Mode

model.train()
model.eval()

Sequential

nn.Sequential(
    nn.LazyLinear(100),
  nn.ReLU()
)

Lazy Layers

automatically detect the input size

Only specify output size

nn.Sequential(
  nn.LazyLinear(1000),
  nn.LazyLinear(10),
  nn.LazyLinear(100)
)

Save Model

torch.save(model, "model.pkl")

View Parameters

for param in model.parameters():
  print(name)

for name, param in model.named_parameters():
    if param.requires_grad:
        print(name, param.data)

Custom Loss Function

class loss(nn.module):
  def forward(self, pred, y):
    error = pred-y
    return torch.mean(
      torch.abs(error)
    )

IDK

Forward pass

model.train()
model.eval()

Backward pass

with torch.set_grad_enabled(True): # turn on history tracking
  # training

with torch.set_grad_enabled(False): # turn off history tracking
  # testing
Last Updated: 2024-05-12 ; Contributors: AhmedThahir

Comments