seppl/pytorch_learn
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

add readme, chapter 03 finished

Browse Source
This commit is contained in:
Joseph Hopfmüller
2022-10-16 23:10:06 +02:00
parent 13a8845ab3
commit b5ab293a9b
5 changed files with 253 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
01_tensors/01_tensors_playground.py
View File

@@ -1 +0,0 @@
pip

166
02_tensors_playground.py Normal file
View File

@@ -0,0 +1,166 @@
import torch
# empty, zeros, ones of different sizes, specify datatype
print('empty, zeros, ones of different sizes, specify datatype')
x = torch.empty(1)
print(x)
x = torch.empty(3)
print(x)
x = torch.zeros(2, 3)
print(x)
x = torch.ones(2,3,4)
print(x)
x = torch.ones(2, 5, dtype=torch.float64)
print(x.dtype)
print()
# from data
print('from data')
x = torch.tensor([2.5, 0.1])
print(x)
print()
#basic ops
print('basic ops')
x = torch.rand(2,2)
y = torch.rand(2,2)
print('add')
z1 = x + y
z2 = torch.add(x,y)
print(x)
print(y)
print(z1)
print(z2)
# in place addition
x.add_(y)
print(x)
print('sub')
z1 = x - y
z2 = torch.sub(x,y)
print(x)
print(y)
print(z1)
print(z2)
# in place addition
x.sub_(y)
print(x)
print('mul')
z1 = x * y
z2 = torch.mul(x,y)
print(x)
print(y)
print(z1)
print(z2)
# in place addition
x.mul_(y)
print(x)
print('div')
z1 = x / y
z2 = torch.div(x,y)
print(x)
print(y)
print(z1)
print(z2)
# in place addition
x.div_(y)
print(x)
print()
#slicing
print('slicing, item')
x = torch.rand(2,3,2)
print(x)
print(x[:,2,:])
print(x[1, 2, 1])
print(x[1, 2, 1].item()) # for single element tensors only
print()
# reshaping
print('reshaping')
x = torch.rand(4,4)
print(x)
y = x.view(16)
print(y)
y = x.view(-1, 8)
print(y)
y = x.view(2, -1)
print(y)
# y = x.view(3, -1) # fails 'shape is invalid'
# print(y)
print()
# transposing
print('transposing')
x = torch.rand(2, 3)
print(x.size())
x = torch.transpose(x, 0, 1)
print(x.size())
x = torch.t(x)
print(x.size())
x = torch.rand(2, 3, 4)
print(f'Original: {x.size()}')
x = torch.transpose(x, 0, 1)
print(f'01: {x.size()}')
x = torch.rand(2, 3, 4)
x = torch.transpose(x, 1, 2)
print(f'12: {x.size()}')
x = torch.rand(2, 3, 4)
x = torch.transpose(x, 0, 2)
print(f'02: {x.size()}')
print()
# numpy
import numpy as np
print('numpy')
a = torch.ones(5)
print(a)
b = a.numpy()
print(b)
print(type(b))
a.add_(1) # vectors/tensors share same memory
print(a)
print(b)
c = np.ones(5)
print(c)
d = torch.from_numpy(c)
print(d)
e = d.to(dtype=torch.float32) # e has its own memory
print(e)
c += 1 # c and d share same memory
print(c)
print(d)
print(e)
# device
print('device')
if torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
print(device)
x = torch.ones(5, device=device)
y = torch.ones(5)
y = y.to(device)
z = x+y
print(z)
z = z.to('cpu') #move tensor back to cpu for conversion into numopy vector
print(z)
a = z.numpy()
print(a)
x = torch.ones(5, requires_grad=True) # enable grad for autograd
print(x)

83
03_autograd.py Normal file
View File

@@ -0,0 +1,83 @@
import torch
x = torch.tensor([1.0,2.0,3.0], requires_grad=True)
print(x)
y = x+2 # a function used in backprop for calculating the gradient is created
# y.retain_grad() # for getting grad of y (a non-leaf tensor)
print(y)
z = y*y*2
z = y.mean()
print(z)
z.backward() # no argument needed because z is scalar -> will calculate the gradient pretty accurately
# print(y.grad)
print(x.grad)
z = y*y*2
print(z)
# z.backward() will fail because z is not scalar -> create vector vor Jacobian-Vector product (JVP)
# you have to specify the step size for the gradient approximation
# (calculation via chain rule Jacobian * vector = gradient vector) vector is size of step for each element -> very small elements approximate the gradient well
v = torch.tensor([0.000000001, 0.000000001, 0.000000001], dtype=torch.float32)
z.backward(v) # pass vector to JVP
print(x.grad)
# prevent operation from being tracked by gradient tracking (requires_grad)
# 3 options
# 1. x.requires_grad_(False) -> turn off requires_grad completely
# 2. x.detach() -> returns new tensor without requires_grad
# 3. with torch.no_grad(): -> lets you do operations without grad tracking temporarily
x = torch.tensor([1.0,2.0,3.0], requires_grad=True)
y = x*x
print(x)
# 1
x.requires_grad_(False)
print(x)
x = torch.tensor([1.0,2.0,3.0], requires_grad=True)
y = x*x
print(x)
# 2
z = x.detach()
print(z)
x = torch.tensor([1.0,2.0,3.0], requires_grad=True)
y = x*x
print(x)
#3
with torch.no_grad():
a = x+2
print(a)
b = x+2
print(b)
# gradients will be summed up! -> empty gradients
#this is a dummy training
weights = torch.ones(4, requires_grad=True)
for epoch in range(3):
model_output = (weights*3).sum()
model_output.backward()
print(weights.grad)
weights.grad.zero_()# clear gradients
#later
optimizer = torch.optim.SGD(weights, lr=0.01) # stochastic gradient descent
optimizer.step()
optimizer.zero_grad() # clear gradients
# RECAP
# turn on gradient tracking for interesting vectors (f(x) = x², f'(x) = ? -> requires_grad=True for x)
# calculate gradient with f.backward(), specify step size for vectors (not needed for scalar functions like mean())
# clear gradients with x.grad.zero_()
# prevent operations from being tracked in the comp graph with one of the 3 options above

0
04_backpropagation.py Normal file
View File

4
README.md Normal file
View File

@@ -0,0 +1,4 @@
# pytorch playground
I'm working through the video [Deep Learning with PyTorch - Full Course](https://www.youtube.com/watch?v=c36lUUr864M) by Python Engineer
the file names match the chapters in the video