Welcome to creating tensors in code! Now that you understand tensor shapes, let's learn how to make them using PyTorch, a popular library for machine learning. PyTorch makes it easy to create and manipulate tensors efficiently.
Engagement Message
Have you heard of PyTorch before?
NumPy arrays and PyTorch tensors are actually very similar—they both represent multi-dimensional data.
You can easily convert a NumPy array to a PyTorch tensor with torch.from_numpy(numpy_array), and back with tensor.numpy(). This makes it simple to move data between scientific Python code (NumPy) and machine learning code (PyTorch).
Engagement Message
Isn't this useful?
The most intuitive way to create tensors is from existing data like lists.
In PyTorch: torch.tensor([1, 2, 3]) creates a 1D tensor from a list.
Engagement Message
What shape would you get from torch.tensor([[1, 2], [3, 4]])?
Sometimes you need tensors filled with specific values. Common patterns include:
Zeros: torch.zeros((3, 4)) creates a 3x4 tensor of zeros
Ones: torch.ones((2, 2)) creates a 2x2 tensor of ones
Engagement Message
Why might you want a tensor filled with zeros when starting a calculation?
Random tensors are crucial for machine learning, especially for initializing model weights.
torch.rand((2, 3)) creates random values between 0 and 1.
torch.randn((2, 3)) creates random values from a normal distribution.
Engagement Message
When might random initialization be better than starting with zeros?
Data types matter for memory and computation speed. Common types include:
torch.float32: 32-bit floating point (most common in ML)torch.int64: 64-bit integerstorch.bool: True/False values
You specify like: torch.zeros((2, 2), dtype=torch.float32)
Engagement Message
What data type would you use for storing pixel intensities (0-255)?
Let's inspect a tensor after creating it from a Python list. This helps you understand both the data and its shape.
This will output:
Engagement Message
Does this make sense?
Type
Swipe Left or Right
Practice Question
Which creation method would you use for each scenario? Match each use case with the best tensor creation approach.
Labels
- Left Label: Zeros/Ones
- Right Label: Random Creation
Left Label Items
- Initializing bias parameters to zero
- Creating a mask to hide certain values
- Starting with a clean slate for accumulation
- Setting up a binary classification target
Right Label Items
- Initializing neural network weights
- Generating synthetic training data
- Adding noise to existing data
- Creating diverse starting conditions
