Confusion regarding tensor.dtypes

[AksharGoyal]

Course video: Learn PyTorch for deep learning in a day. Literally., timestamp: 1:57:00

Hello, I was experimenting with the dtypes and I got this result. I am curious to understand what's the logic behind this. Why am I only getting 4 digits after 0 and why not more after increasing precision?

My code

torch.tensor([0.0102030405], dtype=torch.float16),
torch.tensor([0.0102030405], dtype=torch.float32),
torch.tensor([0.0102030405], dtype=torch.float64)

[mrdbourke]

Hi there,

This is likely due to setting torch.set_printoptions, see: https://pytorch.org/docs/stable/generated/torch.set_printoptions.html#torch.set_printoptions

And see: https://discuss.pytorch.org/t/how-to-print-more-than-4-digits-after-decimal/63405

Example

Default print options.

import torch

float16 = torch.tensor([0.0102030405], dtype=torch.float16)
float32 = torch.tensor([0.0102030405], dtype=torch.float32)
float64 = torch.tensor([0.0102030405], dtype=torch.float64)

print(float16)
print(float32)
print(float64)

Output:

tensor([0.0102], dtype=torch.float16)
tensor([0.0102])
tensor([0.0102], dtype=torch.float64)

Set print options precision to 10:

# Set precision equal to 10
torch.set_printoptions(precision=10)

print(float16)
print(float32)
print(float64)

Output:

tensor([0.0102005005], dtype=torch.float16)
tensor([0.0102030402])
tensor([0.0102030405], dtype=torch.float64)

[AksharGoyal]

This is helpful, thank you so much!