pxalyze 0.1.0

Pxalyze - a single file for your computer vision debugging

Pxalyze

pxalyze - a single file for your computer vision debugging

Installation

pxalyze tries to be very universal so it should work with Python since 3.7 (the lowest tested version)

pip install pxalyze

It requires numpy and optionally opencv for visualizations you can install full version with

pip install pxalyze[opencv]

Usage

pxalyze is a single file library for debugging images as arrays or tensors it is perfect for interactive debugging in REPL environments such as debug consoles or just interactive python shells

The intended usage is the following, but depending on your setup you may choose to import it differently

from pxalyze import *

Alternatively to isolate the namespaces you can also use

import pxalyze as xa

Introspection

what is the method for introspection that helps to quickly analyze arrays or tensors, it returns dict

>>> from pxalyze import *
>>> x = np.random.random((1, 3, 256, 256))
>>> what(x)
{'type': <class 'numpy.ndarray'>, 'min': 1.072805627044815e-05, 'mean': 0.5001992238233383, 'max': 0.9999899398862, 'shape': (1, 3, 256, 256), 'dtype': dtype('float64')}

It works with standard types too, displaying basic information

>>> what([1, 2])
{'type': <class 'list'>, 'len': 2}

>>> what({"a": 0})
{'type': <class 'dict'>, 'keys': ['a']}

You can also use rwhat to recursively introspect lists or dicts of arrays

>>> rwhat([np.array(1), np.array(2)])
[{'type': <class 'numpy.ndarray'>, 'min': 1, 'mean': 1.0, 'max': 1, 'shape': (), 'dtype': dtype('int64')}, {'type': <class 'numpy.ndarray'>, 'min': 2, 'mean': 2.0, 'max': 2, 'shape': (), 'dtype': dtype('int64')}]

When you import * from pxalyze you also get pprint imported as pp so you can use

>>> from pxalyze import *
>>> pp(rwhat([np.array(1), np.array(2)]))
[{'dtype': dtype('int64'),
  'max': 1,
  'mean': 1.0,
  'min': 1,
  'shape': (),
  'type': <class 'numpy.ndarray'>},
 {'dtype': dtype('int64'),
  'max': 2,
  'mean': 2.0,
  'min': 2,
  'shape': (),
  'type': <class 'numpy.ndarray'>}]

Visualization

test and atest are the function that take tensors or arrays and write a visualization on disk

test is a simple wrapper around cv2.imwrite, so it is very limited, but reliable

>>> test(np.random.randint(0, 255, (256, 256, 3)))
True

This will save the noisy image "test.png" to the current folder

If you are saving more than one images for comparison just do

a = np.random.randint(0, 255, (256, 256, 3))
b = np.random.randint(0, 255, (256, 256, 3))

test(a, "a")
test(b, "b")

This saves two files named "test_a.png" and "test_b.png" which is more convenient than type cv2.imwrite("test_a.png", a) each time

atest is more powerful - it can accept wide range of shapes and ranges and visualize

>>> a = np.random.random((256, 256, 3))
>>> b = np.random.random((3, 256, 256))
>>> c = np.random.random((1, 3, 256, 256))
>>> d = np.random.random((10, 3, 256, 256))
>>> e = np.random.random((10, 256, 256, 3)) * 10000

>>> atest(a, "a")
True
>>> atest(b, "b")
True
>>> atest(c, "c")
True
>>> atest(d, "d")
True
>>> atest(e, "e")
True

In case when there is a batch dimension in the tensor atest will try to tile the images into a single one and display them while writing index of each image above it

>>> atest(np.random.random((43, 3, 32, 32)))
True

will produce the following image

Utilities

Use to1 and to255 to minmax normalize arrays

>>> a = np.random.random((1, 2)) * 100 + 25
>>> what(to1(a))
{'type': <class 'numpy.ndarray'>, 'min': 0.0, 'mean': 0.5, 'max': 1.0, 'shape': (1, 2), 'dtype': dtype('float64')}

>>> what(to1(np.ones((1, 2))))
Failed to normalize: min_val == max_val == 1.0 falling back to zeros
{'type': <class 'numpy.ndarray'>, 'min': 0.0, 'mean': 0.0, 'max': 0.0, 'shape': (1, 2), 'dtype': dtype('float64')}

Use imgrid to manually tile the batch of images (this is what atest does)

>>> what(imgrid(np.random.random((3, 3, 25, 25))))
{'type': <class 'numpy.ndarray'>, 'min': 0.00037330399391755087, 'mean': 0.4978226227728666, 'max': 0.9999395359165014, 'shape': (3, 25, 75), 'dtype': dtype('float64')}

You can use lm to conveniently chain function in terminal without the need to match all the brackets

>>> a = np.random.random((10, 3, 16, 16))
>>> what(lm(a, tonp, to1, np.abs, imgrid))
{'type': <class 'numpy.ndarray'>, 'min': 0.0, 'mean': 0.4960013939669902, 'max': 1.0, 'shape': (3, 32, 80), 'dtype': dtype('float64')}

is the same as

what(imgrid(np.abs(to1(tonp(a)))))
{'type': <class 'numpy.ndarray'>, 'min': 0.0, 'mean': 0.4960013939669902, 'max': 1.0, 'shape': (3, 32, 80), 'dtype': dtype('float64')}

but much less brackets!

It actually saved me a lot of manual debugging time since I don't need to waste time on trying to find a single unmatched bracket in the REPL

You can use tonp to conventiently convert torch.Tensors to numpy and not type .detach().cpu().numpy() ever again.

>>> from pxalyze import *
>>> a = torch.randn((4, 3, 16, 16)).to("cuda:0")
>>> what(tonp(a))
{'type': <class 'numpy.ndarray'>, 'min': -3.2309637, 'mean': -0.010580406, 'max': 3.5923967, 'shape': (4, 3, 16, 16), 'dtype': dtype('float32')}

`atest` does is automatically

```python
>>> atest(a)
True

Contributing

Pull requests and issues are welcome! For major changes, please open an issue first to discuss what you would like to change.

License

Apache License 2.0

Versions

This project uses Semantic Versioning - https://semver.org/