💘 Lovely JAX

After all, you are only human.

Source code | ❤️ Lovely Tensors | 💟 Lovely `NumPy` | Discord

Note: I’m pretty new to JAX

If something does not make sense, shoot me an Issue or ping me on Discord and let me know how it’s supposed to work!

Better support for sharded arrays and solid jit/pmap/vmap support coming soon!

Install

pip install lovely-jax

How to use

How often do you find yourself debugging JAX code? You dump an array to the cell output, and see this:

numbers

Array([[[-0.35405433, -0.33692956, -0.4054286 , ..., -0.55955136,
         -0.4739276 ,  2.2489083 ],
        [-0.4054286 , -0.42255333, -0.49105233, ..., -0.91917115,
         -0.8506721 ,  2.1632845 ],
        [-0.4739276 , -0.4739276 , -0.5424266 , ..., -1.0390445 ,
         -1.0390445 ,  2.1975338 ],
        ...,
        [-0.9020464 , -0.8335474 , -0.9362959 , ..., -1.4671633 ,
         -1.2959158 ,  2.2317834 ],
        [-0.8506721 , -0.78217316, -0.9362959 , ..., -1.6041614 ,
         -1.5014129 ,  2.1804092 ],
        [-0.8335474 , -0.81642264, -0.9705454 , ..., -1.6555357 ,
         -1.5527872 ,  2.11191   ]],

       [[-0.19747896, -0.19747896, -0.30252096, ..., -0.47759098,
         -0.37254897,  2.4110641 ],
        [-0.24999997, -0.23249297, -0.33753496, ..., -0.705182  ,
         -0.670168  ,  2.3585434 ],
        [-0.30252096, -0.28501397, -0.39005598, ..., -0.740196  ,
         -0.810224  ,  2.3760502 ],
        ...,
        [-0.42507   , -0.23249297, -0.37254897, ..., -1.0903361 ,
         -1.0203081 ,  2.4285715 ],
        [-0.39005598, -0.23249297, -0.42507   , ..., -1.230392  ,
         -1.230392  ,  2.4110641 ],
        [-0.40756297, -0.28501397, -0.47759098, ..., -1.2829131 ,
         -1.2829131 ,  2.3410363 ]],

       [[-0.67154676, -0.9852723 , -0.88069713, ..., -0.9678431 ,
         -0.68897593,  2.3959913 ],
        [-0.7238344 , -1.0724182 , -0.9678431 , ..., -1.2467101 ,
         -1.0201306 ,  2.3262744 ],
        [-0.82840955, -1.1247058 , -1.0201306 , ..., -1.2641394 ,
         -1.1595641 ,  2.3785625 ],
        ...,
        [-1.229281  , -1.4732897 , -1.3861438 , ..., -1.5081482 ,
         -1.2641394 ,  2.5179958 ],
        [-1.1944225 , -1.4558606 , -1.4210021 , ..., -1.6475817 ,
         -1.4732897 ,  2.4308496 ],
        [-1.229281  , -1.5255773 , -1.5081482 , ..., -1.68244   ,
         -1.5255773 ,  2.3611329 ]]], dtype=float32)

Was it really useful for you, as a human, to see all these numbers?

What is the shape? The size?
What are the statistics?
Are any of the values nan or inf?
Is it an image of a man holding a tench?

import lovely_jax as lj

lj.monkey_patch()

Summary

numbers

Array[196, 196, 3] n=115248 (0.4Mb) x∈[-2.118, 2.640] μ=-0.388 σ=1.073 cpu:0

Better, huh?

numbers[1,:6,1] # Still shows values if there are not too many.

Array[6] x∈[-0.408, -0.232] μ=-0.340 σ=0.075 cpu:0 [-0.250, -0.232, -0.338, -0.408, -0.408, -0.408]

spicy = numbers.flatten()[:12].copy()

spicy = (spicy  .at[0].mul(10000)
                .at[1].divide(10000)
                .at[2].set(float('inf'))
                .at[3].set(float('-inf'))
                .at[4].set(float('nan'))
                .reshape((2,6)))
spicy # Spicy stuff

Array[2, 6] n=12 x∈[-3.541e+03, -1.975e-05] μ=-393.848 σ=1.113e+03 +Inf! -Inf! NaN! cpu:0

jnp.zeros((10, 10)) # A zero array - make it obvious

Array[10, 10] n=100 all_zeros cpu:0

spicy.v # Verbose

Array[2, 6] n=12 x∈[-3.541e+03, -1.975e-05] μ=-393.848 σ=1.113e+03 +Inf! -Inf! NaN! cpu:0
Array([[-3.5405432e+03, -1.9747897e-05,            inf,           -inf,
                   nan, -9.8527229e-01],
       [-4.0542859e-01, -3.0252096e-01, -8.8069713e-01, -4.3967807e-01,
        -3.0252096e-01, -7.7612197e-01]], dtype=float32)

spicy.p # The plain old way

Array([[-3.5405432e+03, -1.9747897e-05,            inf,           -inf,
                   nan, -9.8527229e-01],
       [-4.0542859e-01, -3.0252096e-01, -8.8069713e-01, -4.3967807e-01,
        -3.0252096e-01, -7.7612197e-01]], dtype=float32)

Going `.deeper`

numbers.deeper

Array[196, 196, 3] n=115248 (0.4Mb) x∈[-2.118, 2.640] μ=-0.388 σ=1.073 cpu:0
  Array[196, 3] n=588 x∈[-1.912, 2.411] μ=-0.728 σ=0.519 cpu:0
  Array[196, 3] n=588 x∈[-1.861, 2.359] μ=-0.778 σ=0.450 cpu:0
  Array[196, 3] n=588 x∈[-1.758, 2.379] μ=-0.838 σ=0.437 cpu:0
  Array[196, 3] n=588 x∈[-1.656, 2.466] μ=-0.878 σ=0.415 cpu:0
  Array[196, 3] n=588 x∈[-1.717, 2.448] μ=-0.882 σ=0.399 cpu:0
  Array[196, 3] n=588 x∈[-1.717, 2.431] μ=-0.905 σ=0.408 cpu:0
  Array[196, 3] n=588 x∈[-1.563, 2.448] μ=-0.859 σ=0.416 cpu:0
  Array[196, 3] n=588 x∈[-1.475, 2.431] μ=-0.791 σ=0.463 cpu:0
  Array[196, 3] n=588 x∈[-1.526, 2.429] μ=-0.759 σ=0.499 cpu:0
  ...

# You can go deeper if you need to
numbers[:3,:5,:3].deeper(2)

Array[3, 5, 3] n=45 x∈[-1.316, -0.197] μ=-0.593 σ=0.302 cpu:0
  Array[5, 3] n=15 x∈[-0.985, -0.197] μ=-0.491 σ=0.267 cpu:0
    Array[3] x∈[-0.672, -0.197] μ=-0.408 σ=0.197 cpu:0 [-0.354, -0.197, -0.672]
    Array[3] x∈[-0.985, -0.197] μ=-0.507 σ=0.343 cpu:0 [-0.337, -0.197, -0.985]
    Array[3] x∈[-0.881, -0.303] μ=-0.530 σ=0.252 cpu:0 [-0.405, -0.303, -0.881]
    Array[3] x∈[-0.776, -0.303] μ=-0.506 σ=0.199 cpu:0 [-0.440, -0.303, -0.776]
    Array[3] x∈[-0.916, -0.215] μ=-0.506 σ=0.298 cpu:0 [-0.388, -0.215, -0.916]
  Array[5, 3] n=15 x∈[-1.212, -0.232] μ=-0.609 σ=0.302 cpu:0
    Array[3] x∈[-0.724, -0.250] μ=-0.460 σ=0.197 cpu:0 [-0.405, -0.250, -0.724]
    Array[3] x∈[-1.072, -0.232] μ=-0.576 σ=0.360 cpu:0 [-0.423, -0.232, -1.072]
    Array[3] x∈[-0.968, -0.338] μ=-0.599 σ=0.268 cpu:0 [-0.491, -0.338, -0.968]
    Array[3] x∈[-0.968, -0.408] μ=-0.651 σ=0.235 cpu:0 [-0.577, -0.408, -0.968]
    Array[3] x∈[-1.212, -0.408] μ=-0.761 σ=0.336 cpu:0 [-0.662, -0.408, -1.212]
  Array[5, 3] n=15 x∈[-1.316, -0.285] μ=-0.677 σ=0.306 cpu:0
    Array[3] x∈[-0.828, -0.303] μ=-0.535 σ=0.219 cpu:0 [-0.474, -0.303, -0.828]
    Array[3] x∈[-1.125, -0.285] μ=-0.628 σ=0.360 cpu:0 [-0.474, -0.285, -1.125]
    Array[3] x∈[-1.020, -0.390] μ=-0.651 σ=0.268 cpu:0 [-0.542, -0.390, -1.020]
    Array[3] x∈[-1.003, -0.478] μ=-0.708 σ=0.219 cpu:0 [-0.645, -0.478, -1.003]
    Array[3] x∈[-1.316, -0.513] μ=-0.865 σ=0.336 cpu:0 [-0.765, -0.513, -1.316]

Now in `.rgb` color

The important queston - is it our man?

numbers.rgb

Maaaaybe? Looks like someone normalized him.

in_stats = ( (0.485, 0.456, 0.406),     # mean
             (0.229, 0.224, 0.225) )    # std

# numbers.rgb(in_stats, cl=True) # For channel-last input format
numbers.rgb(in_stats)

It’s indeed our hero, the Tenchman!

`.plt` the statistics

(numbers+3).plt

(numbers+3).plt(center="mean", max_s=1000)

(numbers+3).plt(center="range")

See the `.chans`

# .chans will map values betwen [-1,1] to colors.
# Make our values fit into that range to avoid clipping.
mean = jnp.array(in_stats[0])
std = jnp.array(in_stats[1])
numbers_01 = (numbers*std + mean)
numbers_01

Array[196, 196, 3] n=115248 (0.4Mb) x∈[0., 1.000] μ=0.361 σ=0.248 cpu:0

numbers_01.chans

Grouping

# Make 8 images with progressively higher brightness and stack them 2x2x2.
eight_images = (jnp.stack([numbers]*8) + jnp.linspace(-2, 2, 8)[:,None,None,None])
eight_images = (eight_images
                     *jnp.array(in_stats[1])
                     +jnp.array(in_stats[0])
                ).clip(0,1).reshape(2,2,2,196,196,3)

eight_images

Array[2, 2, 2, 196, 196, 3] n=921984 (3.5Mb) x∈[0., 1.000] μ=0.382 σ=0.319 cpu:0

eight_images.rgb

Sharding

assert jax.__version_info__[0] == 0
from jax.sharding import NamedSharding, Mesh, PartitionSpec as P
from jax.experimental import mesh_utils

# Create a mesh with named axes
devices = mesh_utils.create_device_mesh((4, 2))
mesh = Mesh(devices, axis_names=('y', 'x'))

# Create sharding with PartitionSpec
sharding = NamedSharding(mesh, P('y', 'x'))

x = jax.random.normal(jax.random.PRNGKey(0), (8192, 8192))
y = jax.device_put(x, sharding)

jax.debug.visualize_array_sharding(y)

print(y)

                        
   CPU 0       CPU 1    
                        
                        
   CPU 2       CPU 3    
                        
                        
   CPU 4       CPU 5    
                        
                        
   CPU 6       CPU 7

Array[8192, 8192] n=67108864 (0.2Gb) x∈[-5.420, 5.220] μ=1.508e-05 σ=1.000 S[y,x] 4×2 cpu:0-7

Options | Docs

from lovely_jax import set_config, config

set_config(precision=5, sci_mode=True, color=False)
jnp.array([1., 2, jnp.nan])

Array[3] μ=1.50000e+00 σ=5.00000e-01 NaN! cpu:0 [1.00000e+00, 2.00000e+00, nan]

set_config(precision=None, sci_mode=None, color=None) # None -> Reset to defaults

print(jnp.array([1., 2]))
# Or with config context manager.
with config(sci_mode=True, precision=5):
    print(jnp.array([1., 2]))

print(jnp.array([1., 2]))

Array[2] μ=1.500 σ=0.500 cpu:0 [1.000, 2.000]
Array[2] μ=1.50000e+00 σ=5.00000e-01 cpu:0 [1.00000e+00, 2.00000e+00]
Array[2] μ=1.500 σ=0.500 cpu:0 [1.000, 2.000]

Without `.monkey_patch`

lj.lovely(spicy)

Array[2, 6] n=12 x∈[-3.541e+03, -1.975e-05] μ=-393.848 σ=1.113e+03 +Inf! -Inf! NaN! cpu:0

lj.lovely(spicy, verbose=True)

Array[2, 6] n=12 x∈[-3.541e+03, -1.975e-05] μ=-393.848 σ=1.113e+03 +Inf! -Inf! NaN! cpu:0
Array([[-3.5405432e+03, -1.9747897e-05,            inf,           -inf,
                   nan, -9.8527229e-01],
       [-4.0542859e-01, -3.0252096e-01, -8.8069713e-01, -4.3967807e-01,
        -3.0252096e-01, -7.7612197e-01]], dtype=float32)

lj.lovely(numbers, depth=1)

Array[196, 196, 3] n=115248 (0.4Mb) x∈[-2.118, 2.640] μ=-0.388 σ=1.073 cpu:0
  Array[196, 3] n=588 x∈[-1.912, 2.411] μ=-0.728 σ=0.519 cpu:0
  Array[196, 3] n=588 x∈[-1.861, 2.359] μ=-0.778 σ=0.450 cpu:0
  Array[196, 3] n=588 x∈[-1.758, 2.379] μ=-0.838 σ=0.437 cpu:0
  Array[196, 3] n=588 x∈[-1.656, 2.466] μ=-0.878 σ=0.415 cpu:0
  Array[196, 3] n=588 x∈[-1.717, 2.448] μ=-0.882 σ=0.399 cpu:0
  Array[196, 3] n=588 x∈[-1.717, 2.431] μ=-0.905 σ=0.408 cpu:0
  Array[196, 3] n=588 x∈[-1.563, 2.448] μ=-0.859 σ=0.416 cpu:0
  Array[196, 3] n=588 x∈[-1.475, 2.431] μ=-0.791 σ=0.463 cpu:0
  Array[196, 3] n=588 x∈[-1.526, 2.429] μ=-0.759 σ=0.499 cpu:0
  ...

lj.rgb(numbers, in_stats)

lj.plot(numbers, center="mean")

lj.chans(numbers_01)

Matplotlib integration | Docs

numbers.rgb(in_stats).fig # matplotlib figure

(numbers*0.3+0.5).chans.fig # matplotlib figure

numbers.plt.fig.savefig('pretty.svg') # Save it

!file pretty.svg; rm pretty.svg

pretty.svg: SVG Scalable Vector Graphics image

Add content to existing Axes

fig = plt.figure(figsize=(8,3))
fig.set_constrained_layout(True)
gs = fig.add_gridspec(2,2)
ax1 = fig.add_subplot(gs[0, :])
ax2 = fig.add_subplot(gs[1, 0])
ax3 = fig.add_subplot(gs[1,1:])

ax2.set_axis_off()
ax3.set_axis_off()

numbers_01.plt(ax=ax1)
numbers_01.rgb(ax=ax2)
numbers_01.chans(ax=ax3);

Import hook

Lovely JAX installs an import hook. Set LOVELY_JAX=1, and it will load automatically, no need to modify the code:

Note: You can now set it globally. The installed import hook will trigger only when JAX is imported.

import jax

x = jax.random.normal(jax.random.PRNGKey(0), (4, 16))
print(x)

LOVELY_JAX=1 python test.py

x: Array[4, 16] n=64 x∈[-1.955, 2.180] μ=0.031 σ=0.960 cpu:0

This is especially useful in combination with Better Exceptions:

import jax
import jax.numpy as jnp

x = jax.random.normal(jax.random.PRNGKey(0), (4, 16))
print(f"x: {x}")

w = jax.random.normal(jax.random.PRNGKey(1), (15, 8))
y = jnp.matmul(x, w)  # Dimension mismatch

BETTER_EXCEPTIONS=1 LOVELY_JAX=1 python test.py

x: Array[4, 16] n=64 x∈[-1.955, 2.180] μ=0.031 σ=0.960 cpu:0
Traceback (most recent call last):
  File "/home/xl0/work/projects/lovely-jax/test.py", line 9, in <module>
    y = jnp.matmul(x, w)  # Dimension mismatch
        │          │  └ Array[15, 8] n=120 x∈[-2.746, 2.608] μ=-0.003 σ=1.072 cpu:0
        │          └ Array[4, 16] n=64 x∈[-1.955, 2.180] μ=0.031 σ=0.960 cpu:0
        └ <module 'jax.numpy' from '...'>
  File "...jax/_src/numpy/tensor_contractions.py", line 254, in matmul
    out = lax.dot_general(
TypeError: dot_general requires contracting dimensions to have the same shape, got (16,) and (15,).

Source code | ❤️ Lovely Tensors | 💟 Lovely NumPy | Discord