I/O with Scenes

Φ_Flow comes with a couple of I/O functions to store fields, such as CenteredGrid or StaggeredGrid. These are stored according to the scene format specification which is designed to store sequences.

from phi.flow import *

Writing Data

A scene stores one sequence of data. Batched data can also be stored in a single scene (see below). To create a new scene, we pass the directory and name. An index suffix is automatically appended to the name.

scene = Scene.create('./data', name='sim')
scene

/opt/hostedtoolcache/Python/3.12.7/x64/lib/python3.12/site-packages/phi/field/_scene.py:150: RuntimeWarning: Failed to copy calling script to scene during Scene.create(): [Errno 2] No such file or directory: '/tmp/ipykernel_2835/2238997793.py'
  warnings.warn(f"Failed to copy calling script to scene during Scene.create(): {err}", RuntimeWarning)

'./data/sim_000000'

We can list all scenes in the directory.

Scene.list('./data')

('./data/sim_000000',)

Let's write some data to the scene. Here, we sample random values on a grid according to a smoothness parameter, which we also want to store.

smoothness = math.random_uniform()
data = CenteredGrid(Noise(smoothness=smoothness + 0.5), x=32, y=16)
plot(data)

/opt/hostedtoolcache/Python/3.12.7/x64/lib/python3.12/site-packages/phi/vis/_matplotlib/_matplotlib_plots.py:167: UserWarning: This figure includes Axes that are not compatible with tight_layout, so results might be incorrect.
  plt.tight_layout()  # because subplot titles can be added after figure creation

Fields can be stored using the write method, which creates a new file with the given name and frame index, in this case random_noise_data_000000.npz.

Additional non-field values, like the smoothness, can be added as properties. Any JSON-serializable object can be stored this way.

scene.write(noise_data=data, frame=0)
scene.put_properties(smoothness=smoothness)

Reading Data

Next, let's read the data back from disk. We can reference our scene using Scene.list or Scene.at.

scene = Scene.at('./data', 0)

The properties of the scene are stored as a dict in scene.properties.

scene.properties

{'smoothness': 0.28000545501708984}

Fields can be read using Scene.read given their name and frame index.

plot(scene.read('noise_data', frame=0))

Batched Data

Scenes can also deal with batched data. When batched data is written to a single scene, only one file is created for each stored field.

batched_smoothness = math.random_uniform(batch(batch=4))
batched_data = CenteredGrid(Noise(smoothness=batched_smoothness + 0.5), x=32, y=16)
plot(batched_data, show_color_bar=False)

scene.write(batched_noise_data=batched_data, frame=0)
scene.put_properties(smoothness=batched_smoothness)

scene = Scene.at('./data', 0)
plot(scene.read('batched_noise_data'), show_color_bar=False)

scene.properties['smoothness']

(0.302, 0.464, 0.412, 0.258) along batchᵇ float64

Scenes themselves can also be batched. In that case, a single Scene object represents a batch of actual directories. When writing data to the batched scene, the data are sliced along the dimensions of the scene.

scene_batch = Scene.create('./data', batch(batch=4), name='sim')
scene_batch

/opt/hostedtoolcache/Python/3.12.7/x64/lib/python3.12/site-packages/phi/field/_scene.py:150: RuntimeWarning: Failed to copy calling script to scene during Scene.create(): [Errno 2] No such file or directory: '/tmp/ipykernel_2835/3852168639.py'
  warnings.warn(f"Failed to copy calling script to scene during Scene.create(): {err}", RuntimeWarning)

['./data/sim_000001', './data/sim_000002', './data/sim_000003', './data/sim_000004']

scene_batch.write(batched_noise_data=batched_data, frame=0)
scene_batch.put_properties(smoothness=batched_smoothness)

Now, let's look at the contents of a single scene.

scene = Scene.at('./data', 1)
scene.properties['smoothness']

0.3024602234363556

plot(scene.read('batched_noise_data'))

If the scene has batch dimensions that are not present on the data, the data are tiled automatically, i.e. stored redundantly across the scenes.