refactor defaults and cli (#35)

* avoid repeated nuts code throughout codebase

* refactor and use new function

* sigma seems off by sqrt(2)

* rename

* being careful with defaults

* being careful with defaults and renaming

* renaming

* reorder

* avoid certain defaults for now

* we dont need these ones anymore

* need to remove prior and move to target

* fix test after refactoring

* use typer

* typo

* typo

* fix tests

* rename and use typer

* rename

* more judicious

* step size, no default

* need more arguments now

* rtol change to be appropriate int est

* fix

* high snr is the default

* add typer, but finish in next PR

* rename

* rename

* fix corersponding slurm script

bpd/chains.py

@@ -1,4 +1,11 @@
+from functools import partial
+from typing import Callable
+
+import blackjax
 import jax
+from jax import random
+from jax._src.prng import PRNGKeyArray
+from jax.typing import ArrayLike
 
 
 def inference_loop(rng_key, initial_state, kernel, n_samples: int):
@@ -12,3 +19,36 @@ def one_step(state, rng_key):
     _, (states, infos) = jax.lax.scan(one_step, initial_state, keys)
 
     return (states, infos)
+
+
+def run_inference_nuts(
+    rng_key: PRNGKeyArray,
+    init_positions: ArrayLike,
+    data: ArrayLike,
+    *,
+    logtarget: Callable,
+    n_samples: int,
+    initial_step_size: float,
+    max_num_doublings: int,
+    n_warmup_steps: int = 500,
+    target_acceptance_rate: float = 0.80,
+    is_mass_matrix_diagonal: bool = True,
+):
+    key1, key2 = random.split(rng_key)
+
+    _logtarget = partial(logtarget, data=data)
+
+    warmup = blackjax.window_adaptation(
+        blackjax.nuts,
+        _logtarget,
+        progress_bar=False,
+        is_mass_matrix_diagonal=is_mass_matrix_diagonal,
+        max_num_doublings=max_num_doublings,
+        initial_step_size=initial_step_size,
+        target_acceptance_rate=target_acceptance_rate,
+    )
+
+    (init_states, tuned_params), _ = warmup.run(key1, init_positions, n_warmup_steps)
+    kernel = blackjax.nuts(_logtarget, **tuned_params).step
+    states, _ = inference_loop(key2, init_states, kernel=kernel, n_samples=n_samples)
+    return states.position

bpd/draw.py

@@ -12,10 +12,11 @@ def draw_gaussian(
     g2: float,
     x: float,
     y: float,
-    pixel_scale: float = 0.2,
-    slen: int = 53,
+    *,
+    slen: int,
+    fft_size: int,  # rule of thumb: at least 4 times `slen`
     psf_hlr: float = 0.7,
-    fft_size: int = 256,  # rule of thumb, at least 4 times `slen`
+    pixel_scale: float = 0.2,
 ):
     gsparams = GSParams(minimum_fft_size=fft_size, maximum_fft_size=fft_size)
 
@@ -39,9 +40,10 @@ def draw_gaussian_galsim(
     g2: float,
     x: float,  # pixels
     y: float,
-    pixel_scale: float = 0.2,
-    slen: int = 53,
+    *,
+    slen: int,
     psf_hlr: float = 0.7,
+    pixel_scale: float = 0.2,
 ):
     gal = galsim.Gaussian(flux=f, half_light_radius=hlr)
     gal = gal.shear(g1=e1, g2=e2)

bpd/pipelines/image_ellips.py

@@ -1,14 +1,13 @@
 from functools import partial
 from typing import Callable
 
-import blackjax
 import jax.numpy as jnp
 from jax import Array, random
 from jax import jit as jjit
 from jax._src.prng import PRNGKeyArray
 from jax.scipy import stats
 
-from bpd.chains import inference_loop
+from bpd.chains import run_inference_nuts
 from bpd.draw import draw_gaussian, draw_gaussian_galsim
 from bpd.noise import add_noise
 from bpd.prior import ellip_mag_prior, sample_ellip_prior
@@ -17,7 +16,7 @@
 def get_target_galaxy_params_simple(
     rng_key: PRNGKeyArray,
     shape_noise: float = 1e-3,
-    lf: float = 3.0,
+    lf: float = 6.0,
     hlr: float = 1.0,
     x: float = 0.0,  # pixels
     y: float = 0.0,
@@ -42,29 +41,24 @@ def get_target_images_single(
     rng_key: PRNGKeyArray,
     n_samples: int,
     single_galaxy_params: dict[str, float],
-    psf_hlr: float = 0.7,
-    background: float = 1.0,
-    slen: int = 53,
-    pixel_scale: float = 0.2,
+    *,
+    background: float,
+    slen: int,
 ):
     """In this case, we sample multiple noise realizations of the same galaxy."""
     assert "f" in single_galaxy_params and "lf" not in single_galaxy_params
 
-    noiseless = draw_gaussian_galsim(
-        **single_galaxy_params,
-        pixel_scale=pixel_scale,
-        psf_hlr=psf_hlr,
-        slen=slen,
-    )
+    noiseless = draw_gaussian_galsim(**single_galaxy_params, slen=slen)
     return add_noise(rng_key, noiseless, bg=background, n=n_samples), noiseless
 
 
 # interim prior
 def logprior(
     params: dict[str, Array],
+    *,
+    sigma_e: float,
     flux_bds: tuple = (-1.0, 9.0),
     hlr_bds: tuple = (0.01, 5.0),
-    sigma_e: float = 3e-2,
     sigma_x: float = 1.0,  # pixels
 ) -> Array:
     prior = jnp.array(0.0)
@@ -107,71 +101,27 @@ def logtarget(
     return logprior_fnc(params) + loglikelihood_fnc(params, data)
 
 
-def do_inference(
-    rng_key: PRNGKeyArray,
-    init_positions: dict[str, Array],
-    data: Array,
-    *,
-    logtarget_fnc: Callable,
-    is_mass_matrix_diagonal: bool = False,
-    n_warmup_steps: int = 500,
-    max_num_doublings: int = 5,
-    initial_step_size: float = 1e-3,
-    target_acceptance_rate: float = 0.80,
-    n_samples: int = 100,
-):
-    key1, key2 = random.split(rng_key)
-
-    _logdensity = partial(logtarget_fnc, data=data)
-
-    warmup = blackjax.window_adaptation(
-        blackjax.nuts,
-        _logdensity,
-        progress_bar=False,
-        is_mass_matrix_diagonal=is_mass_matrix_diagonal,
-        max_num_doublings=max_num_doublings,
-        initial_step_size=initial_step_size,
-        target_acceptance_rate=target_acceptance_rate,
-    )
-
-    (init_states, tuned_params), _ = warmup.run(key1, init_positions, n_warmup_steps)
-
-    kernel = blackjax.nuts(_logdensity, **tuned_params).step
-    states, _ = inference_loop(key2, init_states, kernel=kernel, n_samples=n_samples)
-
-    return states.position
-
-
-def pipeline_image_interim_samples(
+def pipeline_image_interim_samples_one_galaxy(
     rng_key: PRNGKeyArray,
     true_params: dict[str, float],
     target_image: Array,
     *,
     initialization_fnc: Callable,
-    sigma_e_int: float = 3e-2,
+    sigma_e_int: float,
     n_samples: int = 100,
     max_num_doublings: int = 5,
     initial_step_size: float = 1e-3,
-    target_acceptance_rate: float = 0.80,
     n_warmup_steps: int = 500,
     is_mass_matrix_diagonal: bool = False,
     slen: int = 53,
-    pixel_scale: float = 0.2,
-    psf_hlr: float = 0.7,
-    background: float = 1.0,
     fft_size: int = 256,
+    background: float = 1.0,
 ):
     k1, k2 = random.split(rng_key)
 
     init_position = initialization_fnc(k1, true_params=true_params, data=target_image)
 
-    _draw_fnc = partial(
-        draw_gaussian,
-        pixel_scale=pixel_scale,
-        slen=slen,
-        psf_hlr=psf_hlr,
-        fft_size=fft_size,
-    )
+    _draw_fnc = partial(draw_gaussian, slen=slen, fft_size=fft_size)
     _loglikelihood = partial(loglikelihood, draw_fnc=_draw_fnc, background=background)
     _logprior = partial(logprior, sigma_e=sigma_e_int)
 
@@ -180,13 +130,12 @@ def pipeline_image_interim_samples(
     )
 
     _inference_fnc = partial(
-        do_inference,
-        logtarget_fnc=_logtarget,
+        run_inference_nuts,
+        logtarget=_logtarget,
         is_mass_matrix_diagonal=is_mass_matrix_diagonal,
         n_warmup_steps=n_warmup_steps,
         max_num_doublings=max_num_doublings,
         initial_step_size=initial_step_size,
-        target_acceptance_rate=target_acceptance_rate,
         n_samples=n_samples,
     )
     _run_inference = jjit(_inference_fnc)

bpd/pipelines/shear_inference.py

@@ -1,56 +1,34 @@
 from functools import partial
 from typing import Callable
 
-import blackjax
-from jax import Array, random
+from jax import Array
 from jax import jit as jjit
 from jax._src.prng import PRNGKeyArray
 from jax.scipy import stats
 
-from bpd.chains import inference_loop
+from bpd.chains import run_inference_nuts
 from bpd.likelihood import shear_loglikelihood
 from bpd.prior import ellip_mag_prior
 
 
-def logtarget_density(g: Array, e_post: Array, loglikelihood: Callable):
+def logtarget_density(g: Array, *, data: Array, loglikelihood: Callable):
+    e_post = data  # comptability with `do_inference_nuts`
     loglike = loglikelihood(g, e_post)
     logprior = stats.uniform.logpdf(g, -0.1, 0.2).sum()
     return logprior + loglike
 
 
-def do_inference(
-    rng_key: PRNGKeyArray,
-    init_g: Array,
-    logtarget: Callable,
-    n_samples: int,
-    n_warmup_steps: int = 500,
-):
-    key1, key2 = random.split(rng_key)
-
-    warmup = blackjax.window_adaptation(
-        blackjax.nuts,
-        logtarget,
-        progress_bar=False,
-        is_mass_matrix_diagonal=True,
-        max_num_doublings=2,
-        initial_step_size=1e-2,
-        target_acceptance_rate=0.80,
-    )
-
-    (init_states, tuned_params), _ = warmup.run(key1, init_g, n_warmup_steps)
-    kernel = blackjax.nuts(logtarget, **tuned_params).step
-    states, _ = inference_loop(key2, init_states, kernel=kernel, n_samples=n_samples)
-    return states.position
-
-
 def pipeline_shear_inference(
     rng_key: PRNGKeyArray,
     e_post: Array,
+    *,
     true_g: Array,
     sigma_e: float,
     sigma_e_int: float,
     n_samples: int,
+    initial_step_size: float,
     n_warmup_steps: int = 500,
+    max_num_doublings: int = 2,
 ):
     prior = partial(ellip_mag_prior, sigma=sigma_e)
     interim_prior = partial(ellip_mag_prior, sigma=sigma_e_int)
@@ -59,13 +37,16 @@ def pipeline_shear_inference(
     _loglikelihood = jjit(
         partial(shear_loglikelihood, prior=prior, interim_prior=interim_prior)
     )
-    _logtarget = partial(logtarget_density, loglikelihood=_loglikelihood, e_post=e_post)
+    _logtarget = partial(logtarget_density, loglikelihood=_loglikelihood)
 
     _do_inference = partial(
-        do_inference,
+        run_inference_nuts,
+        data=e_post,
         logtarget=_logtarget,
         n_samples=n_samples,
         n_warmup_steps=n_warmup_steps,
+        max_num_doublings=max_num_doublings,
+        initial_step_size=initial_step_size,
     )
 
     g_samples = _do_inference(rng_key, true_g)