Explore changing interim prior and impact on calibration (#22)

* add option to run pipelien with sigma_e_int

* results with varying sigma

* more ignore

* be explicit with argument names

* more results

.gitignore

@@ -164,5 +164,6 @@ cython_debug/
 experiments/samples/*.hdf5
 **/slurm-*.out
 **/bash/*.sh
+**/bash/**/*.sh
 **/logs/**
 **log**.txt

bpd/pipelines/toy_ellips.py

@@ -35,7 +35,7 @@ def do_inference(
     init_positions: Array,
     e_obs: Array,
     sigma_m: float,
-    sigma_e: float,
+    initial_step_size: float,
     interim_prior: Callable,
     k: int,
 ):
@@ -51,7 +51,7 @@ def do_inference(
         progress_bar=False,
         is_mass_matrix_diagonal=True,
         max_num_doublings=2,
-        initial_step_size=sigma_e,
+        initial_step_size=initial_step_size,
         target_acceptance_rate=0.80,
     )
 
@@ -86,8 +86,8 @@ def pipeline_toy_ellips_samples(
     _do_inference_jitted = jjit(
         partial(
             do_inference,
-            sigma_e=sigma_e,
             sigma_m=sigma_m,
+            initial_step_size=sigma_e,
             interim_prior=interim_prior,
             k=k,
         )

scripts/vect_toy_shear_gpu.py

@@ -24,6 +24,7 @@
 @click.option("--n-samples-shear", type=int, default=3000, help="shear samples")
 @click.option("--k", type=int, default=100, help="# int. posterior samples per galaxy.")
 @click.option("--shape-noise", type=float, default=1e-3)
+@click.option("--sigma-e-int", type=float, default=2e-3)
 @click.option("--obs-noise", type=float, default=1e-4)
 @click.option("--trim", type=int, default=1)
 def main(
@@ -37,6 +38,7 @@ def main(
     n_samples_shear: int,
     k: int,
     shape_noise: float,
+    sigma_e_int: float,
     obs_noise: float,
     trim: int,
 ):
@@ -57,7 +59,7 @@ def main(
         g1=g1,
         g2=g2,
         sigma_e=shape_noise,
-        sigma_e_int=shape_noise * 2,
+        sigma_e_int=sigma_e_int,
         sigma_m=obs_noise,
         n_samples=n_samples_gals,
         k=k,
@@ -66,7 +68,7 @@ def main(
         pipeline_shear_inference,
         true_g=jnp.array([g1, g2]),
         sigma_e=shape_noise,
-        sigma_e_int=shape_noise * 2,
+        sigma_e_int=sigma_e_int,
         n_samples=n_samples_shear,
     )
     vpipe1 = vmap(pipe1, in_axes=(0,))