Check how muse behaves when we increase the number of latent variables

[Justinezgh]

For these 3 experiments I use the following MUSE algorithm configuration:

theta_start = jnp.array([0.3, 0.8])

prob.solve(
    result=result,
    α=0.2,
    θ_start=theta_start,
    θ_rtol=0,
    z_tol=1e-2,
    progress=True,
    maxsteps=100,
    nsims=100,
    rng=jax.random.PRNGKey(1)
)

simulator configuration:

nb_latent_variables = 100

model = partial(
    lensingLogNormal, 
    N=nb_latent_variables,
    map_size=5,
    gal_per_arcmin2=30,
    sigma_e=0.2,
    model_type='lognormal',
   non_gaussianity=0.03
)

Convergence:

Loss of the last 90 iterations:

---

simulator configuration:

nb_latent_variables = 200

model = partial(
    lensingLogNormal, 
    N=nb_latent_variables,
    map_size=5,
    gal_per_arcmin2=30,
    sigma_e=0.2,
    model_type='lognormal',
   non_gaussianity=0.03
)

Convergence:

Loss of the last 90 iterations:

---

simulator configuration:

nb_latent_variables = 250

model = partial(
    lensingLogNormal, 
    N=nb_latent_variables,
    map_size=5,
    gal_per_arcmin2=30,
    sigma_e=0.2,
    model_type='lognormal',
   non_gaussianity=0.03
)

Convergence:

Loss of the last 90 iterations:

[Justinezgh]

To do

• [x] change the starting point theta (was set to the fiducial params) to check if it can still converge

Update: I changed it to [0.33, 0.83] the results are below

• [ ] reduce the step size to make the score closer to zero

[Justinezgh]

For these 3 experiments I use the following MUSE algorithm configuration:

theta_start = jnp.array([0.33, 0.83])

prob.solve(
    result=result,
    α=0.2,
    θ_start=theta_start,
    θ_rtol=0,
    z_tol=1e-2,
    progress=True,
    maxsteps=100,
    nsims=100,
    rng=jax.random.PRNGKey(1)
)

simulator configuration:

nb_latent_variables = 100

model = partial(
    lensingLogNormal, 
    N=nb_latent_variables,
    map_size=5,
    gal_per_arcmin2=30,
    sigma_e=0.2,
    model_type='lognormal',
   non_gaussianity=0.03
)

Convergence: Loss of the last 90 iterations:

---

simulator configuration:

nb_latent_variables = 200

model = partial(
    lensingLogNormal, 
    N=nb_latent_variables,
    map_size=5,
    gal_per_arcmin2=30,
    sigma_e=0.2,
    model_type='lognormal',
   non_gaussianity=0.03
)

Convergence: Loss of the last 90 iterations:

---

simulator configuration:

nb_latent_variables = 250

model = partial(
    lensingLogNormal, 
    N=nb_latent_variables,
    map_size=5,
    gal_per_arcmin2=30,
    sigma_e=0.2,
    model_type='lognormal',
   non_gaussianity=0.03
)

Convergence:

Loss of the last 90 iterations:

---

simulator configuration:

nb_latent_variables = 500

model = partial(
    lensingLogNormal, 
    N=nb_latent_variables,
    map_size=5,
    gal_per_arcmin2=30,
    sigma_e=0.2,
    model_type='lognormal',
   non_gaussianity=0.03
)

(I don't have the full field contour for this one because 500 * 500 was too much for my GPUs memory)

Convergence:

Loss:

[Justinezgh]

To do

• [x] change the starting point theta (was set to the fiducial params) to check if it can still converge

Update: I changed it to [0.33, 0.83] the results are below

• [ ] reduce the step size to make the score closer to zero

• [ ] understand why when the starting point for the gradient descent is far away the covariance matrix is better approximated.