Smaller network seems to improve quality quite a bit

example/InjectionRecovery.py

@@ -20,44 +20,48 @@ class InjectionRecoveryParser(Tap):
 
     # Noise parameters
     seed: int = 0
-    f_sampling: int  = 2048
+    f_sampling: int  = 4096
     duration: int = 4
     fmin: float = 20.0
     ifos: list[str]  = ["H1", "L1", "V1"]
 
     # Injection parameters
     m1: float = 30.0
     m2: float = 25.0
-    s1_theta: float = 0.
-    s1_phi: float = 0.
-    s1_mag: float = 0.01
-    s2_theta: float = 0.
-    s2_phi: float = 0.
-    s2_mag: float = 0.02
+    s1_theta: float = 0.04
+    s1_phi: float = 0.02
+    s1_mag: float = 0.1
+    s2_theta: float = 0.01
+    s2_phi: float = 0.03
+    s2_mag: float = 0.05
     dist_mpc: float = 400.
     tc: float = 0.
-    phic: float = 0.
-    inclination: float = 0.3
+    phic: float = 0.1
+    inclination: float = 0.5
     polarization_angle: float = 0.7
-    ra: float = 1.1
+    ra: float = 1.2
     dec: float = 0.3
 
     # Sampler parameters
     n_dim: int = 15
     n_chains: int = 500
-    n_loop_training: int = 10
+    n_loop_training: int = 200
     n_loop_production: int = 10
     n_local_steps: int = 300
     n_global_steps: int = 300
     learning_rate: float = 0.001
-    max_samples: int = 50000
+    max_samples: int = 60000
     momentum: float = 0.9
-    num_epochs: int = 500
-    batch_size: int = 50000
+    num_epochs: int = 300
+    batch_size: int = 60000
     stepsize: float = 0.01
     use_global: bool = True
-    keep_quantile: float = 0.1
-    train_thinning: int = 10
+    keep_quantile: float = 0.0
+    train_thinning: int = 1
+    output_thinning: int = 30
+    num_layers: int = 6
+    hidden_size: list[int] = [64,64]
+    num_bins: int = 8
 
     # Output parameters
     output_path: str = "./"
@@ -75,7 +79,7 @@ class InjectionRecoveryParser(Tap):
 
 print("Injection signals")
 
-freqs = jnp.linspace(args.fmin, args.f_sampling/2, args.duration*args.f_sampling//2)
+freqs = jnp.linspace(args.fmin, args.f_sampling/2, args.duration*args.f_sampling)
 
 Mc, eta = ms_to_Mc_eta(jnp.array([args.m1, args.m2]))
 f_ref = args.fmin
@@ -133,8 +137,12 @@ class InjectionRecoveryParser(Tap):
         use_global=args.use_global,
         keep_quantile= args.keep_quantile,
         train_thinning = args.train_thinning,
+        output_thinning = args.output_thinning,
         local_sampler_arg = local_sampler_arg,
         seed = args.seed,
+        num_layers = args.num_layers,
+        hidden_size = args.hidden_size,
+        num_bins = args.num_bins
         )
 
 key, subkey = jax.random.split(key)