Add diagnosis plots for model parameter tuning

borg_velocity/samplers.py

@@ -5,6 +5,8 @@ import jax.numpy as jnp
 import jax
 import blackjax
 from borg_velocity.utils import myprint
+import matplotlib.pyplot as plt
+import os
 
 
 class HMCBiasSampler(borg.samplers.PyBaseSampler):
@@ -352,7 +354,14 @@ class BlackJaxBiasSampler(borg.samplers.PyBaseSampler):
             )
             rng_key, warmup_key, sample_key = jax.random.split(self.rng_key, 3)
             (bj_state, self.parameters), info = warmup.run(warmup_key, self.y, num_steps=self.warmup_nsteps)
-            # x = info.state.position
+            x = info.state.position
+            
+            # Make and save summary figures
+            dirname = os.getcwd()
+            trace_plot(self.attributes, info.state.position, f'{dirname}/tuning_params.png', ncol=4)
+            trace_plot(self.attributes, info.adaptation_state.inverse_mass_matrix, f'{dirname}/tuning_inverse_mass_matrix.png', ncol=4)
+            trace_plot(['stepsize'], info.adaptation_state.step_size, f'{dirname}/tuning_step_size.png', ncol=4)
+
             self.y[:] = bj_state.position
             
             # Save results to the borg state object
@@ -562,3 +571,34 @@ class TransformedBlackJaxBiasSampler(borg.samplers.PyBaseSampler):
 
 
 
+def trace_plot(names, samples, savename, ncol=4):
+    
+    if len(names) == 1:
+        
+        fig, ax = plt.subplots(1, 1, figsize=(4,3))
+        ax.plot(samples[:])
+        ax.set_title(names[0])
+        ax.set_xlabel('Step Number')
+        
+    else:
+   
+        nrow = int(np.ceil(len(names) / ncol))
+
+        fig, axs = plt.subplots(nrow, ncol, figsize=(ncol*4, nrow*3))
+        axs = np.atleast_2d(axs)
+        axs = axs.flatten()
+
+        for i in range(len(names)):
+            axs[i].plot(samples[:,i])
+            axs[i].set_title(names[i])
+            axs[i].set_xlabel('Step Number')
+        for i in range(len(names), len(axs)):
+            axs[i].remove()
+            
+    fig.tight_layout()
+    myprint(f'Saving tuning diagnosis figure: {savename}')
+    fig.savefig(savename, facecolor='white', bbox_inches='tight')
+    fig.clf()
+    plt.close(fig)
+    
+    return

conf/supranta_ini.ini

@@ -47,7 +47,7 @@ epsilon = 0.005
 Nsteps = 20
 refresh = 0.1
 rng_seed = 1
-warmup_nsteps = 100
+warmup_nsteps = 200
 warmup_target_acceptance_rate = 0.7
 mvlam_mua = 0.01
 mvlam_alpha = 0.5

scripts/run_borg.sh

@@ -33,5 +33,5 @@ set -x
 # Just ICs
 INI_FILE=/home/bartlett/fsigma8/borg_velocity/conf/supranta_ini.ini
 cp $INI_FILE ini_file.ini
-# $BORG INIT ini_file.ini
+$BORG INIT ini_file.ini
-$BORG RESUME ini_file.ini
+# $BORG RESUME ini_file.ini