Merge pull request #118 from graphcore-research/visualise_num_error

add num_error visualisation on demand

pyscf_ipu/nanoDFT/nanoDFT.py

@@ -81,8 +81,27 @@ def nanoDFT_iteration(i, vals, opts, mol):
     log["matrices"] = jax.lax.dynamic_update_slice(log["matrices"], H.             reshape(1, 1, N, N), (i, 2, 0, 0))
     log["energy"]   = log["energy"].at[i].set(energy(density_matrix, H_core, diff_JK, E_xc, E_nuc))  # (iterations, 6)
 
+    if opts.vis_num_error is True:
+        import os
+        dir_label = opts.molecule_name
+        num_error_dir = f'num_error/{dir_label}/'
+        os.makedirs(num_error_dir , exist_ok=True)
+
+        def host_callback(data, i):
+            # labels are adjusted to the `data` that will be passed to the callback - keep that in mind when passing different list of tensors
+            labels = ["density_matrix", "V_xc", "diff_JK", "O", "H_core", "L_inv", "E_nuc", "occupancy", "ERI", "grid_weights", "grid_AO", "diis_history", "E_xc", "eigvects", "H"]
+            for l, d  in zip(labels, data):
+                if l == "diis_history" or l == "ERI":
+                    for idx, arr in enumerate(d):
+                        np.savez(f'{num_error_dir}{i}_{l}{idx}.npz', v = np.array(arr))
+                else:
+                    np.savez(f'{num_error_dir}{i}_{l}.npz', v = d)
+
+        jax.debug.callback(host_callback, vals[:-1] + [E_xc, eigvects, H], i)
+
     return [density_matrix, V_xc, diff_JK, O, H_core, L_inv, E_nuc, occupancy, ERI, grid_weights, grid_AO, diis_history, log]
 
+
 def exchange_correlation(density_matrix, grid_AO, grid_weights):
     """Compute exchange correlation integral using atomic orbitals (AO) evalauted on a grid. """
     # Perfectly SIMD parallelizable over grid_size axis.
@@ -546,7 +565,9 @@ def nanoDFT_options(
         ndevices: int = 1, 
         dense_ERI: bool = False,        
         v: bool = False, # verbose 
-        profile: bool = False # if we only want profile exit after IPU finishes. 
+        profile: bool = False, # if we only want profile exit after IPU finishes.
+        vis_num_error: bool = False,
+        molecule_name: str = None
 ):
     """
     nanoDFT
@@ -570,6 +591,10 @@ def nanoDFT_options(
         ao_threshold (float): Zero out grid_AO that are below the threshold in absolute value.
         dense_ERI (bool): Whether to use dense ERI (s1) or sparse symmtric ERI. 
     """
+    if molecule_name is None:
+        # use mol_str as a molecule name (in case it has not been provided)
+        # before mol_str CLI arg is preprocessed and overwritten
+        molecule_name = mol_str
 
     # From a compound name or CID, get a list of its atoms and their coordinates
     mol_str = utils.process_mol_str(mol_str)
@@ -611,6 +636,12 @@ def main():
         nanoDFT_forces = grad(mol, grid_coords, grid_weights, mo_coeff, mo_energy)
         pyscf_E, pyscf_hlgap, pyscf_forces = pyscf_reference(mol_str, opts)
         print_difference(nanoDFT_E, nanoDFT_forces, nanoDFT_logged_E, nanoDFT_hlgap, pyscf_E, pyscf_forces, pyscf_hlgap)
+
+        if opts.vis_num_error is True:
+            from utils import save_plot
+            import sys
+            _plot_title = f"Created with:  python {' '.join(sys.argv)}"
+            save_plot("num_error/", opts.molecule_name, opts.its, _plot_title)
     else:
         # pip install mogli imageio[ffmpeg] matplotlib
         import mogli

pyscf_ipu/nanoDFT/utils.py

@@ -147,4 +147,80 @@ def min_interatomic_distance(mol_str):
     """This computes the minimum distance between atoms."""
     coords = map(itemgetter(1), mol_str) 
     distances = map(lambda x: np.linalg.norm(np.array(x[0]) - np.array(x[1])), combinations(coords, 2))
-    return min(distances)
+    return min(distances)
+
+
+def save_plot(base_data_dir: str, molecule_name: str, iterations: int, _plot_title: str = "Default Title"):
+    import matplotlib.pyplot as plt
+    import matplotlib
+    import os
+    matplotlib.rc('font', **{'family': 'serif', 'serif': ['Computer Modern']})
+
+    import seaborn as sns
+    sns.set_theme()
+    sns.set_style("white")
+
+    data_dir = base_data_dir + molecule_name + '/'
+
+    def prepare(val):
+        val = np.abs(val[val == val])
+        val[np.logical_and(val<1e-15, val!=0)] = 2e-15 # show the ones that go out of plot
+        val[val==0] = 1e-17 # remove zeros.
+        return val
+
+    xticks = []
+    xticklabels = []
+
+    fig, ax = plt.subplots(1, 1, figsize=(14,8))
+    images_subdir = f'{base_data_dir}/tmp_images/num_error/'
+    os.makedirs(images_subdir, exist_ok=True)
+
+    for outer_num, i in enumerate(range(iterations)):
+        skip = 0
+        print(f'figure [{i+1} / {iterations}]\r', end="")
+        plt.cla()
+        plt.title("[Iterations %i] \n"%(i+1) + _plot_title)
+        files = sorted([a for a in os.listdir(data_dir) if "[" not in a and int(a.split("_")[0]) == i and ".jpg" not in a and ".gif" not in a])
+
+        for num, file in enumerate(files):
+            val= np.load(data_dir+file, allow_pickle=True)["v"]
+            shape = val.shape
+            if np.prod(shape) <= 1:
+                skip += 1
+                continue
+
+            val = prepare(val)
+            val = np.sort(val)
+            num_max_dots = 500
+
+            if val.size > num_max_dots: val= val[::int(val.size)//num_max_dots]
+
+            ys = -np.ones(val.shape[0])*(num - skip)
+            ax.plot([1e-15, 1e18], [ys[0], ys[1]], 'C%i-'%(num%10), lw=10, alpha=0.2)
+            ax.plot(val, ys, 'C%io'%(num%10), ms=6, alpha=0.2)
+
+            if i == 0:
+                xticks.append(ys[0])
+                xticklabels.append(file.replace(".npz", "").replace("%i_"%i, ""))
+
+        plt.plot( [10**(-10), 10**(-10)], [0, xticks[-1]], 'C7--', alpha=0.6)
+        plt.plot( [10**(10), 10**10], [0, xticks[-1]], 'C7--', alpha=0.6)
+        plt.plot( [10**(0), 10**0], [0, xticks[-1]], 'C7-', alpha=1)
+
+        for x, label in zip(xticks, xticklabels):
+            ax.text(1e10, x+0.25, label, horizontalalignment='left', size='small', color='black', weight='normal')
+
+        plt.yticks([], [])
+        plt.xscale("log")
+        plt.xlim([10**(-15), 10**18])
+        if i == 0: plt.tight_layout()
+
+        plt.savefig(f'{images_subdir}num_error{outer_num}.jpg')
+
+    import imageio
+    gif_path = f'{base_data_dir}visualize_DFT_num_error_{molecule_name}.gif'
+    writer = imageio.get_writer(gif_path, loop=0, duration=7)
+    for i in range(iterations):
+        writer.append_data(imageio.v2.imread(f'{images_subdir}num_error{i}.jpg'))
+    writer.close()
+    print("Numerical error visualisation saved in", gif_path)