Merge pull request #26 from NREL/float32

fix float 32 json dump and format

.github/workflows/ci.yml

@@ -20,7 +20,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ['3.10']
+        python-version: ['3.11']
     defaults:
       run:
         working-directory: ${{github.workspace}}
@@ -45,7 +45,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ['3.10']
+        python-version: ['3.11']
     defaults:
       run:
         working-directory: ${{github.workspace}}
@@ -65,7 +65,7 @@ jobs:
           python main.py -nosimp -opt -lean -p 0.5 1
           cd ../preProcess
           python makeDataset_spm.py -nosimp -df ../integration_spm -frt 1 -p 0.5 1
-      - name: Test SPM
+      - name: Test SPM training
         run: |
           cd pinn_spm_param/tests
           pytest -v test_spm.py
@@ -82,3 +82,14 @@ jobs:
         run: |
           cd BayesianCalibration_spm
           bash exec_test.sh
+      - name: Test SPM train and post process float 32
+        run: |
+          cd pinn_spm_param
+          bash convert_to_float32.sh
+          cd integration_spm
+          python main.py -nosimp -opt -lean -p 0.5 1
+          cd ../preProcess
+          python makeDataset_spm.py -nosimp -df ../integration_spm -frt 1 -p 0.5 1
+          cd ../tests
+          pytest -v test_spm.py::test_sgd_lbfgs
+          pytest -v test_spm_postProc.py

pinn_spm_param/integration_spm/spm_int.py

@@ -53,17 +53,17 @@ def make_sim_config(t_dom, r_dom):
 
 
 def init_arrays(n_t, n_r):
-    phie = np.zeros(n_t)
-    phis_c = np.zeros(n_t)
-    cs_a = np.zeros((n_t, n_r))
-    cs_c = np.zeros((n_t, n_r))
-    Ds_a = np.zeros(n_r)
-    Ds_c = np.zeros(n_r)
-    rhs_a = np.zeros(n_r)
-    rhs_c = np.zeros(n_r)
-    A = np.zeros((n_r, n_r))
-    B_a = np.zeros(n_r)
-    B_c = np.zeros(n_r)
+    phie = np.zeros(n_t, dtype=np.float64)
+    phis_c = np.zeros(n_t, dtype=np.float64)
+    cs_a = np.zeros((n_t, n_r), dtype=np.float64)
+    cs_c = np.zeros((n_t, n_r), dtype=np.float64)
+    Ds_a = np.zeros(n_r, dtype=np.float64)
+    Ds_c = np.zeros(n_r, dtype=np.float64)
+    rhs_a = np.zeros(n_r, dtype=np.float64)
+    rhs_c = np.zeros(n_r, dtype=np.float64)
+    A = np.zeros((n_r, n_r), dtype=np.float64)
+    B_a = np.zeros(n_r, dtype=np.float64)
+    B_c = np.zeros(n_r, dtype=np.float64)
 
     return {
         "ce": 0,
@@ -98,7 +98,7 @@ def tridiag(ds, dt, dr):
 
 
 def rhs(dt, r, ddr_cs, ds, ddDs_cs, cs, bound_grad):
-    rhs_col = np.zeros(len(r))
+    rhs_col = np.zeros(len(r), dtype=np.float64)
     rhs_col = (
         dt
         * (np.float64(2.0) / np.clip(r, a_min=1e-12, a_max=None))
@@ -259,7 +259,7 @@ def integration(
                     grad_ds_a_cs_a(params["T"], params["R"])
                 )
             else:
-                gradDs_a_cs_a = np.zeros(len(sol["Ds_a"]))
+                gradDs_a_cs_a = np.zeros(len(sol["Ds_a"]), dtype=np.float64)
 
             ddr_csa = np.gradient(
                 sol["cs_a"][i_t - 1, :], r_a, axis=0, edge_order=2
@@ -289,7 +289,8 @@ def integration(
                 params["T"],
                 params["R"],
                 params["cscamax"],
-                deg_ds_c * np.ones(sol["cs_c"][i_t - 1, :].shape),
+                deg_ds_c
+                * np.ones(sol["cs_c"][i_t - 1, :].shape, dtype=np.float64),
             )
             if EXACT_GRAD_DS_CS:
                 gradDs_c_cs_c = grad_ds_c_cs_c(
@@ -302,26 +303,30 @@ def integration(
                 Ds_c_tmp1 = params["D_s_c"](
                     np.clip(
                         sol["cs_c"][i_t - 1, :]
-                        + np.ones(sol["cs_c"].shape[1]) * GRAD_STEP,
+                        + np.ones(sol["cs_c"].shape[1], dtype=np.float64)
+                        * GRAD_STEP,
                         a_min=0,
                         a_max=params["cscamax"],
                     ),
                     params["T"],
                     params["R"],
                     params["cscamax"],
-                    deg_ds_c * np.ones(sol["cs_c"][i_t - 1, :].shape),
+                    deg_ds_c
+                    * np.ones(sol["cs_c"][i_t - 1, :].shape, dtype=np.float64),
                 )
                 Ds_c_tmp2 = params["D_s_c"](
                     np.clip(
                         sol["cs_c"][i_t - 1, :]
-                        - np.ones(sol["cs_c"].shape[1]) * GRAD_STEP,
+                        - np.ones(sol["cs_c"].shape[1], dtype=np.float64)
+                        * GRAD_STEP,
                         a_min=0,
                         a_max=params["cscamax"],
                     ),
                     params["T"],
                     params["R"],
                     params["cscamax"],
-                    deg_ds_c * np.ones(sol["cs_c"][i_t - 1, :].shape),
+                    deg_ds_c
+                    * np.ones(sol["cs_c"][i_t - 1, :].shape, dtype=np.float64),
                 )
                 gradDs_c_cs_c = (
                     (Ds_c_tmp1 - Ds_c_tmp2) / (2 * GRAD_STEP)
@@ -357,7 +362,7 @@ def integration(
             )
             ddr_csa[0] = 0
             ddr_csa[-1] = -sol["j_a"] / sol["Ds_a"][-1]
-            ddr2_csa = np.zeros(n_r)
+            ddr2_csa = np.zeros(n_r, dtype=np.float64)
             ddr2_csa[1 : n_r - 1] = (
                 sol["cs_a"][i_t - 1, : n_r - 2]
                 - 2 * sol["cs_a"][i_t - 1, 1 : n_r - 1]
@@ -392,14 +397,15 @@ def integration(
                 params["T"],
                 params["R"],
                 params["cscamax"],
-                deg_ds_c * np.ones(sol["cs_c"][i_t - 1, :].shape),
+                deg_ds_c
+                * np.ones(sol["cs_c"][i_t - 1, :].shape, dtype=np.float64),
             )
             ddr_csc = np.gradient(
                 sol["cs_c"][i_t - 1, :], r_c, axis=0, edge_order=2
             )
             ddr_csc[0] = 0
             ddr_csc[-1] = -sol["j_c"] / sol["Ds_c"][-1]
-            ddr2_csc = np.zeros(n_r)
+            ddr2_csc = np.zeros(n_r, dtype=np.float64)
             ddr2_csc[1 : n_r - 1] = (
                 sol["cs_c"][i_t - 1, : n_r - 2]
                 - 2 * sol["cs_c"][i_t - 1, 1 : n_r - 1]

pinn_spm_param/util/forwardPass.py

@@ -47,7 +47,7 @@ def var_from_x(x_arr, dummyVal=np.float64(0.0)):
     if x_arr.shape[1] == 2:
         var_full[1] = var_simp[1]
     else:
-        var_full[1] = dummyVal * np.ones(x_arr.shape[0])
+        var_full[1] = dummyVal * np.ones(x_arr.shape[0], dtype=np.float64)
     return var_simp, var_full
 
 
@@ -62,7 +62,7 @@ def rescale_var_list(nn, var_list, dummyVal=np.float64(0.0)):
     else:
         var_list_resc_full[nn.ind_r] = (
             dummyVal
-            * np.ones(var_list_resc_full[nn.ind_t].shape)
+            * np.ones(var_list_resc_full[nn.ind_t].shape, dtype=np.float64)
             / nn.params["rescale_R"]
         )
     return var_list_resc, var_list_resc_full
@@ -276,11 +276,12 @@ def pinn_pred_struct(nn, params_list):
         (n_t * n_r * n_par * n_par, 1),
     ).astype("float64")
     params_r_unr = [
-        np.ones((n_t * n_r, 1)) * entry
+        np.ones((n_t * n_r, 1), dtype=np.float64) * entry
         for ientry, entry in enumerate(params_list)
     ]
     params_r = [
-        np.ones((n_t * n_r, 1)) * nn.rescale_param(entry, ientry)
+        np.ones((n_t * n_r, 1), dtype=np.float64)
+        * nn.rescale_param(entry, ientry)
         for ientry, entry in enumerate(params_list)
     ]
 

pinn_spm_param/util/myNN.py

@@ -3199,6 +3199,17 @@ def prepareLog(self):
         with open(
             os.path.join(self.modelFolder, "config.json"), "w+"
         ) as outfile:
+            # float 32 is not supported by Json
+            for key in self.config:
+                ent_type = str(type(self.config[key]))
+                if "numpy.float" in ent_type and "32" in ent_type:
+                    self.config[key] = float(self.config[key])
+                elif key == "params_min" or key == "params_max":
+                    for ientry, entry in enumerate(self.config[key]):
+                        ent_type = str(type(self.config[key][ientry]))
+                        if "numpy.float" in ent_type and "32" in ent_type:
+                            self.config[key][ientry] = float(entry)
+
             json.dump(self.config, outfile, indent=4, sort_keys=True)
 
         # Make log headers