Bugfix reproducibility & ensemble member order with dask (#347)

* Bugfix: fix random placement of ensemble members in numpy array due to dask multi-threading (#337)

* Bugfix: make STEPS (blending) nowcast reproducable when the seed argument is given (#346)

* Bugfix: make STEPS (blending) nowcast reproducable, independent of number of workers (#346)

* Formatting with black

---------

Co-authored-by: ned <[email protected]>

Author: mpvginde

pysteps/blending/steps.py

@@ -514,6 +514,7 @@ def forecast(
     )
 
     # 2. Initialize the noise method
+    np.random.seed(seed)
     pp, generate_noise, noise_std_coeffs = _init_noise(
         precip,
         precip_thr,
@@ -526,6 +527,7 @@ def forecast(
         noise_stddev_adj,
         measure_time,
         num_workers,
+        seed,
     )
 
     # 3. Perform the cascade decomposition for the input precip fields and
@@ -1662,6 +1664,7 @@ def _init_noise(
     noise_stddev_adj,
     measure_time,
     num_workers,
+    seed,
 ):
     """Initialize the noise method."""
     if noise_method is None:
@@ -1690,6 +1693,7 @@ def _init_noise(
             20,
             conditional=True,
             num_workers=num_workers,
+            seed=seed,
         )
 
         if measure_time:
@@ -1944,7 +1948,6 @@ def _init_random_generators(
     if noise_method is not None:
         randgen_prec = []
         randgen_motion = []
-        np.random.seed(seed)
         for j in range(n_ens_members):
             rs = np.random.RandomState(seed)
             randgen_prec.append(rs)

pysteps/noise/utils.py

@@ -96,39 +96,37 @@ def compute_noise_stddev_adjs(
 
     if dask_imported and num_workers > 1:
         res = []
-    else:
-        N_stds = []
 
+    N_stds = [None] * num_iter
     randstates = []
-    seed = None
+
     for k in range(num_iter):
         randstates.append(np.random.RandomState(seed=seed))
         seed = np.random.randint(0, high=1e9)
 
-    for k in range(num_iter):
-
-        def worker():
-            # generate Gaussian white noise field, filter it using the chosen
-            # method, multiply it with the standard deviation of the observed
-            # field and apply the precipitation mask
-            N = noise_generator(noise_filter, randstate=randstates[k], seed=seed)
-            N = N / np.std(N) * sigma + mu
-            N[~MASK] = R_thr_2
+    def worker(k):
+        # generate Gaussian white noise field, filter it using the chosen
+        # method, multiply it with the standard deviation of the observed
+        # field and apply the precipitation mask
+        N = noise_generator(noise_filter, randstate=randstates[k])
+        N = N / np.std(N) * sigma + mu
+        N[~MASK] = R_thr_2
 
-            # subtract the mean and decompose the masked noise field into a
-            # cascade
-            N -= mu
-            decomp_N = decomp_method(N, F, mask=MASK_)
+        # subtract the mean and decompose the masked noise field into a
+        # cascade
+        N -= mu
+        decomp_N = decomp_method(N, F, mask=MASK_)
 
-            return decomp_N["stds"]
-
-        if dask_imported and num_workers > 1:
-            res.append(dask.delayed(worker)())
-        else:
-            N_stds.append(worker())
+        N_stds[k] = decomp_N["stds"]
 
     if dask_imported and num_workers > 1:
-        N_stds = dask.compute(*res, num_workers=num_workers)
+        for k in range(num_iter):
+            res.append(dask.delayed(worker)(k))
+        dask.compute(*res, num_workers=num_workers)
+
+    else:
+        for k in range(num_iter):
+            worker(k)
 
     # for each cascade level, compare the standard deviations between the
     # observed field and the masked noise field, which gives the correction

pysteps/nowcasts/steps.py

@@ -443,6 +443,7 @@ def f(precip, i):
         precip[i, ~np.isfinite(precip[i, :])] = np.nanmin(precip[i, :])
 
     if noise_method is not None:
+        np.random.seed(seed)
         # get methods for perturbations
         init_noise, generate_noise = noise.get_method(noise_method)
 
@@ -466,6 +467,7 @@ def f(precip, i):
                 20,
                 conditional=True,
                 num_workers=num_workers,
+                seed=seed,
             )
 
             if measure_time:
@@ -543,7 +545,6 @@ def f(precip, i):
     if noise_method is not None:
         randgen_prec = []
         randgen_motion = []
-        np.random.seed(seed)
         for _ in range(n_ens_members):
             rs = np.random.RandomState(seed)
             randgen_prec.append(rs)
@@ -706,7 +707,7 @@ def _check_inputs(precip, velocity, timesteps, ar_order):
 
 
 def _update(state, params):
-    precip_forecast_out = []
+    precip_forecast_out = [None] * params["n_ens_members"]
 
     if params["noise_method"] is None or params["mask_method"] == "sprog":
         for i in range(params["n_cascade_levels"]):
@@ -828,7 +829,7 @@ def worker(j):
 
         precip_forecast[params["domain_mask"]] = np.nan
 
-        precip_forecast_out.append(precip_forecast)
+        precip_forecast_out[j] = precip_forecast
 
     if (
         DASK_IMPORTED