[Fix] restore 'by_epoch' for SchedulerList and fix EPNN (#777)

* restore 'by_epoch' for SchedulerList * fix for epnn
PaddlePaddle · Feb 4, 2024 · eaff0fc · eaff0fc
1 parent 988fd33
commit eaff0fc
Show file tree

Hide file tree

Showing 3 changed files with 19 additions and 6 deletions.
diff --git a/examples/epnn/functions.py b/examples/epnn/functions.py
@@ -226,6 +226,11 @@ def __init__(self, data_state, data_stress, itrain):
         self.data_stress = data_stress
         self.itrain = itrain
 
+    def _cvt_to_ndarray(self, list_dict):
+        for key in list_dict:
+            list_dict[key] = np.asarray(list_dict[key])
+        return list_dict
+
     def get(self, epochs=1):
         # Slow if using BatchSampler to obtain data
         input_dict_train = {
@@ -243,7 +248,7 @@ def get(self, epochs=1):
         label_dict_train = {"dummy_loss": []}
         label_dict_val = {"dummy_loss": []}
         for i in range(epochs):
-            shuffled_indices = paddle.randperm(n=self.data_state.x_train.shape[0])
+            shuffled_indices = np.random.permutation(self.data_state.x_train.shape[0])
             input_dict_train["state_x"].append(
                 self.data_state.x_train[shuffled_indices[0 : self.itrain]]
             )
@@ -256,9 +261,9 @@ def get(self, epochs=1):
             input_dict_train["stress_y"].append(
                 self.data_stress.y_train[shuffled_indices[0 : self.itrain]]
             )
-            label_dict_train["dummy_loss"].append(paddle.to_tensor(0.0))
+            label_dict_train["dummy_loss"].append(0.0)
 
-        shuffled_indices = paddle.randperm(n=self.data_state.x_valid.shape[0])
+        shuffled_indices = np.random.permutation(self.data_state.x_valid.shape[0])
         input_dict_val["state_x"].append(
             self.data_state.x_valid[shuffled_indices[0 : self.itrain]]
         )
@@ -271,7 +276,11 @@ def get(self, epochs=1):
         input_dict_val["stress_y"].append(
             self.data_stress.y_valid[shuffled_indices[0 : self.itrain]]
         )
-        label_dict_val["dummy_loss"].append(paddle.to_tensor(0.0))
+        label_dict_val["dummy_loss"].append(0.0)
+        input_dict_train = self._cvt_to_ndarray(input_dict_train)
+        label_dict_train = self._cvt_to_ndarray(label_dict_train)
+        input_dict_val = self._cvt_to_ndarray(input_dict_val)
+        label_dict_val = self._cvt_to_ndarray(label_dict_val)
         return input_dict_train, label_dict_train, input_dict_val, label_dict_val
 
 
@@ -287,7 +296,7 @@ def __init__(self, dataset_path, train_p=0.6, cross_valid_p=0.2, test_p=0.2):
     def get_shuffled_data(self):
         # Need to set the seed, otherwise the loss will not match the precision
         ppsci.utils.misc.set_random_seed(seed=10)
-        shuffled_indices = paddle.randperm(n=self.x.shape[0])
+        shuffled_indices = np.random.permutation(self.x.shape[0])
         n_train = math.floor(self.train_p * self.x.shape[0])
         n_cross_valid = math.floor(self.cross_valid_p * self.x.shape[0])
         n_test = math.floor(self.test_p * self.x.shape[0])

diff --git a/ppsci/data/__init__.py b/ppsci/data/__init__.py
@@ -89,7 +89,10 @@ def build_dataloader(_dataset, cfg):
         logger.warning(
             "`batch_size` is set to 1 as neither sampler config nor batch_size is set."
         )
-        batch_sampler = None
+        batch_sampler = io.BatchSampler(
+            _dataset,
+            batch_size=cfg["batch_size"],
+        )
 
     # build collate_fn if specified
     batch_transforms_cfg = cfg.pop("batch_transforms", None)

diff --git a/ppsci/optimizer/lr_scheduler.py b/ppsci/optimizer/lr_scheduler.py
@@ -750,6 +750,7 @@ class SchedulerList:
     def __init__(self, scheduler_list: Tuple[lr.LRScheduler, ...]):
         super().__init__()
         self._sch_list = scheduler_list
+        self.by_epoch = False
 
     def step(self):
         for sch in self._sch_list: