small change

Author: x54-729

fastNLP/core/drivers/oneflow_driver/ddp.py

@@ -325,7 +325,7 @@ def broadcast_object(self, obj, src: int = 0, group=None, **kwargs):
             return
         return fastnlp_oneflow_broadcast_object(obj, src, device=self.data_device)
 
-    def all_gather(self, obj, group) -> List:
+    def all_gather(self, obj) -> List:
         r"""
         将 ``obj`` 互相传送到其它所有的 rank 上，其中 ``obj`` 可能是 Tensor，也可能是嵌套结构的 object 。如果不是基础类型的数据，将会尝试通过
         pickle 进行序列化，接收到之后再反序列化。

fastNLP/core/drivers/torch_driver/deepspeed.py

@@ -121,7 +121,7 @@ def __init__(
         parallel_device: Union[List["torch.device"], "torch.device"],
         is_pull_by_torch_run = False,
         fp16: bool = False,
-        deepspeed_kwargs: Dict = None,
+        deepspeed_kwargs: Dict = {},
         **kwargs
     ):
         assert _NEED_IMPORT_DEEPSPEED, "Deepspeed is not imported."

tests/core/drivers/oneflow_driver/test_ddp.py

@@ -541,7 +541,7 @@ def test_save_and_load_model(self, only_state_dict):
                 res1 = driver1.model.evaluate_step(**batch)
                 res2 = driver2.model.evaluate_step(**batch)
 
-                assert oneflow.all(res1["preds"] == res2["preds"])
+                assert oneflow.all(res1["pred"] == res2["pred"])
         finally:
             rank_zero_rm(path)
 
@@ -635,9 +635,10 @@ def test_save_and_load_with_bucketedbatchsampler(self, device, only_state_dict,
 
                 left_x_batches.update(batch["x"].reshape(-1, ).tolist())
                 left_y_batches.update(batch["y"].reshape(-1, ).tolist())
+                batch = driver1.move_data_to_device(batch)
                 res1 = driver1.model.evaluate_step(**batch)
                 res2 = driver2.model.evaluate_step(**batch)
-                assert oneflow.all(res1["preds"] == res2["preds"])
+                assert oneflow.all(res1["pred"] == res2["pred"])
 
             assert len(left_x_batches) + len(already_seen_x_set) == len(self.dataset) / num_replicas
             assert len(left_x_batches | already_seen_x_set) == len(self.dataset) / num_replicas
@@ -727,9 +728,10 @@ def test_save_and_load_with_randomsampler(self, device, only_state_dict, fp16):
 
                 left_x_batches.update(batch["x"].reshape(-1, ).tolist())
                 left_y_batches.update(batch["y"].reshape(-1, ).tolist())
+                batch = driver1.move_data_to_device(batch)
                 res1 = driver1.model.evaluate_step(**batch)
                 res2 = driver2.model.evaluate_step(**batch)
-                assert oneflow.all(res1["preds"] == res2["preds"])
+                assert oneflow.all(res1["pred"] == res2["pred"])
 
             assert len(left_x_batches) + len(already_seen_x_set) == len(self.dataset) / num_replicas
             assert len(left_x_batches | already_seen_x_set) == len(self.dataset) / num_replicas

tests/core/drivers/oneflow_driver/test_dist_utils.py

@@ -81,7 +81,7 @@ def test_tensor_object_transfer_tensor(device):
 def test_fastnlp_oneflow_all_gather():
     local_rank = int(os.environ["LOCAL_RANK"])
     obj = {
-        "tensor": oneflow.full(size=(2, ), value=local_rank, dtype=oneflow.int).cuda(),
+        "tensor": oneflow.full((2, ), local_rank, oneflow.int).cuda(),
         "numpy": np.full(shape=(2, ), fill_value=local_rank),
         "bool": local_rank % 2 == 0,
         "float": local_rank + 0.1,
@@ -91,8 +91,8 @@ def test_fastnlp_oneflow_all_gather():
         },
         "list": [local_rank]*2,
         "str": f"{local_rank}",
-        "tensors": [oneflow.full(size=(2, ), value=local_rank, dtype=oneflow.int).cuda(),
-                    oneflow.full(size=(2, ), value=local_rank, dtype=oneflow.int).cuda()]
+        "tensors": [oneflow.full((2, ), local_rank, oneflow.int).cuda(),
+                    oneflow.full((2, ), local_rank, oneflow.int).cuda()]
     }
     data = fastnlp_oneflow_all_gather(obj)
     world_size = int(os.environ["WORLD_SIZE"])
@@ -118,7 +118,7 @@ def test_fastnlp_oneflow_broadcast_object():
     local_rank = int(os.environ["LOCAL_RANK"])
     if os.environ["LOCAL_RANK"] == "0":
         obj = {
-            "tensor": oneflow.full(size=(2, ), value=local_rank, dtype=oneflow.int).cuda(),
+            "tensor": oneflow.full((2, ), local_rank, oneflow.int).cuda(),
             "numpy": np.full(shape=(2, ), fill_value=local_rank, dtype=int),
             "bool": local_rank % 2 == 0,
             "float": local_rank + 0.1,
@@ -128,8 +128,8 @@ def test_fastnlp_oneflow_broadcast_object():
             },
             "list": [local_rank] * 2,
             "str": f"{local_rank}",
-            "tensors": [oneflow.full(size=(2, ), value=local_rank, dtype=oneflow.int).cuda(),
-                        oneflow.full(size=(2, ), value=local_rank, dtype=oneflow.int).cuda()]
+            "tensors": [oneflow.full((2, ), local_rank, oneflow.int).cuda(),
+                        oneflow.full((2, ), local_rank, oneflow.int).cuda()]
         }
     else:
         obj = None

tests/core/metrics/test_accuracy_torch.py

@@ -8,8 +8,6 @@
 import pytest
 import numpy as np
 
-from sklearn.metrics import accuracy_score as sklearn_accuracy
-
 from fastNLP.core.dataset import DataSet
 from fastNLP.core.metrics.accuracy import Accuracy
 from fastNLP.core.metrics.metric import Metric
@@ -21,6 +19,10 @@
     from torch.multiprocessing import Pool, set_start_method
 else:
     from fastNLP.core.utils.dummy_class import DummyClass as set_start_method
+try:
+    from sklearn.metrics import accuracy_score as sklearn_accuracy
+except:
+    pass
 
 set_start_method("spawn", force=True)