refine python interface for bfloat16

python/paddle/fluid/data_feeder.py

@@ -67,6 +67,9 @@ def convert_dtype(dtype):
             # however, jointly supporting python2 and python3, (as well as python4 maybe)
             # may still be a long-lasting problem.
             return str(dtype)
+        # NOTE(zhangbo): Now numpy not support bfloat, and paddle use uint16 to carry the data of bfloat16, and there binary is consistent.
+        if dtype in ['bfloat16']:
+            return 'uint16'
 
     raise TypeError(
         "dtype must be any of [bool, float16, uint16, float32, float64, int8, int16, "

python/paddle/fluid/dygraph/amp/auto_cast.py

@@ -235,7 +235,7 @@ def amp_guard(enable=True,
     level = level.upper()
     if not (level in ['O0', 'O1', 'O2']):
         raise ValueError(
-            "level should be O0, O1 or O2. O0 represents fp32 train mode, O1 represents AMP train mode, O2 represents pure fp16 train mode."
+            "level should be O0, O1 or O2. O0 represents fp32 train mode, O1 represents AMP train mode, O2 represents pure fp16/bf16 train mode."
         )
 
     # check amp_dtype: float16 or bfloat16

python/paddle/fluid/layers/tensor.py

@@ -230,7 +230,7 @@ def cast(x, dtype):
         x(Tensor): An input N-D Tensor with data type bool, float16,
             float32, float64, int32, int64, uint8.
         dtype(np.dtype|core.VarDesc.VarType|str): Data type of the output:
-            bool, float16, float32, float64, int8, int32, int64, uint8.
+            bool, float16, float32, float64, int8, int32, int64, uint8, bfloat16.
 
     Returns:
         Tensor: A Tensor with the same shape as input's.
@@ -245,6 +245,8 @@ def cast(x, dtype):
     """
     if in_dygraph_mode():
         if not isinstance(dtype, core.VarDesc.VarType):
+            if dtype == 'bfloat16':
+                dtype = 'uint16'
             dtype = convert_np_dtype_to_dtype_(dtype)
         out = _C_ops.cast(x, 'in_dtype', x.dtype, 'out_dtype', dtype)
         return out

python/paddle/tensor/to_string.py

@@ -247,7 +247,7 @@ def to_string(var, prefix='Tensor'):
     data = _format_tensor(
         np_var, summary, indent=indent, max_width=max_width, signed=signed)
 
-    return _template.format(
+    _template = _template.format(
         prefix=prefix,
         shape=var.shape,
         dtype=convert_dtype(var.dtype),
@@ -256,6 +256,11 @@ def to_string(var, prefix='Tensor'):
         indent=' ' * indent,
         data=data)
 
+    if var.dtype == paddle.fluid.core.VarDesc.VarType.BF16:
+        _template = _template + "\nthis data real dtype is bfloat16, you can use paddle.cast to cast it to float32 and get it's float data."
+
+    return _template
+
 
 def eager_tensor_to_string(tensor, prefix='Tensor'):
     indent = len(prefix) + 1