transformer目录，按照pytorch风格整合进nn.module (#79)

* py:
1.transformer目录，按照pytorch风格整合进nn.module

Author: miaobyte

front/py/deepx/nn/functional/changeshape.py

@@ -1,7 +1,15 @@
 from typing import Union
 from deepx import Tensor
-from .leaffunc_changeshape import reshape,indexselect, concat,broadcastTo
+from .leaffunc_changeshape import reshape,indexselect, concat,broadcastTo,permute
 from .leaffunc_init import newtensor,arange
+
+
+def transpose(t:Tensor,dim0:int,dim1:int,out:Union[Tensor,str]='')->Tensor:
+    dimorder = list(range(t.ndim))
+    dimorder[dim0],dimorder[dim1]=dimorder[dim1],dimorder[dim0]
+    return permute(t,tuple(dimorder),out)
+
+
 def squeeze(t:Tensor,dim:int)->Tensor:
     assert isinstance(dim,int)
     assert isinstance(t,Tensor)

front/py/deepx/nn/functional/leaffunc_changeshape.py

@@ -9,7 +9,6 @@ def reshape(t:Tensor,shape:tuple[int,...],out:Union[Tensor,str]='')->Tensor:
     for i in shape:
         assert isinstance(i,int) and i>0
 
-    outtensor=out
     if isinstance(out,str) or out is None:
         outshape=shape
         outtensor=newtensor(outshape,dtype=t.dtype,name=out)
@@ -40,11 +39,6 @@ def permute(t:Tensor,
     rtf_transpose(t,dimorder,outtensor,defaultauthor['transpose'])
     return outtensor
 
-def transpose(t:Tensor,out:Union[Tensor,str]='')->Tensor:
-    dimorder = list(range(t.ndim))
-    dimorder[-1],dimorder[-2]=dimorder[-2],dimorder[-1]
-    return permute(t,tuple(dimorder),out)
-
 
 
 def concat(tensors:Union[list[Tensor],tuple[Tensor,...]],dim:int,out:Union[Tensor,str]='')->Tensor:

front/py/deepx/nn/functional/rtf.py

@@ -37,7 +37,7 @@ def A_op_C(op:str,a:Tensor,out:Tensor,author='miaobyte'):
     ir=DeepxIR(op, args, returns,author)
     send(ir)
 
-def A_b1_b2_op_C(op:str,a:Tensor,b1:tuple[int],b2:bool,out:Tensor,author='miaobyte'):
+def A_b1_b2_op_C(op:str,a:Tensor,b1:tuple[int,...],b2:bool,out:Tensor,author='miaobyte'):
     args=[Param.tensor(a),Param.vector(b1,'int32'),Param.varbool(b2)]
     returns=[Param.tensor(out)]
     ir=DeepxIR(op, args, returns,author)

front/py/deepx/nn/functional/rtf_changeshape.py

@@ -2,14 +2,14 @@
 from deepx.nn.deepxir import DeepxIR,Param
 from deepx.scheduler import send
 
-def rtf_reshape(t:Tensor,shape:tuple[int],out:Tensor,author='miaobyte'):
+def rtf_reshape(t:Tensor,shape:tuple[int,...],out:Tensor,author='miaobyte'):
     args=[Param.tensor(t),Param.vector(shape,'int32')]
     returns=[Param.tensor(out)]
     ir=DeepxIR("reshape", args, returns,author)
     send(ir)
 
 
-def rtf_transpose(t:Tensor,dimorder:tuple[int],out:Tensor,author='miaobyte'):
+def rtf_transpose(t:Tensor,dimorder:tuple[int,...],out:Tensor,author='miaobyte'):
     args=[Param.tensor(t),Param.vector(dimorder,'int32')]
     returns=[Param.tensor(out)]
     ir=DeepxIR("transpose", args, returns,author)
@@ -22,7 +22,7 @@ def rtf_concat(tensors:tuple[Tensor],dim:int,out:Tensor,author='miaobyte'):
     send(ir)
 
 
-def rtf_broadcastTo(t:Tensor,new_shape:tuple[int],out:Tensor,author='miaobyte'):
+def rtf_broadcastTo(t:Tensor,new_shape:tuple[int,...],out:Tensor,author='miaobyte'):
     args=[Param.tensor(t),Param.vector(new_shape,'int32')]
     returns=[Param.tensor(out)]
     ir=DeepxIR("broadcastTo", args, returns,author)

front/py/deepx/nn/functional/rtf_reduce.py

@@ -1,18 +1,18 @@
 from deepx.tensor import Tensor
 from .rtf import A_b1_b2_op_C
 
-def rtf_sum(a:Tensor,dim:tuple[int],keepdim:bool,out: Tensor, author:str='miaobyte')->Tensor:
+def rtf_sum(a:Tensor,dim:tuple[int,...],keepdim:bool,out: Tensor, author:str='miaobyte')->Tensor:
     A_b1_b2_op_C("sum",a,dim,keepdim,out,author)
 
 
-def rtf_prod(a:Tensor,dim:tuple[int],keepdim:bool,out:Tensor, author:str='miaobyte')->Tensor:
+def rtf_prod(a:Tensor,dim:tuple[int,...],keepdim:bool,out:Tensor, author:str='miaobyte')->Tensor:
     A_b1_b2_op_C("prod",a,dim,keepdim,out,author)
 
 
-def rtf_reducemax(a:Tensor,dim:tuple[int],keepdim:bool,out:Tensor, author:str='miaobyte')->Tensor:
+def rtf_reducemax(a:Tensor,dim:tuple[int,...],keepdim:bool,out:Tensor, author:str='miaobyte')->Tensor:
     A_b1_b2_op_C("reducemax",a,dim,keepdim,out,author)
 
 
-def rtf_reducemin(a:Tensor,dim:tuple[int],keepdim:bool,out:Tensor, author:str='miaobyte')->Tensor:
+def rtf_reducemin(a:Tensor,dim:tuple[int,...],keepdim:bool,out:Tensor, author:str='miaobyte')->Tensor:
     A_b1_b2_op_C("reducemin",a,dim,keepdim,out,author)
 

front/py/deepx/nn/modules/conv.py front/py/deepx/nn/modules/conv/conv.pyfront/py/deepx/nn/modules/conv.py renamed to front/py/deepx/nn/modules/conv/conv.py

@@ -41,7 +41,7 @@ def reset_parameters(self) -> None:
 
     def forward(self, input: Tensor) -> Tensor:
         #`y = xA^T + b`
-        y=input @ self.weight.T
+        y=input @ self.weight.mT
         oldshape=y.shape
         if self.bias is not None:
             y.reshape_(tuple(y.shape[1:]))

front/py/deepx/nn/modules/linear.py

@@ -5,7 +5,7 @@
     "silu":swish_fn,
 }
 
-class LlamaMLP(Module):
+class MLP(Module):
     def __init__(self, config:dict):
         super().__init__()
         # 输入层大小