linear:linear验证

Author: miaobyte

excuter/common/src/deepx/mem/mem.hpp

@@ -133,6 +133,10 @@ namespace deepx::mem
         template <typename T>
         shared_ptr<Tensor<T>> gettensor(const string &name) const
         {
+            if (mem.find(name)== mem.end())
+            {
+                throw std::runtime_error("tensor not found: " + name);
+            }
             auto ptr = mem.at(name);
             return std::static_pointer_cast<Tensor<T>>(ptr);
         }

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

@@ -13,7 +13,7 @@ def __init__(self, in_features, out_features, bias=True,dtype:str="float32"):
             self.bias = None
 
     def forward(self, input):
-        output=input.matmul_(self.weight.T)
+        output=input.matmul(self.weight.T)
         if self.bias is not None:
             output=output+self.bias
         return output

front/py/deepx/nn/modules/module.py

@@ -3,9 +3,10 @@
                     Any, List, overload)
 from collections import OrderedDict
 from deepx import Tensor
-
+from deepx.autograd import Graph
 class Module:  
     def __init__(self, name: Optional[str] = None):
+        self._graph=Graph.get_default()
         self._name = name or self._generate_default_name()
         self._parent: Optional[Module] = None
         self._modules: OrderedDict[str, Module] = OrderedDict()
@@ -20,6 +21,10 @@ def _generate_default_name(self) -> str:
         self.__class__._instance_counter += 1
         return f"{base_name}_{count}"
 
+    @property
+    def graph(self):
+        return self._graph
+    
     def __setattr__(self, name: str, value: Any) -> None:
         if not name.startswith('_'):
             if isinstance(value, Module):

front/py/deepx/tensor/tensor.py

@@ -109,6 +109,11 @@ def __truediv__(self, other):
     def __matmul__(self, other):
         return self.matmul(other)
 
+    #自动转置最后两个维度，适用于二维矩阵
+    @property
+    def T(self) -> str:
+        return self.transpose(1,0)
+
     def __repr__(self) -> str:
         from deepx.nn.functional import printtensor
         s=printtensor(self)

front/py/examples/3_module/1_linear.py

@@ -1,15 +1,19 @@
 from deepx.nn.modules import Linear, Module
-
+from deepx import Tensor
 class Net(Module):
     def __init__(self):
         super().__init__()
-        self.fc1 = Linear(10, 5)
-        self.fc2 = Linear(5, 1)
+        self.fc1 = Linear(64, 32)
+        self.fc2 = Linear(32, 2)
     def forward(self, x):
         x = self.fc1(x)
         x = self.fc2(x)
         return x
 
 net = Net()
+input=Tensor(shape=[1,64])
+out=net.forward(input)
+net.graph.to_dot().render('linear.dot',format='svg')
+
 for name, param in net.named_parameters():
     print(f"{name}: {param.shape}")

front/py/examples/3_module/linear.dot

@@ -0,0 +1,13 @@
+// Computational Graph
+digraph {
+	rankdir=TB
+	node [shape=record]
+	125633308838112 [label="tensor_1
+(5, 10)" color=skyblue fillcolor=aliceblue fontname="Sans-Serif" labeljust=l shape=box style=filled]
+	125633308837296 [label="tensor_2
+(5,)" color=skyblue fillcolor=aliceblue fontname="Sans-Serif" labeljust=l shape=box style=filled]
+	125633308838160 [label="tensor_3
+(1, 5)" color=skyblue fillcolor=aliceblue fontname="Sans-Serif" labeljust=l shape=box style=filled]
+	125633308838400 [label="tensor_4
+(1,)" color=skyblue fillcolor=aliceblue fontname="Sans-Serif" labeljust=l shape=box style=filled]
+}