Merge pull request #88 from lucasdelimanogueira/tmp

Tmp

norch/csrc/cuda.cu

@@ -24,9 +24,8 @@ __host__ void cpu_to_cuda(Tensor* tensor, int device_id) {
 
     tensor->data = data_tmp;
 
-    const char* device_str = "cuda";
-    tensor->device = (char*)malloc(strlen(device_str) + 1);
-    strcpy(tensor->device, device_str); 
+    tensor->device = (char*)malloc(strlen("cuda") + 1);
+    strcpy(tensor->device, "cuda"); 
 }
 
 __host__ void cuda_to_cpu(Tensor* tensor) {
@@ -37,9 +36,8 @@ __host__ void cuda_to_cpu(Tensor* tensor) {
 
     tensor->data = data_tmp;
 
-    const char* device_str = "cpu";
-    tensor->device = (char*)malloc(strlen(device_str) + 1);
-    strcpy(tensor->device, device_str); 
+    tensor->device = (char*)malloc(strlen("cpu") + 1);
+    strcpy(tensor->device, "cpu"); 
 }
 
 __host__ void free_cuda(float* data) {

norch/csrc/tensor.h

@@ -13,6 +13,8 @@ typedef struct {
 extern "C" {
     Tensor* create_tensor(float* data, int* shape, int ndim, char* device);
     void delete_tensor(Tensor* tensor);
+    void delete_strides(Tensor* tensor);
+    void delete_device(Tensor* tensor);
     float get_item(Tensor* tensor, int* indices);
     Tensor* add_tensor(Tensor* tensor1, Tensor* tensor2);
     Tensor* sum_tensor(Tensor* tensor, int axis, bool keepdims);

norch/nn/loss.py

@@ -11,7 +11,7 @@ class Loss(Module, ABC):
 
     def forward(self, predictions, labels):
         raise NotImplementedError
-    
+            
     def __call__(self, *inputs):
         return self.forward(*inputs)
     

norch/optim/optimizers/sgd.py

@@ -9,6 +9,7 @@ class SGD(Optimizer):
         self.momentum = momentum
         self._cache = {'velocity': [p.zeros_like() for (_, _, p) in self.parameters]}
 
+    
     def step(self):
         for i, (module, name, _) in enumerate(self.parameters):
             parameter = getattr(module, name)

norch/tensor.py

@@ -26,10 +26,10 @@ class Tensor:
             
             self.shape = shape.copy()
             
-            self.data_ctype = (ctypes.c_float * len(data))(*data.copy())
-            self.shape_ctype = (ctypes.c_int * len(shape))(*shape.copy())
-            self.ndim_ctype = ctypes.c_int(len(shape))
-            self.device_ctype = device.encode('utf-8')
+            self._data_ctype = (ctypes.c_float * len(data))(*data.copy())
+            self._shape_ctype = (ctypes.c_int * len(shape))(*shape.copy())
+            self._ndim_ctype = ctypes.c_int(len(shape))
+            self._device_ctype = device.encode('utf-8')
 
             self.ndim = len(shape)
             self.device = device
@@ -47,15 +47,11 @@ class Tensor:
             Tensor._C.create_tensor.restype = ctypes.POINTER(CTensor)
             
             self.tensor = Tensor._C.create_tensor(
-                self.data_ctype,
-                self.shape_ctype,
-                self.ndim_ctype,
-                self.device_ctype
+                self._data_ctype,
+                self._shape_ctype,
+                self._ndim_ctype,
+                self._device_ctype
             )
-
-            del self.data_ctype
-            del self.shape_ctype
-            del self.device_ctype
         
         else:
             self.tensor = None,
@@ -85,7 +81,21 @@ class Tensor:
         return flat_data, shape
     
     def __del__(self):
-        if self.tensor is not None:
+            
+        if hasattr(self, '_data_ctype') and self._data_ctype is not None:
+            # tensor created by user (ctypes) will be deleted by python garbage collector
+            # only strides need to be deallocated manually because it is created inside C code
+            Tensor._C.delete_strides.argtypes = [ctypes.POINTER(CTensor)]
+            Tensor._C.delete_strides.restype = None
+            Tensor._C.delete_strides(self.tensor)
+
+            Tensor._C.delete_device.argtypes = [ctypes.POINTER(CTensor)]
+            Tensor._C.delete_device.restype = None
+            Tensor._C.delete_device(self.tensor)
+
+        elif self.tensor is not None:
+            # tensor created during operations must be deallocated
+
             Tensor._C.delete_tensor.argtypes = [ctypes.POINTER(CTensor)]
             Tensor._C.delete_tensor.restype = None
             Tensor._C.delete_tensor(self.tensor)