singa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From GitBox <...@apache.org>
Subject [GitHub] [singa] nudles commented on a change in pull request #662: CUDNN LSTM
Date Thu, 09 Apr 2020 14:16:03 GMT
nudles commented on a change in pull request #662: CUDNN LSTM
URL: https://github.com/apache/singa/pull/662#discussion_r406234132
 
 

 ##########
 File path: python/singa/autograd.py
 ##########
 @@ -3239,95 +3239,176 @@ def __init__(
             bidirectional (bool): If True, becomes a bidirectional RNN. 
                 Default: False
         """
-        self.nonlinearity = nonlinearity
-
-        Wx_shape = (input_size, hidden_size)
-        self.Wx = []
-        for i in range(4):
-            w = Tensor(shape=Wx_shape, requires_grad=True, stores_grad=True)
-            w.gaussian(0.0, 1.0)
-            self.Wx.append(w)
-
-        Wh_shape = (hidden_size, hidden_size)
-        self.Wh = []
-        for i in range(4):
-            w = Tensor(shape=Wh_shape, requires_grad=True, stores_grad=True)
-            w.gaussian(0.0, 1.0)
-            self.Wh.append(w)
-
-        Bx_shape = (hidden_size,)
-        self.Bx = []
-        for i in range(4):
-            b = Tensor(shape=Bx_shape, requires_grad=True, stores_grad=True)
-            b.set_value(0.0)
-            self.Bx.append(b)
-
-        self.Bh = []
-        for i in range(4):
-            b = Tensor(shape=Bx_shape, requires_grad=True, stores_grad=True)
-            b.set_value(0.0)
-            self.Bh.append(b)
-
-        self.params = self.Wx + self.Wh + self.Bx + self.Bh
+        self.backend = backend
+        if backend == "singa":
+            self.nonlinearity = nonlinearity
+
+            Wx_shape = (input_size, hidden_size)
+            self.Wx = []
+            for i in range(4):
+                w = Tensor(shape=Wx_shape, requires_grad=True, stores_grad=True)
+                w.gaussian(0.0, 1.0)
+                self.Wx.append(w)
+
+            Wh_shape = (hidden_size, hidden_size)
+            self.Wh = []
+            for i in range(4):
+                w = Tensor(shape=Wh_shape, requires_grad=True, stores_grad=True)
+                w.gaussian(0.0, 1.0)
+                self.Wh.append(w)
+
+            Bx_shape = (hidden_size,)
+            self.Bx = []
+            for i in range(4):
+                b = Tensor(shape=Bx_shape, requires_grad=True, stores_grad=True)
+                b.set_value(0.0)
+                self.Bx.append(b)
+
+            self.Bh = []
+            for i in range(4):
+                b = Tensor(shape=Bx_shape, requires_grad=True, stores_grad=True)
+                b.set_value(0.0)
+                self.Bh.append(b)
+
+            self.params = self.Wx + self.Wh + self.Bx + self.Bh
+        elif backend == "cudnn":
+            if not singa.USE_CUDA:
+                raise Exception("Could not use cudnn without cuda compiled.\n")
+            if not inputs:
+                raise Exception("Input is required for init cudnn LSTM.\n")
+
+            CUDNN_LSTM_MODE = 2
+
+            cinputs = singa.VecTensor()
+            [cinputs.append(i.data) for i in inputs]
+            self.rnn_handle = singa.CudnnRNNHandle(cinputs, input_size,
+                                                   hidden_size, CUDNN_LSTM_MODE)
+        else:
+            raise Exception("Unsupported backend %s.\n" % backend)
+
+    def cpp_vec_tensor_to_py_tensor(self, cpp_vec_tensor):
+        py_tensors = list()
+        for cTensor in cpp_vec_tensor:
+            new_t = tensor.Tensor()
+            new_t.data = cTensor
+            new_t.shape = tuple(new_t.data.shape())
+            new_t.device = new_t.data.device()
+            new_t.dtype = new_t.data.data_type()
+            py_tensors.append(new_t)
+        return py_tensors
 
     def __call__(self, xs, h0_c0):
-        # xs: a tuple or list of input tensors
-        # h0_c0: a tuple of (h0, c0)
-        h0, c0 = h0_c0
-        if not isinstance(xs, list):
-            xs = list(xs)
-        inputs = xs + list((h0, c0))
-        self.device_check(*inputs)
-        # self.device_check(inputs[0], *self.params)
-        self.device_check(inputs[0], *(self.Wx + self.Wh + self.Bx + self.Bh))
-        batchsize = xs[0].shape[0]
-        out = []
-        h, c = self.step_forward(xs[0], h0, c0, self.Wx, self.Wh, self.Bx,
-                                 self.Bh)
-        out.append(h)
-        for x in xs[1:]:
-            assert x.shape[0] == batchsize
-            h, c = self.step_forward(x, h, c, self.Wx, self.Wh, self.Bx,
+        if self.backend == "singa":
 
 Review comment:
   LSTM is a Layer subclass.
   It should call lstm function to do the forward by passing the parameters to it.
   lstm function creates _LSTM operation, which implements the forward() and backward().
   
   Pls check the _Conv2d(Operation), conv2d() function and Conv2d(Layer).

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services

Mime
View raw message