singa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From GitBox <...@apache.org>
Subject [GitHub] [singa] dcslin commented on a change in pull request #662: CUDNN LSTM
Date Wed, 22 Apr 2020 04:04:18 GMT

dcslin commented on a change in pull request #662:
URL: https://github.com/apache/singa/pull/662#discussion_r412652761



##########
File path: python/singa/autograd.py
##########
@@ -3330,6 +3330,94 @@ def step_forward(self, x, h, c, Wx, Wh, Bx, Bh):
         return hout, cout
 
 
+class _RNN(Operation):
+    """ RNN operation with c++ backend
+    """
+    def __init__(self, handle):
+        assert singa.USE_CUDA is True, "Not able to run without CUDA"
+        super(_RNN, self).__init__()
+        self.handle = handle
+
+    def forward(self, x, W):
+        # TODO: CPU forward
+
+        # GPU forward
+        if training:
+            y = singa.GpuRNNForwardTraining(x, W, self.handle)
+            self.inputs = (x, W, y)
+        else:
+            y = singa.GpuRNNForwardInference(x, W, self.handle)
+
+        return y
+
+    def backward(self, dy):
+        assert training is True and hasattr(
+            self, "inputs"), "Please set training as True before do BP. "
+
+        # TODO: CPU backward
+
+        # GPU backward
+        dx = singa.GpuRNNBackwardx(self.inputs[2], dy, self.inputs[1], self.handle)
+        dW = singa.GpuRNNBackwardW(self.inputs[0], self.inputs[2], self.handle)
+        return dx, dW
+
+class RNN_direct(Layer):
+    """ `RNN_direct` class implements with c++ backend and run the operation
+          directly on cuDNN
+
+        While `RNN` class implements with high level singa API
+    """
+    def __init__(self, input_size, hidden_size, rnn_mode="lstm"):
+        """
+            Args:
+                input_size: input feature dim
+                hidden_size: hidden feature dim
+                rnn_mode: accepted value: "vanilla", "tanh", "relu",  "lstm", "gru"
+        """
+        assert singa.USE_CUDA is True, "Not able to run without CUDA"
+
+        self.rnn_mode = rnn_mode
+        self.input_size = input_size
+        self.hidden_size = hidden_size
+
+        # TODO: CPU parameter
+
+        # GPU parameter
+        # cudnn_rnn_mode: 0 - RNN RELU, 1 - RNN TANH, 2 - LSTM, 3 - GRU
+        if self.rnn_mode == "lstm":
+            self.cudnn_rnn_mode = 2
+        elif self.rnn_mode == "vanilla" or self.rnn_mode == "tanh":
+            self.cudnn_rnn_mode = 1
+        elif self.rnn_mode == "relu":
+            self.cudnn_rnn_mode = 0
+        elif self.rnn_mode == "gru":
+            self.cudnn_rnn_mode = 3
+
+    def __call__(self, x):
+        if not hasattr(self, "handle"):
+            cpp_x = singa.VecTensor()
+            [cpp_x.append(i.data) for i in x]
+
+            # TODO: CPU handle
+
+            # GPU handle
+            self.handle = singa.CudnnRNNHandle(cpp_x, self.input_size, self.hidden_size,
self.cudnn_rnn_mode)

Review comment:
       In the cudnn settings, x is a list of {batchsize, featuresize} with length `seq_length`,
and batchsize might vary. And also there are `seq_length` of `CudnnTensorDescriptor` which
can have different batch size.
   
   We might allow only fixed value of batch size in one batch of input. Then we could remove
`x` from constructor.
   
   Side note 1. current Conv2d is using similar design, `x` in constructor
   2. native rnn implementation currently has similar API, x is a list of {batchsize, featuresize}
with length `seq_length`




----------------------------------------------------------------
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



Mime
View raw message