[incubator-tvm] branch master updated: [ONNX] Skip ADD inside Gemm op when vector is zero (#5697)

[masahi]

This is an automated email from the ASF dual-hosted git repository.

masahi pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-tvm.git


The following commit(s) were added to refs/heads/master by this push:
     new 879158a  [ONNX] Skip ADD inside Gemm op when vector is zero (#5697)
879158a is described below

commit 879158a07158f85bc4bb63127ac0226aab744532
Author: Balint Cristian <cristian.bal...@gmail.com>
AuthorDate: Sat May 30 04:10:22 2020 +0300

    [ONNX] Skip ADD inside Gemm op when vector is zero (#5697)
---
 python/tvm/relay/frontend/onnx.py | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/python/tvm/relay/frontend/onnx.py 
b/python/tvm/relay/frontend/onnx.py
index ea1ac90..be88683 100644
--- a/python/tvm/relay/frontend/onnx.py
+++ b/python/tvm/relay/frontend/onnx.py
@@ -462,6 +462,10 @@ class Gemm(OnnxOpConverter):
         inputs[0] = _op.nn.batch_flatten(inputs[0])
         out = _op.nn.dense(_expr.const(alpha) * inputs[0],
                            inputs[1], units=channels)
+        # skip (beta * C) if zero
+        C_array = params[inputs[2].name_hint].asnumpy()
+        if (beta == 0.0) or np.array_equal(C_array, np.array([0])):
+            return out
         return _op.nn.bias_add(out, _expr.const(beta) * inputs[2])