Created by: zlinker

System information -PaddlePaddle version: 1.7

Description sequence_softmax here should not accumulate exp values directly, which can cause overflow.

here is a demo:

def network():
    data = fluid.layers.data("data", shape=[-1, 1], dtype="float32")
    output = fluid.layers.sequence_softmax(data)
    return output

seq_softmax = network()
data = fluid.create_lod_tensor(np.array([1, 2, 3, 4000], dtype="float32"), [[4]], fluid.CPUPlace())
exe = fluid.Executor(fluid.CPUPlace())
exe.run(fluid.default_startup_program())
result = exe.run(feed={"data": data}, fetch_list=[seq_softmax], return_numpy=False)
print(np.array(result)

output :

[ 0.  0.  0. nan]