Unreasonable output of dropout_op with GPU kernel #8654
Description
Bugs may exist in the GPU kernel of dropout_op.
I just found that dropout_op often outputs tensor with all 0s even with a big size. I test the dropout_op with GPU by simple codes (the dropout_prob is 0.1 and the input data is 1s with shape [64, 32, 512]), and print the max and min of the output tensor 10 times as following：
1.0 1.0 0.0 0.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.0 0.0 0.0 0.0 1.0 1.0 The test codes are as following:
import numpy as np import paddle.v2 as paddle import paddle.fluid as fluid def program(): x = fluid.layers.data(name='x', shape=[64, 32, 512], dtype='float32', append_batch_size=False) out = fluid.layers.dropout(x, dropout_prob=0.1, is_test=False) return out def main(): #place = fluid.CPUPlace() place = fluid.CUDAPlace(0) exe = fluid.Executor(place) out = program() data_input = {} in_tensor = fluid.LoDTensor() in_tensor.set(np.ones([64, 32, 512], dtype="float32"), place) data_input['x'] = in_tensor for i in range(10): out_ = exe.run(fluid.framework.default_main_program(), feed=data_input, fetch_list=[out])[0] print np.max(out_), np.min(out_) if __name__ == "__main__": main()Running the code with CPU gets the proper outputs.
1.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 1.0 0.0 Metadata
Metadata
Assignees
Labels
No labels