when i convert bert model to tensorRT model, it show this error message
[b]ERROR: bert/encoder/layer_0/attention/self/query/MatMul: at least three non-batch dimensions are required for input
ERROR: UffParser: Parser error: bert/encoder/layer_0/attention/self/query/BiasAdd: The input to the Scale Layer is required to have a minimum of 3 dimensions.
[2019-05-21 14:04:03,466] [logo] [FATAL] [1233] [1250000] [] [140166035814592] /home/huiyuan.lb/project/model-deploy-support/app/tensorrt-uff/tool/UffToTRTModel.cpp:164,parseUffModel error
run error[/b]
this is my code:
input tensor info
[
{
"name": "bert/embeddings/LayerNorm/batchnorm/add_1",
"dims": [1, 128, 768],
"format" : "kNC"
},
{
"name" : "bert/encoder/mul",
"dims": [1, 128, 128],
"format" : "kNC"
}
]
output tensor name
bert/encoder/Reshape_13
this is my uff model path:
http://alg-misc.cn-hangzhou.oss.aliyun-inc.com/huiyuan%2Ftmp%2Fbert.uff
there are some custom operators,i have impl it with plugins。