Tensorrt: ERROR: builtin_op_importers.cpp:2179 In function importPad: [8] Assertion failed: inputs.at(1).is_weights()

[HelloAlone]

Description

My current workflow is pb -> onnx -> tensorrt. Thanks to @jignparm , (refer https://github.com/onnx/tensorflow-onnx/issues/994) I finally converted the original pb to onnx model. But I received a importPad error when converting onnx to tensorrt engine. It seems is a TensorRT issue according to jignparm.

[V] [TRT] ModelImporter.cpp:125: preprocessor/PadV2 [Pad] inputs: [preprocessor/resize/ResizeBilinear:0 -> (-1, -1, -1, -1)], [preprocessor/PadV2__135:0 -> (8)], [preprocessor/PadV2/constant_values:0 -> ()],
ERROR: builtin_op_importers.cpp:2179 In function importPad:
[8] Assertion failed: inputs.at(1).is_weights()

Environment

TensorRT Version: 7.1.3.4 GA
GPU Type: GTX 1080Ti
Nvidia Driver Version: 440.33.01
CUDA Version: 10.2
CUDNN Version: 8
Operating System + Version: Ubuntu 16.04
Python Version (if applicable): 3.7
TensorFlow Version (if applicable): 1.15
Onnx Version (if applicable): 1.6.0
TensorFlow-onnx Version (if applicable): 1.6.2

Relevant Files

The onnx file can be download at https://yadi.sk/d/NW5no-n0HtNECw. The original pb file can be download at https://yadi.sk/d/PN3tDlm6GzwSpw

Steps To Reproduce

1. edit the pb and set the initializer to a type TRT can handle (float64->float32):

import tensorflow as tf
from tensorflow.python.compiler.tensorrt import trt_convert as trt
import numpy as np
from PIL import Image
with tf.Session() as sess:
    # First deserialize your frozen graph:
    with tf.gfile.GFile('models/rjcd_13_0.1421_0.1290.pb', 'rb') as f:
        frozen_graph = tf.GraphDef()
        frozen_graph.ParseFromString(f.read())
    with tf.Graph().as_default() as graph:
        tf.import_graph_def(frozen_graph, name='')
    new_model = tf.GraphDef()
    with tf.Session(graph=graph) as sess:
        for n in sess.graph_def.node:
            if n.name == 'resize_information_computer/truediv/Cast' or n.name == 'resize_information_computer/truediv/Cast_1':
                nn = new_model.node.add()
                nn.CopyFrom(n)
                nn.attr['DstT'].CopyFrom(tf.AttrValue(type=tf.float32.as_datatype_enum))
            elif n.name == 'resize_information_computer/truediv':
                nn = new_model.node.add()
                nn.CopyFrom(n)
                nn.attr['T'].CopyFrom(tf.AttrValue(type=tf.float32.as_datatype_enum))
            elif n.name == 'resize_information_computer/ToFloat':
                nn = new_model.node.add()
                nn.CopyFrom(n)
                nn.attr['SrcT'].CopyFrom(tf.AttrValue(type=tf.float32.as_datatype_enum))
            else:
                nn = new_model.node.add()
                nn.CopyFrom(n)
    with tf.gfile.GFile('models/rjcd_13_0.1421_0.1290_new.pb', mode='wb') as f:
        f.write(new_model.SerializeToString())

1. convert pb to onnx

python -m tf2onnx.convert --graphdef rjcd_13_0.1421_0.1290_new.pb --output rjcd_13_0.1421_0.1290.onnx --inputs image_input:0,max_detections:0,iou_threshold:0,score_threshold:0 --outputs output/Squeeze:0 --opset 11 --fold_const

1. convert onnx to tensorrt engine

trtexec --onnx=rjcd_13_0.1421_0.1290.onnx --explicitBatch

[XinnWang]

i met the same problem. have you solved it ?

[zccyman]

have same problem.

[lrh454830526]

Have you solved the problem?

[hewumars]

I met the same problem.have you solved it ?
ERROR: builtin_op_importers.cpp:436 In function importConv: Assertion failed: inputs.at(2).is_weights()
TRT version:7.1.3
ONNX IR version: 0.0.7
Opset version: 12
Producer name: tf2onnx
Producer version: 1.6.3

[hyc-xyz]

same problem

While parsing node number 223 [Pad -> "544"]:
ERROR: builtin_op_importers.cpp:2220 In function importPad:
[8] Assertion failed: inputs.at(1).is_weights()

[cognitiveRobot]

Getting the same error. Any update?

ERROR: builtin_op_importers.cpp:2179 In function importPad:
[8] Assertion failed: inputs.at(1).is_weights()
[09/28/2020-13:43:43] [E] Failed to parse onnx file
[09/28/2020-13:43:43] [E] Parsing model failed
[09/28/2020-13:43:43] [E] Engine creation failed
[09/28/2020-13:43:43] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec # /usr/src/tensorrt/bin/trtexec --onnx=model_r50_frcnn_fpn.onnx --saveEngine=model_r50_frcnn_fpn.trt

[mk-nvidia]

Thanks for reporting, we'll take a look.

[simutisernestas]

Facing same issue. @mk-nvidia any progress ?

[SteveMacenski]

Anyone find a solution to this?

[batrlatom]

internet if flooding with the requests for solution :] . this would help to convert maskrnn, fasterrcnn, retinanet and many other networks.

[TomWildenhain-Microsoft]

The assertion suggests that TensorRT requires the pads to be constant (known at model load time). In the ONNX model, this is not the case:

The pads are computed by this logic:

This is partially be because the input size itself is unknown:

There are 3 possible ways to get this model to work:

1. Update TensorRT to support non-const pads (as per ONNX spec)
2. Update tf2onnx to convert non-const pads to a composition of other ops with the same behavior (Const of shape, concat) IF TensorRT is the conversion target. May cause perf issues
3. Choose fixed image dims for the input and reconvert the model using those dims. I haven't tried it, but the calculations should fold away into constants.

[adjagetia]

Solved it using onnx-simplifier