Converting TF1 .meta weights to onnx yields to ONNX parsing error.

[MargauxMasson]

Describe the bug Converting TF1 .meta weights to ONNX yields to ONNX parsing error.

Urgency Moderate

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): NAME="Ubuntu" VERSION="20.04.3 LTS (Focal Fossa)" ID=ubuntu ID_LIKE=debian PRETTY_NAME="Ubuntu 20.04.3 LTS" VERSION_ID="20.04"
• Tensorflow Version: 1.15.5 // also tried with tf 2.4.0
• Python version: Python 3.8.10
• Using tensorflow=1.15.5, onnx=1.10.2, tf2onnx=1.10.0/5cd3b5, tensorrt=8.0.3.4

To Reproduce Describe steps/code to reproduce the behavior. Please upload/link the model you are trying to convert if possible.

python -m tf2onnx.convert --checkpoint path-to-model.meta --output model-onnx-opset15-tf1-test.onnx --inputs inputs/input_img:0 --outputs resnet18-linknet/logits/BiasAdd:0 --opset 15 && polygraphy surgeon sanitize model-onnx-opset15-tf1-test.onnx --fold-constants --output model-onnx-opset15-tf1-test-folded.onnx && polygraphy run model-onnx-opset15-tf1-test-folded.onnx --trt

The input shape should be (512, 1024, 8) and the output shape is (512, 1024, 53). It is a semantic segmentation model (LinkNet). All the steps work smoothly until trying to run the ONNX model.

Error (same with opsets 15,14,13,12):

[TensorRT] WARNING: onnx2trt_utils.cpp:362: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32. [TensorRT] ERROR: ModelImporter.cpp:720: While parsing node number 570 [Neg -> "Neg__1238:0"]: [TensorRT] ERROR: ModelImporter.cpp:721: --- Begin node --- [TensorRT] ERROR: ModelImporter.cpp:722: input: "Sub__1222:0" output: "Neg__1238:0" name: "Neg__1238" op_type: "Neg"

[TensorRT] ERROR: ModelImporter.cpp:723: --- End node --- [TensorRT] ERROR: ModelImporter.cpp:725: ERROR: onnx2trt_utils.cpp:2031 In function unaryHelper: [8] Assertion failed: validUnaryType && "This version of TensorRT does not support the given operator with the given input data type." [E] In node 570 (unaryHelper): UNSUPPORTED_NODE: Assertion failed: validUnaryType && "This version of TensorRT does not support the given operator with the given input data type." [!] Could not parse ONNX correctly

I tried with TF1 and TF2 but both end up with this error.

Any help would be appreciated as converting these TF1 weights to ONNX is a crucial step to switch to using ONNX/TensorRT inference.

Thank you!

[fatcat-z]

TensorRT doesn't support INT64. Is it possible to change the weights from INT64 to INT32 for your case?

[fatcat-z]

It's been over 3 months, so closing this. Feel free to open a new one if the issue still exists.