Added the ability to extrapolate ZeroPadding1D, ZeroPadding2D, and ZeroPadding3D.

[PINTO0309]

Issue Type

Feature Request

OS

Ubuntu

OS architecture

x86_64, aarch64, armv7, armv6, Other

Programming Language

C++, Python

Framework

OpenVINO, PyTorch, ONNX, TensorFlow, TensorFlowLite, TensorFlow.js, TensorRT, TF-TRT, CoreML, Myriad Inference Engine

Download URL for ONNX / OpenVINO IR

• ResNet-18 2D - TransposeConv3D

1. https://github.com/PINTO0309/PINTO_model_zoo/tree/main/202_stereoDNN
2. https://github.com/NVIDIA-AI-IOT/redtail/tree/master/stereoDNN

Description

Add the ability to extrapolate ZeroPadding layers with the --weight_replacement_config option.

Relevant Log Output

ERROR: Exception encountered when calling layer "tf.math.add_63" (type TFOpLambda).

Dimensions must be equal, but are 65 and 66 for '{{node tf.math.add_63/Add}} = AddV2[T=DT_FLOAT](Placeholder, Placeholder_1)' with input shapes: [1,65,129,64], [1,66,130,64].

Call arguments received:
  • x=tf.Tensor(shape=(1, 65, 129, 64), dtype=float32)
  • y=tf.Tensor(shape=(1, 66, 130, 64), dtype=float32)
  • name=None
ERROR: model_path  : saved_model_resnet18_2d_257x513/openvino/FP32/resnet18_2d_257x513.xml
ERROR: weights_path: saved_model_resnet18_2d_257x513/openvino/FP32/resnet18_2d_257x513.bin
ERROR: layer_id    : 309
ERROR: input_layer0 layer_id=288: KerasTensor(type_spec=TensorSpec(shape=(1, 65, 129, 64), dtype=tf.float32, name=None), name='tf.nn.elu_38/Elu:0', description="created by layer 'tf.nn.elu_38'")
ERROR: input_layer1 layer_id=308: KerasTensor(type_spec=TensorSpec(shape=(1, 66, 130, 64), dtype=tf.float32, name=None), name='tf.math.add_60/Add:0', description="created by layer 'tf.math.add_60'")
ERROR: The trace log is below.
Traceback (most recent call last):
  File "/usr/local/bin/openvino2tensorflow", line 1019, in convert
    tf_layers_dict[layer_id] = tf.math.add(tmp_layers[0], tmp_layers[1])
  File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/util/traceback_utils.py", line 153, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/usr/local/lib/python3.8/dist-packages/keras/layers/core/tf_op_layer.py", line 107, in handle
    return TFOpLambda(op)(*args, **kwargs)
  File "/usr/local/lib/python3.8/dist-packages/keras/utils/traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
ValueError: Exception encountered when calling layer "tf.math.add_63" (type TFOpLambda).

Dimensions must be equal, but are 65 and 66 for '{{node tf.math.add_63/Add}} = AddV2[T=DT_FLOAT](Placeholder, Placeholder_1)' with input shapes: [1,65,129,64], [1,66,130,64].

Call arguments received:
  • x=tf.Tensor(shape=(1, 65, 129, 64), dtype=float32)
  • y=tf.Tensor(shape=(1, 66, 130, 64), dtype=float32)
  • name=None

Source code for simple inference testing code

No response