Onnx2tf

- Corrected the decision condition to enable `DepthwiseConv2D`
- [[japan_PP-OCRv3] For min_value and max_value in Clip.py 14](https://github.com/PINTO0309/onnx2tf/issues/14)
- [japan_PP-OCRv3_rec_infer.onnx.zip](https://github.com/PINTO0309/onnx2tf/files/10474146/japan_PP-OCRv3_rec_infer.onnx.zip)

Solves the problem of TensorFlow runtime abort during inference when `DepthwiseConv2D` is generated with different values of strides in the row and column directions, such as `[1,2]` or `[2,1]` strides. `DepthwiseConv2D` seems to work correctly only when all strides have the same value. In this case, `GroupConvolution` is generated.

- e.g. TensorFlow (Keras) - `DepthwiseConv2D` error message

Current implementation only supports equal length strides in the row and column dimensions. [Op:DepthwiseConv2dNative]

Call arguments received by layer "tf.compat.v1.nn.depthwise_conv2d_3" " f"(type TFOpLambda):
• input=tf.Tensor(shape=(3, 18, 82, 64), dtype=float32)
• filter=tf.Tensor(shape=(3, 3, 64, 1), dtype=float32)
• strides=['1', '2', '1', '1']
• padding='VALID'
• rate=['1', '1']
• name=None
• data_format=None
• dilations=None


What's Changed
* Corrected the decision condition to enable `DepthwiseConv2D` by PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/144


**Full Changelog**: https://github.com/PINTO0309/onnx2tf/compare/1.5.27...1.5.28

1. Improved full validation algorithm
2. Problem solved where validation could fail for a sequence of tensors of the same shape.
[Bug in onnx_tf_tensor_validation 139](https://github.com/PINTO0309/onnx2tf/issues/139)
4. Validation speed is now approximately five times faster.
5. The more complex the structure of the model, the greater the improvement in validation speed.
6. When converting `pidnet_S_cityscapes_192x320.onnx`, I noticed that the output size of `AveragePool` is different between ONNX and TensorFlow, but I will separate the correspondence from this pull request. Maybe this pull request will improve it. [MaxPool padding insertion logic update 138](https://github.com/PINTO0309/onnx2tf/pull/138)

INFO: onnx_op_type: AveragePool onnx_op_name: /spp/scale2/scale2.0/AveragePool
INFO: input_name.1: /layer5/layer5.1/Add_output_0 shape: [1, 512, 3, 5] dtype: float32
INFO: output_name.1: /spp/scale2/scale2.0/AveragePool_output_0 shape: [1, 512, 1, 2] dtype: float32
INFO: tf_op_type: AveragePooling2D
INFO: input.1.x: name: tf.math.add_81/Add:0 shape: (1, 3, 5, 512) dtype: <dtype: 'float32'>
INFO: input.2.pool_size: val: [9, 9]
INFO: input.3.strides: val: [4, 4]
INFO: input.4.padding: val: [[0, 0], [4, 4], [4, 4], [0, 0]]
INFO: output.1.output: name: tf.concat_1/concat:0 shape: (1, 2, 2, 512) dtype: <dtype: 'float32'>

7. Changed `Skipped` indication to `Skipped (Deleted or Shape Unmatched)`.

- sample

onnx2tf -i pidnet_S_cityscapes_192x320.onnx -cotof -cotoa 1e-4

- Before
- After

What's Changed
* Improved full validation algorithm by PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/143


**Full Changelog**: https://github.com/PINTO0309/onnx2tf/compare/1.5.26...1.5.27

- Improved error messages
- Before
- After

What's Changed
* Improved error messages by PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/142


**Full Changelog**: https://github.com/PINTO0309/onnx2tf/compare/1.5.25...1.5.26

- Fix `np.ndarray` of weights to be converted to tensor with `tf.convert_to_tensor`
- This modification significantly reduces the size bloat of the generated `saved_model` Protocol Buffer file.

What's Changed
* Fix `np.ndarray` of weights to be converted to tensor with `tf.convert_to_tensor` by PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/141


**Full Changelog**: https://github.com/PINTO0309/onnx2tf/compare/1.5.24...1.5.25

- Reduces the frequency of `MatMul` shape unmatch errors
- [When Onnx Matmul inputs have different dimension 133](https://github.com/PINTO0309/onnx2tf/issues/133)
- The next development step will utilize dummy inference to automatically correct for inference errors.

What's Changed
* Reduces the frequency of `MatMul` shape unmatch errors by PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/140


**Full Changelog**: https://github.com/PINTO0309/onnx2tf/compare/1.5.23...1.5.24

- Improve processing of `Gather` to eliminate `tf.where`
- [[MobileBERT] Add --optimization_for_gpu_delegate option 132](https://github.com/PINTO0309/onnx2tf/issues/132)
- [I cannot run a transformer model with token-level output on accelerated hardware in TF Lite 59232](https://github.com/tensorflow/tensorflow/issues/59232)

- Before
- After