Paddle2ONNX v0.9.4

Experimental Feature

• A new version of Paddle2ONNX is integrated in v0.9.4, which will make conversion faster and also optimize the converted onnx model. You can enable the new feature with flag --enable_dev_version=True #669

Bug Fix

• Fix bug of shape inference while exporting OCR model #656
• Fix bug of box_coder under some onnx opset_version #667
• Fix bug of unsqueeze under some onnx opset_version #652

New Operators

• Support mish、eye and one_hot_v2 #663 #655

Paddle2ONNX v0.9.2

New Operators and Functions

• Add enable_auto_update_opset choice, default is True, related pull request #635
• flip related pull request #594
• set_value related pull request #634
• unfold related pull request #647

Bug Fix

• Update the documentation of OP's development guide . #602
• Fix fill_any_like convert bugs, related pull request #624
• Fix prelu convert bugs, related pull request #626
• Fix unsqueeze convert bugs, related pull request #637
• Fix box_coder 、multiclass_nms 、yolo_box 、softmax_with_cross_entropy 、unstack 、expand_as_v2 convert bugs, related pull request #639

Paddle2ONNX v0.9.1

New Feature

• Update the implementation of dynamic to static . #394

New Operators and Functions

• tan、softshrink、 tanh_shrink、 thresholded_relu、 unique、 where、 masked_select、bicubic_interp_v2、scatter、scatter_nd_add、meshgrid related pull request #392
• linspace related pull request #470
• gausian_random related pull request #478
• gather_nd related pull request #479
• Roll related pull request #513

Bug Fix

• Fix lookup_table_v2、elementwise_mod、isnan convert bugs, related pull request #423
• Fix clip convert bugs, related pull request #436
• Fix reduce_all、reduce_any convert bugs, related pull request #440
• Fix flatten_contiguous_range convert bugs, related pull request #441
• Fix reduce_min/max convert bugs, related pull request #442
• Fix batch_norm convert bugs, related pull request #449
• Fix gather 、matmul_v2 convert bugs, related pull request #450
• Fix logsoftmax 、softmax 、logsumexp convert bugs, related pull request #453
• Fix squeeze convert bugs, related pull request #454
• Fix mean convert bugs, related pull request #459
• Fix pad2d 、pad3d convert bugs, related pull request #460
• Fix Pow 、prelu convert bugs, related pull request #463
• Fix multiclass_nms convert bugs, related pull request #467
• Fix thresholded_relu convert bugs, related pull request #477
• Fix softmax_with_cross_entropy convert bugs, related pull request #483
• Fix argsort convert bugs, related pull request #486
• Fix concat 、size 、stack convert bugs, related pull request #487
• Fix split convert bugs, related pull request #491
• Fix slice/stride_slice convert bugs, related pull request #494
• Fix uniform_random convert bugs, related pull request #495
• Fix Squeeze convert bugs, related pull request #496
• Fix unsqueeze convert bugs, related pull request #497
• Fix Expand_v2 convert bugs, related pull request #498
• Fix concat convert bugs, related pull request #510
• Fix Unstack convert bugs, related pull request #514
• Fix flatten_contiguous_range convert bugs, related pull request #532
• Fix matmul 、split convert bugs, related pull request #563
• Fix pool2d convert bugs, related pull request #576

Contributors

• Thanks to @AndPuQing for contributing 11 OPs in this release.

Paddle2ONNX v0.9

New Feature

• Support define input shape dict. #383

New Operators and Functions

• isfinite_v2、acos、asin、atan、ceil、cos、cosh、arg_min、elu 、hard_shrink related pull request #357
• logical_not、logical_or、erf、isinf_v2、isnan、floor_mod、isnan_v2、less_than 、log2、logsumexp related pull request #358
• selu、logical_xor、log10、log1p、reduce_all、reduce_any、round、sin、rsqrt 、sign、logsigmoid related pull request #362
• sinh、brelu、mv、dot、dist、softsign、log_softmax、argsort、index_select 、size、pad1d、hardtanh related pull request #363
• conv3d 、pool3d related pull request #405
• Add onnx and onnxruntime version requirement prompts. #335
• Add the op development guide for Paddle2ONNX. #368
• Add nms plugin for openvino. #386

Bug Fix

• pool2d operator, related pull request #330
• resize operator, related pull request #348
• pad3 operator, related pull request #349
• concat operator, related pull request #377
• argmax operator, Fix bugs in argmax op when dtype is int32, related pull request #393
• padding operator, related pull request #400
• unsqueeze2 operator, related pull request #405
• When exporting onnx, an error occurred when including a register_buffer. related pull request #356

Contributors

• Thanks to @AndPuQing for contributing 43 OPs in this release.
• Thanks to @SamDM for solving the problem of exporting models with register_buffer to ONNX in this release.
• Thanks to @franva for suggesting and fixing some small bugs for us in this release.

Paddle2ONNX v0.8

New Feature

• Automatic correction of opset version
• Support the fallback selection of custom op

New Operator Support

• softplus related pull request #252
• elementwise_mod
• elementwise_floordiv
• p_norm
• depthwise_conv2d_transpose
• group_norm
• top_k related pull request #293

Bug Fix

• clip operator, related pull request #280
• fill_constant_batch_size_like operator, related pull request #318
• multiclass_nms operator, fix issue of #315, related pull request #318
• top_k_v2 operator, related pull request #252
• expand_v2 operator, related pull request #252
• assign_value operator, related pull request #252
• slice operator, related pull request #268
• sum operator, related pull request #293

Paddle2ONNX v0.7

Bug Fix

• Resize operator, fix issue of #241, related pull request #272
• 'PRelu` operator, related pull request #266

New Operator Support

• PixelShuffle operator, fix issue of #260, related pull request #261

Paddle2ONNX v0.6

New Model Support

Support 2.0 model toolkit

• PaddleDetection: 8 models structure supported
• PaddleSeg: All the models are supported
• PaddleClas: All the models are supported
• PaddleOCR: Server and mobile ocr models are supported

New operators Support

• softplus
• elementwise_mod
• elementwise_floordiv
• p_norm
• depthwise_conv2d_transpose
• group_norm

Bug Fix

• fill_constant_batch_size_like
• multiclass_nms
• top_k_v2
• expand_v2
• assign_value
• deformable_conv
• prelu

paddle2onnx 0.5

This upgrade is mainly to support more official featured models, including PPYOLO, PPOCR, FasterRCNN, ERNIE and ViT. In addition, we have a more comprehensive coverage test for PaddleSeg and PaddleClas . For more detailed upgrade, refer to the following list:

• Support PPYOLO, FasterRCNN, PPOCR, and ERNIE. Full support to convert PaddleClas models, including ViT (Vision Transformer). Please refer to docs/en/model_zoo.md
• Optimize the prompt of OpMapper support status.
• fix bug the incorrect input order of the model, such as yolov3.
• When pool is adative and ksize is [1,1], use gobal_pool to map adative_pool.
• Add a lot of English documents.
• Add tutorial for converting dynamic computational graph to ONNX.

paddle2onnx 0.4

• support paddle dygraph model convert to ONNX.
• add op_mapper register function
• support convert Paddle model(not combined)
• add simple tutorial
• add op_list

Paddle - ONNX Convertor

Now support for convert Paddle Model to ONNX Opset 9, 10, 11. please use --onnx_opset to set opset_version of onnx you want to convert.

Add support for Paddle operators:

• swish
• floor
• uniform_random
• abs
• instance_norm
• clip
• tanh
• log
• pad2d
• prelu