tf2onnx-1.9.3

Summary

• added support for opset-15
• added support for tf-2.7
• misc fixes

Changes since tf2onnx-1.9.2

• remove 2.6rc0 override by @guschmue in #1683
• Implement GRU rewriter for tf2 by @TomWildenhain-Microsoft in #1688
• Fix bug in reverseV2 for 1D tensors by @TomWildenhain-Microsoft in #1691
• Handle Abs in TransposeOptimizer by @mateusztabaka in #1699
• Fix tf1 loop variable parsing for duplicated Enter nodes by @TomWildenhain-Microsoft in #1697
• Tom/keras gru by @TomWildenhain-Microsoft in #1692
• Implement symmetric padding mode by @TomWildenhain-Microsoft in #1698
• Fix loop optimizer when scan output is also loop-carried-dependent by @TomWildenhain-Microsoft in #1701
• Fix bug in reading hash tables in freeze_session by @TomWildenhain-Microsoft in #1709
• Increase unit test timeout for windows job that often fails by @TomWildenhain-Microsoft in #1711
• Fix RandomUniform/RandomNormal rewriters for non-const inputs by @TomWildenhain-Microsoft in #1710
• Fix usage of custom_ops, custom_op_handlers, and custom_rewriter ags by @TomWildenhain-Microsoft in #1708
• Fix nightly tests by @TomWildenhain-Microsoft in #1712
• Fix nightly (explicit_pads not supported in old tf versions) by @TomWildenhain-Microsoft in #1717
• Tom/tfjs bugfixes by @TomWildenhain-Microsoft in #1723
• Pin onnxruntime-extensions until issue is fixed by @TomWildenhain-Microsoft in #1733
• hardswish support for tflite by @guschmue in #1735
• add opset-15 by @guschmue in #1731
• Optimize einsum when node[1].is_const() by @xadupre in #1739
• replace depreciated docker image by @guschmue in #1750
• Add Keras LSTM support by @q-ycong-p in #1752
• Bug-fix to skip keras LSTM test for TF2.1 by @q-ycong-p in #1754
• tf-2.6.0 -> tf-2.6.1 to avoid build issues by @guschmue in #1759
• Fix the Broken CI by @leimao in #1761
• Fix TF1-Keras Dilated Conv ONNX Exported by @leimao in #1744
• tf-2.7 by @guschmue in #1762
• #1763: fix output shape of flatten rewriter by @guschmue in #1766
• #1749: Fix fused_batch_norm 5d NDHWC input reshape convert by @hwangdeyu in #1769
• [Test ONNX 1.10.2] Verify whether 1.10.2 from TestPyPI solve the crash issue by Shape-15 by @jcwchen in #1753

Full Changelog: v1.9.2...v1.9.3

A huge thank you to our community for this release!

leimao,mateusztabaka,q-ycong-p

tf2onnx-1.9.2

Summary

• support for tf-2.6 (required some change for models >2GB and tflite)
• experimental support for tensorflow.js
• support for tf.signal.rfft
• we added all unit tests from keras-onnx to prevent regressions when transitioning from keras-onnx to tf2onnx
• misc fixes

Changes since tf2onnx-1.9.1

• tf-2.6 to ci, reflect tf-2.6 in readme 1680 (guschmue)
• Fix issues with tflite on tf2.6 1681 (TomWildenhain-Microsoft)
• Don't copy/split transpose unless further optimization is likely 1679 (TomWildenhain-Microsoft)
• Implement RaggedTensorToTensor for tensors with dense (uniform) dims 1676 (TomWildenhain-Microsoft)
• Fix conversion of OneHot for dtypes unsupported by ORT 1675 (TomWildenhain-Microsoft)
• Fix bug in topk casts affecting other input consumers 1674 (TomWildenhain-Microsoft)
• Fix Where/Select op for bool dtypes 1673 (TomWildenhain-Microsoft)
• Fix freezing of ResourceGather op 1672 (TomWildenhain-Microsoft)
• Fix script about benchmarking humpback_whale model 1671 (xadupre)
• Implement VALUE_ROWIDS format for ragged to tensor 1664 (TomWildenhain-Microsoft)
• Improve detection of NaN values in Select op 1663 (TomWildenhain-Microsoft)
• tf-2.6.0rc2 1662 (guschmue)
• Fix keras nightly applications 1661 (TomWildenhain-Microsoft)
• Implement RandomShuffle op 1658 (TomWildenhain-Microsoft)
• Implement reduce all/any for non-const axes 1657 (TomWildenhain-Microsoft)
• Improve lookup table loading for from_keras conversion 1654 (TomWildenhain-Microsoft)
• Fix implementation of skip_tf_versions 1653 (TomWildenhain-Microsoft)
• Fix bug in einsum optimizer reading attributes 1652 (TomWildenhain-Microsoft)
• Update RFFT2D to handle 3D arrays 1649 (xadupre)
• support for AsString, StringToNumber 1648 (guschmue)
• Handle missing tensorflowjs package in unit tests 1647 (TomWildenhain-Microsoft)
• Update script to test tflite models from tfhub 1644 (xadupre)
• add provider to run_pretrained_models 1642 (guschmue)
• add inception_v4 1641 (guschmue)
• Skip onnx checker for new onnx version for IR3 bug 1638 (TomWildenhain-Microsoft)
• [WIP] Fix tfhub scripts (enformer, ...) 1637 (xadupre)
• Skip onnx checker for new onnx version for IR3 bug 1636 (TomWildenhain-Microsoft)
• Move check earlier to fix loop in transpose optimizer 1631 (TomWildenhain-Microsoft)
• Add PRelu to Transpose optimizer 1630 (TomWildenhain-Microsoft)
• Fix TFJS conversion for old TFJS models 1629 (TomWildenhain-Microsoft)
• Add conversion of _FusedDepthwiseConv2dNative and Prelu for TFJS 1628 (TomWildenhain-Microsoft)
• Add variable to make it possible to test weekly onnx package 1627 (TomWildenhain-Microsoft)
• Fix tfhub scripts checking discrepencies on resnet 1626 (xadupre)
• Fix tfjs quantized weight parsing for fp16 quant 1625 (TomWildenhain-Microsoft)
• Tests for onnx 1.9.101 1624 (TomWildenhain-Microsoft)
• Add pretrained model tests for tfjs 1622 (TomWildenhain-Microsoft)
• RFFT2D for TFLite 1621 (TomWildenhain-Microsoft)
• Implement fused tfjs ops 1620 (TomWildenhain-Microsoft)
• Tom/fused tfjs ops 1619 (TomWildenhain-Microsoft)
• Trim down regular keras2onnx tests 1618 (TomWildenhain-Microsoft)
• Fix Roll op for negative shifts 1616 (TomWildenhain-Microsoft)
• Add TFJS conversion 1614 (TomWildenhain-Microsoft)
• Work on tfjs support [WIP] 1612 (TomWildenhain-Microsoft)
• Fix nightly build by moving tf2.4.1 to Python3.7 1611 (TomWildenhain-Microsoft)
• Make keras2onnx tests deterministic 1608 (TomWildenhain-Microsoft)
• Make keras2onnx tests use templates 1605 (TomWildenhain-Microsoft)
• fix state in azuredevops 1604 (guschmue)
• Move unit tests from regular to nightly 1602 (TomWildenhain-Microsoft)
• Refactor tflite/tf logic out of tfonnx into tf_utils and tflite_utils 1599 (TomWildenhain-Microsoft)
• Add windows keras2onnx applications tests 1598 (TomWildenhain-Microsoft)
• Fix converter for rfft when fft_length < shape[1], implements fft2d 1596 (xadupre)
• Refactor tfonnx.py 1595 (TomWildenhain-Microsoft)
• Add script to benchmark bert and mobilenet 1594 (xadupre)
• Move keras2onnx unit tests inside tests dir 1591 (TomWildenhain-Microsoft)
• Tom/keras2onnx application tests 1590 (TomWildenhain-Microsoft)

A huge thank you to our contributors for this release!

tf2onnx-1.9.1

Summary

Fixes, support for opset-14, improved keras support

Changes since 1.8.5

• Implement Round for lower opsets 1589 (TomWildenhain-Microsoft)
• Fix bug in CropAndResize for empty tensors 1588 (TomWildenhain-Microsoft)
• Add yamnet to the list of tfhub models tested 1587 (xadupre)
• fix opset typo 1585 (guschmue)
• Implement rewriter for LSTM nodes in tf2 1584 (TomWildenhain-Microsoft)
• add opset-14, move ci pipeline to test opset-8 -> opset-14 1583 (guschmue)
• Add pure keras unit tests for Windows 1582 (TomWildenhain-Microsoft)
• Add target for converting to channels last format 1581 (TomWildenhain-Microsoft)
• Adds scripts to benchmark tfhub models and check discrepencies 1580 (xadupre)
• Removes uses of onnxruntime when optimizing einsum 1579 (xadupre)
• Add Windows keras2onnx tests 1578 (TomWildenhain-Microsoft)
• Fix issue with tf1 loops with tensor array read last pattern 1577 (TomWildenhain-Microsoft)
• add pure keras tests from keras2onnx 1574 (TomWildenhain-Microsoft)
• Fix timeout issues on nightly 1572 (TomWildenhain-Microsoft)
• Make keras2onnx tests deterministic 1571 (TomWildenhain-Microsoft)
• Add warning for non-cpp protobuf 1570 (TomWildenhain-Microsoft)
• Const fold Ifs and remove keras_learning_phase inputs 1569 (TomWildenhain-Microsoft)
• Fix nightly CI 1568 (TomWildenhain-Microsoft)
• Misc fixes for adding keras2onnx tests 1567 (TomWildenhain-Microsoft)
• Fix link to tutorials/keras-resnet50.ipynb in README.md 1566 (iizukak)
• Fix nightly pipeline for tf1.13 1565 (TomWildenhain-Microsoft)
• Fix nightly build for tf2.1 1564 (TomWildenhain-Microsoft)
• Added tests from keras2onnx 1563 (TomWildenhain-Microsoft)
• Prevent tf1 loop ops from const folding 1561 (TomWildenhain-Microsoft)
• Hack to support > 2GB tensors 1558 (TomWildenhain-Microsoft)
• Implement conversion of set intersection, union, difference 1556 (TomWildenhain-Microsoft)
• Const fold Concat 1554 (TomWildenhain-Microsoft)
• Add constant folding for Split op 1553 (TomWildenhain-Microsoft)
• Fix Conv groups for unknown input dims 1552 (TomWildenhain-Microsoft)
• Tom/ort 1.8 1551 (TomWildenhain-Microsoft)
• ort -> 1.8 1550 (guschmue)
• [DON'T MERGE] check pretrained models 1548 (xadupre)
• Replace onnxruntime_customops by onnxruntime_extensions 1547 (xadupre)
• Fix subtle bug in ConvTranspose 1542 (TomWildenhain-Microsoft)
• better way to get kwargs 1541 (guschmue)
• add script to run optimizer on onnx files 1538 (guschmue)
• Use structured inputs by default from CLI 1534 (TomWildenhain-Microsoft)
• Remove captured inputs when using CLI 1533 (TomWildenhain-Microsoft)
• Implement conversion of CTCGreedyDecoder 1530 (TomWildenhain-Microsoft)
• Made from_keras work for tf1 1529 (TomWildenhain-Microsoft)
• Fixed bug in transpose optimizer for mul of conv with multiple consumers 1528 (TomWildenhain-Microsoft)
• Add 1 more model and update script 1527 (TomWildenhain-Microsoft)
• Make rewriters use get_tensor when needed to avoid subtle bugs 1526 (TomWildenhain-Microsoft)
• Fix Conv for unknown input rank 1524 (TomWildenhain-Microsoft)
• Fix dtype in SegmentSum 1523 (TomWildenhain-Microsoft)
• Tom/more models 1522 (TomWildenhain-Microsoft)
• Convert EnsureShape to Identity 1521 (TomWildenhain-Microsoft)
• More group C models 1520 (TomWildenhain-Microsoft)
• Remove from future import ..., only support python 3 1519 (xadupre)
• Add support for register custom ops 1518 (zerollzeng)
• Fix Gemm rewriter for bias node with multiple outputs 1517 (TomWildenhain-Microsoft)
• more HF models 1516 (guschmue)
• Added more group A and C tfhub models 1515 (TomWildenhain-Microsoft)
• fix licence classifier 1514 (guschmue)
• Explicit padding for const shape ConvTranspose 1513 (calvinmccarter-at-lightmatter)
• Add hash table support and frozen graph repairing for from_keras 1512 (TomWildenhain-Microsoft)

A huge thank you to our contributors for this release!

zerollzeng,calvinmccarter-at-lightmatter,iizukak

tf2onnx-1.8.5

Summary

Fixes, support for tf-2.5, improved keras support

Changes since tf2onnx-1.8.4

• Pin flatbuffers version 1511 (TomWildenhain-Microsoft)
• Keep inputs of placeholder with default when removing unused nodes 1510 (TomWildenhain-Microsoft)
• Add option to run_pretrained_models to save ort/tf profile 1509 (TomWildenhain-Microsoft)
• add huggingface tutorial 1508 (guschmue)
• Implement conversion of RaggedToVariant and RaggedFromVariant in loops 1503 (TomWildenhain-Microsoft)
• Add getting started sample code 1502 (TomWildenhain-Microsoft)
• Implement DivNoNan conversion 1499 (TomWildenhain-Microsoft)
• Fix bug in conv rewriter if intermediate node outputs are used 1498 (TomWildenhain-Microsoft)
• Add optimizer to push q and dq ops into place 1497 (TomWildenhain-Microsoft)
• Fix bug in transpose optimizer where node is deleted twice 1496 (TomWildenhain-Microsoft)
• Remove double reshape 1495 (TomWildenhain-Microsoft)
• Unit test for from_function with non-tensor inputs in spec 1492 (TomWildenhain-Microsoft)
• Filter structured inputs to only include TensorSpecs 1490 (TomWildenhain-Microsoft)
• Improve keras output order detection 1489 (TomWildenhain-Microsoft)
• Correct min opset of CombinedNMS to be opset 12 1488 (TomWildenhain-Microsoft)
• Add hack for legacy keras 1486 (TomWildenhain-Microsoft)
• Implement RandomStandardNormal conversion 1484 (TomWildenhain-Microsoft)
• Fix bug with BatchMatMulV2 when it has the same input twice 1483 (TomWildenhain-Microsoft)
• add huggingface/electra 1482 (guschmue)
• Fix errors in tf freezing with AssignVariableOp 1481 (TomWildenhain-Microsoft)
• Reorder keras outputs if possible 1480 (TomWildenhain-Microsoft)
• Fix bug in while shapes 1479 (TomWildenhain-Microsoft)
• Fix bug that renamed subgraph i/o twice 1478 (TomWildenhain-Microsoft)
• Improve keras API 1477 (TomWildenhain-Microsoft)
• Fix Slice that is Identity 1476 (TomWildenhain-Microsoft)
• Fix Slice conversion for nodes with multiple outputs 1473 (TomWildenhain-Microsoft)
• Make conv dilations rewriter work for dynamic pads 1470 (TomWildenhain-Microsoft)
• add more huggingface models for testing 1469 (guschmue)
• Match SpaceToBatch->Conv->BatchToSpace pattern more generally 1468 (TomWildenhain-Microsoft)
• Skip unregistering optional inputs/outputs that are empty 1467 (TomWildenhain-Microsoft)
• Handle DecodeError in tf_loader 1466 (calvinmccarter-at-lightmatter)
• Add ReduceMean/ReduceMax -> GlobalAveragePool/GlobalMaxPool optimizer 1465 (TomWildenhain-Microsoft)
• Improve perf of SegmentSum 1463 (TomWildenhain-Microsoft)
• Tom/tf 2.5 1462 (TomWildenhain-Microsoft)
• Optimize SpaceToBatchND sequence to DepthwiseConv2dNative dilations 1461 (TomWildenhain-Microsoft)
• [WIP] add tf-2.5 to ci 1460 (guschmue)
• Tom/batch to space crop 1459 (TomWildenhain-Microsoft)
• Add support for Explicit padding 1458 (TomWildenhain-Microsoft)
• Switch Einsum to Matmul when 2nd inp is constant 1457 (TomWildenhain-Microsoft)
• Remove the OneHot from SparseSoftmaxCrossEntropyWithLogits for opset >= 11 1454 (TomWildenhain-Microsoft)
• Tom/enable transpose any rank 1453 (TomWildenhain-Microsoft)
• Check shapes and dtypes in optimizer tests 1452 (TomWildenhain-Microsoft)
• Fix MaxpoolWithArgmax 1451 (TomWildenhain-Microsoft)
• Enable string tests in CI pipeline 1450 (TomWildenhain-Microsoft)
• Fix TensorListStack and Loop shapes 1448 (TomWildenhain-Microsoft)
• Skip deepcopies for large models 1446 (TomWildenhain-Microsoft)
• Fix bug in casting Log inputs 1445 (TomWildenhain-Microsoft)
• respect target argument 1444 (guschmue)
• Add support for range_given attr of QuantizeDequantizeV2 1442 (TomWildenhain-Microsoft)
• Verify dtypes against onnx type inference 1441 (TomWildenhain-Microsoft)
• Tests for onnx package 1440 (TomWildenhain-Microsoft)
• Make transpose optimizer push through unsqueeze 1436 (TomWildenhain-Microsoft)
• add onnx conv2d op group attr 1435 (BBuf)
• Prepare to enable Transpose optimizer for transposes of any rank 1433 (TomWildenhain-Microsoft)
• Fix bug in reshape_optimizer and disable silent error catching in opt… 1431 (TomWildenhain-Microsoft)
• Fix random uniform rewriter for scalar shapes 1429 (TomWildenhain-Microsoft)
• Make unit tests aggressively verify graph shape information 1427 (TomWildenhain-Microsoft)
• Make unit tests verify graph shape information [1426](https://github.com/onnx/tensor...

tf2onnx-1.8.4

Summary

• new, simplified python api (see here)
• support for structured outputs so output names of keras model are the same as keras output names
• examples for tflite and keras

Changes since tf2onnx-1.8.3

• raise error if signature is missing 1406 (guschmue)
• Added conversion for DenseBincount 1401 (TomWildenhain-Microsoft)
• fix broken link 1397 (guschmue)
• Fix raw_data to no longer have __EXTERNAL when data stored externally 1392 (TomWildenhain-Microsoft)
• [WIP] Support for WordpieceTokenizer (part of Bert) 1390 (xadupre)
• Added debugging advice to GitHub issues template 1389 (TomWildenhain-Microsoft)
• Added TensorRT as a target and fixed Resize op conversion for TensorRT 1388 (TomWildenhain-Microsoft)
• Fixed bug in transpose optimizer for Mul when rank=5 1387 (TomWildenhain-Microsoft)
• Enhance BatchToSpaceND and SpaceToBatchND handlers 1386 (mateusztabaka)
• Fixed bug in depthwise conv when kernel is shared between two nodes 1384 (TomWildenhain-Microsoft)
• Add argmax and reduction ops to transpose optimizer 1383 (TomWildenhain-Microsoft)
• Fix bug where unknown shape overwrites loop iter and cond shapes 1382 (TomWildenhain-Microsoft)
• Allow tflite to cut graph if inputs/outputs are provided 1381 (TomWildenhain-Microsoft)
• Added rewriter for specific pattern in a 1P model 1380 (TomWildenhain-Microsoft)
• Tom/combined non max suppression 1376 (TomWildenhain-Microsoft)
• Tom/tflite flexops 1374 (TomWildenhain-Microsoft)
• Bug with Dequantize with no quantization params 1373 (TomWildenhain-Microsoft)
• Convert tflite PRelu op 1372 (TomWildenhain-Microsoft)
• Add partial support for operator FFT 1371 (xadupre)
• cleanup huggingface tests 1370 (guschmue)
• Fix parsing of tflite string tensors and other tflite fixes 1369 (TomWildenhain-Microsoft)
• fix nightly ci - large model support needs 2.1 or better 1368 (guschmue)
• Improve reliability of reading tflite models 1366 (TomWildenhain-Microsoft)
• add int support for pool ops 1365 (guschmue)
• fix python api issue for large models 1363 (guschmue)
• add unittest for --tflite command line 1362 (guschmue)
• Tom/more co models 1361 (TomWildenhain-Microsoft)
• fix --tflite command line 1360 (guschmue)
• fix --tflite command line 1359 (guschmue)
• Added more models for Co 1358 (TomWildenhain-Microsoft)
• Tests to validate ORT release 1.7.0.dev202102184 1357 (TomWildenhain-Microsoft)
• Add flags for cutting/renaming IO 1355 (TomWildenhain-Microsoft)
• Fix RandomUniformInt for scalar values 1354 (TomWildenhain-Microsoft)
• new api 1353 (guschmue)
• First step to rename IO 1352 (TomWildenhain-Microsoft)
• Handle Mul broadcastable const inputs other than (1,) or (N,) in Tran… 1351 (mateusztabaka)
• Full conversion of MatrixBandPart 1350 (TomWildenhain-Microsoft)

A huge thank you to our contributors for this release!

mateusztabaka

tf2onnx-1.8.3

Summary

Final touches for tflite support, support for more tf ops, fixes

Changes since 1.8.2

• Convert RandomUniformInt 1347 (TomWildenhain-Microsoft)
• Implement conversions for AdjustSaturation and AdjustContrastv2 1346 (TomWildenhain-Microsoft)
• Fix typo in README.md 1345 (mrry)
• onnx requires all inputs for bn to be the same T 1344 (guschmue)
• Implement conversions of StringUpper and StringLower 1343 (TomWildenhain-Microsoft)
• Handle ReduceSum in TransposeOptimizer 1342 (mateusztabaka)
• Use simple_handler for Reciprocal,Sqrt in TransposeOptimizer 1340 (mateusztabaka)
• Handle Mul as Square in TransposeOptimizer 1339 (mateusztabaka)
• fix tf.roll with axes=-1 1338 (guschmue)
• Add models for Co semester 1336 (TomWildenhain-Microsoft)
• Fix make_const for strings 1335 (TomWildenhain-Microsoft)
• Implement conversion of RaggedGather 1333 (TomWildenhain-Microsoft)
• Implement RaggedTensorToTensor conversion 1332 (TomWildenhain-Microsoft)
• Optimize Transpose->Pad when 'pads' input is not a constant 1331 (mateusztabaka)
• Optimize Transpose->Slice regardless of Slice's axes 1330 (mateusztabaka)
• Add support for GatherV2 batch_dims attr 1329 (TomWildenhain-Microsoft)
• Fix bug in run_pretrained_models for large models 1327 (TomWildenhain-Microsoft)
• Make merge optimizer faster 1326 (TomWildenhain-Microsoft)
• Added more tflite models to pretrained tests 1325 (TomWildenhain-Microsoft)
• Disambiguate a unit test method 1323 (xadupre)
• Tom/tflite dequantize 1322 (TomWildenhain-Microsoft)
• Implement const dequantize pushing for per-axis dequantization 1321 (TomWildenhain-Microsoft)
• Fix bug in depthwise conv dilations with padding=same 1320 (TomWildenhain-Microsoft)
• fix StridedSlice for ellipsis+newaxis 1319 (guschmue)
• Implement non-standard nms for tflite 1318 (TomWildenhain-Microsoft)
• Added optimization for dequantized constant folding 1316 (TomWildenhain-Microsoft)
• Make TransposeOptimizer to optimize 5D transposes 1315 (mateusztabaka)
• Extend transpose optimizer for quantize and dequantize 1314 (TomWildenhain-Microsoft)
• Add tflite support to run_pretrained_models.py 1313 (TomWildenhain-Microsoft)
• Fixed bug in shapes for tfltie postprocess 1312 (TomWildenhain-Microsoft)
• Make tflite consider [] to designate unknown rank 1311 (TomWildenhain-Microsoft)
• Tom/sentencetokenize 1309 (TomWildenhain-Microsoft)
• Add conversion for SentencepieceTokenizeOp 1308 (TomWildenhain-Microsoft)
• Fix common bug where node inputs are assumed to come from output[0] 1307 (TomWildenhain-Microsoft)
• Fix bug in raggedToSparse and add casts to Unique for non-int64 ints 1306 (TomWildenhain-Microsoft)

A huge thank you to our contributors for this release!

mateusztabaka

tf2onnx-1.8.2

Summary

Fixes, initial support for tflite->onnx conversion

Changes since 1.8.1

• test_unsorted_segment_ops needs tf-2.3 or better 1305 (guschmue)
• Tests for tflite postprocess 1302 (TomWildenhain-Microsoft)
• Allow -1 to indicate unknown dimensions for shape override 1300 (TomWildenhain-Microsoft)
• Support for bicubic resize 1299 (TomWildenhain-Microsoft)
• Enabled tflite support and tests 1297 (TomWildenhain-Microsoft)
• use test.pypi.org via ONNX_TEST env variable 1296 (guschmue)
• Added conversion of TFLite_Detection_PostProcess 1293 (TomWildenhain-Microsoft)
• Fix 2 bugs to avoid extranious warnings 1292 (TomWildenhain-Microsoft)
• trim ci pipeline 1291 (guschmue)
• Optimization for tflite loops 1289 (TomWildenhain-Microsoft)
• transpose optimizer wrongly removes a transpose 1288 (guschmue)
• Added support for mutable hash tables and fixed bug in making string … 1282 (TomWildenhain-Microsoft)
• Added code for parsing tflite customops 1281 (TomWildenhain-Microsoft)
• Fixed bug in tflite relu6 1280 (TomWildenhain-Microsoft)
• Implemented conversion of RaggedToSparse 1276 (TomWildenhain-Microsoft)
• Minor bugfixes to support tflite 1275 (TomWildenhain-Microsoft)
• Moved tflite folder to tf2onnx 1273 ([TomWildenhain-Microsoft](ht
  tps://github.com/TomWildenhain-Microsoft))
• [LICENSE] remove leftover MIT licenses 1272 (szha)
• [LICENSE] add ALv2 licenses 1271 (szha)
• Added handlers for tflite 1270 (TomWildenhain-Microsoft)
• updated supported ops doc 1269 (guschmue)
• Added support for PlaceholderWithDefault ops with computed defaults 1268 (TomWildenhain-Microsoft)
• Added @tfl_op decorator to handlers 1267 (TomWildenhain-Microsoft)

A huge thank you to our contributors for this release!

szha

tf2onnx-1.8.1

Summary

opset-13 support, bug fixes

Changes since tf2onnx-1.7.2

• Created tflite utils for parsing tflite graphs 1265 (TomWildenhain-Microsoft)
• Fix issue with LSTM state variables initialization 1264 (buddhapuneeth)
• Fixed bug in reading table info from saved models 1263 (TomWildenhain-Microsoft)
• Reenabled placeholderwithdefault tests 1262 (TomWildenhain-Microsoft)
• Fixed bugs in squeeze/unsqueeze for string ops 1261 (TomWildenhain-Microsoft)
• Added support for converting RaggedRange 1256 (TomWildenhain-Microsoft)
• Fixed bug in Transpose optimizer mul handler for consts with multiple… 1255 (TomWildenhain-Microsoft)
• Added a workaround for the TF bug that tensor.numpy() doesn't always … 1250 (TomWildenhain-Microsoft)
• Fixed conversion of Batch Norm when training=true 1249 (TomWildenhain-Microsoft)
• fix sparse_softmax_cross_entropy bug when logits is negative infinity 1248 (liangkaihuan)
• Added flatbuffer parse code for tflite 1247 (TomWildenhain-Microsoft)
• Simplified TA get item to avoid shape bug 1244 (TomWildenhain-Microsoft)
• Set UT to test opset 13 1243 (TomWildenhain-Microsoft)
• More bug fixes for opset 13 support 1241 (TomWildenhain-Microsoft)
• add support dynamic max_iterations in loop 1240 (wangqiaoshi)
• Add support for multiple scan outputs 1238 (TomWildenhain-Microsoft)
• Bug fixes for opset 13 support 1237 (TomWildenhain-Microsoft)
• Fixed resize op for opset 13 1236 (TomWildenhain-Microsoft)
• Fix quant and dequant for opset 13 1235 (TomWildenhain-Microsoft)
• Made unittests require optimizers to not fail 1234 (TomWildenhain-Microsoft)
• Update split op for opset13 and add test cases 1233 (TomWildenhain-Microsoft)
• fix for tf-2.4, ort does not support all topk dtypes 1232 (guschmue)
• support for tf-2.4 1231 (guschmue)
• Added recursive search for table info of saved models 1228 (TomWildenhain-Microsoft)
• more complete nms support, fix efficientdet tutorial 1225 (guschmue)
• Fix bug in reading hash table initializers for tables with names othe… 1223 (TomWildenhain-Microsoft)
• [WIP] Support operators Squeeze/Unsqueeze for opset 13 (v2) 1221 (xadupre)
• Updated ExpandDims handler for opset 13 unsqueeze 1218 (TomWildenhain-Microsoft)
• [WIP] Support operators Squeeze/Unsqueeze for opset 13 1217 (xadupre)
• Tom/squeeze unsqueeze 1216 (TomWildenhain-Microsoft)
• Improve messaging for RandomUniform 1215 (TomWildenhain-Microsoft)
• Created GraphBuilder method for ReduceSum to prepare for opset 13 1213 (TomWildenhain-Microsoft)
• [WIP] Opset 13 support 1211 (TomWildenhain-Microsoft)
• Added covid-twitter-bert to perf testing script 1209 (TomWildenhain-Microsoft)
• Added universal-sentence-encoder-qa to perf testing script 1208 (TomWildenhain-Microsoft)
• Allow specifying domain of custom ops 1206 (TomWildenhain-Microsoft)
• Improved perf of SegmentSum 1205 (TomWildenhain-Microsoft)
• Added conversion for SparseReshape 1204 (TomWildenhain-Microsoft)
• Added conversion for Bincount op 1202 (TomWildenhain-Microsoft)
• Don't fold constants if the result is way larger than inputs 1201 (TomWildenhain-Microsoft)
• Avoid using Where op in Select and SelectV2 conversions 1199 (TomWildenhain-Microsoft)
• Added two universal sentence encoder models to perf testing list 1198 (TomWildenhain-Microsoft)
• Changed tf_utils pass1 to mostly use attribute type not name to determine how to convert. 1196 (TomWildenhain-Microsoft)
• Added WikiWords models to perf testing yaml 1195 (TomWildenhain-Microsoft)
• Added NNLM models to perf testing yaml 1194 (TomWildenhain-Microsoft)
• Updated run_pretrained_models script to support hash tables and string ops 1193 (TomWildenhain-Microsoft)
• Implemented conversion for StringEqual 1192 (TomWildenhain-Microsoft)
• Implemented conversions for all segment ops 1191 (TomWildenhain-Microsoft)
• Remove references to non-variable resources from graph before freezing 1190 (TomWildenhain-Microsoft)
• Added optimization for TableLookup when key is const 1189 (TomWildenhain-Microsoft)
• Allow Reduce ops to have negative axis if opset >= 11 1188 (TomWildenhain-Microsoft)
• Added conversions for SparseSegment ops 1187 (TomWildenhain-Microsoft)
• Added conversions for string ops using contrib ops 1185 (TomWildenhain-Microsoft)
• Allow conversion of Select op when condition has rank 1 and input has… 1184 (TomWildenhain-Microsoft)
• Added conversion for LookupTableSizeV2 1183 (TomWildenhain-Microsoft)
• Added conversion for Roll op 1182 (TomWildenhain-Microsoft)
• Implemented conversio...

tf2onnx-1.7.2

Summary

Fixes and changes to how we use onnx shape inference needed for onnx-1.8

Changes since v1.7.1

• fix shape inference for onnx-1.8 1165 (guschmue)
• Fix parsing of Sparse ops 1164 (TomWildenhain-Microsoft)
• Added conversions for Segment ops 1161 (TomWildenhain-Microsoft)
• Implemented conversion of DynamicStitch 1160 (TomWildenhain-Microsoft)
• Added support for DynamicPartition 1159 (TomWildenhain-Microsoft)
• Update pipeline run_pretrained_models to use tf 2.3.0 1155 (TomWildenhain-Microsoft)
• Fixed regression for SelectV2 when shape is unknown 1154 (TomWildenhain-Microsoft)
• Added output_frozen_graph flag and marked --fold_const as deprecated 1151 (TomWildenhain-Microsoft)
• Fixed bug with ConvTranspose 1150 (TomWildenhain-Microsoft)
• Enabled test_hashtable_lookup for tf2 1149 (TomWildenhain-Microsoft)
• Set aggressive inlining to true 1148 (TomWildenhain-Microsoft)
• Improve ConvTranspose shape inference to work with unknown batch dim 1147 (TomWildenhain-Microsoft)
• Fixed half_pixel_centers for resize_nearest_neighbor 1146 (TomWildenhain-Microsoft)
• Added constant folding for Slice nodes following a shape node of know… 1145 (TomWildenhain-Microsoft)
• Resize align corners attribute 1144 (NikolasMarkou)
• Fixed bug in some rewriters where ops list aliased to g._nodes 1143 (TomWildenhain-Microsoft)
• Added support for TensorScatterUpdate 1142 (TomWildenhain-Microsoft)
• Improved detection of conversion errors due to large models 1141 (TomWildenhain-Microsoft)
• Added more attributes to tf ignored list 1140 (TomWildenhain-Microsoft)
• Fixed bug with order of args when make large models 1139 (TomWildenhain-Microsoft)
• more tests 1138 (guschmue)
• Remove special characters from model zip for large model format 1137 (TomWildenhain-Microsoft)
• add more tests 1136 (guschmue)
• Added perf test model yaml file 1133 (TomWildenhain-Microsoft)
• Added new input_get functions to run_pretrained_models.py 1128 (TomWildenhain-Microsoft)
• Don't add identity before output if output is also a graph input 1127 (TomWildenhain-Microsoft)

A huge thank you to our contributors for this release!

NikolasMarkou

tf2onnx-1.7.1

Summary

large model support, conversion performance improvements and fixes.

Changes since v1.6.3

• Fixed error in computing shape of unpack op 1125 (TomWildenhain-Microsoft)
• Moves unused class attribute to function local variable. 1124 (xadupre)
• Added cast to float before log 1123 (TomWildenhain-Microsoft)
• cast log to float 1122 (TomWildenhain-Microsoft)
• Add cast after Size if expected type is not int64 1119 (TomWildenhain-Microsoft)
• Fixed bug in run_pretrained_models.py for models with subgraphs 1118 (TomWildenhain-Microsoft)
• Fixes #595, support operator InvertPermutation 1117 (xadupre)
• Better error message when it fails due to RFFT 1116 (xadupre)
• Add support for operators RFFT, ComplexAbs 1114 (xadupre)
• Rescaled get_beach and added get_zeros_int32 and get_zeros_int64 1113 (TomWildenhain-Microsoft)
• Fix --concrete_func for multiple inputs in signature 1110 (TomWildenhain-Microsoft)
• Added support for converting NonMaxSuppression in opset 10 with dynam… 1109 (TomWildenhain-Microsoft)
• Update run_pretrained_models.py to support large models 1107 (TomWildenhain-Microsoft)
• fix bn fuse for fp16 1106 (guschmue)
• Added constant folding using TF for large models 1105 (TomWildenhain-Microsoft)
• Add an end2end example with tf.keras 1103 (xadupre)
• Hash tensor tensor values in merge_duplicated_nodes to increase conversion speed 1102 (TomWildenhain-Microsoft)
• Made einsum op convert equation string to lower case 1099 (TomWildenhain-Microsoft)
• Fix issue with model esrgan-tf2_1 1098 (xadupre)
• reflect support for python 3.8, tf-2.3 1097 (guschmue)
• Added f to the list of TF function attributes 1094 (TomWildenhain-Microsoft)
• Added support for converting large models 1090 (TomWildenhain-Microsoft)
• Add cast to same type before equal operator 1089 (bedapisl)
• Added graph methods for saving using the external data storage format 1088 (TomWildenhain-Microsoft)
• Created compress_graph_def function 1087 (TomWildenhain-Microsoft)
• Rename coverter into converter 1086 (xadupre)
• Fix the missing axis when LogSoftmax is in graph 1084 (peterjc123)
• Add functionality for QDQ per channel 1081 (peri044)
• Add a script to profile the conversion of a model 1077 (xadupre)
• Feature: Added optimization step to remove upsample layers with all ones in scale 1074 (NikolasMarkou)
• #1070 replacing Reshape and Transpose operators to the kernel that are invalid in tensorrt with just a reshaped tensor 1071 (NikolasMarkou)
• Fix/gemm rewriter bias add 1069 (phager90)
• fix a NCHW pb converting bug in pad_rewriter 1063 (charlieguo0307)
• Faster insertion of operator cast, missing replacement in #1059 1061 (xadupre)
• Faster insertion of operator cast 1059 (xadupre)
• Changes type instead of name in operator clip 1058 (xadupre)
• Replace sys.maxsize by np.iinfo(np.int64).max 1057 (xadupre)
• Use the same syntax to replace an node input 1056 (xadupre)
• Removes unnecessary print 1055 (xadupre)
• Perf gain in tf_utils.py, more efficient error messages, faster comparisons 1054 (xadupre)
• Minor perf gain in graph_matcher.py 1053 (xadupre)- Added brackets to fix error in QueueDequeueUpToV2 1050 (TomWildenhain-Microsoft)
• [WIP] Improve performance by adding forward indexes, second version. 1049 (xadupre)
• Implemented QueueDequeueUpToV2 1047 (TomWildenhain-Microsoft)
• Fixed bug in padding calculation for padding='SAME' when dilations>1 1046 (TomWildenhain-Microsoft)
• Conv3DBackpropInputV2 1045 (ralovich)
• Adding Constant Folding for Reshape Nodes to Optimizer 1042 (phager90)
• Added support for Conv3DBackpropInputV2 1041 (TomWildenhain-Microsoft)
• [WIP] Improve performance by adding forward indexes. 1035 (xadupre)
• fix label offset in tutorial 1026 (guschmue)
• [WIP] Improve performance by adding identity nodes 1023 (xadupre)
• enable tf-2.3 in ci pipeline 1022 (guschmue)

A huge thank you to our contributors for this release!

bedapisl, peterjc123, peri044, NikolasMarkou, phager90, charlieguo0307, ralovich, phager90