Technical Updates 🤖Adjust the Trace API interface, encapsulate the transform pass, and add fuse, calibrate, quantize and export methods; Trace API adds dataset, statistic binding, and accuracy eval interface; circom adapts alexnet, dense, repairs padding shape, concat operator and other issues; Fix dropout + tuplegetitem tvm running error, fix use too much shared data video memory error; ZKML supports shufflenet, densenet, vgg, squeezenet, and the accuracy error of the original floating poin...

Technical Updates 🤖Adjust the Trace API interface, encapsulate the transform pass, and add fuse, calibrate, quantize and export methods; Trace API adds dataset, statistic binding, and accuracy eval interface; circom adapts alexnet, dense, repairs padding shape, concat operator and other issues; Fix dropout + tuplegetitem tvm running error, fix use too much shared data video memory error; ZKML supports shufflenet, densenet, vgg, squeezenet, and the accuracy error of the original floating poin...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)ZKML adds TVM Runtime Inference reasoning simulation test quantized accuracy, and debugs the loss of accuracy;ZKML adjusts the data set mount interface and output result analysis class;ZKML adjusts the restore logic of the quantitative model output results;Fix the problem of input_data not passing mrt transformer in model2circom;ZKML adjusts the data set mount, and increases the data set input in the calibrated pass to improv...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)ZKML adds TVM Runtime Inference reasoning simulation test quantized accuracy, and debugs the loss of accuracy;ZKML adjusts the data set mount interface and output result analysis class;ZKML adjusts the restore logic of the quantitative model output results;Fix the problem of input_data not passing mrt transformer in model2circom;ZKML adjusts the data set mount, and increases the data set input in the calibrated pass to improv...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)ZKML trains RNN language sequence prediction model;ZKML subgraph interface storage is optimized, and parent and child traces share node pointers (hash);ZKML investigates accuracy parameters and tests the impact of parameters on the quantitative model;ZKML Make config class for quantization procedure(with statement);ZkRollupThe test ran boojum's sha256 circuit. It succeeded (very quickly) when the preimage

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)ZKML trains RNN language sequence prediction model;ZKML subgraph interface storage is optimized, and parent and child traces share node pointers (hash);ZKML investigates accuracy parameters and tests the impact of parameters on the quantitative model;ZKML Make config class for quantization procedure(with statement);ZkRollupThe test ran boojum's sha256 circuit. It succeeded (very quickly) when the preimage

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM) & Model RepresentationTool(MRT) R&DMRT adjusts the implementation of Inference, uses TVM underlying Pass to perform sampling and output intermediate results, and especially judges some unrealized operators;MRT solves the problem of Circom Shape adaptation;MRT code refactoring, abstract parent class and virtual function;ZkRollupInvestigate the test run of cuda bellman;Investigate plonk, lookup and plonkup and test the impleme...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM) & Model RepresentationTool(MRT) R&DMRT adjusts the implementation of Inference, uses TVM underlying Pass to perform sampling and output intermediate results, and especially judges some unrealized operators;MRT solves the problem of Circom Shape adaptation;MRT code refactoring, abstract parent class and virtual function;ZkRollupInvestigate the test run of cuda bellman;Investigate plonk, lookup and plonkup and test the impleme...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM) & Model RepresentationTool(MRT) ZKML adds new adaptation to CVM, using codegen to generate operator json data ZKML develops circomlib adaptation code ZkRollup Implement zkx (zk for cortex in plonk) type-related libraries (block, tx, op, aggregate, gas, etc.), and pass related single tests. (Compared with zion, the new types: mint_nft, withdraw_nft, swap and change pubkey can be authorized by onchain functions, and support ei...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM) & Model RepresentationTool(MRT) ZKML adds new adaptation to CVM, using codegen to generate operator json data ZKML develops circomlib adaptation code ZkRollup Implement zkx (zk for cortex in plonk) type-related libraries (block, tx, op, aggregate, gas, etc.), and pass related single tests. (Compared with zion, the new types: mint_nft, withdraw_nft, swap and change pubkey can be authorized by onchain functions, and support ei...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)Complete the function of coco, voc, cifar10 dataset loading;circom generator RightShift and MulScalar parameter alignment. Type annotations, Symbol type confusion fixes;Generate graph optimization, lose invalid nodes after topological sorting. scalar is integrated into op as an attribute, the original parameter is still output to input.json, fix circom compilation repair;Increase NLP model training, long sequence numerical pr...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)Complete the function of coco, voc, cifar10 dataset loading;circom generator RightShift and MulScalar parameter alignment. Type annotations, Symbol type confusion fixes;Generate graph optimization, lose invalid nodes after topological sorting. scalar is integrated into op as an attribute, the original parameter is still output to input.json, fix circom compilation repair;Increase NLP model training, long sequence numerical pr...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)MRT adds BN + Dense, and quantitative support for BN equivalent conversion;MRT adds Dropout support and automatically fuse TupleGetItem;Fix the symbol name problem: cause circom code compilation error (including '%', '.', etc. cause the variable name to be ambiguous during compilation);MRT organizes existing work, codes, deletes invalid codes, writes documents, including TEST environment preparation, runni...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)MRT adds BN + Dense, and quantitative support for BN equivalent conversion;MRT adds Dropout support and automatically fuse TupleGetItem;Fix the symbol name problem: cause circom code compilation error (including '%', '.', etc. cause the variable name to be ambiguous during compilation);MRT organizes existing work, codes, deletes invalid codes, writes documents, including TEST environment preparation, runni...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)ZKML optimizes ResNet model accuracy, improving accuracy by 14%ZKML adjusts the Trace API and adds interfaces for data set binding, preprocessing, and postprocessing information.ZKML completes the data set ImageNet, coco and other interfaces, and downloads more data setsZKML investigates ezkl, runs through the process and compares ZKML Circom benchmark performanceZkRollupImplementing the opcode of zkcvmTry to implement the ex...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)ZKML optimizes ResNet model accuracy, improving accuracy by 14%ZKML adjusts the Trace API and adds interfaces for data set binding, preprocessing, and postprocessing information.ZKML completes the data set ImageNet, coco and other interfaces, and downloads more data setsZKML investigates ezkl, runs through the process and compares ZKML Circom benchmark performanceZkRollupImplementing the opcode of zkcvmTry to implement the ex...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)The location and parameters of the logs interface in the design and test model quantification processFix bugs in zkml, and adjust and optimize the codeImplement the interface of image input_data to scale and round separatelyAfter the model is converted to circom, modify input.json and set the input valueResnet18 resnet34 resnet50 resnet101 resnet152 format converted successfullyZkRollupImplement the js sdk that interacts with...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)The location and parameters of the logs interface in the design and test model quantification processFix bugs in zkml, and adjust and optimize the codeImplement the interface of image input_data to scale and round separatelyAfter the model is converted to circom, modify input.json and set the input valueResnet18 resnet34 resnet50 resnet101 resnet152 format converted successfullyZkRollupImplement the js sdk that interacts with...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM) & Model RepresentationTool(MRT)ZKML adjusts Trace API, fixes load/store issues, and adds force option;ZKML adjusts the ShapeAdapter pass in ZKSection9, adds the function of eliminating Batch in CircomGenerator, and adapts to the circom deep learning library;ZKML adjusted the circom deep learning library API and modified some calculation logic;ZKML exports JSON data in CVM format, and adds operators such as conv and relu;ZkRo...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM) & Model RepresentationTool(MRT)ZKML adjusts Trace API, fixes load/store issues, and adds force option;ZKML adjusts the ShapeAdapter pass in ZKSection9, adds the function of eliminating Batch in CircomGenerator, and adapts to the circom deep learning library;ZKML adjusted the circom deep learning library API and modified some calculation logic;ZKML exports JSON data in CVM format, and adds operators such as conv and relu;ZkRo...

Technical Updates 🤖MRT/CVMMRT has added support for the Detection model, with quantization already implemented for models such as Yolo, SSD, etc.Implemented automatic Spliter/Merger algorithm logic to eliminate the problem of duplicates in segmented subgraphs (increased reference counting).Added constant pass fusion for operators oneslike and zeroslike.Added adaptation and quantization logic for operators shape of, reverse, maximum, and minimum.Adapted Circom for the detection model, added i...

Technical Updates 🤖MRT/CVMMRT has added support for the Detection model, with quantization already implemented for models such as Yolo, SSD, etc.Implemented automatic Spliter/Merger algorithm logic to eliminate the problem of duplicates in segmented subgraphs (increased reference counting).Added constant pass fusion for operators oneslike and zeroslike.Added adaptation and quantization logic for operators shape of, reverse, maximum, and minimum.Adapted Circom for the detection model, added i...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)ZKML circom adapts conv2d padding: insert pad2d op before conv2d in symbol;Circom modified conv2d inputs & shape to adapt to the calculation of upper and lower nodes in the topology map;Circom fixes the problem that stride=1 is hard-coded to 1, adds stride support and derivation, and requires the stride of xy to be consistent;ZKML adds support for cvm_right_shift, cvm_clip, etc., and converts bias_add to conv operator;ZkRollu...

Technical Updates 🤖Machine-Learning(ML), CortexVirtual Machine(CVM)ZKML circom adapts conv2d padding: insert pad2d op before conv2d in symbol;Circom modified conv2d inputs & shape to adapt to the calculation of upper and lower nodes in the topology map;Circom fixes the problem that stride=1 is hard-coded to 1, adds stride support and derivation, and requires the stride of xy to be consistent;ZKML adds support for cvm_right_shift, cvm_clip, etc., and converts bias_add to conv operator;ZkRollu...