Modelpublic

nilcars/tensorflow_tensorflow_model

sourceHugging Faceupdated 2y agoView on Hugging Face

0likes5downloads

Model Card

SetFit with sentence-transformers/all-mpnet-base-v2

This is a SetFit model that can be used for Text Classification. This SetFit model uses sentence-transformers/all-mpnet-base-v2 as the Sentence Transformer embedding model. A LogisticRegression instance is used for classification.

The model has been trained using an efficient few-shot learning technique that involves:

1.Fine-tuning a Sentence Transformer with contrastive learning.
2.Training a classification head with features from the fine-tuned Sentence Transformer.

Model Details

Model Description

—Model Type: SetFit
—Sentence Transformer body: sentence-transformers/all-mpnet-base-v2
—Classification head: a LogisticRegression instance
—Maximum Sequence Length: 384 tokens
—Number of Classes: 3 classes

Model Sources

—Repository: SetFit on GitHub
—Paper: Efficient Few-Shot Learning Without Prompts
—Blogpost: SetFit: Efficient Few-Shot Learning Without Prompts

Model Labels

Label	Examples
question	<ul><li>"Parse output of `mobile_ssd_v2_float_coco.tflite` ### Issue type\n\nSupport\n\n### Have you reproduced the bug with TensorFlow Nightly?\n\nNo\n\n### Source\n\nsource\n\n### TensorFlow version\n\nv2.11.1\n\n### Custom code\n\nYes\n\n### OS platform and distribution\n\nLinux Ubuntu 20.04\n\n### Mobile device\n\nAndroid\n\n### Python version\n\nNo response\n\n### Bazel version\n\n6.2.0\n\n### GCC/compiler version\n\n12\n\n### CUDA/cuDNN version\n\nNo response\n\n### GPU model and memory\n\nNo response\n\n### Current behavior?\n\nI'm trying to use the model mobile_ssd_v2_float_coco.tflite on a C++ application, I'm able to execute the inference and get the results.\r\n\r\nBased on the Netron app I see that its output is:\r\n\r\n\r\nBut I couldn't find an example code showing how to parse this output.\r\n\r\nI tried to look into https://github.com/tensorflow/tensorflow/issues/29054 and https://github.com/tensorflow/tensorflow/issues/40298 but the output of the model is different from the one provided here.\r\n\r\nDo you have any example code available in Java, Python, or even better in C++ to parse this model output?\n\n### Standalone code to reproduce the issue\n\n```shell\nNo example code is available to parse the output of mobile_ssd_v2_float_coco.tflite.\n``\n\n\n### Relevant log output\n\n_No response_"</li><li>'Tensorflow Lite library is crashing in WASM library at 3rd inference <details><summary>Click to expand!</summary> \r\n \r\n ### Issue Type\r\n\r\nSupport\r\n\r\n### Have you reproduced the bug with TF nightly?\r\n\r\nYes\r\n\r\n### Source\r\n\r\nsource\r\n\r\n### Tensorflow Version\r\n\r\n2.7.0\r\n\r\n### Custom Code\r\n\r\nYes\r\n\r\n### OS Platform and Distribution\r\n\r\nEmscripten, Ubuntu 18.04\r\n\r\n### Mobile device\r\n\r\n_No response_\r\n\r\n### Python version\r\n\r\n_No response_\r\n\r\n### Bazel version\r\n\r\n_No response_\r\n\r\n### GCC/Compiler version\r\n\r\n_No response_\r\n\r\n### CUDA/cuDNN version\r\n\r\n_No response_\r\n\r\n### GPU model and memory\r\n\r\n_No response_\r\n\r\n### Current Behaviour?\r\n\r\n`shell\r\nHello! I have C++ code that I want to deploy as WASM library and this code contains TFLite library. I have compiled TFLite library with XNNPack support using Emscripten toolchain quite easy, so no issue there. I have a leight-weight convolution+dense model that runs perfectly on Desktop, but I am starting having problems in the browser.\r\n\r\nIn 99% of cases I have an error on the third inference:\r\n\r\nUncaught RuntimeError: memory access out of bounds\r\n\r\nThrough some trivial debugging I have found out that the issue comes from _interpreter->Invoke() method. Does not matter if I put any input or not, I just need to call Invoke() three times and I have a crash.\r\n\r\nFirst thing first: I decided to add more memory to my WASM library by adding this line to CMake:\r\n\r\nSET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -s TOTAL_STACK=134217728 -s TOTAL_MEMORY=268435456")\r\nSET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -s TOTAL_STACK=134217728 -s TOTAL_MEMORY=268435456")\r\n\r\n128 MB and 256 MB in total for 1 MB model - I think this is more than enough. And on top of that, I am allowing Memory Growth. But unfortunately, I have exactly the same issue.\r\n\r\nI am beating on this problem for 2 weeks straight and at this stage I have no clue how to fix it. Also I have tried to set custom allocation using TfLiteCustomAllocation but in this case I have a crash on the very first inference. I guess I was not using it right, but unfortunately I couldn\'t find even one tutorial describing how to apply custom allocation in TFLite.\r\n\r\nI said that I have a crash in 99% of cases. There was one time when WASM library worked and inference worked as well. It happens just randomly once, and I couldn\'t reproduce it anymore.\r\n``\r\n\r\n\r\n### Standalone code to reproduce the issue\r\n\r\n```shell\r\nHere is the code that does TFLite inference\r\n\r\n\r\n#include <cstdlib>\r\n#include "tflitemodel.h"\r\n#include <iostream>\r\n\r\n#include "tensorflow/lite/interpreter.h"\r\n#include "tensorflow/lite/util.h"\r\n\r\nnamespace tracker {\r\n\r\n#ifdef EMSCRIPTEN\r\n\tvoid TFLiteModel::init(std::stringstream& stream) {\r\n\r\n\t\tstd::string imgstr = stream.str();\r\n\t\tstd::vector<char> imgmodeldata(imgstr.size());\r\n\t\tstd::copy(imgstr.begin(), imgstr.end(), imgmodeldata.begin());\r\n\r\n\t\tmodel = tflite::FlatBufferModel::BuildFromBuffer(imgstr.data(), imgstr.size());\r\n#else\r\n\tvoid TFLiteModel::init(const std::string& path) {\r\n\t\tmodel = tflite::FlatBufferModel::BuildFromFile(path.cstr());\r\n\r\n#endif\r\n\r\n\t\ttflite::ops::builtin::BuiltinOpResolver resolver;\r\n\t\ttflite::InterpreterBuilder(_model, resolver)(&_interpreter);\r\n\r\n\t\t_interpreter->AllocateTensors();\r\n\r\n\t\t/for (int i = 0; i < interpreter->tensorssize(); i++) {\r\n\t\t\tTfLiteTensor* tensor = interpreter->tensor(i);\r\n\r\n\t\t\tif (tensor->allocationtype == kTfLiteArenaRw		tensor->allocationtype == kTfLiteArenaRwPersistent) {\r\n\r\n\t\t\t\tint alignedbytes = tensor->bytes + (tflite::kDefaultTensorAlignment - tensor->bytes % tflite::kDefaultTensorAlignment) % tflite::kDefaultTensorAlignment;\r\n\r\n\t\t\t\tTfLiteCustomAllocation customAlloc;\r\n\t\t\t\tint result = posix_memalign(&customAlloc.data, tflite::kDefaultTensorAlignment, tensor->bytes);\r\n\t\t\t\tif (result != 0		customAlloc.data == NULL) {\r\n\t\t\t\t\tstd::cout << "posixmemalign does not work!\\\\n";\r\n\t\t\t\t}\r\n\r\n\t\t\t\tTfLiteStatus st = interpreter->SetCustomAllocationForTensor(i, customAlloc);\r\n\t\t\t\tstd::cout << "status = " << st << std::endl;\r\n\t\t\t\tif (tensor->bytes % tflite::kDefaultTensorAlignment != 0) {\r\n\t\t\t\t\tstd::cout << "bad! i " << i << ", size " << tensor->bytes << std::endl;\r\n\t\t\t\t}\r\n\t\t\t\tallocations.pushback(customAlloc);\r\n\t\t\t}\r\n\t\t}\r\n\t\texit(0);/\r\n\t}\r\n\r\n\tvoid TFLiteModel::forward(const cv::Mat& img_input, const std::vector<float>& lms_input) {\r\n\r\n\t\tfloat modelin = interpreter->typedinputtensor<float>(0);\r\n\t\tstd::memcpy(modelin, imginput.data, imginput.total() * imginput.elemSize());\r\n\r\n\t\tfloat lms_in = _interpreter->typed_input_tensor<float>(1);\r\n\t\tstd::memcpy(lms_in, lms_input.data(), sizeof(float) lmsinput.size());\r\n\t\t\r\n\t\tinterpreter->Invoke();\r\n\t}\r\n\r\n\tfloat TFLiteModel::out() {\r\n\t\treturn _interpreter->typed_output_tensor<float>(0);\r\n\t}\r\n\r\n\tstd::vector<int> TFLiteModel::getOutputShape() const {\r\n\t\tTfLiteTensor outtensor = interpreter->outputtensor(0);\r\n\t\tTfLiteIntArray dims = outtensor->dims;\r\n\r\n\t\tstd::vector<int> sh;\r\n\t\tfor (int i = 0; i < dims->size; i++) {\r\n\t\t\tsh.push_back(dims->data[i]);\r\n\t\t}\r\n\r\n\t\treturn sh;\r\n\t}\r\n}\r\n```\r\n\r\n\r\n### Relevant log output\r\n\r\n_No response_</details>'</li><li>'error: \'tf.Conv2D\' op is neither a custom op nor a flex op ### 1. System information\r\n\r\n- OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 20.04\r\n- TensorFlow installation (pip package or built from source): pip package\r\n- TensorFlow library (version, if pip package or github SHA, if built from source): v2.10\r\n\r\n### 2. Code\r\nCode for conversion\r\n```\r\nconverter = tf.lite.TFLiteConverter.from_saved_model(f\'savedmodel/decoder\')\r\ntflite_model = converter.convert()\r\n\r\n# save the model\r\nwith open(f\'{name}.tflite\', \'wb\') as f:\r\n f.write(tflite_model)\r\n```\r\nCode for the model\r\n```\r\nlatent = keras.layers.Input((n_h, n_w, 4))\r\ndecoder = Decoder()\r\ndecoder = keras.models.Model(latent, decoder(latent))\r\n```\r\n```\r\nclass Decoder(keras.Sequential):\r\n def __init__(self):\r\n super().__init__(\r\n [\r\n keras.layers.Lambda(lambda x: 1 / 0.18215 x),\r\n PaddedConv2D(4, 1),\r\n PaddedConv2D(512, 3, padding=1),\r\n ResnetBlock(512, 512),\r\n AttentionBlock(512),\r\n ResnetBlock(512, 512),\r\n ResnetBlock(512, 512),\r\n ResnetBlock(512, 512),\r\n ResnetBlock(512, 512),\r\n keras.layers.UpSampling2D(size=(2, 2)),\r\n PaddedConv2D(512, 3, padding=1),\r\n ResnetBlock(512, 512),\r\n ResnetBlock(512, 512),\r\n ResnetBlock(512, 512),\r\n keras.layers.UpSampling2D(size=(2, 2)),\r\n PaddedConv2D(512, 3, padding=1),\r\n ResnetBlock(512, 256),\r\n ResnetBlock(256, 256),\r\n ResnetBlock(256, 256),\r\n keras.layers.UpSampling2D(size=(2, 2)),\r\n PaddedConv2D(256, 3, padding=1),\r\n ResnetBlock(256, 128),\r\n ResnetBlock(128, 128),\r\n ResnetBlock(128, 128),\r\n tfa.layers.GroupNormalization(epsilon=1e-5),\r\n keras.layers.Activation("swish"),\r\n PaddedConv2D(3, 3, padding=1),\r\n ]\r\n )\r\n```\r\n\r\n### 3. Failure after conversion\r\nconversion fails\r\n\r\n\r\n### 5. (optional) Any other info / logs\r\n[error.log](https://github.com/tensorflow/tensorflow/files/10302790/error.log)\r\n``\r\nSome ops are not supported by the native TFLite runtime, you can enable TF kernels fallback using TF Select. See instructions: https://www.tensorflow.org/lite/guide/ops_select \r\nTF Select ops: Conv2D\r\nDetails:\r\n\ttf.Conv2D(tensor<?x?x?x?xf32>, tensor<1x1x512x512xf32>) -> (tensor<?x?x?x512xf32>) : {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}\r\n\ttf.Conv2D(tensor<?x?x?x?xf32>, tensor<3x3x128x128xf32>) -> (tensor<?x?x?x128xf32>) : {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}\r\n\ttf.Conv2D(tensor<?x?x?x?xf32>, tensor<3x3x128x3xf32>) -> (tensor<?x?x?x3xf32>) : {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}\r\n\ttf.Conv2D(tensor<?x?x?x?xf32>, tensor<3x3x256x128xf32>) -> (tensor<?x?x?x128xf32>) : {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}\r\n\ttf.Conv2D(tensor<?x?x?x?xf32>, tensor<3x3x256x256xf32>) -> (tensor<?x?x?x256xf32>) : {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}\r\n\ttf.Conv2D(tensor<?x?x?x?xf32>, tensor<3x3x512x256xf32>) -> (tensor<?x?x?x256xf32>) : {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}\r\n\ttf.Conv2D(tensor<?x?x?x?xf32>, tensor<3x3x512x512xf32>) -> (tensor<?x?x?x512xf32>) : {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}\r\n``\r\nAccording to the error message, I suspect that it can not recognize the input shape. But as you can see on the above code, input is specified for the functional API for` decoder `model. \r\n(FYI, The inference code is called with` predictonbatch `method. I found out other model with` predictonbatch `is converted successfully, but that model doesn\'t contain` conv2d `block inside. Can using` predictonbatch `together with` conv2d `be a problem?)\r\n\r\nI\'m sure` conv2d` is on the allowlist for TFLite operators. Any suggestions for this problem? Thank you.'</li></ul>
feature	<ul><li>'tf.keras.optimizers.experimental.AdamW only support constant weightdecay <details><summary>Click to expand!</summary> \n \n ### Issue Type\n\nFeature Request\n\n### Source\n\nsource\n\n### Tensorflow Version\n\n2.8\n\n### Custom Code\n\nNo\n\n### OS Platform and Distribution\n\nNo response\n\n### Mobile device\n\nNo response\n\n### Python version\n\nNo response\n\n### Bazel version\n\nNo response\n\n### GCC/Compiler version\n\nNo response\n\n### CUDA/cuDNN version\n\nNo response\n\n### GPU model and memory\n\nNo response\n\n### Current Behaviour?\n\n```shell\ntf.keras.optimizers.experimental.AdamW only supports constant weight decay. But usually we want the weightdecay value to decay with learning rate schedule.\n```\n\n\n### Standalone code to reproduce the issue\n\n```shell\nThe legacy tfa.optimizers.AdamW supports callable weight_decay, which is much better.\n``\n\n\n### Relevant log output\n\n_No response_</details>'</li><li>'RFE tensorflow-aarch64==2.6.0 build ? System information\r\n TensorFlow version (you are using): 2.6.0\r\n- Are you willing to contribute it (Yes/No): Yes\r\n\r\nDescribe the feature and the current behavior/state.\r\n\r\nBrainchip Akida AKD1000 SNN neuromorphic MetaTF SDK support 2.6.0 on x86_64. They claim support for aarch64, but when creating a virtualenv it fails on aarch64 due to lacking tensorflow-aarc64==2.6.0 build.\r\n\r\nWill this change the current api? How?\r\n\r\nNA\r\n\r\nWho will benefit with this feature?\r\n\r\nCustomer of Brainchip Akida who run on Arm64 platforms.\r\n\r\nAny Other info.\r\n\r\nhttps://doc.brainchipinc.com/installation.html\r\n\r\n\r\n'</li><li>"How to calculate 45 degree standing position of body from camera in swift (Pose estimation) <details><summary>Click to expand!</summary> \n \n ### Issue Type\n\nFeature Request\n\n### Source\n\nsource\n\n### Tensorflow Version\n\npod 'TensorFlowLiteSwift', '~> 0.0.1-nightly', :subspecs => ['CoreML', 'Metal']\n\n### Custom Code\n\nYes\n\n### OS Platform and Distribution\n\n_No response_\n\n### Mobile device\n\n_No response_\n\n### Python version\n\n_No response_\n\n### Bazel version\n\n_No response_\n\n### GCC/Compiler version\n\n_No response_\n\n### CUDA/cuDNN version\n\n_No response_\n\n### GPU model and memory\n\n_No response_\n\n### Current Behaviour?\n\n``shell\nHow to calculate 45 degree standing position of body from camera in swift.\n```\n\n\n### Standalone code to reproduce the issue\n\n```shell\nHow to calculate 45 degree standing position of body from camera in swift using the body keypoints. (Pose estimation)\n```\n\n\n### Relevant log output\n\nNo response</details>"</li></ul>
bug	<ul><li>'Abort when running tensorflow.python.ops.genarrayops.depthtospace ### Issue type\n\nBug\n\n### Have you reproduced the bug with TensorFlow Nightly?\n\nNo\n\n### Source\n\nbinary\n\n### TensorFlow version\n\n2.11.0\n\n### Custom code\n\nYes\n\n### OS platform and distribution\n\n22.04\n\n### Mobile device\n\nNo response\n\n### Python version\n\n3.9\n\n### Bazel version\n\nNo response\n\n### GCC/compiler version\n\nNo response\n\n### CUDA/cuDNN version\n\nnvidia-cudnn-cu11==8.6.0.163, cudatoolkit=11.8.0\n\n### GPU model and memory\n\nNo response\n\n### Current behavior?\n\nDue to very large integer argument\n\n### Standalone code to reproduce the issue\n\n```shell\nimport tensorflow as tf\r\nimport os\r\nimport numpy as np\r\nfrom tensorflow.python.ops import gen_array_ops\r\ntry:\r\n arg_0_tensor = tf.random.uniform([3, 2, 3, 4], dtype=tf.float32)\r\n arg_0 = tf.identity(arg_0_tensor)\r\n arg_1 = 2147483647\r\n arg_2 = "NHWC"\r\n out = gen_array_ops.depth_to_space(arg_0,arg_1,arg_2,)\r\nexcept Exception as e:\r\n print("Error:"+str(e))\r\n\r\n```\n```\n\n\n### Relevant log output\n\n``shell\n023-08-13 00:23:53.644564: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.\r\n2023-08-13 00:23:54.491071: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:981] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2023-08-13 00:23:54.510564: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:981] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2023-08-13 00:23:54.510736: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:981] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2023-08-13 00:23:54.511051: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX2 FMA\r\nTo enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\r\n2023-08-13 00:23:54.511595: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:981] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2023-08-13 00:23:54.511717: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:981] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2023-08-13 00:23:54.511830: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:981] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2023-08-13 00:23:54.572398: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:981] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2023-08-13 00:23:54.572634: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:981] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2023-08-13 00:23:54.572791: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:981] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\r\n2023-08-13 00:23:54.572916: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1613] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 153 MB memory: -> device: 0, name: NVIDIA GeForce GTX 1660 Ti, pci bus id: 0000:01:00.0, compute capability: 7.5\r\n2023-08-13 00:23:54.594062: I tensorflow/compiler/xla/stream_executor/cuda/cuda_driver.cc:735] failed to allocate 153.88M (161349632 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY: out of memory\r\n2023-08-13 00:23:54.594484: I tensorflow/compiler/xla/stream_executor/cuda/cuda_driver.cc:735] failed to allocate 138.49M (145214720 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY: out of memory\r\n2023-08-13 00:23:54.600623: F tensorflow/core/framework/tensor_shape.cc:201] Non-OK-status: InitDims(dim_sizes) status: INVALID_ARGUMENT: Expected a non-negative size, got -2\r\nAborted\r\n\r\n``\n``\n'</li><li>"float8 (both e4m3fn and e5m2) missing from numbertype ### Issue Type\r\n\r\nBug\r\n\r\n### Have you reproduced the bug with TF nightly?\r\n\r\nNo\r\n\r\n### Source\r\n\r\nbinary\r\n\r\n### Tensorflow Version\r\n\r\n2.12.0\r\n\r\n### Custom Code\r\n\r\nYes\r\n\r\n### OS Platform and Distribution\r\n\r\nmacOS-13.2.1-arm64-arm-64bit\r\n\r\n### Mobile device\r\n\r\n_No response_\r\n\r\n### Python version\r\n\r\n3.9.6\r\n\r\n### Bazel version\r\n\r\n_No response_\r\n\r\n### GCC/Compiler version\r\n\r\n_No response_\r\n\r\n### CUDA/cuDNN version\r\n\r\n_No response_\r\n\r\n### GPU model and memory\r\n\r\n_No response_\r\n\r\n### Current Behaviour?\r\n\r\nFP8 datatypes are missing from kNumberTypes `in` tensorflow/core/framework/types.h`, and also missing from` TFCALLFLOATTYPES(m)` in `tensorflow/core/framework/registertypes.h`. This causes simple ops (like slice, transpose, split, etc.) to raise NotFoundError.\r\n\r\n### Standalone code to reproduce the issue\r\n\r\n```python\r\nimport tensorflow as tf\r\nfrom tensorflow.python.framework import dtypes\r\n\r\na = tf.constant([[1.2345678, 2.3456789, 3.4567891], [4.5678912, 5.6789123, 6.7891234]], dtype=dtypes.float16)\r\nprint(a)\r\n\r\na_fp8 = tf.cast(a, dtypes.float8_e4m3fn)\r\nprint(a_fp8)\r\n\r\nb = a_fp8[1:2] # tensorflow.python.framework.errors_impl.NotFoundError\r\nb = tf.transpose(a_fp8, [1, 0]) # tensorflow.python.framework.errors_impl.NotFoundError\r\n```\r\n\r\n\r\n### Relevant log output\r\n\r\n``\r\ntensorflow.python.framework.errors_impl.NotFoundError: Could not find device for node: {{node StridedSlice}} = StridedSlice[Index=DT_INT32, T=DT_FLOAT8_E4M3FN, begin_mask=0, ellipsis_mask=0, end_mask=0, new_axis_mask=0, shrink_axis_mask=0]\r\nAll kernels registered for op StridedSlice:\r\n device='XLA_CPU_JIT'; Index in [DT_INT32, DT_INT16, DT_INT64]; T in [DT_FLOAT, DT_DOUBLE, DT_INT32, DT_UINT8, DT_INT16, 930109355527764061, DT_HALF, DT_UINT32, DT_UINT64, DT_FLOAT8_E5M2, DT_FLOAT8_E4M3FN]\r\n device='CPU'; T in [DT_UINT64]\r\n device='CPU'; T in [DT_INT64]\r\n device='CPU'; T in [DT_UINT32]\r\n device='CPU'; T in [DT_UINT16]\r\n device='CPU'; T in [DT_INT16]\r\n device='CPU'; T in [DT_UINT8]\r\n device='CPU'; T in [DT_INT8]\r\n device='CPU'; T in [DT_INT32]\r\n device='CPU'; T in [DT_HALF]\r\n device='CPU'; T in [DT_BFLOAT16]\r\n device='CPU'; T in [DT_FLOAT]\r\n device='CPU'; T in [DT_DOUBLE]\r\n device='CPU'; T in [DT_COMPLEX64]\r\n device='CPU'; T in [DT_COMPLEX128]\r\n device='CPU'; T in [DT_BOOL]\r\n device='CPU'; T in [DT_STRING]\r\n device='CPU'; T in [DT_RESOURCE]\r\n device='CPU'; T in [DT_VARIANT]\r\n device='CPU'; T in [DT_QINT8]\r\n device='CPU'; T in [DT_QUINT8]\r\n device='CPU'; T in [DT_QINT32]\r\n device='DEFAULT'; T in [DT_INT32]\r\n [Op:StridedSlice] name: strided_slice/\r\n``\r\n\r\n``\r\ntensorflow.python.framework.errors_impl.NotFoundError: Could not find device for node: {{node Transpose}} = Transpose[T=DT_FLOAT8_E4M3FN, Tperm=DT_INT32]\r\nAll kernels registered for op Transpose:\r\n device='XLA_CPU_JIT'; Tperm in [DT_INT32, DT_INT64]; T in [DT_FLOAT, DT_DOUBLE, DT_INT32, DT_UINT8, DT_INT16, 930109355527764061, DT_HALF, DT_UINT32, DT_UINT64, DT_FLOAT8_E5M2, DT_FLOAT8_E4M3FN]\r\n device='CPU'; T in [DT_UINT64]\r\n device='CPU'; T in [DT_INT64]\r\n device='CPU'; T in [DT_UINT32]\r\n device='CPU'; T in [DT_UINT16]\r\n device='CPU'; T in [DT_INT16]\r\n device='CPU'; T in [DT_UINT8]\r\n device='CPU'; T in [DT_INT8]\r\n device='CPU'; T in [DT_INT32]\r\n device='CPU'; T in [DT_HALF]\r\n device='CPU'; T in [DT_BFLOAT16]\r\n device='CPU'; T in [DT_FLOAT]\r\n device='CPU'; T in [DT_DOUBLE]\r\n device='CPU'; T in [DT_COMPLEX64]\r\n device='CPU'; T in [DT_COMPLEX128]\r\n device='CPU'; T in [DT_BOOL]\r\n device='CPU'; T in [DT_STRING]\r\n device='CPU'; T in [DT_RESOURCE]\r\n device='CPU'; T in [DT_VARIANT]\r\n [Op:Transpose]\r\n``"</li><li>"My customized OP gives incorrect outputs on GPUs since` tf-nightly 2.13.0.dev20230413 ### Issue type\n\nBug\n\n### Have you reproduced the bug with TensorFlow Nightly?\n\nYes\n\n### Source\n\nbinary\n\n### TensorFlow version\n\n2.13\n\n### Custom code\n\nYes\n\n### OS platform and distribution\n\nfedora 36\n\n### Mobile device\n\n_No response_\n\n### Python version\n\n3.11.4\n\n### Bazel version\n\n_No response_\n\n### GCC/compiler version\n\n_No response_\n\n### CUDA/cuDNN version\n\n_No response_\n\n### GPU model and memory\n\n_No response_\n\n### Current behavior?\n\nI have a complex program based on TensorFlow with several customized OPs. These OPs were created following https://www.tensorflow.org/guide/create_op. Yesterday TF 2.13.0 was released, but after I upgraded to 2.13.0, I found that one of my customized OP gives incorrect results on GPUs and still has the correct outputs on CPUs.\r\n\r\nThen I tested many tf-nightly `versions and found that` tf-nightly 2.13.0.dev20230412 `works but` tf-nightly 2.13.0.dev20230413` fails. So the situation is shown in the following table:\r\n	version	CPU	GPU	\r\n	--------	---------	-----------	\r\n	tensorflow 2.12.0	Correct	Correct	\r\n	tensorflow 2.13.0	Correct	Incorrect	\r\n	tf-nightly 2.13.0.dev20230412	Correct	Correct	\r\n	tf-nightly 2.13.0.dev20230413	Correct	Incorrect	\r\n\r\nI'd like to know what changed between April 12th and 13th related to the customized OPs. This can be a breaking change to downstream applications or an internal bug. Thanks!\r\n\r\nHere is a quick link for commits between April 12th and 13th:\r\nhttps://github.com/tensorflow/tensorflow/commits/master?before=525da8a93eca846e32e5c41eddc0496b25a2ef5b+770\r\n\n\n### Standalone code to reproduce the issue\n\n```shell\nIndeed, the reason is still unclear to me, so it is hard to create a minimal example.\r\n\r\nThe code of our customized OPs is https://github.com/deepmodeling/deepmd-kit/blob/37fd8d193362f91c925cf7c2f3a58b97dc921b27/source/op/prod_force_multi_device.cc#L49-L166\n```\n\n\n### Relevant log output\n\nNo response"</li></ul>

Uses

Direct Use for Inference

First install the SetFit library:

bash

pip install setfit

Then you can load this model and run inference.

python

from setfit import SetFitModel

# Download from the 🤗 Hub
model = SetFitModel.from_pretrained("setfit_model_id")
# Run inference
preds = model("Data init API for TFLite Swift <details><summary>Click to expand!</summary> 
 
 ### Issue Type

Feature Request

### Source

source

### Tensorflow Version

2.8+

### Custom Code

No

### OS Platform and Distribution

_No response_

### Mobile device

_No response_

### Python version

_No response_

### Bazel version

_No response_

### GCC/Compiler version

_No response_

### CUDA/cuDNN version

_No response_

### GPU model and memory

_No response_

### Current Behaviour?

The current Swift API only has init functions from files on disk unlike the Java (Android) API which has a byte buffer initializer. It'd be convenient if the Swift API could initialize Interpreters from Data.



### Standalone code to reproduce the issue

No code. This is a feature request



### Relevant log output

_No response_</details>")

<!--

Downstream Use

List how someone could finetune this model on their own dataset. -->

<!--

Out-of-Scope Use

List how the model may foreseeably be misused and address what users ought not to do with the model. -->

<!--

Bias, Risks and Limitations

What are the known or foreseeable issues stemming from this model? You could also flag here known failure cases or weaknesses of the model. -->

<!--

Recommendations

What are recommendations with respect to the foreseeable issues? For example, filtering explicit content. -->

Training Details

Training Set Metrics

Training set	Min	Median	Max
Word count	5	353.7433	6124

Label	Training Sample Count
bug	200
feature	200
question	200

Training Hyperparameters

—batch_size: (16, 2)
—num_epochs: (1, 1)
—max_steps: -1
—sampling_strategy: oversampling
—num_iterations: 20
—bodylearningrate: (2e-05, 1e-05)
—headlearningrate: 0.01
—loss: CosineSimilarityLoss
—distancemetric: cosinedistance
—margin: 0.25
—endtoend: False
—use_amp: False
—warmup_proportion: 0.1
—seed: 42
—evalmaxsteps: -1
—loadbestmodelatend: False

Training Results

Epoch	Step	Training Loss	Validation Loss
0.0007	1	0.1719	-
0.0067	10	0.2869	-
0.0133	20	0.2513	-
0.02	30	0.1871	-
0.0267	40	0.2065	-
0.0333	50	0.2302	-
0.04	60	0.1645	-
0.0467	70	0.1887	-
0.0533	80	0.1376	-
0.06	90	0.1171	-
0.0667	100	0.1303	-
0.0733	110	0.121	-
0.08	120	0.1126	-
0.0867	130	0.1247	-
0.0933	140	0.1764	-
0.1	150	0.0401	-
0.1067	160	0.1571	-
0.1133	170	0.0186	-
0.12	180	0.0501	-
0.1267	190	0.1003	-
0.1333	200	0.0152	-
0.14	210	0.0784	-
0.1467	220	0.1423	-
0.1533	230	0.1313	-
0.16	240	0.0799	-
0.1667	250	0.0542	-
0.1733	260	0.0426	-
0.18	270	0.047	-
0.1867	280	0.0062	-
0.1933	290	0.0085	-
0.2	300	0.0625	-
0.2067	310	0.095	-
0.2133	320	0.0262	-
0.22	330	0.0029	-
0.2267	340	0.0097	-
0.2333	350	0.063	-
0.24	360	0.0059	-
0.2467	370	0.0016	-
0.2533	380	0.0025	-
0.26	390	0.0033	-
0.2667	400	0.0006	-
0.2733	410	0.0032	-
0.28	420	0.0045	-
0.2867	430	0.0013	-
0.2933	440	0.0011	-
0.3	450	0.001	-
0.3067	460	0.0044	-
0.3133	470	0.001	-
0.32	480	0.0009	-
0.3267	490	0.0004	-
0.3333	500	0.0006	-
0.34	510	0.001	-
0.3467	520	0.0003	-
0.3533	530	0.0008	-
0.36	540	0.0003	-
0.3667	550	0.0023	-
0.3733	560	0.0336	-
0.38	570	0.0004	-
0.3867	580	0.0003	-
0.3933	590	0.0006	-
0.4	600	0.0008	-
0.4067	610	0.0011	-
0.4133	620	0.0002	-
0.42	630	0.0004	-
0.4267	640	0.0005	-
0.4333	650	0.0601	-
0.44	660	0.0003	-
0.4467	670	0.0003	-
0.4533	680	0.0006	-
0.46	690	0.0005	-
0.4667	700	0.0003	-
0.4733	710	0.0006	-
0.48	720	0.0001	-
0.4867	730	0.0002	-
0.4933	740	0.0002	-
0.5	750	0.0002	-
0.5067	760	0.0002	-
0.5133	770	0.0016	-
0.52	780	0.0001	-
0.5267	790	0.0005	-
0.5333	800	0.0004	-
0.54	810	0.0039	-
0.5467	820	0.0031	-
0.5533	830	0.0008	-
0.56	840	0.0003	-
0.5667	850	0.0002	-
0.5733	860	0.0002	-
0.58	870	0.0002	-
0.5867	880	0.0001	-
0.5933	890	0.0004	-
0.6	900	0.0002	-
0.6067	910	0.0008	-
0.6133	920	0.0005	-
0.62	930	0.0005	-
0.6267	940	0.0002	-
0.6333	950	0.0001	-
0.64	960	0.0002	-
0.6467	970	0.0007	-
0.6533	980	0.0002	-
0.66	990	0.0002	-
0.6667	1000	0.0002	-
0.6733	1010	0.0002	-
0.68	1020	0.0002	-
0.6867	1030	0.0002	-
0.6933	1040	0.0004	-
0.7	1050	0.0076	-
0.7067	1060	0.0002	-
0.7133	1070	0.0002	-
0.72	1080	0.0001	-
0.7267	1090	0.0002	-
0.7333	1100	0.0001	-
0.74	1110	0.0365	-
0.7467	1120	0.0002	-
0.7533	1130	0.0002	-
0.76	1140	0.0003	-
0.7667	1150	0.0002	-
0.7733	1160	0.0002	-
0.78	1170	0.0004	-
0.7867	1180	0.0001	-
0.7933	1190	0.0001	-
0.8	1200	0.0001	-
0.8067	1210	0.0001	-
0.8133	1220	0.0002	-
0.82	1230	0.0002	-
0.8267	1240	0.0001	-
0.8333	1250	0.0001	-
0.84	1260	0.0002	-
0.8467	1270	0.0002	-
0.8533	1280	0.0	-
0.86	1290	0.0002	-
0.8667	1300	0.032	-
0.8733	1310	0.0001	-
0.88	1320	0.0001	-
0.8867	1330	0.0001	-
0.8933	1340	0.0003	-
0.9	1350	0.0001	-
0.9067	1360	0.0001	-
0.9133	1370	0.0001	-
0.92	1380	0.0001	-
0.9267	1390	0.0001	-
0.9333	1400	0.0001	-
0.94	1410	0.0001	-
0.9467	1420	0.0001	-
0.9533	1430	0.031	-
0.96	1440	0.0001	-
0.9667	1450	0.0003	-
0.9733	1460	0.0001	-
0.98	1470	0.0001	-
0.9867	1480	0.0001	-
0.9933	1490	0.0001	-
1.0	1500	0.0001	-

Framework Versions

—Python: 3.10.12
—SetFit: 1.0.3
—Sentence Transformers: 3.0.1
—Transformers: 4.39.0
—PyTorch: 2.3.0+cu121
—Datasets: 2.20.0
—Tokenizers: 0.15.2

Citation

BibTeX

bibtex

@article{https://doi.org/10.48550/arxiv.2209.11055,
    doi = {10.48550/ARXIV.2209.11055},
    url = {https://arxiv.org/abs/2209.11055},
    author = {Tunstall, Lewis and Reimers, Nils and Jo, Unso Eun Seo and Bates, Luke and Korat, Daniel and Wasserblat, Moshe and Pereg, Oren},
    keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
    title = {Efficient Few-Shot Learning Without Prompts},
    publisher = {arXiv},
    year = {2022},
    copyright = {Creative Commons Attribution 4.0 International}
}

<!--

Glossary

Clearly define terms in order to be accessible across audiences. -->

<!--

Model Card Authors

Lists the people who create the model card, providing recognition and accountability for the detailed work that goes into its construction. -->

<!--

Model Card Contact

Provides a way for people who have updates to the Model Card, suggestions, or questions, to contact the Model Card authors. -->