Intel® Transfer Learning Tool v0.7.0

New Features

• Gaudi Integration for Pytorch TLT Image Anomaly Detection
• Gaudi Integration for Pytorch TLT Image Classification (Disease Prediction)
• Enabled 58 torchvision models for compatibility with TLT Anomaly Detection
• Distributed fine tuning example with Slurm
• Enabled bfloat16 automatic option to Pytorch image classification models for training and inference

Bug Fixes / Improvements

• Updated TLT installer for better dependency management
• Added security context to Kubernetes Llama2 fine tuning template to allow running as a non-root user
• Updated the Hugging Face token to be volume mounted in to the container when fine tuning with Kubernetes
• Updated the Llama2 fine tuning Docker image to use Python 3.10 and updated library versions

Validated Configuration

• Ubuntu 20.04/22.04 LTS
• Intel® Optimization for TensorFlow 2.14.0
• Intel® Extension for PyTorch 2.2.0
• PyTorch 2.2.0
• TorchVision 0.17.0
• Intel® Neural Compressor 2.4.1
• TensorFlow Hub 0.15.0

Known Issues

• TLT distributed module currently has an issue with horovod installation due to torch version conflict.
• dinov2 models are enabled for image anomaly detection use case, but not currently compatible for training and inference yet

GitHub pages

https://intelai.github.io/transfer-learning/v0.7.0/

Intel® Transfer Learning Tool v0.6.0

New Features

• Support for LLM Text Generation in TLT API and CLI with custom datasets, including support for 5 new LLM models from HuggingFace
• New notebooks for PyTorch text generation and HuggingFace image classification
• Added ability to enable auto mixed precision during evaluation and prediction for image classification as well as TensorFlow text classification

Bug Fixes / Improvements

• Bug fixes for CLI text classification evaluation command and metrics
• Improvements to documentation for setup and installation of notebook requirements
• Bug fix for PyTorch image classification batch not shuffling properly
• Fixes for the grouped file and dataset columns in the e2e multimodal notebook
• Fixed batching and accuracy bugs in TensorFlow Multinode and PyTorch Image Classification
• Fixed minor bug involving Torchvision model weights

New Intel® Transfer Learning Tool API tutorial Notebooks:

• Text Generation Instruction Tuning LLMs Transfer Learning with PyTorch

New Native Framework Notebooks:

• Large Language Model Instruction Tuning with PyTorch
• Image Classification Transfer Learning with HuggingFace

Validated Configuration

• Ubuntu 20.04/22.04 LTS
• Intel® Optimization for TensorFlow 2.12.0
• Intel® Extension for PyTorch 1.13.100
• PyTorch 1.13.1
• TorchVision 0.14.1
• Intel® Neural Compressor 2.1.1
• TensorFlow Hub 0.15.0

GitHub pages

https://intelai.github.io/transfer-learning/v0.6.0/

Intel® Transfer Learning Tool v0.5.0

New Features

• Support for Hugging Face text classification models using TensorFlow
• Support for Keras model preprocessors
• Disributed training Dockerfiles and Helm chart
• Optimized CPU settings for TensorFlow and PyTorch
• Intel® Neural Compressor 2.1.1 API migrations
• Virtualenv support for multinode training with PyTorch
• Added overwrite option for quantization and optimize_graph APIs
• Improved AI Kit compatibility
• Updated vision workflow
• Improved design and usability of documentation

New Notebooks:

• Two performance comparison notebooks for image classification and text classification

Bug fixes:

• multi-class text classification with TensorFlow

Validated configuration

• Ubuntu 20.04/22.04 LTS
• Intel® Optimization for TensorFlow 2.12.0
• Intel® Extension for PyTorch 1.13.100
• PyTorch 1.13.1
• TorchVision 0.14.1
• Intel® Neural Compressor 2.1.1
• TensorFlow Hub 0.13.0

Known issues

• The CLI's benchmarking command fails for some models, such as TensorFlow text classification. Recommend using the API for advanced workflows.

GitHub pages

https://intelai.github.io/transfer-learning/v0.5.0/

Intel® Transfer Learning Tool v0.4.0

New Features

• Support for multi-node training with TensorFlow and Horovod
• Improved SimSiam feature extraction for Anomaly Detection
• Added multi-class datasets to Text Classification Notebooks
• Separate model downloader
• Quantization for TensorFlow Text Classification
• User defined splits for custom image classification models
• Anomaly Detection Workflow
• Vision Workflow
• CutPaste feature extraction for Anomaly Detection
• Added support for PyTorch Hub models
• Added support for Keras Applications
• Improved freeze/unfreeze and list trainable parameters features
• Improved documentation and list of supported models in Models.md

Featured Notebooks:

• Intel® Transfer Learning Tool API tutorial Notebooks:

  • Image Classification Transfer Learning with PyTorch
  • Image Classification Transfer Learning with TensorFlow
  • Text Classification fine-tuning with PyTorch
  • Text Classification fine-tuning with TensorFlow
  • Image Anomaly Detection Transfer Learning with PyTorch

• Intel® Transfer Learning Tool API end-to-end Notebooks:

  • Image Anomaly Detection with PyTorch
  • Document-Level Sentiment Analysis using PyTorch
  • Medical Imaging Classification using TensorFlow
  • Multimodal Cancer Detection using TensorFlow and PyTorch
  • Remote Sensing Image Scene Classification using TensorFlow

• Native Framework Notebooks:

  • Transfer Learning for Image Classification using PyTorch
  • Transfer Learning for Image Classification using TensorFlow
  • Text Classification fine-tuning with PyTorch
  • Text Classification fine-tuning with TensorFlow
  • Question Answering fine-tuning using TensorFlow

Validated configuration

• Ubuntu 20.04/22.04 LTS
• Intel® Optimization for TensorFlow 2.11.0
• Intel® Extension for PyTorch 1.13.0
• PyTorch 1.13.1
• TorchVision 0.14.1
• Intel® Neural Compressor 2.0
• TensorFlow Hub 0.12.0

Known issues

• Multimodal Cancer Detection end-to-end pipeline notebook fails during quantization

Intel® Transfer Learning Tool v0.3.0

Initial release of the Intel® Transfer Learning Tool (TLT). TLT makes transfer learning workflows easier and faster by providing a command-line interface (CLI) and a Python library (API) that leverages public model hubs, Intel optimized deep learning frameworks, and your custom dataset to efficiently generate new deep learning models optimized for deployment on Intel CPUs.

Features

• Low-code API and no-code CLI for:
  • Image classification transfer learning using Intel® Optimization for TensorFlow
    • Support for 19 models from TF Hub
    • Support for custom datasets and the TensorFlow datasets catalog
    • Post-training quantization and benchmarking using Intel® Neural Compressor, when using custom datasets
    • FP32 graph optimization using Intel® Neural Compressor
    • Reduced training time with auto-mixed precision on Intel® third or fourth generation Xeon processors
  • Image classification transfer learning using PyTorch and Intel® Extension for PyTorch
    • Support for 60 models from Torchvision
    • Support for custom datasets and select Torchvision datasets
    • Post-training quantization using Intel® Neural Compressor, when using custom datasets
  • Text Classification transfer learning using Intel Optimization for TensorFlow
    • Support for 26 models from TF Hub
    • Support for custom datasets and the TensorFlow datasets catalog
    • Post-training quantization and benchmarking using Intel® Neural Compressor, when using custom datasets
    • Reduced training time with auto-mixed precision on Intel® third or fourth generation Xeon processors
  • Text Classification transfer learning using PyTorch and Intel® Extension for PyTorch
    • Support for 4 models from HuggingFace
    • Support for custom datasets and select Hugging Face datasets
    • Post-training quantization using Intel® Neural Compressor, when using custom datasets
  • Anomaly Detection transfer learning using PyTorch
    • Support for 5 models from Torchvision
    • Support for custom datasets
• Additional Features:
  • Dataset scaling, cropping, batching, splitting, and augmentation
  • APIs for training, prediction, and evaluation
  • Export model for deployment or resume training from checkpoints
  • Reproducible experiments
  • Support for user-provided pre-trained models
  • Model customization options:
    • Additional dense layers
    • Selective layer freezing and unfreezing for PyTorch models
  • Advanced training options:
    • Early stopping
    • Custom learning rate and learning rate decay
    • Auto-evaluation at the end of each training epoch
    • Configurable optimizers & loss functions

New Notebooks:

• Tutorials:

  • BERT Binary Text Classification with TF Hub
  • BERT Binary Text Classification with PyTorch
  • Image Classification with TensorFlow
  • Image Classification with TensorFlow using Graph Optimization
  • Image Classification with PyTorch

• End-to-End notebooks:

  • Document-Level Sentiment Analysis (SST2) using PyTorch
  • Medical Imaging Classification (Colorectal histology) using TensorFlow
  • Remote Sensing Image Scene Classification (Resisc) using TensorFlow
  • Multimodal Cancer Detection using TensorFlow and PyTorch
  • Anomaly Detection using PyTorch

Validated configuration

• Ubuntu 20.04 LTS
• Intel® Optimization for TensorFlow 2.10.0
• Intel® Extension for PyTorch 1.13.0
• PyTorch 1.13.1
• Intel® Neural Compressor 1.14.2
• TensorFlow Hub 0.12.0
• Torchvision 0.14.1