Targeting the IPU from TensorFlow 1ΒΆ
- 1. Introduction
- 2. Tutorial
- 3. Targeting the Poplar XLA device
- 4. Compiling and pre-compiling executables
- 5. Training a model
- 6. Efficient IPU I/O
- 7. Example using IPUEstimator
- 8. Example using IPUPipelineEstimator
- 9. Distributed training
- 10. Half-precision floating point and stochastic rounding
- 11. IPU-optimised operations
- 12. IPU Outlined Functions
- 13. Writing custom operations
- 14. IPU host embeddings
- 15. IPU embedded application runtime
- 16. Retrieving information about compilation and execution
- 17. IPU TensorFlow Addons
- 18. TensorFlow API changes
- 19. TensorFlow Python API
- 19.1. Operations and utilities related to the Graphcore IPU
- 19.2. Compiler interface
- 19.3. Scoping contexts
- 19.4. Infeed queue
- 19.5. Outfeed queue
- 19.6. General utilities
- 19.7. Configuration utilities
- 19.8. Looping utilities
- 19.9. Distributed training
- 19.10. Horovod
- 19.11. Serving utilities
- 19.12. Datasets
- 19.13. Estimators
- 19.14. Keras layers
- 19.15. Operators
- 19.15.1. Control flow operations.
- 19.15.2. Custom operations
- 19.15.3. Functional operators
- 19.15.4. Image operations
- 19.15.5. Graphcore utility operations
- 19.15.6. IPU specific maths operations
- 19.15.7. Pipelining operators
- 19.15.8. Popnn primitive neural network operators
- 19.15.9. Popnn normalization operators
- 19.15.10. Popnn recurrent neural network operators
- 19.15.11. Popops all to all and all gather operators
- 19.15.12. Popops cross replica operators
- 19.15.13. Popops embedding operators
- 19.15.14. Popops reduce scatter operator
- 19.15.15. Popops within replica operators
- 19.15.16. Poprand operators
- 19.15.17. Utility operations to be used in replicated mode
- 19.15.18. Slicing operators
- 19.15.19. Statistics operators
- 19.15.20. Embedded application runtime
- 19.16. Optimisers
- 19.17. Sharding
- 20. TensorFlow operators supported by the IPU
- 21. IPU TensorFlow Addons API changes
- 22. IPU TensorFlow Addons Python API
- 23. Resources
- 24. Trademarks & copyright