Predictive synthesis of API-centric code

Abstract

Today's programmers, especially data science practitioners, make heavy use of data-processing libraries (APIs) such as PyTorch, Tensorflow, NumPy, and the like. Program synthesizers can provide significant coding assistance to this community of users; however program synthesis also can be slow due to enormous search spaces.In this work, we examine ways in which machine learning can be used to accelerate enumerative program synthesis. We present a deep-learning-based model to predict the sequence of API functions that would be needed to go from a given input to a desired output, both being numeric vectors. Our work is based on two insights. First, it is possible to learn, based on a large number of input-output examples, to predict the likely API function needed. Second, and importantly, it is also possible to learn to compose API functions into a sequence, given an input and the desired final output, without explicitly knowing the intermediate values.We show that we can speed up an enumerative synthesizer by using predictions from our model variants. These speedups significantly outperform previous ways (e.g. Deep-Coder) in which researchers have used ML models in enumerative synthesis.CCS Concepts: • Software and its engineering → Programming by example; Automatic programming; API languages.

C2TACO: Lifting Tensor Code to TACO

“…Early work looked at generating Excel commands from a few examples [31]. The same concept and has been used for other tasks [21,73], including generating PyTorch or TensorFlow code from tensor inputs [46,59]. TF-Coder [59] takes as input a single userprovided example to generate equivalent TensorFlow code using type constraints and bottom-up enumerative synthesis.…”

“…TF-Coder [59] takes as input a single userprovided example to generate equivalent TensorFlow code using type constraints and bottom-up enumerative synthesis. Alternative schemes [16,46] use deep learning models trained on IO samples to guide the generation of code.…”

mlirSynth: Automatic, Retargetable Program Raising in Multi-Level IR Using Program Synthesis

“…The use of input/output examples to synthesize high-level code has been explored in a number of projects [32], [58], [21], [20]. It has been used to generate pytorch or tensor-flow code from tensor inputs [49], [43]. TF-coder [49] uses type-constraints and equivalences to efficiently apply enumerative program search while [43] uses a DeepCoder [11] style predictive model to guide code generation.…”

“…It has been used to generate pytorch or tensor-flow code from tensor inputs [49], [43]. TF-coder [49] uses type-constraints and equivalences to efficiently apply enumerative program search while [43] uses a DeepCoder [11] style predictive model to guide code generation. AutoPandas [13] uses a more powerful graph neural network based model to guide the generation of Panda code.…”