Differentiable Program Synthesis

---

Introduction

We propose a differentiable approach for synthesizing differentiable programs. More details on Differentiable Synthesis of Program Architectures.

Requirements

• Python 3.6+
• PyTorch 1.4.0+
• scikit-learn 0.22.1+
• Numpy
• tqdm

Code Structures

• train_nas.py includes the main function to synthesize differentiable programs.
• eval_test.py is used for evaluation. Given a synthesized program, it calculates the program's F1 score and accuracy on a dataset.
• dsl_*.py defines the parameterized domain-specific languages for the benchmarks.
• dsl/ includes the semantics of the predefined domain-specific languages.
• algorithms/ contains the main synthesis algorithm (nas.py).
• program_graph.py defines our algorithm's learning representation.
• utils/ includes useful code for data loading, loss functions, etc.

Run the Code

Program Search

• To search a programmatic classifier for the Crim13 dataset, run sh train_nas.sh;
• To search a programmatic classifier for the Fly-vs-fly dataset, run sh train_nas_fly.sh;
• To search a programmatic classifier for the Basketball dataset, run sh train_nas_basketball.sh;
• To search a programmatic classifier for the SK152 dataset, run sh train_nas_sk152.sh.

Program Evaluation

To evaluate a synthesized program on its test dataset, use sh eval_test.sh.

Comparision against NEAR

To run NEAR with our released datasets, please download its source code and follow the instructions from here.

Data

All the datasets can be downloaded from here.

More Details about dPads's Command-line Arguments for Sequence Classification

--input_type and --output_type defines if the program input/output data is over "list" or "atom". "list" refers to sequences of data and "atom" refers to values.

--input_size and --output_size defines the dimension of each frame in an input (output) sequence.

--num_label defines the number of classification categories for a sequence classification task.

--lossfxn defines the loss function for training. The choices include "crossentropy", "bcelogits" and "softf1".

--max_depth defines the maximum depth of the abstract syntax tree of any synthesized program.

--symbolic_epochs defines the number of epochs used for program search on top of a program derivation graph.

--neural_epochs defines the number of epochs used for program selection from a trained program derivation graph.

--train_valid_split defines the ratio to split the training dataset for synthesizing program structures and program parameters.

--batch_size defines the batch size for training.

--node_share and --graph_unfold. Please refer to the ablation study section of our paper.

Experiment Results

We report dPad's results on the four datesets, averaged over 5 random seeds (0,1000,2000,3000,4000):

	Crim13	Fly-vs-fly	Basketball	SK152
F1 Scores	0.458	0.887	0.945	0.337
Accuracy	0.812	0.853	0.939	0.337
Time (mins)	147.87	348.25	174.68	162.70