Synthesizing Programs for Images using Reinforced Adversarial Learning
Authors: Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, S. M. Ali Eslami, Oriol Vinyals
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our approach on three real-world and one synthetic image dataset. The first, MNIST (Le Cun et al., 1998), is regarded as a standard sanity check for newly proposed generative models. It contains 70,000 examples of handwritten digits, of which 10,000 constitute a test set. In the second set of experiments, we train an agent to generate the strokes for a given target digit, and we compare two kinds of rewards discussed in Section 3.3: fixed ℓ2-distance and the discriminator score. The results are summarized in Figure 8a (blue curves). |
| Researcher Affiliation | Collaboration | 1Montreal Institute for Learning Algorithms, Montr eal, Canada 2Deep Mind, London, United Kingdom. |
| Pseudocode | No | The paper describes the SPIRAL architecture and training process in textual descriptions and diagrams (Figure 2), but it does not include formal pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions "A video of the agent can be found at https://youtu.be/i Syvw Awa7vk." but does not provide any concrete access to the source code for the methodology described. |
| Open Datasets | Yes | We validate our approach on three real-world and one synthetic image dataset. The first, MNIST (Le Cun et al., 1998)... The second dataset, OMNIGLOT (Lake et al., 2015)... Since both MNIST and OMNIGLOT represent a restricted line drawing domain, we diversify our set of experiments by testing the proposed method on CELEBA (Liu et al., 2015). |
| Dataset Splits | No | The paper mentions for MNIST: "It contains 70,000 examples of handwritten digits, of which 10,000 constitute a test set." However, it does not explicitly provide training/validation/test splits with percentages, sample counts, or specific predefined split details for all datasets or a general splitting methodology for reproducibility. |
| Hardware Specification | No | The paper mentions the use of GPUs ("a separate GPU for adversarial training", "off-policy GPU learner") but does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'libmypaint' and 'Mu Jo Co' environments, but it does not specify version numbers for these or any other software dependencies, libraries, or frameworks used in the experiments. |
| Experiment Setup | No | The paper describes aspects of the training process, such as using A2C, a Wasserstein discriminator, and the number of steps per episode (20 steps). However, it does not provide specific hyperparameter values like learning rates, batch sizes, number of epochs, or optimizer-specific settings for reproducibility. |