Optimization Planning for 3D ConvNets
Authors: Zhaofan Qiu, Ting Yao, Chong-Wah Ngo, Tao Mei
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on seven public video recognition benchmarks demonstrate the advantages of our proposal. |
| Researcher Affiliation | Collaboration | 1JD AI Research, Beijing, China 2School of Computing and Information Systems, Singapore Management University, Singapore. |
| Pseudocode | No | The paper describes methods and processes but does not include any labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | The experiments are conducted on HMDB51 (Kuehne et al., 2011), UCF101 (Soomro et al., 2012), Activity Net (Caba Heilbron et al., 2015), SS-V1 (Goyal et al., 2017), SS-V2, Kinetics-400 (Carreira & Zisserman, 2017) and Kinetics-600 (Carreira et al., 2018) datasets. |
| Dataset Splits | Yes | For optimization planning, the original training set of each dataset is split into two parts for learning the network weights and validating the performance, respectively. We construct this internal validation set with the same size as the original validation/test set. |
| Hardware Specification | Yes | The time cost for grid search/optimization planning is reported with 8 NVidia Titan V GPUs in parallel. |
| Software Dependencies | No | The paper mentions that the network training is implemented on 'PyTorch framework' but does not provide a specific version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | For optimization planning, we set the number of the choices for both input clip length Nl and learning rate Nr as 3, and utilize the extended transition graph introduced in Section 3.2. The candidate values of input clip length {l1, l2, l3} and learning rate {r1, r2, r3} for each dataset are summarized in Table 2. Specifically, on SS-V1, SS-V2, Kinetics-400 and Kinetics-600 datasets, the base learning rate is set as 0.04 and the dropout ratio is fixed as 0.5. |