Spherical CNNs
Authors: Taco S. Cohen, Mario Geiger, Jonas Köhler, Max Welling
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 EXPERIMENTS In a first sequence of experiments, we evaluate the numerical stability and accuracy of our algorithm. In a second sequence of experiments, we showcase that the new cross-correlation layers we have introduced are indeed useful building blocks for several real problems involving spherical signals. Our examples for this are recognition of 3D shapes and predicting the atomization energy of molecules. |
| Researcher Affiliation | Academia | Taco S. Cohen University of Amsterdam Mario Geiger EPFL Jonas Köhler University of Amsterdam Max Welling University of Amsterdam & CIFAR |
| Pseudocode | No | The paper describes algorithms verbally but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/jonas-koehler/s2cnn. |
| Open Datasets | Yes | We created two instances of this dataset: one in which each digit is projected on the northern hemisphere and one in which each projected digit is additionally randomly rotated. (Spherical MNIST) The SHREC17 task (Savva et al., 2017) contains 51300 3D models taken from the Shape Net dataset (Chang et al., 2015) In the QM7 task (Blum and Reymond, 2009; Rupp et al., 2012) the atomization energy of molecules has to be predicted from geometry and charges. |
| Dataset Splits | No | The paper mentions training and testing, but does not specify validation splits or percentages for any dataset. |
| Hardware Specification | No | The paper mentions memory usage but does not provide specific hardware details like CPU/GPU models or types of computing resources used for experiments. |
| Software Dependencies | No | The paper mentions 'Py Torch code' but does not provide specific version numbers for PyTorch or any other software dependencies required for reproducibility. |
| Experiment Setup | Yes | Architecture and Hyperparameters As a baseline model, we use a simple CNN with layers conv-Re LU-conv-Re LU-FC-softmax, with filters of size 5 5, k = 32, 64, 10 channels, and stride 3 in both layers ( 68K parameters). We compare to a spherical CNN with layers S2conv-Re LU-SO(3)conv-Re LUFC-softmax, bandwidth b = 30, 10, 6 and k = 20, 40, 10 channels ( 58K parameters). Our network consists of an initial S2conv-BN-Re LU block followed by two SO(3)conv-BN-Re LU blocks. ... We used 50, 70, and 350 features for the S2 and the two SO(3) layers, respectively. Further, in each layer we reduce the resolution b, from 128, 32, 22 to 7 in the final layer. We use a deep Res Net style S2CNN. Each Res Net block is made of S2/SO(3)conv-BN-Re LU-SO(3)conv-BN after which the input is added to the result. We share weights among atoms making filters permutation invariant, by pushing the atom dimension into the batch dimension. In each layer we downsample the bandwidth, while increasing the number of features F. |