Compositional Zero-Shot Learning via Fine-Grained Dense Feature Composition

Authors: Dat Huynh, Ehsan Elhamifar

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on four popular datasets of Deep Fashion, AWA2, CUB, and SUN, showing that our method significantly improves the state of the art.
Researcher Affiliation Academia Dat Huynh Northeastern University huynh.dat@northeastern.edu Ehsan Elhamifar Northeastern University eelhami@ccs.neu.edu
Pseudocode Yes Algorithm 1 Composing Dense Features
Open Source Code No The paper does not provide concrete access to source code for the described methodology.
Open Datasets Yes We conduct experiments on four popular datasets: Deep Fashion [4], AWA2 [68], CUB [69], and SUN [70].
Dataset Splits Yes We follow the data splits of [2] for Deep Fashion and of [68] for AWA2, CUB, and SUN.
Hardware Specification Yes We implement our framework in Py Torch and optimize it using RMSprop[74] with the default setting, learning rate of 0.0001 and batch size of 50 having an equal number of samples per class. We pre-train DAZLE on seen classes and use it to compose dense features for at most 2000 and 4000 iterations, respectively, on a NVIDIA V100 GPU.
Software Dependencies No The paper mentions 'Py Torch' and 'RMSprop' but does not specify version numbers for these or any other software dependencies.
Experiment Setup Yes We implement our framework in Py Torch and optimize it using RMSprop[74] with the default setting, learning rate of 0.0001 and batch size of 50 having an equal number of samples per class. We pre-train DAZLE on seen classes and use it to compose dense features for at most 2000 and 4000 iterations, respectively... To prevent seen class bias, we add a margin of 1 to unseen class scores and 1 to seen class scores... We experiment in two settings: i) using pre-trained Image Net features (pre-trained setting) and ii) fine-tuning the Res Net backbone on each dataset... We use the feature map of the last convolutional layer whose size is 7 7 2048... To measure the robustness of our method, we fix the hyperparameters at T = 5, k = 5, b = 50 (T = 10, k = 10, b = 50) for the pretrained (fine-tuned) setting on all datasets.