Skip to content

Latest commit

 

History

History
40 lines (36 loc) · 2.48 KB

hyperparams.md

File metadata and controls

40 lines (36 loc) · 2.48 KB
Raw

‼️ Hyperparameters to set for implementing the paper

This file explains the hyperparameters to set in the scripts (definitions here)

  • manualSeed : random seed for code reproducibility

  • cls_weight : to decide how much weight to give to the classification loss, i.e., $\alpha_2$ in Eq. 6 of the paper (NOTE: cls_weight_unseen is not used in our work)

  • nclass_all : total number of classes (including background)

  • syn_num : number of features to synthesize per object class. An ablation study of its effect on model performance can be seen in Figure 2(b) of the paper

  • val_every : validation rule

  • cuda : setting the GPU flag

    Regarding the WGAN network

  • netG_name : generator network

  • netD_name : discriminator network

  • nepoch : total epochs to train

  • ngh : linear layer neurons for generator network (ablations can be done to check for possible effects of network strength by varying the number of neurons)

  • ndh : linear layer neurons for discriminator network (ablations can be done to check for possible effects of network strength by varying the number of neurons)

  • lambda1 : for the gradient penalty trick of Wasserstein GANs (default value taken from the original NIPS 2017 paper)

  • critic_iter : number of critic (discriminator) iterations per generator iteration for Wasserstein GANs (default value taken from the original NIPS 2017 paper)

  • nz : dimension of the random noise vector

  • gan_epoch_budget : random pick subset of features to train GAN

  • lr : learning rate for training GAN

  • lr_step : to decay the learning rate every lr_step epochs

  • lz_ratio : mode seeking loss weight, i.e., $\alpha_3$ in our paper

  • regressor_lamda : $\alpha_4$ in our paper

  • triplet_lamda : $\alpha_5$ in our paper

  • pretrain_regressor : The MLP for maintaining cyclic consistency

Others

  • dataset : name of dataset
  • batch_size : batch size for training
  • attSize : dimension of semantic vector
  • resSize : dimension of visual embedding
  • lr_cls : learning rate for training classifier
  • pretrain_classifier : output of step 2
  • class_embedding : class-attribute matrix
  • dataroot : root directory for data
  • testsplit/trainsplit : features extracted for test/train sets
  • classes_split : seen/unseen split of classes
  • outname : output results here