ZSL

Baseline

(author: Chen Zhuo)
DeViSE: "Devise: A deep visual-semantic embedding model" (pytorch)
CONSE: "Zero-shot learning by convex combination of semantic embeddings" (Matlab)
SAE: "Semantic autoencoder for zero-shot learning" (pytorch)
SYNC: "Synthesized classifiers for zero-shot learning" (Matlab)

---

(author: Geng Yuxia)
GCNZ: "Zero-shot recognition via semantic embeddings and knowledge graphs" (python2 + tensorflow)
DGP: "Rethinking knowledge graph propagation for zero-shot learning" (python3 + pytorch)
GAZSL: "A generative adversarial approach for zero-shot learning from noisy texts" (pytorch)
LisGAN: "Leveraging the invariant side of generative zero-shot learning" (pytorch)

Setting

DeViSE, CONSE, SAE and SYNC run with two semantic embeddings (w2v: word embedding; g2v: trained kg embedding)
the dimension of w2v is 500, and g2v is 100.
w2v download: https://github.com/pujols/zero-shot-learning

GCNZ and DGP 's input is word embedding;
GAZSL, LisGAN and KG-GAN also run with w2v and g2v.

Experiment Class Split (for GCNZ, DGP, LisGAN, GAZSL, KG-GAN)

Exp1: original animal classes subset (seen:398, unseen:485)
Exp9: "animal" subset ImNet-A (seen:25, unseen:55)
Exp10: "other" subset ImNet-O (seen:10, unseen:25)

Run Command

GCNZ

construct graph:
python io_graph.py --mtr_exp_name Exp9 --exp_name Exp9_2555

prepare graph input:
python io_train_sample.py --mtr_exp_name Exp9 --exp_name Exp9_2555

train:
python train_predict_gcn.py --mtr_exp_name Exp9 --exp_name Exp9_2555

test:
python test_gcn.py --mtr_exp_name Exp9 --exp_name Exp9_2555 --feat 900
test (gzsl):
python test_gcn.py --mtr_exp_name Exp9 --exp_name Exp9_2555 --feat 900 --gzsl

DGP

prepare graph:
python make_induced_graph.py --mtr_exp_name Exp9 --exp_name Exp9_2555

train:
python train_predict_gpm.py --mtr_exp_name Exp9 --exp_name Exp9_2555

test:
python test_gpm.py --mtr_exp_name Exp9 --exp_name Exp9_2555 --pred 400

test (gzsl):
python test_gpm.py --mtr_exp_name Exp9 --exp_name Exp9_2555 --pred 400 --gzsl

KG_GAN

w2v: python gan.py --ExpName Exp9 --SemEmbed w2v
g2v: python gan.py --ExpName Exp9 --SemFile g2v.mat --SemSize 100 --NoiseSize 100

gzsl: python gan.py --ExpName Exp9 --SemFile g2v.mat --SemSize 100 --NoiseSize 100 --GZSL