hanzy1996 / ce-gzsl Goto Github PK

Hello, I would like to know how to generate the file sent_splits.mat or where to get the file of other datasets, e.g., AwA2, SUN, APY.
Thank you so much!

你好！
请问现在能麻烦你提供一下你训练时使用的所有超参数吗
超参数太影响实验结果了！几乎达不到论文里的结果

Following your dataset link, i can't find the FLO datatset. Can you provide it?

感谢你们分享的工作。你好，我想请问一下训练你们所提出的GZSL框架需要多大显存的GPU，对应的训练时间是多长？谢谢！

Hi @Hanzy1996 , the flower dataset that I found on the internet contains only data splits and images.

Could you share the link to the text embeddings? Or ideally, a link to the FLO dataset you used in your experiments?

请问实验设置为什么是without finetuning? 通常不是具有finetune的实验效果更好吗？

期待回复~感谢！

285行移动到223行

Hi @Hanzy1996 , did you use text embeddings for the CUB dataset? If so, is it really possible to compare your results with TF-VAEGAN since they used original attributes?

I notice that you use sent attributes of CUB dataset, but we usually use att attributes on CUB, like CLSWGAN/TF-VAEGAN and so on. Could you please provide your result when using att attributes for fair comparison?

How long does it usually take for a sample training run? I'm currently running this model on CUB using the given example parameters and it is taking a significant runtime (hours for 10s of iterations). Why is this significantly slower than the baselines compared against, like TF-VAEGAN? Is there a way to improve the runtime?

Hi Han，

Could you please tell me what the epoch of each dataset should be set to?Thanks.

Can you provide the best parameters for each dataset?

Much appreciation for your interest!

We have tested our method on CUB using the attribute semantic descriptor, but the results are not as good (U=52.4, S=60.9, H=56.4) as using the sentence semantic descriptor. You can also replace the sent with the attribute in the codes.
Besides, we find that some other methods, like CADA-VAE[1], cycle-CLSWGAN[2], etc., also use the sentence semantic descriptor for CUB. And our results on CUB can still outperform theirs.

Best wishes!

[1] Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders. Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. CVPR2019.
[2] Multi-modal Cycle-consistent Generalized Zero-Shot Learning. Rafael Felix, Vijay Kumar B G, Ian Reid, and Gustavo Carneiro. ECCV2018

Originally posted by @Hanzy1996 in #4 (comment)

Hello, the H I got on SUN by using the parameters you provided and the method in the paper is 2% less than the result in your paper, I wonder if you have encountered such a situation. Could you please provide SUN's training parameters again?
--dataset SUN
--class_embedding att
--syn_num 100
--batch_size 1024
--attSize 102
--nz 102
--embedSize 2048
--outzSize 512
--nhF 1024
--ins_weight 0.01
--cls_weight 0.01
--ins_temp 0.1
--cls_temp 0.1
--manualSeed 4115
--nclass_all 717
--nclass_seen 645
--gpus 1
--lr_decay_epoch 100
--lr 0.00005‘

The file res101.mat which seems to be necessary is not in your repository. Could you provide it in order to reproduce that?
Thank you

作者您好，我在运行SUN 数据集时很容易出现NAN的情况，但在其他数据集上都不会有这个问题，请问是模型本身的问题（比如说WGAN训练不稳定），还是SUN数据集本身有脏数据呢？

I want to train with my own dataset, but I don't know how to make a dataset similar to that in this article. I'd like to know how to make res101.mat and att.splits.mat files in XLSA file?

I feel the attribute of unseen classes are not used in the training stage. In other word, in the training stage, there is not synthetic feature generated from Generator by using attribute of unseen classes.

Hi @Hanzy1996 , I've implemented a Keras port of your code. What's the best way to share it with the community?