Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

关于bottom up and top down模型的问题 #31

Closed
JimLee4530 opened this issue Jan 22, 2018 · 26 comments
Closed

关于bottom up and top down模型的问题 #31

JimLee4530 opened this issue Jan 22, 2018 · 26 comments

Comments

@JimLee4530
Copy link

你好,
bottom up and top down的论文里说 60k iterations 训9个小时就能到达cider 120.1 的效果。
但我将你代码模型的参数修改成论文中的参数,并将attention lstm那块的输入修改成每个bounding box 的image feature的均值。但是也达不到论文里的效果。
想问问你,以你的经验来看,觉得会是什么原因呢?

@ruotianluo
Copy link
Owner

跟batch size有关。然后这个repo用的festure不是文章中用的feature。

@coldmanck
Copy link

coldmanck commented Jan 31, 2018

@JimLee4530

I think you can try the pretrained features provided by the author here.

@JimLee4530
Copy link
Author

thanks ,It achieve 1.20 Cider score on Karpathy's test split after self-critical training.

@HrsPythonix
Copy link

您好,请问您方便能分享一下训练时的具体参数嘛,比如rnn的hidden unit num,还有batch size等等

@JimLee4530
Copy link
Author

我用的都是论文里提供的参数。batch size为100, rnn_hide_size为1000, embedding 为1000. attention hide size为512

@HrsPythonix
Copy link

你的max epoch设置的是30? 然后最后没有Cider optimization的结果Bleu_1有0.770吗?

@JimLee4530
Copy link
Author

JimLee4530 commented May 19, 2018

对的,max epoch设置的是30, 当时的结果如下,Bleu_1好像没到0.770,未经过Cider optimization的结果如下:
{'CIDEr': 1.1534126230436987,
'Bleu_4': 0.3712410192810052,
'Bleu_3': 0.4753132784679994,
'Bleu_2': 0.6087557246854488,
'Bleu_1': 0.7656615739972887,
'ROUGE_L': 0.57175632862905756,
'METEOR': 0.2799153984204049}

@HrsPythonix
Copy link

好的 非常感谢

@deepylt
Copy link

deepylt commented Jun 5, 2018

@JimLee4530 你好,请问一下,我用跟你一致的参数设置,即--batch_size 100 --rnn_size 1000 --input_encoding_size 1000 --input_encoding_size 512, 但是学习率与scheduled samping 设置与ruotian的readme中的默认设置一致(adam, learning_rate: 5e-4 开始step下降, ss:0-0.25),但是训练30epoch之后最高的CIDEr分数是1.11,比你跑的1.15低挺多的。我注意到top-down论文中是采用sgd更新策略,学习率从0.01下降,训练60k次迭代。请问下你有修改其他参数设置吗?

@JimLee4530
Copy link
Author

我的设置你一样,并没有修改其他的。你的特征用了论文里的feature了吗?

@deepylt
Copy link

deepylt commented Jun 5, 2018

嗯是的,用了每个图片36个box的特征,请问你是用的10-100 adapt的还是36 fix的特征?

@JimLee4530
Copy link
Author

我的是10~100 adapt的。

@deepylt
Copy link

deepylt commented Jun 6, 2018

我试试10-100的,谢谢!

@yangxuntu
Copy link

@JimLee4530 请问您是否可以给一下cider是1.20的详细的参数设置,我按照上面所说的参数设置最后发现并没有跑到1.20,只能跑到1.12多。我觉得应该是我的学习速率设置的不对,也有可能是eval时的参数设置的有问题。
所以希望您能够展示一下1.20的超参,谢谢了!

@JimLee4530
Copy link
Author

JimLee4530 commented Jun 26, 2018

python train.py --id topdown --caption_model topdown --input_json data/cocotalk.json --input_fc_dir data/cocotalk_fc --input_att_dir faster_rcnn_pool5 --input_label_h5 data/cocotalk_label.h5 --batch_size 100 --learning_rate 5e-4 --learning_rate_decay_start 0 --scheduled_sampling_start 0 --checkpoint_path log_topdown --save_checkpoint_every 6000 --val_images_use 5000 --max_epochs 30 --rnn_size 1000 --input_encoding_size 1000 --att_feat_size 2048 --att_hid_size 512

这是我设置的参数,其他的都是opt文件里默认的。

接下来是 eval
python eval.py --dump_images 0 --num_images 5000 --model log_topdown/model-best.pth --infos_path log_topdown/infos_topdown-best.pkl --language_eval 1 --beam_size 5
eval的参数应该会去读在 infos文件里读训练的参数。beam size 我设成5了。和论文一样的。不过换成3和2,结果差不多。

@yangxuntu
Copy link

@JimLee4530 非常感谢!!!

@ivy94419
Copy link

@JimLee4530 @yangxuntu @deepylt Can somebody share the 10-100 adapt or 30 fixed features? I download it from Chrome browser, but the speed is too slow and it often interrupts.

@yangxuntu
Copy link

I download all these features about 1-2 days, be patient...

@ivy94419
Copy link

@yangxuntu It always interrupts and redownload, it will start from the begin......

@yangxuntu
Copy link

I do not meet such problem...
Maybe you need to use a vpn or something else.

@ivy94419
Copy link

@yangxuntu what browser or tool you used for download? I have vpn in my computer

@yangxuntu
Copy link

I do not use any vpn. I just slowly download the features. And after one to two days, I finish downloading them.

@y657250418
Copy link

y657250418 commented Apr 3, 2019

@ruotianluo @JimLee4530 请问有使用cider优化后的topdown模型在online(https://competitions.codalab.org/competitions/3221)上测评过吗?
我训练时得分很高
{
'CIDEr': 1.2122694918211574,
'Bleu_4': 0.3716883181404758,
'Bleu_3': 0.49495639150868476,
'Bleu_2': 0.6458257633211044,
'Bleu_1': 0.8147943927215487,
'ROUGE_L': 0.5806054907761705,
'METEOR': 0.2812694786562275
}

但是在online上测试时得分很低
{'reflen': 413361, 'guess': [425299, 384795, 344291, 303787], 'testlen': 425299, 'correct': [278690, 123972, 48991, 19223]}
ratio: 1.02888032495
Bleu_1: 0.655
Bleu_2: 0.459
Bleu_3: 0.311
Bleu_4: 0.209
computing METEOR score...
METEOR: 0.214
computing Rouge score...
ROUGE_L: 0.466
computing CIDEr score...
CIDEr: 0.691

是不是在测试输出json的文件时需要使用test2014的10-100的特征?
我是直接加载模型的参数没有更改,命令如下:
python eval.py --model log_td/model.pth --infos_path log_td/infos_td.pkl --image_folder data/test2014/ --dump_images 0 --coco_json image_info_test2014.json

想请问一下知道原因吗?感恩~~~

@JimLee4530
Copy link
Author

首先 online 上的得分确实会低的。其次是你test的命令, 有用到beam search 吗?用了beam search 会高点。

@y657250418
Copy link

@JimLee4530 使用了beam search,beam_size=2,想请问一下您用原始模型在online上的得分大概是多少?方便分享一下吗?感谢

@yangxuntu
Copy link

yangxuntu commented Apr 15, 2019 via email

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

8 participants