leftthomas / cgd Goto Github PK

Pretrained models seem to be broken. Can you please upload them? Thanks!

i appreciate for your code!

i am trying to train isc dataset with CGD.
but, in case of isc dataset, default parameters and hyper parameters make poor results.
i used default parameters except batch_size(=32)

following CSV is my results. it dose not be improved each epoch.
please let me know what i am missing.

Thanks~~

epoch,train_loss,train_accuracy,test_recall@1,test_recall@2,test_recall@4,test_recall@8
1,8.402177921616206,0.06961633663366336,0.028133351588621736,0.03516668803058565,0.04923336673527956,0.10550006991252303
2,8.584146980601963,0.050278465346534656,0.29540020041167736,0.45013362541794777,0.8158671669661999,1.5332676470279694
3,8.797435745744423,0.038675742574257425,0.0211000136914663,0.028133351588621736,0.04923336673527956,0.0844000547658652
4,8.711510207983526,0.019337871287128713,0.028133351588621736,0.04923336673527956,0.0703333760611713,0.1406667521223426
5,,0.0270730198019802,0.014066675794310868,0.014066675794310868,0.03516668803058565,0.0703333760611713
6,,0.01547029702970297,0.014066675794310868,0.014066675794310868,0.03516668803058565,0.0703333760611713
7,,0.0,0.014066675794310868,0.014066675794310868,0.03516668803058565,0.0703333760611713
8,,0.01547029702970297,0.014066675794310868,0.014066675794310868,0.03516668803058565,0.0703333760611713
9,,0.06188118811881188,0.014066675794310868,0.014066675794310868,0.03516668803058565,0.0703333760611713
10,,0.03094059405940594,0.014066675794310868,0.014066675794310868,0.03516668803058565,0.0703333760611713
11,,0.03094059405940594,0.014066675794310868,0.014066675794310868,0.03516668803058565,0.0703333760611713
12,,0.01547029702970297,0.014066675794310868,0.014066675794310868,0.03516668803058565,0.0703333760611713
13,,0.03094059405940594,0.014066675794310868,0.014066675794310868,0.03516668803058565,0.0703333760611713

I've got an issue that I found out the result of Resnet-50 got 79.3 on SOP dataset, while the orginal paper had 84.3, what happen on this differences? Did you change some hyperparameters, or just ran on 20 epchos which were not enough.

Strange.
I have similar images 's1.jpg', 's2.jpg'

Case 1:

model_state_file = 'sop_uncropped_resnet50_SG_1536_0.1_0.5_0.1_128_model.pth'
gd_config = 'SG'
num_classes = 11318
state_dict = torch.load(model_state_file, map_location=torch.device('cpu'))
net = Model('resnet50', gd_config, 1536, num_classes)
net.load_state_dict(state_dict)
net.eval()

Loss (Equlidean) = 0.48

Case 2:
I change gd_config to 'GS'

model_state_file = 'sop_uncropped_resnet50_SG_1536_0.1_0.5_0.1_128_model.pth'
gd_config = 'GS'
num_classes = 11318
state_dict = torch.load(model_state_file, map_location=torch.device('cpu'))
net = Model('resnet50', gd_config, 1536, num_classes)
net.load_state_dict(state_dict)
net.eval()

Loss (Equlidean) = 0.14

For other products - the same.

hello may I know your structure of dataset？

There is a problem to download weights from baidu oustside China.
I was downloaded weights with some tricks.
I uploaded also it to mail.ru cloud

https://cloud.mail.ru/public/kwZE/7aRRZRPLX

Sorry, interface have russian language, but it possible download weights without registration.
Please add to readme following line:
Alternate download link for pretrained models on datasets 'Standard Online Products' and 'In-shop Clothes' available for download outside China: https://cloud.mail.ru/public/kwZE/7aRRZRPLX

Hi, The pretrained model links are broken. Do you plan to share the working links?

Hi. I got an error when running train.py with torch.autograd.detect_anomaly() as follows:
"RuntimeError: Function 'PowBackward0' returned nan values in its 0th output"

This is related to line 27 of model.py. Could you check this please. Thank you for your support.

I have tried but i failed...

I trained the network on isc dataset, but the training loss became nan at the very beginning of the training process.

I know it's ridiculous. I apologize for disturbing you.

I can't train properly after replacing backone with seresnet50
can you add backone of seresnet50?

Are there any pretrained models available to download?

How do I train CGD on an Unlabeled dataset?

As shown in the screenshot, the recall is very low as 0.01% for In-Shop Dataset. I have downloaded the dataset at /home/xxx/Documents/CGD/dataset/isc/Img/img/MEN WOMEN . Did I do something wrong?