confusezius / deep-metric-learning-baselines Goto Github PK

Unable to use in windows, have a solution?

It seems common practice to multiply the learning rate of the embedding layer by 10x or so in the literature. I do not see that in the code - maybe I missed it?

Hi, @Confusezius ,

Thanks for your excellent code and for sharing it!

When running

python3.6 Standard_Training.py --gpu 0 --savename resnet_inshop_margin_dist  --dataset in-shop --n_epochs 40 --tau 25 --loss marginloss --sampling distance

the following issue raises and I've put the dataset under Datasets directory and the data is fine.

Getting pretrained weights...
Done.
MARGINLOSS Setup for RESNET50 with DISTANCE sampling on IN-SHOP complete with #weights: 23770304
Traceback (most recent call last):
  File "Standard_Training.py", line 161, in <module>
    dataloaders      = data.give_dataloaders(opt.dataset, opt)
  File "/home/code/Deep-Metric-Learning-Baselines/datasets.py", line 50, in give_dataloaders
    datasets = give_InShop_datasets(opt)
  File "/home/code/Deep-Metric-Learning-Baselines/datasets.py", line 280, in give_InShop_datasets
    train_dataset     = BaseTripletDataset(train_image_dict, opt,   samples_per_class=opt.samples_per_class)
  File "/home/code/Deep-Metric-Learning-Baselines/datasets.py", line 398, in __init__
    transforms.RandomHorizontalFlip(0.5)])
TypeError: object() takes no parameters

So, could u shed me some light? Thanks!

I wanted to know how we could save the model weights and use it for inference on some test data ?
I tried using torch.save() and torch.load() after all completion of all epochs, but the results are not similar as to automatic eval on test split and quite bad.
Could you please provide some inputs on this ?

I wanna ensure if the Function recover_closest_one_dataset(line 339 from auxiliaries.py ) which is to recover closest Example Image has relation with each training process（specifically means Kmean' process）, Because i found during every testing epoch, the feature_matrix_all seems to be the same,（line 279 from evaluate.py ），and the process of getting one sample' closest samples is achived by faiss.IndexFlatL2 then .add .search. So it means i can get all examples' three closest samples after training the first epoch not after every training epoch?
Looking forward to your reply.

Hi,
Have you tested these deep metric learning methods on VehicleID? What about the results?

Hi there, I'm new in this field and I tried to use my own dataset similar to the CAR dataset structure. I'm running on Windows 10 and I couldn't install faiss , i tried using auxiliaries_nofaiss.py for all the imports, is there any solution to this? the model is able to run for 1 epoch, however, the evaluation part is taking very long time, the report is saying 119 GB, and giving me the error as below:

Any help from anyone would be appreciated. Thanks a lot.

GPU:0, dataset:rp2k, arch:resnet50, embed_dim:128, embed_init:default
loss:fastap, sampling:None, samples_per_class:0, resize256:False
bs:32, lr:1e-05, fc_lr_mul:0, decay:0.0004, gamma:0.3, tau:[20], bnft:False
Running with learning rates 1e-05...
Epoch (Train) 0: Mean Loss [0.0136]: 100%|█████████████████████████████████████████| 4605/4605 [27:19<00:00, 2.92it/s]
Computing Evaluation Metrics...: 100%|█████████████████████████████████████████████| 5585/5585 [07:55<00:00, 13.33it/s]
Traceback (most recent call last):
File "Standard_Training.py", line 365, in
main()
File "Standard_Training.py", line 344, in main
eval.evaluate(opt.dataset, LOG, save=True, **eval_params)
File "C:\Users\user\Documents\DeepMetricLearningBaselines\evaluate.py", line 57, in evaluate
ret = evaluate_one_dataset(LOG, **kwargs)
File "C:\Users\user\Documents\DeepMetricLearningBaselines\evaluate.py", line 279, in evaluate_one_dataset
F1, NMI, recall_at_ks, feature_matrix_all = aux.eval_metrics_one_dataset(model, dataloader, device=opt.device, k_vals=opt.k_vals, opt=opt)
File "C:\Users\user\Documents\DeepMetricLearningBaselines\auxiliaries_nofaiss.py", line 239, in eval_metrics_one_dataset
k_closest_points = squareform(pdist(feature_coll)).argsort(1)[:, :int(np.max(k_vals)+1)]
File "C:\Users\user\anaconda3\envs\kunhe\lib\site-packages\scipy\spatial\distance.py", line 1985, in pdist
dm = np.empty((m * (m - 1)) // 2, dtype=np.double)
MemoryError: Unable to allocate 119. GiB for an array with shape (15967113051,) and data type float64

@

Thanks for your code and recent researches.

The question is, it seems like code presented in https://github.com/Confusezius/Deep-Metric-Learning-Baselines/blob/master/googlenet.py is for Inception-BN, not googlenet. Am I missing something?

Not an issue of the code, just wanted to point out this dataset is available for download:
The DeepFashion page has a Google Drive link, and following that link you can find the retrieval dataset:
https://drive.google.com/drive/folders/0B7EVK8r0v71pVDZFQXRsMDZCX1E

I believe there's a bug in BaseTripletDataset.transform: during training, images are not resized to 256x256 before taking crops.

Interestingly, this did not seem to affect results by much, at least when I tested it in Online Products with Margin loss...

In lines 150 and 152, it seems like the method 'semihardsampling' acts like MS mining('Multi Similarity Loss with General Pair Weighting', CVPR2019), not semi-hard sampling.

neg_mask = np.logical_and(neg,d<d[np.where(pos)[0]].max())
pos_mask = np.logical_and(pos,d>d[np.where(neg)[0]].min())

In terms of similarity, these lines seem equivalent to MS mining with $\epsilon = 0$ to me.
Could you please check it again?

Hi Author,
May I know which part I should look into if I want to make this an inference mode?

I can modify the code myself. Hope to get some guidance in terms of the part /modules to look into . I'm a bit confused because the whole module is complex.

Use Case:
Input 2 similar / different images, output similarity score or a MATCH/Not MATCH result.

Regards,
Alex

I cloned the repo, ran the install.sh script and tried running the Standard_Training.py file, but as soon as the network starts training, I get the error

RuntimeError: CUDA out of memory. Tried to allocate 42.88 MiB (GPU 0; 7.79 GiB total capacity; 6.35 GiB already allocated; 46.38 MiB free; 601.50 KiB cached)

I am using a 2070 Super GPU with 8 GB of vram. Is this too little to run the baseline, or do you think there is some other problem?

Thank you in advance

Hi
Thank you for your work and code!

I tested on CUB and CAR and it worked well but when I used SOP dataset, it seemed as it got stuck in the evaluation process, right inside the calculation of kmeans. The test values of Epoch 0 never got out and the program just got stuck there. It was the same for faiss or no_faiss setting.

Did I miss anything or could you suggest some hints to solve it?

Cheers

File "Deep-Metric-Learning-Baselines/netlib.py", line 21, in <module> import pretrainedmodels as ptm ModuleNotFoundError: No module named 'pretrainedmodels'

I run the code then get the problem, Thank you very much for your help.

Hi, thanks for the nice code. I noticed that your re-implementation of the Proxy-NCA is much higher than that from their original paper on CUB-200, e.g. R@1= 64.0 vs R@1= 49.

I understand that the original paper uses Inception, and you train a ResNet50, are the performance boost all from architecture ? Or is there any other implementation differences ?

did not find your 'pretrainedmodels' module in your code.

Hi
just wondering which version of in-shop dataset did you use? Since there are two versions in the page (one normal and one high resolution) and the high resolution one needs a password.

Thanks

thanks for your contribution

hello, I want to reproduce your result on Cars196, but when i download the dataset from the official website, the organization of the dataset doesn't like the one you provide in the readme. Could you please provide a hint or the code of how to handle these data. Thanks :)

Hello sir,

In the file 'auxiliaries.py', line 391:

sample_idxs = np.random.choice(np.arange(len(gallery_feature_matrix_all)), n_image_samples)

If I understand your code correctly, should 'gallery_feature_matrix_all' be 'query_feature_matrix_all'?

Thanks

Such as Market1501, DukeMTMC Re-ID, CUHK, MSMT17 and etc.

Minor suggestion: renaming Sampler to TripletSampler (or TupleSampler, etc.), so that it will not be confused with torch.utils.data.Sampler :)

sorry, cannot find the 'main.py' code..
thanks