I am trying to train the model and replicate results for the CIFAR10 dataset. I downloaded the dataset as per instructions and followed steps for training as in the attached file -
trainingStepsCIFAR10.md. It contains command and config used (if different from the default config in repo) along with output snippets.

MOCO is getting trained. I think cluster heads are not getting trained as accuracy is not increasing from 26%. I also tried a lower learning rate of 0.0005 but it didn't help. Finding locally consistent examples fails to find any consistent example at ratio_confident=0.99 (provided default) and only 6770 at ratio_confident=0.9.

Is this normal? If not, can you provide me some direction on where the issue might be present and how it can be resolved?

Detailed logs of training cluster head are in file: spice-self-logs.txt There are few errors of broken pipe. I think those are due to the usage of Jupyter for training and reconnections with Jupyter server. It shouldn't affect the convergence of the model.

dear author, image size of my own datasets is 224*224. in train_moco.py, how should I change learning rate and batch_size?

when I ran with batch_size=128 and lr remained , running loss is about 9.5.

    targets_i.append(gt_cluster_labels[h][img_idx_i].to(cfg.gpu, non_blocking=True))

    loss_dict = model(imgs_i, target=targets_i, forward_type="loss")

    loss = sum(loss for loss in loss_dict.values())
    loss_mean = loss / num_heads

    optimizer.zero_grad()
    loss_mean.backward()
    optimizer.step()

i want to train my own data WITHOUT ground truth labels. but it seems this step need labels?

Dear authors of SPICE,

I am very impressed by your work. Just a quick question: how the semantic prediction matrix P is obtained from the features?

Thanks!

as the title said, can i train this model and lable those data?they are completely without any class labels, just want to assign them with a specific class(no need for meaning of classes)

Trained the feature model with custom dataset , and did the stage 2 training using train_self_v2.py script,

But in the state 2 training evaluation metrics the accuracy was always -1 (never improved on 56 epochs), it was because the clusters created were not equal to the number of ground truth labels , but the training was completed.

After when running the local_consistency.py script the model output from the local_consistency forward type output was none

idx_select, labels_select = model(feas_sim=feas_sim, scores=scores, forward_type="local_consistency")

Tried with different ratio_confident and score_th values , but the values of idx_select, labels_select are tensor([ ], dtype=torch.int64) tensor([ ], dtype=torch.int64)

And so the reliable_labels are not created.

Hi,

I want to plot the embeddings' tsne. I am using feas_moco_512_l2.npy as the embeddings and datasets/stl10/stl10_binary/fold_indices.txt as the labels. The tsne plot looks a bit random though I can see different clusters... It seems that "fold_indices.txt" is not what I think it is. I've also noticed that there are some indices that correspond to multiple different classes (for example index 4555 shows up for both class '1' and class '8' in "fold_indices.txt").

Any ideas?

Thanks!

Hello Sir,

I saw targets on CIFAR10.py Does SPICE need targets(label for data on supervised learning)? If so, why do you need targets on your process?

Thanks, Edward Cho.

UserWarning: NVIDIA GeForce RTX 3090 with CUDA capability sm_86 is not compatible with the current PyTorch installation. The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_61 sm_70 sm_75 compute_37. If you want to use the NVIDIA GeForce RTX 3090 GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

Hi, Thanks for sharing your code. I want to do an evaluation on cifar100 I used the following net and weight but I got a mismatch error while loading state_dict.

--load_path ./model_zoo/model_cifar100_cls_res18.pth --net resnet18_cifar and --load_path ./model_zoo/model_cifar100.pth - -net WideResNet could you please tell me which net and model I should use for CIFAR100? Thanks

When trying to run your code I get the following error, do you have any suggestions as to what I might be doing wrong?

(base) anders@CenturionUbuntu:~/PycharmProjects/Spice/SPICE$ python tools/eval_self.py --config-file configs/cifar10/eval.py --weight model_zoo/model_cifar10.pth.tar --all 1 Traceback (most recent call last): File "tools/eval_self.py", line 12, in import torch File "/home/anders/anaconda3/lib/python3.8/site-packages/torch/init.py", line 188, in _load_global_deps() File "/home/anders/anaconda3/lib/python3.8/site-packages/torch/init.py", line 141, in _load_global_deps ctypes.CDLL(lib_path, mode=ctypes.RTLD_GLOBAL) File "/home/anders/anaconda3/lib/python3.8/ctypes/init.py", line 381, in init self._handle = _dlopen(self._name, mode) OSError: /home/anders/anaconda3/lib/python3.8/site-packages/torch/lib/../../../../libcufft.so.10: invalid ELF header

Hi ! Great work.

To allow training MoCo with CIFAR datasets. I suggest you change row 184 in 'tools/train_moco.py'.

from:

     model = moco.builder.MoCo(
        base_model,
        args.moco_dim, args.moco_k, args.moco_m, args.moco_t, args.mlp)

to:

     model = moco.builder.MoCo(
        base_model,
        args.moco_dim, args.moco_k, args.moco_m, args.moco_t, args.mlp, args.img_size)

Another question, Do you perform your final evaluation (SPICE-semi) on the resulting 'model_best.pth' or 'latest_model.pth' ? I haven't saw any reference for it. Thanks !

It seems like you have prepared arguments to run the script on single GPU (--multiprocessing-distributed and --gpu), but then I run into "Only DistributedDataParallel is supported." exception. Or when I run the script only with "--gpu 0" as argument, I run into error originating in spawn .py code.

I only have access to a computer with single Nvidia 3090 GPU and want to test your model as part of my Master's thesis. Does it mean that I would need to implement a version for 1 GPU myself, or am I missing something in your implementation?

Hello Sir,

I think that Clustering should do not have 'the targets(labels)'. But I found targets in your code (cifar.py).

So I want to remove codes that related to targets(labels). How to change your code?? Is it possible?

Thanks, Edward Cho

Hello Sir,

As I mentioned at Title, I was tried to train your code.

When I tried train according to your help-site(https://github.com/niuchuangnn/SPICE/blob/main/train.md), I met some error.

4. Determine reliable images

python tools/local_consistency.py --config-file ./configs/cifar10/eval.py --embedding ./results/cifar10/embedding/feas_moco_512_l2.npy

This is error message.

Traceback (most recent call last):
  File "tools/local_consistency.py", line 141, in <module>
    main()
  File "tools/local_consistency.py", line 131, in main
    acc = calculate_acc(labels_select, gt_labels_select)
  File "./spice/utils/evaluation.py", line 21, in calculate_acc
    assert len(np.unique(ypred)) == len(np.unique(y))
AssertionError

So I checked labels_select and gt_labels_select,

len(labels_select) : 10
tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
len(gt_labels_select) : 10
[8 5 8 5 3 4 1 5 1 3]

What I should do for this problem??

Thanks, Edward Cho.

Hi, I have a question to ask. After I follow the steps to install, start training python tools/train_moco.py --img_size 32 --moco-k 12800 --arch resnet18_cifar --save_folder ./results/cifar10/moco_res18_cls --resume ./results/cifar10/moco_res18_cls/checkpoint_last.pth.tar --data_type cifar10 --data ./datasets/cifar10 --all 0in training tutorial

Below is the error:

(u) C:\Users\Kelly>cd SPICE

(u) C:\Users\Kelly\SPICE>python tools/train_moco.py --img_size 32 --moco-k 12800 --arch resnet18_cifar --save_folder ./results/cifar10/moco_res18_cls --resume ./results/cifar10/moco_res18_cls/checkpoint_last.pth.tar --data_type cifar10 --data ./datasets/cifar10 --all 0 Use GPU: 0 for training Traceback (most recent call last): File "tools/train_moco.py", line 453, in main() File "tools/train_moco.py", line 145, in main mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, args)) File "C:\Users\Kelly.conda\envs\u\lib\site-packages\torch\multiprocessing\spawn.py", line 240, in spawn return start_processes(fn, args, nprocs, join, daemon, start_method='spawn') File "C:\Users\Kelly.conda\envs\u\lib\site-packages\torch\multiprocessing\spawn.py", line 198, in start_processes while not context.join(): File "C:\Users\Kelly.conda\envs\u\lib\site-packages\torch\multiprocessing\spawn.py", line 160, in join raise ProcessRaisedException(msg, error_index, failed_process.pid) torch.multiprocessing.spawn.ProcessRaisedException:

-- Process 0 terminated with the following error: Traceback (most recent call last): File "C:\Users\Kelly.conda\envs\u\lib\site-packages\torch\multiprocessing\spawn.py", line 69, in _wrap fn(i, *args) File "C:\Users\Kelly\SPICE\tools\train_moco.py", line 170, in main_worker dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url, File "C:\Users\Kelly.conda\envs\u\lib\site-packages\torch\distributed\distributed_c10d.py", line 602, in init_process_group default_pg = _new_process_group_helper( File "C:\Users\Kelly.conda\envs\u\lib\site-packages\torch\distributed\distributed_c10d.py", line 727, in _new_process_group_helper raise RuntimeError("Distributed package doesn't have NCCL " "built in") RuntimeError: Distributed package doesn't have NCCL built in

How do I need to solve thanks Kelly