你好，请问是否有实验，transformer 分支不进行loss 监督的消融实验呢？从模型结构来看，conv 层的特征已经融合了transformer 特征，如果只监督conv 分支，结果如何呢？我这边在检测DBnet的backbone 中使用了conformer 结构，但是最后只监督了conv 这个分支，结果没有优于原来，可能也和其他变量相关。

When just do inference using the model, I found that the model don't do load_state_dict after torch.load The related code is from line 301 to line 322 in './Conformer-main/main.py'

Thanks for your nice work! Here I encountered a question about training from scratch for custom data, the error message is shown as the following:

D:\dl\Conformer-main>python main.py --model Conformer_small_patch16 --data-set IMNET --batch-size 4 --lr 0.001 --num_workers 0 --data-path ./datasets/test/ --output_dir ./output/test/ --epochs 10 Not using distributed mode Namespace(aa='rand-m9-mstd0.5-inc1', batch_size=4, clip_grad=None, color_jitter=0.4, cooldown_epochs=10, cutmix=1.0, cutmix_minmax=None, data_path='./datasets/test/', data_set='IMNET', decay_epochs=30, decay_rate=0.1, device='cuda', dist_url='env://', distributed=False, drop=0.0, drop_block=None, drop_path=0.1, epochs=10, eval=False, evaluate_freq=1, finetune='', inat_category='name', input_size=224, lr=0.001, lr_noise=None, lr_noise_pct=0.67, lr_noise_std=1.0, min_lr=1e-05, mixup=0.8, mixup_mode='batch', mixup_prob=1.0, mixup_switch_prob=0.5, model='Conformer_small_patch16', model_ema=True, model_ema_decay=0.99996, model_ema_force_cpu=False, momentum=0.9, num_workers=0, opt='adamw', opt_betas=None, opt_eps=1e-08, output_dir='./output/test/', patience_epochs=10, pin_mem=True, recount=1, remode='pixel', repeated_aug=True, reprob=0.25, resplit=False, resume='', sched='cosine', seed=0, smoothing=0.1, start_epoch=0, train_interpolation='bicubic', warmup_epochs=5, warmup_lr=1e-06, weight_decay=0.05, world_size=1) Creating model: Conformer_small_patch16 number of params: 37673424 Start training Traceback (most recent call last): File "main.py", line 375, in main(args) File "main.py", line 335, in main set_training_mode=args.finetune == '' # keep in eval mode during finetuning File "D:\dl\Conformer-main\engine.py", line 30, in train_one_epoch for samples, targets in metric_logger.log_every(data_loader, print_freq, header): File "D:\dl\Conformer-main\utils.py", line 157, in log_every header, total_time_str, total_time / len(iterable))) ZeroDivisionError: float division by zero

Kindly for help, thanks!

Hi,Gay. I do not understand why LayerNorm in down-sampler FCU and BatchNorm in up-sampler FCU are used to regularize features.Is there any special meaning？

Nice work! And I have some questions here. The "small" pretrained model posted on home page is conformer_small_patch16.pth. And I haven't found the patch32 one which is the pretrained model in your faster RCNN's config. May I ask this checkpoint file, please? And Can you provide the configuration file of Base and Large model? Thank you very much!

{"train_lr": 0.0006003999999998758, "train_loss_0": 5.2468776138196835, "train_loss_1": 5.2314488993825545, "test_loss": 12.697948946271623, "test_loss_0": 6.487931878226144, "test_loss_1": 6.210017108917237, "test_acc1": 0.0, "test_acc1_head1": 0.0, "test_acc1_head2": 0.0, "epoch": 4, "n_parameters": 77557992}

Dear author, I saw the picture of feature analysis in your paper（Figure 4）. I think the visualization effect is very good. Can you provide the code here? Thank you.

1. Could you interpret the last_fusion in ConvTransBlock? I did not find a specific explanation in your paper.
2. In ConvBlock, could you explain what is x_t in the forward function?

Hi, thanks for sharing your nice work with us.

When I was opening Google Drive link of Conformer tiny patch16, I can only get the base one.

Could you update the link?

Thank you!

Happy Chinese New Year. I tried to use the command "python main.py --model Conformer_tiny_patch16 --data-set CIFAR --batch-size 512 --lr 0.001 --num_workers 4 --data-path ../data --output_dir ./output/Conformer_tiny_patch16_batch512_lr1e3_epoch500_cifar100 --epochs 500 --input-size 32" to train Conformer on the CIFAR100 dataset with one tesla v100 GPU . But only got 45.89% accuracy. Would you mind providing the training parameters and top-1 accuracy on CIFAR 100 dataset?

batch size=128, initial lr =0.001 Epoch: [0] [ 1070/14862] eta: 1:11:01 lr: 0.000001 loss_0: 5.6751 (5.6817) loss_1: 5.6846 (5.7059) time: 0.3053 data: 0.0002 max mem: 4992

when i use 4 3090Ti to train the conformer-S model, with batchsize=128, epoch=300 as the sh file you showed. 1.it may take around 9 days, is this the normal speed? 2.and what can i do to shorten training time, like with larger batchsize, 3.and it would have some bad influence or not?

Hi! Thank you for sharing your work~ There is one place in the paper that I don't quite understand. You mentioned Rotation Invariance and Scale Invariance. But I don't see a detailed proof of them. Is this conclusion just based on experimental comparisons, or is there a theoretical basis I haven't noticed? Thank you.

I saw the object detection result in the paper, but the input image size is (800,1333), I think may cause the model take up more space. Is there any code to achieve object detection?