mboudiaf / tim Goto Github PK

Hi,
In #7 I've said I plan to adapt this method to another aluminum dataset.
I've already done that, yielding a rather great result(5 way 1 shot with accuracy 0.5171, 5 way 5 shot with accuracy 0.6927).
There is a question: in the backbone training stage, the highest accuracy it achieved is about 0.64(below 0.6927), is this abnormal?
And another interesting thing is, when I use backbone trained with the aluminum dataset to evaluate on NEU-CLS dataset, it achieved an incredible accuracy of [0.7010, 0.8141]!!!
Best

@mboudiaf . I can not download dataset files by using the python file ‘download_data.py’ in the file directory ‘./scripts/downloads‘’, maybe the dataset google drive download link is unavailable, can you provide the new dataset google drive download link to me? Thank you very much.

HI ,
I try to run "Download_data.py" and "Download_models.py" for some resources,but always show:
" requests.exceptions.ConnectionError: HTTPSConnectionPool(host='docs.google.com', port=443): Max retries exceeded with url: /uc?export=download&id=15MFsig6pjXO7vZdo-1znJoXtHv4NY-AF (Caused by NewConnectionError('<requests.packages.urllib3.connection.VerifiedHTTPSConnection object at 0x0000014B42FF0208>: Failed to establish a new connection: ",but i can open "https://docs.google.com",excuse for this URL is right?

Hi, when I want to evaluate Domain shift (Table 2. in paper), it is needed to provide an appropriate pre-trained model (for example "checkpoints/mini2cub/softmax/resnet18") but, if we want to train the appropriate model from scratch, how it is possible to train it for the cross domain?
I did not find the script.

Hello, I have a trivial problem about library visdom_logger.
I have never seen this library, and also I did not find it in the PYPI.
Could you please tell what it is and where I can download it?
Thank you!

I wish you a merry Christmas. On this beautiful day, I would like to ask you a question about tuning parameters. When I run the training file according to your steps, for example, when I run resnet18.sh (because I will report an error, I only modify / src / datasets)/ ingredient.py：num_workers = 0，After training, the accuracy of 1-shot 0.3776 and 5-shot 0.5026 can only be achieved in the mini dataset. Is there any other parameter adjustment skills

I have been having some issues downloading tieredImagenet from the link provided it says forbidden network. Could you kindly change the settings so that I can download it please since I am trying to use it for my research

when i run training code：bash scripts/train/resnet18.sh，
have a problem;

ERROR - FSL training - Failed after 0:04:19!
Traceback (most recent calls WITHOUT Sacred internals):
File "/home/data/qinyanfei/code/TIM-master/src/main.py", line 118, in main
if (epoch) % trainer.meta_val_interval == 0:
AttributeError: 'Trainer' object has no attribute 'meta_val_interval'

can you answer me？

Thank you for releasing code for such a creative method! However I've faced some problems when reproducing the results.

From what I've known, TIM is one of the most s-o-t-a few-shot methods using WRN as backbone at present. But the pretraining strategy shown in the code seems problematic, as I could only train a model that gets a 34.99% accuracy in a 16-way 1-shot task on miniImageNet, comparing to ~44% with the strategy used in SIB or EPNet. It has been pointed out that the quality of pretrained backbone will significantly influence the performance of the method, and since TIM outperforms SIB and EPNet a lot with no extra fine-tuning stage, this becomes puzzling.

This the training strategy for (miniImageNet, WRN) I found in your code:
inital LR=0.1, optimizer: SGD w/ nesterov momentum=0.9 , weight decay=1e-4
N_epoch=90
LR schedule: multistep, LR*=0.1@epoch 45&67
Label smoothing 0.1
Data augment: Color jitter

Did I miss something? If not, could you tell me why you didn't adopt a better pretrained backbone to further improve the results? Thanks a lot.

Hello, I can't open the iCloud link to go download data and checkpoints. Can you please provide a new link, thank you very much!

ERROR - FSL training - Failed after 0:04:19!
Traceback (most recent calls WITHOUT Sacred internals):
File "/home/data/qinyanfei/code/TIM-master/src/main.py", line 118, in main
if (epoch) % trainer.meta_val_interval == 0:
AttributeError: 'Trainer' object has no attribute 'meta_val_interval'

I want to adapt this dataset to Ali Aluminum Dataset.
Now I've made my own split file, written my own al_tianchi.sh like:

python3 -m src.main \
		with dataset.path="/path/to/al_tianchi" \
		visdom_port=8097 \
		dataset.split_dir="split/al_tianchi" \
		ckpt_path="checkpoints/al_tianchi/softmax/resnet18" \
		dataset.batch_size=128 \
		dataset.jitter=True \
		model.arch='resnet18' \
		model.num_classes=10 \
		optim.scheduler="multi_step" \
		epochs=90 \
		trainer.label_smoothing=0.1

and trained my own resnet18 model(as I understand, the backbone).
What should I do next to test TIM on this dataset?
Sorry for my bad English, and sorry for my disturbing.
Best Wishes for You :)

Traceback (most recent call last):
File "/home/dell/anaconda3/lib/python3.6/runpy.py", line 193, in _run_module_as_main
"main", mod_spec)
File "/home/dell/anaconda3/lib/python3.6/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/home/data/qinyanfei/code/TIM-master/src/main.py", line 8, in
from visdom_logger import VisdomLogger
ModuleNotFoundError: No module named 'visdom_logger'

Is the module missing？

Hey Malik,

I am trying to use your model in my work and I was wondering what are the correct hyperparameters you are using in your experiments. In the paper you say that you use 1000 iterations and for the Adam optimizers the suggested ones is the paper, I assume that the lr is the one that pytorch uses 1e-3. However in your tim.py code under the config() you use lr = 1e-4 I was wondering if that is correct