clarifai / few-shot-ctm Goto Github PK

The res-net used in code is from pytorch official webiste,which pre-trained from image-net. Is it fair compared to other models?

I run all the possible config in the code you provide with a 5way-1shot setting on mini-imagenet, but result is far from you reported in the paper.
And when i run with resnet backbone, the code seems to download pretrain model from pytorch official website, so I doubt the you really get the 62.05 without augmentation in 5way-1shot?

It seems many people think it's hard to got the result.....

when i run the code with " python main.py --gpu_id 0 1",I got this error :
AttributeError: 'DataParallel' object has no attribute 'forward_CTM'
so how to use multi graphics card to run it

I just run as the readme writes
python main.py --yaml_file configs/demo/mini/20way_1shot.yaml
and i got this error
Traceback (most recent call last): File "main.py", line 204, in <module> main() File "main.py", line 28, in main opts = Config(config['options']['ctrl.yaml_file'], config['options']) File "/home/few-shot-ctm/core/config.py", line 111, in __init__ merge_cfg_from_file(yaml_file, self) File "/home/few-shot-ctm/tools/general_utils.py", line 278, in merge_cfg_from_file _merge_a_into_b(yaml_cfg, _config) File "/hom/few-shot-ctm/tools/general_utils.py", line 334, in _merge_a_into_b _merge_a_into_b(v, b[k], stack=stack_push) File "/home/few-shot-ctm/tools/general_utils.py", line 324, in _merge_a_into_b raise KeyError('Non-existent config key: {}'.format(full_key)) KeyError: 'Non-existent config key: ctmnet.pred_source'
python version 3.6.8 pytorch version 1.2.0
Any idea?

Thanks for your detailed code. If i understand right, as paper 3.2.4 said,

the combination is an element wise multiplication of .....

but in core.model, the implement is torch.matmul whatever the self.dnet_supp_manner is. There is a distinct difference both in calculation and instinctive explanation. If i misunderstand sth, please tell me.

by the way, in my understanding, self.dnet_supp_manner = 1 corresponds to option1 in paper 3.2.4, self.dnet_supp_manner = 3 corresponds to option 2, self.dnet_supp_manner = 2 corresponds to confusion of option 1 and option2.

After taking a look at everyone's questions, I would like to ask, is it necessary to run the code?

Until now, the author still has not provided the pre-training codes of the model. A reasonable guess is that he used 80 classes (64 + 16) to pre-trained the model. So, the proposed method may not actually improve performance, just a work of pre-trained the model on meta-training and meta-testing set.

As I can see in the code, when I use ResNet18 as backbone, the Concentrator, Projector and Reshaper are same which is

self.main_component = nn.Sequential(self._make_layer(Bottleneck, out_size*2, 3, stride=1), 
                                    self._make_layer(Bottleneck, out_size, 2, stride=1))
self.projection = nn.Sequential(self._make_layer(Bottleneck, out_size*2, 3, stride=1),
                                    self._make_layer(Bottleneck, out_size, 2, stride=1))
self.reshaper = nn.Sequential(self._make_layer(Bottleneck, out_size*2, 3, stride=1),
                                    self._make_layer(Bottleneck, out_size, 2, stride=1))

Am I right? If so, it's inconsistent with the paper because the Projector and Reshaper should be simple CNN as described. Besides, the model capacity will be much larger than before. May I misunderstand the code?

i just run the command line python main.py --yaml_file configs/demo/mini/20way_1shot.yaml，and change self.ctrl.device = 'cuda' , ctrl.device = 'cuda'.The code is almost running 48 hours in one GPU. And just get

Evaluating at epoch 216, step 212, with eval_length 600 ... (be patient)
Current accuracy is 0.1530 < previous best accuracy is 0.1740 (ep55, iter212)
Can you give me some constructive advice?

Why a such simple method can be implemented in a such complicated way, can you release a simpler version of project that can reproduce the result in paper ?