mmfewshot is an open source few shot learning toolbox based on PyTorch

When trying to reproduce results of meta-rcnn and TFA, under 1 shot setting of split1, I find that reproduced results of meta-rcnn is much higher, which is confusing.In paper of meta-rcnn(this 19.9 is the result i want to get):

In paper of TFA:

Result in paper shows that result of split1 under 1 shot setting is 19.9. But my results is much higher: base training : mAP is 76.2 finetunning : all class is 47.40, novel class is 38.80, base class is 50.53 Which is much higher than results in paper. This is confusing. Besides, in the README.md of meta-rcnn, results are even higher:

under split1 1 shot setting, the results of TFA I get is 40.4 which is basically the same as the paper report.

Could you please kindly answer my questions?

https://github.com/open-mmlab/mmfewshot/blob/486c8c2fd7929880eab0dfcd73a3dd3a512ddfbe/configs/detection/base/datasets/nway_kshot/base_voc.py#L106

Hi, thanks for your great work in fsod. I want to know why the value of samples_per_gpu is not 15 instead of 16 for voc base training. Hope you can help me.

我的coco数据目录是这样的： data --coco ----annotations ----train2014 ----val2014 --few_shot_ann ----coco ------benchmark_10shot -------- ... 当我运行fsce下的coco预训练config时，会报错：no such file or directory: 'data/few_shot_ann/coco/annotaions/train.json' 请问这个train.json是哪里来的，预训练的标签不是应该调用coco文件夹下的annotations吗？ 另外我在data preparation找到一个trainvalno5k.json和5k.json，请问是这两个json文件吗？ 期待您的回答！

Traceback (most recent call last): File "/home/lbc/miniconda3/envs/mmfewshot/lib/python3.7/runpy.py", line 193, in _run_module_as_main "__main__", mod_spec) File "/home/lbc/miniconda3/envs/mmfewshot/lib/python3.7/runpy.py", line 85, in _run_code exec(code, run_globals) File "/home/lbc/mmfewshot-main/tools/detection/misc/initialize_bbox_head.py", line 289, in <module> main() File "/home/lbc/mmfewshot-main/tools/detection/misc/initialize_bbox_head.py", line 278, in main args) File "/home/lbc/mmfewshot-main/tools/detection/misc/initialize_bbox_head.py", line 169, in random_init_checkpoint new_weight[:prev_cls] = pretrained_weight[:prev_cls] RuntimeError: The expanded size of the tensor (21) must match the existing size (54) at non-singleton dimension 0. Target sizes: [21, 1024]. Tensor sizes: [54, 1024]

The process of fsce on my own coco format datasets is:

1. Base Training : ckpt(step1)
2. step two: ues the best val pth of step 1 for train? python3.7 -m tools.detection.misc.initialize_bbox_head --src1 ./work_dirs/fsce_r101_fpn_coco_base-training/best_bbox_mAP_iter_105000.pth --method random_init --save-dir ./work_dirs/fsce_r101_fpn_coco-split1_base-training

Motivation

When the length of the last row_data is less than 10 and greater than 0, the row_data will not be printed

Modification

When the last row_data is not empty, add to table_data

Currently, the demo script (classification) takes only one sample in the support set. It uses the process_support_images() method to forward the support set. How to modify this in order to allow for more than one sample in the support set?

One idea could be to place another set of support images in a different folder and then forward that as well. Then the model.before_forward_support() method can be modified if it resets the features. For e.g. for meta_baseline_head, it is resetting saved features.

Then (again for meta_baseline), meta_baseline_head.before_forward_query would also have to be modified since it is replacing the self.mean_support_feats with the mean of the new support set.

Would these two changes in this case be enough to adapt for a few-shot instead of a one-shot inference?

According to the document, the following errors occur during training. I don't know how to solve them. Has anyone encountered them. TypeError: init() got an unexpected keyword argument 'persistent_workers'

Hi, Great Job.

I have some questions about the training process of the matching network(classification)

• In this line, https://github.com/open-mmlab/mmfewshot/blob/31583cccb8ef870c9e688b1dc259263b73e58884/configs/classification/matching_net/mini_imagenet/matching-net_conv4_1xb105_mini-imagenet_5way-1shot.py?_pjax=%23js-repo-pjax-container%2C%20div%5Bitemtype%3D%22http%3A%2F%2Fschema.org%2FSoftwareSourceCode%22%5D%20main%2C%20%5Bdata-pjax-container%5D#L28 You use num_shots=5 for training 5-way-1-shot, is this a bug?
• The batch size shown in the result table is 64, I would like to know whether this number is the training batch size or test batch size?
• How many gaps between the meta-val and meta-test split in your experiment?
  • In the log of matching_net 5-way-1-shot, the max accuracy is about 51%, while the test result is 53%, does it means there exists ~2 points between two sets?

Thanks, Best

The error occurs when running demo_metric_classifier_1shot_inference with a custom trained NegMargin model. The meta_test_head is None. Testing the model with dist_test works as expected though. I am not sure why it didn't save the meta test head. A comment here says that it is only built and run on testing. I am not sure what that means though.

The model config is the same as the standard in other config files:

model = dict(
    type='NegMargin',
    backbone=dict(type='Conv4'),
    head=dict(
        type='NegMarginHead',
        num_classes=6,
        in_channels=1600,
        metric_type='cosine',
        margin=-0.01,
        temperature=10.0),
    meta_test_head=dict(
        type='NegMarginHead',
        num_classes=6,
        in_channels=1600,
        metric_type='cosine',
        margin=0.0,
        temperature=5.0))

Otherwise, the config file itself is similar to other neg_margin config files for the cube dataset.

I found that the “frozen_parameters” parameter is used in many detection models, but I have not found where this parameter is used in the relevant source code. Which part of the source code should I see?

在mmfewshot/detection/datasets/coco.py/FewShotCocoDefaultDataset 中的coco_benchmark固定了数据集路径为f'data/few_shot_ann/coco/benchmark_{shot}shot/full_box_{shot}shot_{class_name}_trainval.json'。但是我的few_shot_ann路径和上面不同，并且FewShotCocoDefaultDataset没有办法接受数据集路径的参数，希望可以增加此参数

Traceback (most recent call last): File "/root/mmfewshot/./tools/detection/train.py", line 236, in main() File "/root/mmfewshot/./tools/detection/train.py", line 225, in main train_detector( File "/root/mmfewshot/mmfewshot/detection/apis/train.py", line 48, in train_detector data_loaders = [build_dataloader(ds, **train_loader_cfg) for ds in dataset] File "/root/mmfewshot/mmfewshot/detection/apis/train.py", line 48, in data_loaders = [build_dataloader(ds, **train_loader_cfg) for ds in dataset] File "/root/mmfewshot/mmfewshot/detection/datasets/builder.py", line 311, in build_dataloader data_loader = TwoBranchDataloader( TypeError: init() got an unexpected keyword argument 'persistent_workers' Killing subprocess 9272 Traceback (most recent call last): File "/opt/conda/envs/pytorch1.8/lib/python3.9/runpy.py", line 197, in _run_module_as_main return _run_code(code, main_globals, None, File "/opt/conda/envs/pytorch1.8/lib/python3.9/runpy.py", line 87, in _run_code exec(code, run_globals) File "/opt/conda/envs/pytorch1.8/lib/python3.9/site-packages/torch/distributed/launch.py", line 340, in main() File "/opt/conda/envs/pytorch1.8/lib/python3.9/site-packages/torch/distributed/launch.py", line 326, in main sigkill_handler(signal.SIGTERM, None) # not coming back File "/opt/conda/envs/pytorch1.8/lib/python3.9/site-packages/torch/distributed/launch.py", line 301, in sigkill_handler raise subprocess.CalledProcessError(returncode=last_return_code, cmd=cmd) subprocess.CalledProcessError: Command '['/opt/conda/envs/pytorch1.8/bin/python', '-u', './tools/detection/train.py', '--local_rank=0', 'configs/detection/mpsr/voc/split1/mpsr_r101_fpn_2xb2_voc-split1_base-training.py', '--launcher', 'pytorch']' returned non-zero exit status 1.

Since few shot dataset is just for finetuning the model and the test.py won't save the change of the model, where should I put my fewshot dataset? training set or validation set? In that way, I could use the pth file to predict my images in the demo.py?

We meet problem when training on classification models. We test several times, the code is blocked on this line of command in classification.api.train

When I was studying the process of small sample learning, I came across tiff images in the data set. At this point, there is a problem with the dataset loading, would like to ask if you can add a tiff format read method.