ankitdhall / learning_embeddings Goto Github PK

Hi, @ankitdhall
how to fix the bug?

python network/ethec_experiments.py --experiment_name exp_test --experiment_dir exp --image_dir ETHEC_dataset_v0.1/ETHEC_dataset/IMAGO_build_test_resized/ --n_epochs 1 --model resnet18 --loss multi_level --set_mode train
PyTorch Version: 2.1.0+cpu
Torchvision Version: 0.16.0+cpu
Traceback (most recent call last):
File "network/ethec_experiments.py", line 466, in
ETHEC_train_model(args)
File "network/ethec_experiments.py", line 247, in ETHEC_train_model
repo = git.Repo(search_parent_directories=True)
File "C:\Users\omnisky.conda\envs\learning_embeddings\lib\site-packages\git\repo\base.py", line 267, in init
raise InvalidGitRepositoryError(epath)
git.exc.InvalidGitRepositoryError: E:\learning_embeddings

Greetings,
I notice in the ethec_experiments.py of the Adam1X branch, both ETHECExperiment() and ETHEC2D() need a CIFAR10 object. May I ask what is CIFAR10 used here? If the aiming dataset is ETHEC1.0, why do we still need to define the label map of CIFAR10?
Sincerely thanks for any possible response.

When running

--experiment_name
test1
--experiment_dir
test1_dir
--image_dir
<path_to>/ETHEC_dataset_v0.2/ETHEC_dataset/IMAGO_build_test_resized
--n_epochs
10
--model
resnet50
--loss
multi_level
--set_mode
train 

it crashes with:

Traceback (most recent call last):
  File "learning_embeddings/network/ethec_experiments.py", line 473, in <module>
    ETHEC_train_model(args)
  File "learning_embeddings/network/ethec_experiments.py", line 361, in ETHEC_train_model
    sampler=None if arguments.class_weights else WeightedResampler(
  File "learning_embeddings/network/experiment.py", line 231, in __init__
    label_ids = [dataset[ind]['leaf_label'] for ind in range(len(dataset))]
  File "learning_embeddings/network/experiment.py", line 231, in <listcomp>
    label_ids = [dataset[ind]['leaf_label'] for ind in range(len(dataset))]
  File "learning_embeddings/data/db.py", line 3789, in __getitem__
    'labels': torch.from_numpy(self.labelmap.get_one_hot(sample['family'], sample['subfamily'], sample['genus'],
  File "learning_embeddings/data/db.py", line 1071, in get_one_hot
    retval[self.genus_specific_epithet[specific_epithet] + self.levels[0] + self.levels[1] + self.levels[2]] = 1
KeyError: 'mannii'

What am i doing wrong ?

Can you share the exact train, test, val json files that you used . The ones with the dataset seems to be incompatible with the code.

There are still few issues in the code:

1. The split that are given are sort of corrupt as there are many images that have image name in the 'image_path' and 'image_path' as empty.
2. After this, while running ethec_experiments.py , the train set and everything is made by db.py, so if labelmap = ETHECLabelMap, then there are 550 specific_epithets and 560 genus_specific_epithet. So, the issue comes is that the train set and the labeling of the last level is done with the help of specific epithets but decode_children that is used while training, decode the last level label by using genus_specific_epithet.
3. If oder_embeddings.py is run then the first issue is ETHECDB doesnot take input with_images, if that is resolved, the training is stucked after the first epoch.

Please see for these issues and make the code easily reproducible