ainimal/aini_modules

A PyTorch Computer Vision (CV) module library for building n-D networks flexibly ~

aini_modules's Issues

problem, 3D weights cannot be loaded

Thank you for providing a lot of 3D pre-training weights. When I loaded the weights of VC3D_kenshohara, I ran the following code and found that all weights were Non-Pretrained keys: 318. Is there any problem?

import torch
from wama_modules.thirdparty_lib.VC3D_kenshohara.wide_resnet import generate_model
from wama_modules.utils import load_weights
m = generate_model()
pretrain_path = r"wideresnet-50-kinetics.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights)
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
keys ( Current model,C )  318 odict_keys(['conv1.weight', 'bn1.weight', 'bn1.bias', 's_tracked'])
keys ( Pre-trained  ,P )  267 dict_keys(['module.conv1.weight', 'module.bn1.weight...er4.2.convle.fc.bias'])
keys (   In C &   In P )  0 dict_keys([])
keys ( NoIn C &   In P )  267 dict_keys(['module.co...odule.layer4.2.bas'])
keys (   In C & NoIn P )  318 dict_keys(['conv1.weig...es_tracked'])
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Pretrained keys : 0 dict_keys([])
Non-Pretrained keys: 318 dict_keys(['conv1.weight', 'bn1.we...tches_tracked'])
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

pretrained weights of Radimagenet

the original radimagenet provides 4 networks's weights, but i only find one (resnet) in your repo?
https://github.com/BMEII-AI/RadImageNet

pretrained weights of the costom networks like wm.encoder

I notice that all the pretrained weights in this repo are from other repo and you just collect them.
Do you have some pretrained weights of your code, like wm.encoder?

word embedding weights?

https://github.com/WAMAWAMA/wama_modules/blob/main/demo/multi_label/Demo_CVPR2019_MultiLabel_ML_GCN.py#L129

hi, tks for your multi-labl codes, and could you share this weights cache=r'C:\git\.vector_cache' below?

# get word embedding
glove = vocab.GloVe(name="6B", dim=300, cache=r'C:\git\.vector_cache')

whats the meaning of Support multi-class per label?

Thanks for your code in simple style b(￣▽￣)d, but I cannot under stand the "Support multi-class per label" in the table of multi-label network structures, whats the meaning of it ?

Need more demos, such as multi-model structures?

Wow, thank you for the demo you provided. They look very concise.
BTW, do you have any plans to implement some multi-modal models? For example, the 'co-attention model' mentioned in this paper
An Empirical Study of Training End-to-End Vision-and-Language Transformers

Label_category_dict

Hi,
I'm interested in running a 3D VGG on a CT dataset (single channel).
I want to classify the images into 3 classes (0, 1, or 3).
I did not understand how the label_category_dict works.
How to set the number of classes?

Do you have a snippet of code for this scenario?

Thank you.
Best,

Paolo