Computer Vision , AI

Deep equilibrium model에 대한 정리 일반적인 deep neural network를 수식의 형태로 표현해보면 다음과 같이 표현할 수 있을 것이다.sigma: activation functionW_i: weights of i-th layerb_i: bias of i-th layerz_i : latent vector of i-th layer각 annotation이 다음과 같을 때 i+1번째 layer의 latent vector의 값은 이전인 i번째 latent vector가 W_i를 통과하여 bias b_i를 더해준 뒤 activation function을 거쳐 nonlinearity를 확보한 형태라고 할 수 있을 것이다.  위의 일반적인 neural network를 넘어서 weight-ti..

cannot import name 'container_abcs' from 'torch._six'timm=0.3.2 version에서 발생한 문제 torch._six에 container_abcs가 없는 torch 버전을 사용해서 그렇다.collections.abc를 container_abcs로 import 해오는 것으로 해결할 수 있는 경우가 있다.   # from torch._six import container_abcsimport collections.abc as container_abcs

# mmcv 2.0 이하 버전에 해당 우선 깔려 있는 mmcv를 pip으로 지운다.pip uninstall mmcvorpip uninstall mmcv-full 이후 버전을 명시하여 새로 mmcv를 설치한다.pip install mmcv-full==1.4.4 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10/index.html https://download.openmmlab.com/mmcv/dist/cu113/torch1.10/index.html download.openmmlab.com pip으로 설치하되 뒤에 버전 명시를 해주어야함 버전 명시 예시https://download.openmmlab.com/mmcv/dist/cu{쿠다버전}/torch{..

state_dict()에는 requires_grad에 대한 정보가 안담기는 것으로 보인다. model.prompt_embeddings.requires_grad의 return 값은 True여도model.state_dict()['prompt_embeddings'].requires_grad의 return 값은 False로 나온다.

보호되어 있는 글입니다.

● Summary: Simple method for Image editing with a diffusion model only using CLIP [CLS] token embedding ● Approach highlight Image editing without labels using only the detection model Crop the original image and augment the image for CLIP embedding Only use [CLS] token to prevent the model from just doing copy-and-paste Classifier free sampling for image identity (scale factor) ● Main Results ●..

● Summary: a single framework for Image composition (color harmonization, geometric correction, shadow generation) with no label ● Approach highlight Self-supervised learning: segment object from the original image and mask that portion Content adaptor for object identity: image-to-text embedding using CLIP embedding (to use a diffusion model designed for text embedding) Diffusion with the maske..

● Summary: Zero-shot i mage translation using cross-attention map guidance ● Approach highlight Noise regularization for image inversion: to ensure Gaussian noise Cross-attention map guidance: Allows you to edit only the parts you want while maintaining the overall context of the original image ● Main Results ● Discussion Is it really a zero-shot setup? (using CLIPBLIP)

● Summary: hierarchically structured behavior and long-horizon coordination for RL ● Approach highlight Hierarchically structured behavior Imitation for low level ex)run Reinforcement learning for Drill ex) kick, dribble Distillation for single player Multi player reinforcement learning ● Main Results: ● Discussion Limitation of simple reward only goal score) Too heavy model

● Summary: Text to Video generation with Text Image Data ● Approach highlight Text-to-Image Model: DALLE 2 architecture Spatiotemporal layers: U-Net based spatiotemporal diffusion decoder makes a frame from noise Frame interpolation network ● Main Results: ● Discussion How to generate temporal frames from the spatiotemporal decoder How to learn the relationship between text and action that can o..