Recent background programs have been used in many areas. Background programs, which are used in a variety of areas, from Web servers to system monitoring tools, are based on the concept of a process called daemon.
This session is intended to be a user of the *nix family operating system and has been viewed using services or daemons. It's for people who want to make it.
Therefore, the presentation will be accompanied by the following knowledge:
Required:
- Experience in using the *nix operating system (linux, BSD, MacOS)
- *nix File System Concepts
- Python basic grammar
- Python Packaging
Select:
- Basic understanding of the process
This document discusses generative adversarial networks (GANs). It introduces GANs as a framework where a generative model and discriminative model compete against each other, with the generative model trying to produce fake samples to fool the discriminative model and the discriminative model trying to distinguish between real and fake samples. The document outlines sections on the introduction to GANs, how adversarial nets work, theoretical results, and experiments with GANs.
Generating sequences with recurrent neural networksheedaeKwon
This document discusses recurrent neural networks for generating sequences and summarizes several applications. It introduces RNNs and their limitations in storing old information. Then it describes long short-term memory networks as a solution and how they can be used for text prediction at the character level and handwriting prediction and synthesis using datasets like IAM-OnDB. Experiments are discussed for handwriting prediction.
Fully convolutional networks for semantic segmentation heedaeKwon
This document discusses using fully convolutional networks (FCNs) for semantic segmentation. It introduces FCNs as an adaptation of classifiers to dense prediction tasks like segmentation by replacing fully connected layers with convolutional layers to preserve spatial information. The key aspects covered are related work on dense prediction with convolutional networks, how FCNs adapt classifiers for segmentation by removing fully connected layers and using backward strided convolutions for upsampling to produce dense pixel-level predictions, and presents results of the FCN approach.
Feature pyramid networks for object detection heedaeKwon
This document discusses feature pyramid networks for object detection. It introduces feature pyramid networks which use a bottom-up pathway to generate feature maps at multiple scales from a convolutional neural network and a top-down pathway that combines high-level and low-level semantic information. It then describes applying feature pyramid networks to region proposal networks and Fast/Faster R-CNN models for object detection and presents experimental results on using feature pyramid networks for region proposal and object detection.
This document describes a new model called Attention is All You Need that uses attention mechanisms without recurrent or convolutional layers. It introduces a model architecture that uses multi-head attention and feed-forward networks along with techniques like dropout and label smoothing. The model is evaluated on machine translation tasks using WMT 2014 English-German and English-French datasets and outperforms existing sequence models.
The document discusses squeeze-and-excitation networks, which are simple and computationally lightweight neural networks that introduce a slight increase in model complexity. Squeeze-and-excitation blocks work by squeezing global spatial information into a channel descriptor and then exciting each channel by a gating mechanism conditioned on the descriptor. The document describes related work on deeper architectures and attention mechanisms and provides details on experiments training squeeze-and-excitation networks on image datasets including error rates, optimizers, and learning rates used.
Perceptual losses for real time style transfer and super-resolutionheedaeKwon
This document proposes using perceptual loss functions rather than per-pixel loss for real-time style transfer and super-resolution. It introduces using feature reconstruction and style reconstruction losses calculated using pretrained neural networks to better capture perceptual differences between the output and ground truth images. The document outlines the method and provides experimental results demonstrating its effectiveness on tasks of style transfer and single-image super-resolution.
Recent background programs have been used in many areas. Background programs, which are used in a variety of areas, from Web servers to system monitoring tools, are based on the concept of a process called daemon.
This session is intended to be a user of the *nix family operating system and has been viewed using services or daemons. It's for people who want to make it.
Therefore, the presentation will be accompanied by the following knowledge:
Required:
- Experience in using the *nix operating system (linux, BSD, MacOS)
- *nix File System Concepts
- Python basic grammar
- Python Packaging
Select:
- Basic understanding of the process
This document discusses generative adversarial networks (GANs). It introduces GANs as a framework where a generative model and discriminative model compete against each other, with the generative model trying to produce fake samples to fool the discriminative model and the discriminative model trying to distinguish between real and fake samples. The document outlines sections on the introduction to GANs, how adversarial nets work, theoretical results, and experiments with GANs.
Generating sequences with recurrent neural networksheedaeKwon
This document discusses recurrent neural networks for generating sequences and summarizes several applications. It introduces RNNs and their limitations in storing old information. Then it describes long short-term memory networks as a solution and how they can be used for text prediction at the character level and handwriting prediction and synthesis using datasets like IAM-OnDB. Experiments are discussed for handwriting prediction.
Fully convolutional networks for semantic segmentation heedaeKwon
This document discusses using fully convolutional networks (FCNs) for semantic segmentation. It introduces FCNs as an adaptation of classifiers to dense prediction tasks like segmentation by replacing fully connected layers with convolutional layers to preserve spatial information. The key aspects covered are related work on dense prediction with convolutional networks, how FCNs adapt classifiers for segmentation by removing fully connected layers and using backward strided convolutions for upsampling to produce dense pixel-level predictions, and presents results of the FCN approach.
Feature pyramid networks for object detection heedaeKwon
This document discusses feature pyramid networks for object detection. It introduces feature pyramid networks which use a bottom-up pathway to generate feature maps at multiple scales from a convolutional neural network and a top-down pathway that combines high-level and low-level semantic information. It then describes applying feature pyramid networks to region proposal networks and Fast/Faster R-CNN models for object detection and presents experimental results on using feature pyramid networks for region proposal and object detection.
This document describes a new model called Attention is All You Need that uses attention mechanisms without recurrent or convolutional layers. It introduces a model architecture that uses multi-head attention and feed-forward networks along with techniques like dropout and label smoothing. The model is evaluated on machine translation tasks using WMT 2014 English-German and English-French datasets and outperforms existing sequence models.
The document discusses squeeze-and-excitation networks, which are simple and computationally lightweight neural networks that introduce a slight increase in model complexity. Squeeze-and-excitation blocks work by squeezing global spatial information into a channel descriptor and then exciting each channel by a gating mechanism conditioned on the descriptor. The document describes related work on deeper architectures and attention mechanisms and provides details on experiments training squeeze-and-excitation networks on image datasets including error rates, optimizers, and learning rates used.
Perceptual losses for real time style transfer and super-resolutionheedaeKwon
This document proposes using perceptual loss functions rather than per-pixel loss for real-time style transfer and super-resolution. It introduces using feature reconstruction and style reconstruction losses calculated using pretrained neural networks to better capture perceptual differences between the output and ground truth images. The document outlines the method and provides experimental results demonstrating its effectiveness on tasks of style transfer and single-image super-resolution.
Spatial transformer networks are a dynamic mechanism that can select the most relevant regions of images using spatial transformations. They consist of a localization network that outputs transformation parameters, a parameterized sampling grid that applies attention or affine transformations, and differentiable image sampling to extract image patches. Experiments show they improve performance on tasks like distorted MNIST classification, street view house number recognition, and fine-grained bird species classification by focusing on discriminative regions.
Learning to learn by gradient descent by gradient descentheedaeKwon
This document discusses using gradient descent to learn update rules for optimization problems. It proposes learning the optimizer itself by treating it as a differentiable function with its own parameters. The goal is to learn update rules rather than using hand-designed rules. The method and results sections are missing details on how this is implemented and evaluated.
Grad cam visual explanations from deep networks via gradient-based localizati...heedaeKwon
Grad-CAM is a technique to produce visual explanations for predictions from convolutional neural networks by generating localization heatmaps highlighting the important regions in the image for a specific prediction. It works by taking the gradients of any target concept (like the class score for a particular image classification task) flowing into the final convolutional layer and projecting back onto the feature maps to produce a coarse localization map highlighting the important regions in the image for predicting the concept. The technique does not require any modifications to the network architecture or training procedures. Results show Grad-CAM can provide visual explanations for CNN predictions by highlighting important regions in the input image.
This document discusses convolutional neural networks and provides details on GoogLeNet, an architecture that utilized deeper convolutions to achieve state-of-the-art results in image classification. It covers related work on CNNs and the Network in Network model. The architectural details section explains innovations in GoogLeNet, while the training methodology discusses hyperparameters like the optimizer, momentum, learning rate, and data splits. Results showed a 4% decrease in error rate using stochastic gradient descent.
Learning deep features for discriminative localizationheedaeKwon
Class activation mapping is a technique that uses global average pooling to visualize important regions in images that CNNs use to identify objects. It works by applying global average pooling to activation maps of the last convolutional layer to obtain the importance of each region for predicting the class. The technique was proposed to localize objects for weakly supervised tasks and help understand what CNNs learn from images.
Image net classification with deep convolutional neural networksheedaeKwon
The document summarizes research on using deep convolutional neural networks for image classification on the ImageNet dataset. It describes collecting a large dataset, using techniques like ReLU activation, local response normalization, overlapping pooling, dropout, and training across multiple GPUs. The results showed the CNN approach enabled more powerful models for image classification compared to previous methods.
This document discusses neural image caption generation using attention mechanisms. It introduces image caption generators that previously lost information using high-level representations or needed powerful mechanisms when using low-level representations. It then describes using an encoder-decoder model with CNN and RNN, and explores two types of attention mechanisms: stochastic "hard" attention and deterministic "soft" attention to better generate image captions while preserving important information.
This document discusses very deep convolutional networks for large-scale image recognition. It describes network configurations that use 3x3 convolutional filters with max pooling layers and fully connected layers. The networks have 11 or 19 weight layers and use 1x1 convolutional filters to introduce nonlinearity. Classification experiments on ImageNet data with over 1 million training images achieve top-1 and top-5 error rates.
The document proposes an end-to-end neural image caption generation system that combines aspects of state-of-the-art vision and language models. It uses a CNN encoder and RNN decoder trained to maximize the probability of a target sequence of words given an input image. Experiments use techniques like pre-trained CNNs and word embeddings to prevent overfitting. Results show the effect of dataset size on generalization, and evaluate the model's ability to transfer learning across datasets by measuring BLEU scores. Human evaluations and analyses of the model's word embeddings are also presented.
Spatial transformer networks are a dynamic mechanism that can select the most relevant regions of images using spatial transformations. They consist of a localization network that outputs transformation parameters, a parameterized sampling grid that applies attention or affine transformations, and differentiable image sampling to extract image patches. Experiments show they improve performance on tasks like distorted MNIST classification, street view house number recognition, and fine-grained bird species classification by focusing on discriminative regions.
Learning to learn by gradient descent by gradient descentheedaeKwon
This document discusses using gradient descent to learn update rules for optimization problems. It proposes learning the optimizer itself by treating it as a differentiable function with its own parameters. The goal is to learn update rules rather than using hand-designed rules. The method and results sections are missing details on how this is implemented and evaluated.
Grad cam visual explanations from deep networks via gradient-based localizati...heedaeKwon
Grad-CAM is a technique to produce visual explanations for predictions from convolutional neural networks by generating localization heatmaps highlighting the important regions in the image for a specific prediction. It works by taking the gradients of any target concept (like the class score for a particular image classification task) flowing into the final convolutional layer and projecting back onto the feature maps to produce a coarse localization map highlighting the important regions in the image for predicting the concept. The technique does not require any modifications to the network architecture or training procedures. Results show Grad-CAM can provide visual explanations for CNN predictions by highlighting important regions in the input image.
This document discusses convolutional neural networks and provides details on GoogLeNet, an architecture that utilized deeper convolutions to achieve state-of-the-art results in image classification. It covers related work on CNNs and the Network in Network model. The architectural details section explains innovations in GoogLeNet, while the training methodology discusses hyperparameters like the optimizer, momentum, learning rate, and data splits. Results showed a 4% decrease in error rate using stochastic gradient descent.
Learning deep features for discriminative localizationheedaeKwon
Class activation mapping is a technique that uses global average pooling to visualize important regions in images that CNNs use to identify objects. It works by applying global average pooling to activation maps of the last convolutional layer to obtain the importance of each region for predicting the class. The technique was proposed to localize objects for weakly supervised tasks and help understand what CNNs learn from images.
Image net classification with deep convolutional neural networksheedaeKwon
The document summarizes research on using deep convolutional neural networks for image classification on the ImageNet dataset. It describes collecting a large dataset, using techniques like ReLU activation, local response normalization, overlapping pooling, dropout, and training across multiple GPUs. The results showed the CNN approach enabled more powerful models for image classification compared to previous methods.
This document discusses neural image caption generation using attention mechanisms. It introduces image caption generators that previously lost information using high-level representations or needed powerful mechanisms when using low-level representations. It then describes using an encoder-decoder model with CNN and RNN, and explores two types of attention mechanisms: stochastic "hard" attention and deterministic "soft" attention to better generate image captions while preserving important information.
This document discusses very deep convolutional networks for large-scale image recognition. It describes network configurations that use 3x3 convolutional filters with max pooling layers and fully connected layers. The networks have 11 or 19 weight layers and use 1x1 convolutional filters to introduce nonlinearity. Classification experiments on ImageNet data with over 1 million training images achieve top-1 and top-5 error rates.
The document proposes an end-to-end neural image caption generation system that combines aspects of state-of-the-art vision and language models. It uses a CNN encoder and RNN decoder trained to maximize the probability of a target sequence of words given an input image. Experiments use techniques like pre-trained CNNs and word embeddings to prevent overfitting. Results show the effect of dataset size on generalization, and evaluate the model's ability to transfer learning across datasets by measuring BLEU scores. Human evaluations and analyses of the model's word embeddings are also presented.
2. lists
what is anconda?
Virtual environmen
Difference between AI ,machine learning and deep learning
A.I.
Machine learning
Deep learning
kaggle competition
데이콘
돌아보기
plan
3. ananconda
-라이브러리들을 쉽게 설치하고 관리할 수 있게 해주는
도구
- 밴드를 하고싶다
보컬,드럼 등 A단톡방 초대
- 농구를 하고싶다
센터 가드 등을 B단톡방 초대
=> 관리가 쉽다.
- 밴드를 하고싶다
보컬,드럼 등 A단톡방 초대
- 농구를 하고싶다
센터 가드 등을 A단톡방 초대
=> 관리가 어렵다
4. Virture enviroment
특정공간에서만 사용하는 제한된 공간
왜 필요합니까?
다수의 인원 작업시 프로젝트에 마다 다른 버전이 필요
so, pyhton,pip 등을 하나의 환경에서 설치시
= 비효율 + 번거로움 => why?
=> ex) 1클라 2클라 3클라 작업
1버전 설치 => 1클라 프로젝트 1버전 =>1버전 삭제 =>
2버전 설치 => 2클라 프로젝트 2버전 => 2버전 삭제 =>
3바전 설치 => 3클라 프로젝트 3버전
위와 같은 번거로움을 제거해주는 “가상환경”
5. Difference between AI ,machine leaning and
deep learning
A.I가 뭐야?
Machine
learning이뭐야?
deep learning
이뭐야?
알파고
알파고
알파고
9. kaggle competition
Conway's Reverse Game of Life 2020
Python ,Pytorch
켜져있거나/꺼저있는 세포보드로 구성
=>어떻게 발전하는지 관찰
규칙 : 인구과잉 , 정체 , 인구 부족 , 번식
알고리즘 작동방식
목적 : 시작보드 예측
10. 데이콘
운동 동작 분류 AI 경진대회
사용 언어 : Python , R
-train_features.csv (1875000, 8)
id 별 600 time 간 동작 데이터
id 3125개 x 600 time =1875000 데이터
-train_labels.csv (3125, 3)
id 별 동작과 동작 label(61개)
목표 : 다음 동작 예측