An easy way to build PyTorch datasets. Modularly build datasets and automatically cache processed results

Overview

EasyDatas

An easy way to build PyTorch datasets. Modularly build datasets and automatically cache processed results

Installation

pip install git+https://github.com/SymenYang/EasyDatas

Usage

Find files in disk

from EasyDatas.Prefab import ListFile, RecursionFiles, SpanFiles
from EasyDatas.Prefab import Chain

# Type 1: Find files recursively
# Example:
RFiles = RecursionFiles({
    "path" : path_to_root,
    "pattern" : ".*\.npy",
    "files" : True, # default to be true
    "dirs" : False # default to be true
})
RFiles.resolve()
print(len(RFiles)) # Total num of npy files in path_to_root
print(RFiles[0]) # {"path" : "/xxxx/xxxx/xxxx.npy"(pathlib.Path object)}

# Or Type 2: Hierarchically find files
HFiles = Chain([
    ListFile({
        "path" : path_to_root,
        "pattern" : ".*",
        "files" : False, # default to be true
    }),
    SpanFiles({
        "pattern" : ".*\.npy"
        "dirs" : False # default to be true
    })
])
HFiles.resolve()
print(len(HFiles)) # Total num of npy in files in path_to_root's depth-one sub-dir
print(HFiles[0]) # {"path" : "path_to_root/xxxx/xxxx.npy"(pathlib.Path object)}

ListFile, RecursionFiles, SpanFiles will output files/dirs in the dictionary order

Load files to memory

from EasyDatas.Prefab import LoadData, NumpyLoad,NumpyLoadNPY
#Type 1: use numpy.load to load a npy format file
LoadChain = Chain([
    RFiles, # defined in the previous section. Or any other EasyDatas module providing path
    NumpyLoadNPY({
        "data_name" : "data" # default to be "data"
    })
])
LoadChain.resolve()
print(len(loadChain)) # The same with RFiles
print(LoadChain[0]) # {"data" : np.ndarray}

# Type 2: write your own codes to load
import numpy as np
LoadChainCustom = Chain([
    HFiles,
    LoadData({
        "data_name" : "custom_data" # default to be "data"
        },
        function = lambda x : np.loadtxt(str(x))
    )
])
LoadChainCustom.resolve()
print(len(LoadChainCustom)) # The same with HFiles
print(LoadChainCustom[0]) # {"custom_data" : np.ndarray}

# The custom LoadData could be replaced by NumpyLoad module.

Preprocessing

from EasyDatas.Prefab import Picker, ToTensor
from EasyDatas.Core import Transform, CachedTransform

class customTransform1(CachedTransform): 
    # Cached Transform will process all datas and cache the results in disk.
    def custom_init(self):
        self.times = self.get_attr("times", 2) # default value is 2

    def deal_a_data(self, data : dict):
        data["data"] = data["data"] * self.times
        return data


class customTransform2(Transform): 
    # Non-cached transform will process a data when it is been needed.
    def deal_a_data(self, data : dict):
        data["data"] = data["data"] + 1
        return data


TrainDataset = Chain([
    LoadChain,
    Picker(
        pick_func = lambda data,idx,total_num : idx <= 0.8 * total_num
    ),
    customTransform1({
        "times" : 3
    }),
    customTransform1(),
    customTransform2(),
    ToTensor()
])
TrainDataset.resolve()
print(len(TrainDataset)) # 0.8 * len(LoadChain)
print(TrainDataset[0]) # {"data" : torch.Tensor with (raw data * 3 * 2 + 1) }

# Or we can write all of them in one chain and only resolve once
TrainDataset = Chain([
    RecursionFiles({
        "path" : path_to_root,
        "pattern" : ".*\.npy",
        "dirs" : False # default to be true
    }),
    NumpyLoadNPY({
        "data_name" : "data" # default to be "data"
    }),
    Picker(
        pick_func = lambda data,idx,total_num : idx <= 0.8 * total_num
    ),
    customTransform1({
        "times" : 3
    }),
    customTransform1(),
    customTransform2(),
    ToTensor()
])
TrainDataset.resolve()
print(len(TrainDataset)) # 0.8 * len(LoadChain)
print(TrainDataset[0]) # {"data" : torch.Tensor with (raw data * 3 * 2 + 1) }

All EasyDatas modules are the child of torch.utils.data.Dataset. Thus we can directly put them into a dataloader

About caches

An EasyDatas module will store caches only if the args["need_cache"] is True. The defualt setting is False. Cache will be save in the args["cache_root"] path, which is set to CWD in default. The cache name will contain two parts. The first is about the module's args when it was created, the second is about the module's previous modules cache name. All the information are encoded to a string and EasyDatas will use that string to determine whether there is a valid cache for this module instance. Therefore, if one module's args have been changed, all modules' cache after this module will be recomputed.

Custom cache name

One can override name_args(self) function to change the properties that need to be considerd into cache name. The default implementation is:

class EasyDatasBase
    ...
    def name_args(self):
            """
        Return args dict for getting cache file's name
        Default to return all hashable values in self.args except cache_root
        """
        ret = {}
        for key in self.args:
            if isinstance(self.args[key],collections.Hashable):
                if key == "cache_root":
                    continue
                ret[key] = self.args[key]
        return ret
    ...

Processing Datas

All EasyDatas module have two functions to deal datas. The first is deal_datas and the second is deal_a_data. In default, deal_datas will send all datas to deal_a_data one-by-one and collect the return value as the output of this module. In most situation, customizing deal_a_data is safe, clear and enough. But in some other situation, we want to deal all datas by our own, we could override deal_datas function. There are two useful functions in EasyDatasBase class that will be helpful in deal_datas, which are self.get()and self.put()

class EasyDatasBases:
    def get(self,idx = None,do_copy = True) -> dict|None:
        pass

    def put(self,data_dict : dict,idx = -1) -> None:
        pass

If idx is not provided, get will automaticaly get datas from previous module one-by-one. If it meets the end, it will return None. A module with no previous module could not use get function. If the do_copy is set to False, it will directly return previous module's data, which is a reference. Otherwise, it will deep copy the data and return.
put function will automaticaly put datas in to return and cache list. if idx is provided, the data_dict will be put in to the position. The total number of datas will be count automaticaly in put function.
Besides, in deal_a_data function, one can use put functions and return None for increasing the data items.

Other modules

There are some other modules that are not introduced beyond.

EasyDatas.Core.EasyDatasBase

Defined base functions, logging and default processing

EasyDatas.Core.RawDatas

Base class for ListFile, RecursionFiles. RawDatas needs no previous dataset and the deal_datas function needs to be overrided

EasyDatas.Core.Merge

Merge multiple EasyDatas modules by merge their data dict. The modules need to have the same length.

# assume A is an EasyDatas module with A[0] == {"data_1" : xxx}
# assume B is an EasyDatas module with B[0] == {"data_2" : xxx}
M = Merge([A,B])
print(len(M)) # The same with A and B
print(M[0]) # {"data_1" : xxx, "data_2" : xxx}

EasyDatas.Core.Stack

Stack multiple EasyDatas modules by combine their items.

# assume A is an EasyDatas module with A[0] == {"data_1" : xxx} and len(A) = 1000
# assume B is an EasyDatas module with B[0] == {"data_2" : xxx} and len(B) = 500
S = Stack([A,B])
print(len(S)) # 1500 which is len(A) + len(B)
print(S[999]) # {"data_1" : xxx}
print(S[1000]) # {"data_2" : xxx}

In most cases, Stack are used to stack modules which have same data format.

Owner
Ximing Yang
Fudan University
Ximing Yang
AirPose: Multi-View Fusion Network for Aerial 3D Human Pose and Shape Estimation

AirPose AirPose: Multi-View Fusion Network for Aerial 3D Human Pose and Shape Estimation Check the teaser video This repository contains the code of A

Robot Perception Group 41 Dec 05, 2022
Auditing Black-Box Prediction Models for Data Minimization Compliance

Data-Minimization-Auditor An auditing tool for model-instability based data minimization that is introduced in "Auditing Black-Box Prediction Models f

Bashir Rastegarpanah 2 Mar 24, 2022
Image data augmentation scheduler for albumentations transforms

albu_scheduler Scheduler for albumentations transforms based on PyTorch schedulers interface Usage TransformMultiStepScheduler import albumentations a

19 Aug 04, 2021
LBK 26 Dec 28, 2022
🙄 Difficult algorithm, Simple code.

🎉TensorFlow2.0-Examples🎉! "Talk is cheap, show me the code." ----- Linus Torvalds Created by YunYang1994 This tutorial was designed for easily divin

1.7k Dec 25, 2022
Reducing Information Bottleneck for Weakly Supervised Semantic Segmentation (NeurIPS 2021)

Reducing Information Bottleneck for Weakly Supervised Semantic Segmentation (NeurIPS 2021) The implementation of Reducing Infromation Bottleneck for W

Jungbeom Lee 81 Dec 16, 2022
[ICCV2021] Learning to Track Objects from Unlabeled Videos

Unsupervised Single Object Tracking (USOT) 🌿 Learning to Track Objects from Unlabeled Videos Jilai Zheng, Chao Ma, Houwen Peng and Xiaokang Yang 2021

53 Dec 28, 2022
[NeurIPS'20] Self-supervised Co-Training for Video Representation Learning. Tengda Han, Weidi Xie, Andrew Zisserman.

CoCLR: Self-supervised Co-Training for Video Representation Learning This repository contains the implementation of: InfoNCE (MoCo on videos) UberNCE

Tengda Han 271 Jan 02, 2023
The official PyTorch implementation of Curriculum by Smoothing (NeurIPS 2020, Spotlight).

Curriculum by Smoothing (NeurIPS 2020) The official PyTorch implementation of Curriculum by Smoothing (NeurIPS 2020, Spotlight). For any questions reg

PAIR Lab 36 Nov 23, 2022
DEMix Layers for Modular Language Modeling

DEMix This repository contains modeling utilities for "DEMix Layers: Disentangling Domains for Modular Language Modeling" (Gururangan et. al, 2021). T

Suchin 43 Nov 11, 2022
[CVPR2021] De-rendering the World's Revolutionary Artefacts

De-rendering the World's Revolutionary Artefacts Project Page | Video | Paper In CVPR 2021 Shangzhe Wu1,4, Ameesh Makadia4, Jiajun Wu2, Noah Snavely4,

49 Nov 06, 2022
Python Implementation of algorithms in Graph Mining, e.g., Recommendation, Collaborative Filtering, Community Detection, Spectral Clustering, Modularity Maximization, co-authorship networks.

Graph Mining Author: Jiayi Chen Time: April 2021 Implemented Algorithms: Network: Scrabing Data, Network Construbtion and Network Measurement (e.g., P

Jiayi Chen 3 Mar 03, 2022
Computing Shapley values using VAEAC

Shapley values and the VAEAC method In this GitHub repository, we present the implementation of the VAEAC approach from our paper "Using Shapley Value

3 Nov 23, 2022
Runtime type annotations for the shape, dtype etc. of PyTorch Tensors.

torchtyping Type annotations for a tensor's shape, dtype, names, ... Turn this: def batch_outer_product(x: torch.Tensor, y: torch.Tensor) - torch.Ten

Patrick Kidger 1.2k Jan 03, 2023
PyTorch code of "SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks"

SLAPS-GNN This repo contains the implementation of the model proposed in SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks

60 Dec 22, 2022
basic tutorial on pytorch

Quick Tutorial on PyTorch PyTorch Basics Linear Regression Logistic Regression Artificial Neural Networks Convolutional Neural Networks Recurrent Neur

7 Sep 15, 2022
CC-GENERATOR - A python script for generating CC

CC-GENERATOR A python script for generating CC NOTE: This tool is for Educationa

Lêkzï 6 Oct 14, 2022
Unofficial pytorch implementation of the paper "Dynamic High-Pass Filtering and Multi-Spectral Attention for Image Super-Resolution"

DFSA Unofficial pytorch implementation of the ICCV 2021 paper "Dynamic High-Pass Filtering and Multi-Spectral Attention for Image Super-Resolution" (p

2 Nov 15, 2021
An integration of several popular automatic augmentation methods, including OHL (Online Hyper-Parameter Learning for Auto-Augmentation Strategy) and AWS (Improving Auto Augment via Augmentation Wise Weight Sharing) by Sensetime Research.

An integration of several popular automatic augmentation methods, including OHL (Online Hyper-Parameter Learning for Auto-Augmentation Strategy) and AWS (Improving Auto Augment via Augmentation Wise

45 Dec 08, 2022
Official repo for BMVC2021 paper ASFormer: Transformer for Action Segmentation

ASFormer: Transformer for Action Segmentation This repo provides training & inference code for BMVC 2021 paper: ASFormer: Transformer for Action Segme

42 Dec 23, 2022