I am looking for the complete taxonomy of PapersWithCode: areas-tasks-subtasks. PaperswithCode website: https://paperswithcode.com/ PaperswithCode API: https://paperswithcode.com/api/v1/docs/
I already tried to use the PapersWithCode-API Here's a python example of what I requested hoping to built the area-task-subtasks mapping.
import request
area_id = 'computer-vision'
q = f'https://paperswithcode.com/api/v1/areas/{area_id}/tasks/?page=1&items_per_page=500'
res = requests.get(q).json()
Output:
[{'id': 'aesthetics-quality-assessment',
'name': 'Aesthetics Quality Assessment',
'description': 'Automatic assessment of aesthetic-related subjective ratings.'},
{'id': 'user-constrained-thumbnail-generation',
'name': 'User Constrained Thumbnail Generation',
'description': 'Thumbnail generation is the task of generating image thumbnails from an input image.\r\n\r\n<span style="color:grey; opacity: 0.6">( Image credit: [User Constrained Thumbnail Generation using Adaptive Convolutions](https://arxiv.org/pdf/1810.13054v3.pdf) )</span>'},
{'id': 'sensor-fusion',
'name': 'Sensor Fusion',
'description': '**Sensor Fusion** is the broad category of combining various on-board sensors to produce better measurement estimates. These sensors are combined to compliment each other and overcome individual shortcomings.\r\n\r\n\r\n<span class="description-source">Source: [Real Time Dense Depth Estimation by Fusing Stereo with Sparse Depth Measurements ](https://arxiv.org/abs/1809.07677)</span>'},
{'id': 'lip-sync-1',
'name': 'Constrained Lip-synchronization',
'description': 'This task deals with lip-syncing a video (or) an image to the desired target speech. Approaches in this task work only for a specific (limited set) of identities, languages, speech/voice. See also: Unconstrained lip-synchronization - https://paperswithcode.com/task/lip-sync'},
{'id': 'online-multi-object-tracking',
'name': 'Online Multi-Object Tracking',
'description': 'The goal of **Online Multi-Object Tracking** is to estimate the spatio-temporal trajectories of multiple objects in an online video stream (i.e., the video is provided frame-by-frame), which is a fundamental problem for numerous real-time applications, such as video surveillance, autonomous driving, and robot navigation.\r\n\r\n\r\n<span class="description-source">Source: [A Hybrid Data Association Framework for Robust Online Multi-Object Tracking ](https://arxiv.org/abs/1703.10764)</span>'},
{'id': 'cross-domain-few-shot',
'name': 'Cross-Domain Few-Shot',
'description': ''}, ...]
I checked the entire response and there is no information on whether each task has any parent or child task.
As pointed out by @JYL the main resource to use can be found at https://github.com/paperswithcode/paperswithcode-data.
From there it is possible to retrieve Task-Subtask information from the "Evaluation Tables".
I was able to reconstruct the tree using the following python code:
### retrieving all tasks hierarchy
import pandas as pd
import json
import gzip
with gzip.open('data/evaluation-tables.json.gz', 'r') as fin:
eval_tables = json.loads(fin.read().decode('utf-8'))def expand_tasks_tree(subtasks,parent,root,level):
global index_dict
r_tmp = []
for subtask in subtasks:
task = subtask['task']
if not task in index_dict.keys():
index_dict[task] = max(index_dict.values())+1
r_tmp += [{'level':level,'root':root,'parent':parent,'task':task,'id':index_dict[task],'parent_id':index_dict[parent]}]
try: r_tmp += expand_tasks_tree(subtask['subtasks'],task,root,level+1)
except: print(subtask)
return r_tmp
index_dict = {'root':0}
eval_all = [{'task':item['categories'][0] if item['categories'] else 'uncategorized','subtasks':[item]} for item in eval_tables]
res = expand_tasks_tree(eval_all,'root','',0)pd.DataFrame.from_records(res).drop_duplicates(['level','root','parent','task','id','parent_id'])This results in the following dataframe:
