auto-model-development

Introduction

This project mainly contains four components:

• collecting(done)
  • model: model cards and model names on huggingface
  • dataset: dataset descriptions from paperswithcode
  • task: task descriptions from paperswithcode
• mapping(not done)
  • dataset to task(done)
  • model to dataset(not done)
  • build (model, dataset, task) tuples(not done)
• graph(not done)
  • produce models' fx graphs with concrete trace tools
  • produce xml graphs with traced fx graphs
• producing samples(not done)
  • a sample is a tuple of (model, dataset, task)
  • model is a xml graph of a model.
  • dataset is a dataset description. The dataset is used to train/evaluate the model.
  • task is a task description. The task is one of the tasks that the dataset is used on.

Usage

More details can be found in README files under each folder. Most of codes need to be specified proper paths before running.

• collect_evaluations: early-stage python scripts used for parsing and collecting evaluation results of models on huggingface. They are probably not useful now.
• collect_scripts: python scripts used for collecting model cards, dataset descriptions, task descriptions and dataset-task mapping.
• mapping: codes used for collecting model to dataset mapping and build (model, dataset, task) tuples.
• graph: concrete_trace_utils for producing fx graphs and fxgraph_to_seq for producing xml graphs.
• generate_samples: uses the (model, dataset, task) tuples to concatenate xml graphs, dataset descriptions and task descriptions to produce samples.

Storage

The prepared data is stored on MSRAGPUM19. Paths are listed below:

• collecting
  • model cards: /data/data0/v-junliang/DNNGen/auto_model_dev/model_cards, one file for each model, the file name is the model name, the file content is the markdown content of the model card.
  • model names: /home/v-junliang/DNNGen/auto_model_dev/graph/huggingface_model_names, each file is a model name set.
  • dataset descriptions: /data/data0/v-junliang/DNNGen/auto_model_dev/dataset_desc, one file for each dataset, the file name is the dataset index name(used in url) of paperswithcode, the file contains the dataset formal name and the description.
  • task descriptions: /data/data0/v-junliang/DNNGen/auto_model_dev/task_desc, one file for each task, the file name is the task index name of paperswithcode, the file contains the task formal name and the description.
• mapping
  • dataset to task: /data/data0/v-junliang/DNNGen/auto_model_dev/dataset_to_task, one file for each dataset, the file name is the dataset index name of paperswithcode, the file contains tasks that the dataset is used on.
  • huggingface
    • model to dataset: /data/data0/v-junliang/DNNGen/auto_model_dev/mapping/huggingface/model_to_datasets, one file for each model, the file contains datasets it the model used on(according to fuzzywuzzy, not sure for correctness).
    • triples: each file is under the folder structure {task}/{dataset}/{model}, the file content is a (model name, dataset index name, task index name) tuple.
• graph
  • it's better to use the latest tracer to produce new xml graphs that are wanted.
  • already produced xml graphs(each of the below folders contains only a part of the whole models, i.e., there are many models not traced):
    • 1.0: /data/data0/v-junliang/DNNGen/graph2seq1.0, xml graphs are not folded and have arguments information for module node.
    • 2.0: /data/data0/v-junliang/DNNGen/graph2seq2.0, not folded, and with shape propagation and dce(dce here may not be correct).
    • 3.0: /data/data0/v-junliang/DNNGen/graph2seq3.0, folded to the first module level.
    • 4.0: internblob v-junliang xml, unfolded, without shape propagation, without dce.

Unsolved

• mapping
  • huggingface
    • model to dataset: mapping/tried_models is the set of tried models, but it is not complete. And there is some error running the model_to_dataset.py script.
    • build (model, dataset, task) tuples: waiting for the above one.
  • mmdetection
    • model to dataset: not started.
    • build (model, dataset, task) tuples: not started.
• execute concrete trace for all models
• generate samples for both huggingface and mmdetection.