Skip to content

Latest commit

 

History

History
121 lines (95 loc) · 4.48 KB

README.md

File metadata and controls

121 lines (95 loc) · 4.48 KB
Raw

Overview

This tool allows to work collect and visualize easily scripts runned with Hydra. With it you can sort runs and experiment per tag, hyperparameters, group, etc. You can also export and visualize them by different ways using streamlit and plotly.

Installation

  1. Clone the repo via: git clone https://github.com/csem/experiment_tracker.git
  2. install it with: pip install -e experiment_tracker/

Getting started

Try out the example

Populate the outputs folder

Once you have installed the package you can try out the example by running:

python examples/training.py

This will create a folder called outputs. You can also try out different options by running:

python examples/python examples/training.py
.py --help

Whenever you run the example you will get a new folder in outputs with the current date and time. Inside this folder you will find the results of your experiment with the configuration you have specified. If you want to populate quickly the outputs folder with some results you can run:

source examples/populate_outputs.sh

Visualize the results

Now you have some results in your outputs folder you can visualize them by running:

streamlit run examples/streamlit_example.py

Collect your results

Let's assume you have run an experiment "my_experiment" of 3 runs (see below for the terminology) on 2021-01-26 at 17:07:39. And that, for each run you are saving a npy file and a log file.

my_experiment/
└── 2021-01-26
    └── 17-07-39
        ├── 0
        │   ├── .hydra
        │   │   ├── config.yaml
        │   │   ├── hydra.yaml
        │   │   └── overrides.yaml
        │   ├── lstm_finetuning.log
        │   └── metric.npy
        ├── 1
        │   ├── .hydra
        │   │   ├── config.yaml
        │   │   ├── hydra.yaml
        │   │   └── overrides.yaml
        │   ├── lstm_finetuning.log
        │   └── metric.npy
        ├── 2
        │   ├── .hydra
        │   │   ├── config.yaml
        │   │   ├── hydra.yaml
        │   │   └── overrides.yaml
        │   ├── lstm_finetuning.log
        │   └── metric.npy
        └── multirun.yaml

The following code snippets will return a pd.DataFrame containing the parameters for each and the path to all files with npy suffix.

import experiment_tracker
df = experiment_tracker.load_project(base_path="experiments", query_string="*.ckpt", logic="multirun") # logic can either be "multirun" [if hydra is run with -m flag] or "singlerun"

You can also specify "filter_hms" argument in order to collect only day in a specific day

df = experiment_tracker.load_project(base_path="final_results/train_wc_uncond", query_string="results.json", filter_hms="2022-10-09")

if you want to return also the path to your log files you can pass a list instead of a string as parameter to query_string

from experiment_tracker import process
loader = process.file_loader.FileLoader(query_string=["*.npy","*.log"])
df = loader.load_project("my_experiment")

The returned pd.DataFrame contains the paths of your results relative to the folder where you run your code the parameters, from .hydra/config, of each run. The pd.DataFrame has hierchical structure for rows.

  • Rows:
    • Level 0: Date: The date of your experiment (i.e. 2021-01-26 17-07-39)
    • Level 1: Run: The integer representing the value of your run (0 to 3)
  • Columns:
    • Level 0: Parameters: Hydra config variables + [collected_path_0,..,collected_path_n]

You can know more about hierchical pandas dataframe here: https://pandas.pydata.org/pandas-docs/stable/user_guide/advanced.html

Once you have the pandas dataframe you can do whataver you want with it! Here some examples:

  1. Going through each run and load the npy file in a list, with the same order as the pd.DataFrame.
results = [np.load(x,allow_pickle=True) for x in  df.loc[:,collected_path_0]]
  1. Get all parameters for run 1
df.loc[ ('2021-01-26 17-07-39', 1),:]

Terminology:

  • Project: A collection of one or more experiments. (Each Project has one or more days sub-folders, and then one or more time sub-folders)
  • Experiment: A collection of one or more runs. Each experiment contains one or more runs subfolders
  • Run: A training of a learning algorithm plus its performance evaluation