zeionara / traice

Tiny yet useful tool for consistent model training logs generation

traice

Tiny yet useful tool for consistent model training logs generation.

Installation

To install through pip use the following command:

pip install traice

The tool requires only pandas package to be installed. However, there is environment.yml file which can be used for the same environment which is used for developing the tool:

conda env create -f environment.yml

Usage

The tool may be used as follows (see examples/dummy.py):

from random import seed, uniform
from time import time, sleep

from traice import Traicer

traicer = Traicer()


def train_step():
    sleep(uniform(0, 1))


seed(17)

init_timestamp = time()

for i in range(1, 5):
    start_timestamp = time()
    train_step()
    traicer.push(i, uniform(0, 1 / i), (time_ := time()) - start_timestamp, time_ - init_timestamp)

print(traicer.df)

Essentially, it accumulates all push arguments in a list which is then converted to a dataframe. The example produces the following log (the last two columns may differ a bit):

   epoch      loss      time  cumulative_time
0      1  0.806691  0.522609         0.522609
1      2  0.144813  0.961565         1.484184
2      3  0.234740  0.767061         2.251254
3      4  0.027541  0.661659         2.912921

Testing

To run test execute the following statement in your terminal:

python -m unittest discover test