Skip to content

Latest commit

 

History

History
74 lines (53 loc) · 3.04 KB

logging.md

File metadata and controls

74 lines (53 loc) · 3.04 KB
Raw

Logging System

Core Features

Colored & Unified Logging & Printing

Built upon rich, we aim to provide a comprehensive and easy-to-access logging system where you don't have to deal with confusing logging levels. We simply want to print something to screen and the log and print functions are built just for that.

# Print this message, record where and when it was printed, make the path blue
log(f'Video generated: {blue(output_path)}')

Call External Function with Grace

The run and read function will cause os.system but record your commands and its output to the console. Note that if peak performance is required, try running the command directly in Python like replacing os.system('mv xxx yyy') with os.rename('xxx', 'yyy'). There will be a notable throughput difference.

# Call ffmpeg using the `run` interface
def generate_video(...):
    ...
    run(cmd)
    return output

Loss Printing

For easier logging and monitoring of the training process, we create and maintain a rich table for the loss information. The table will be updated every time you call display_table and will be updated on screen according to the frequency you set by replacing the global variable live.

# Render table to screen
log_stats = dotdict(psnr=32.0, ssim=1.0, loss=0.0)
display_table(log_stats)  # render dict as a table (live, console, table)

Progress Bar

We also provide a progress bar for you to monitor the progress of your training. We speicifically modified the progress bar to have similar APIs like tqdm, while at the same time print messages with the global console object. Since all printing actions are managed by the same console, the logging, loss table printing and progress bar update will not interfere with each other.

for i in tqdm(range(10), back=2, desc='Progress', disable=False):
    log(f'index: {i}')

Exception Traceback & Debugging

The integrated way for debugging a EasyVolcap application is to use the command-line debugger pdbr. By placing a breakpoint() statement at the line you want the interpreter to stop at, you can run the original command anywhere (like on a remote server with only a terminal).

Another useful trick is to wrap your main function with the decorator catch_throw, which we've monkey-patched to stop if some unsolved exception was raised in that function. All main entry points of EasyVolcap are wrapped with this decorator.

@catch_throw
def main():
    ...

Using EasyVolcap's Logging Elsewhere

We specifically designed the logging system of EasyVolcap to be light-weight and flexible enough. To use it elsewhere, like in an external script, simply do:

from easyvolcap.utils.console_utils import *

This will create a global console variable along with a set of convenience functions. See some of my latest scripts for examples (depth_fusion.py, volume_fusion).