Table of Contents
Shortcuts

How to

Log Analysis

MMPose provides tools/analysis_tools/analyze_logs.py to analyze the training log. The log file can be either a json file or a text file. The json file is recommended, because it is more convenient to parse and visualize.

Currently, the following functions are supported:

  • Plot loss/accuracy curves

  • Calculate training time

Plot Loss/Accuracy Curves

The function depends on seaborn, please install it first by running pip install seaborn.

log_curve

python tools/analysis_tools/analyze_logs.py plot_curve ${JSON_LOGS} [--keys ${KEYS}] [--title ${TITLE}] [--legend ${LEGEND}] [--backend ${BACKEND}] [--style ${STYLE}] [--out ${OUT_FILE}]

Examples:

  • Plot loss curve

    python tools/analysis_tools/analyze_logs.py plot_curve log.json --keys loss_kpt --legend loss_kpt

  • Plot accuracy curve and export to PDF file

    python tools/analysis_tools/analyze_logs.py plot_curve log.json --keys acc_pose --out results.pdf

  • Plot multiple log files on the same figure

    python tools/analysis_tools/analyze_logs.py plot_curve log1.json log2.json --keys loss_kpt --legend run1 run2 --title loss_kpt --out loss_kpt.png

Calculate Training Time

python tools/analysis_tools/analyze_logs.py cal_train_time ${JSON_LOGS} [--include-outliers]

Examples:

python tools/analysis_tools/analyze_logs.py cal_train_time log.json

The result is as follows:

-----Analyze train time of hrnet_w32_256x192.json-----
slowest epoch 56, average time is 0.6924
fastest epoch 1, average time is 0.6502
time std over epochs is 0.0085
average iter time: 0.6688 s/iter

Get Model Params & FLOPs

MMPose provides tools/analysis_tools/get_flops.py to get model parameters and FLOPs.

python tools/analysis_tools/get_flops.py ${CONFIG_FILE} [--shape ${INPUT_SHAPE}] [--cfg-options ${CFG_OPTIONS}]

Description of all arguments:

CONFIG_FILE : The path of a model config file.

--shape: The input shape to the model.

--input-constructor: If specified as batch, it will generate a batch tensor to calculate FLOPs.

--batch-size：If --input-constructor is specified as batch, it will generate a random tensor with shape (batch_size, 3, **input_shape) to calculate FLOPs.

--cfg-options: If specified, the key-value pair optional cfg will be merged into config file.

Example:

python tools/analysis_tools/get_flops.py configs/body_2d_keypoint/topdown_heatmap/coco/td-hm_hrnet-w32_8xb64-210e_coco-256x192.py

We will get the following results:

==============================
Input shape: (1, 3, 256, 192)
Flops: 7.7 GFLOPs
Params: 28.54 M
==============================

Note

This tool is still experimental and we do not guarantee that the number is absolutely correct. Some operators are not counted into FLOPs like GN and custom operators.