Tencent / PocketFlow

An Automatic Model Compression (AutoMC) framework for developing smaller and faster AI applications.
https://pocketflow.github.io
Other
2.79k stars 490 forks source link

How to show the difference between the compressed and uncompressed model,and the difference between accuracys #288

Open itracy19 opened 5 years ago

itracy19 commented 5 years ago

Dear Partners,

When i use amc to compress a model,i want to check the detail effect,so,please give some advice,thanks a lot。Commands i run are as follows:

  1. run "./scripts/run_local.sh nets/resnet_at_ilsvrc12_run.py" to generate model in ./models/

  2. run "./scripts/run_local.sh nets/resnet_at_cifar10_run.py --learner channel --cp_prune_option auto --cp_preserve_ratio 0.5" to prune model with amc,and get models as follows:

/PocketFlow$ ll models total 36240 drwxrwxr-x 2 yuanxuepeng yuanxuepeng 4096 5月 9 20:40 ./ drwxrwxr-x 19 yuanxuepeng yuanxuepeng 4096 5月 10 11:24 ../ -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 1095656 5月 7 17:19 best_model.ckpt.data-00000-of-00001 -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 3648 5月 7 17:19 best_model.ckpt.index -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 2978537 5月 7 17:19 best_model.ckpt.meta -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 613 5月 9 20:40 checkpoint -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 1095656 5月 9 20:24 model.ckpt-60000.data-00000-of-00001 -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 3648 5月 9 20:24 model.ckpt-60000.index -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 4398582 5月 9 20:24 model.ckpt-60000.meta -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 1095656 5月 9 20:29 model.ckpt-70000.data-00000-of-00001 -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 3648 5月 9 20:29 model.ckpt-70000.index -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 4398582 5月 9 20:29 model.ckpt-70000.meta -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 1095656 5月 9 20:33 model.ckpt-80000.data-00000-of-00001 -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 3648 5月 9 20:33 model.ckpt-80000.index -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 4398582 5月 9 20:33 model.ckpt-80000.meta -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 1095656 5月 9 20:37 model.ckpt-90000.data-00000-of-00001 -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 3648 5月 9 20:37 model.ckpt-90000.index -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 4398582 5月 9 20:37 model.ckpt-90000.meta -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 1095656 5月 9 20:40 model.ckpt-97656.data-00000-of-00001 -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 3648 5月 9 20:40 model.ckpt-97656.index -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 4398582 5月 9 20:40 model.ckpt-97656.meta -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 1095656 5月 7 15:03 original_model.ckpt.data-00000-of-00001 -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 3648 5月 7 15:03 original_model.ckpt.index -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 338201 5月 7 15:03 original_model.ckpt.meta -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 1095656 5月 7 16:40 pruned_model.ckpt.data-00000-of-00001 -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 3648 5月 7 16:40 pruned_model.ckpt.index -rw-rw-r-- 1 yuanxuepeng yuanxuepeng 2961706 5月 7 16:40 pruned_model.ckpt.meta

  1. Then,i wonder how to check the difference between original_model and pruned_model from model size to accurracy.

Thanks