p95768021
/
InfiniTensor
forked from jiuyuan/InfiniTensor
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Add: save optime result

This commit is contained in:
Liyan Zheng 2022-11-02 17:38:08 +08:00
parent eb993f7829
commit 1b1fc2585b
2 changed files with 7 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
python/infinitensor/evalute_onnx.py
View File

@ -475,4 +475,4 @@ if __name__ == "__main__":
# parser.add_argument("--output", help="Output file") # parser.add_argument("--output", help="Output file")
args = parser.parse_args() args = parser.parse_args()
import_onnx(args.model, args.bs) import_onnx(args.model, args.bs)
print_result() print_result(args.model)

8
python/infinitensor/rules.py
View File

@ -1,6 +1,8 @@
import pandas as pd import pandas as pd
import numpy as np import numpy as np
from operator_timer import * from operator_timer import *
from datetime import datetime
pd.options.display.float_format = '{:,.3f}'.format pd.options.display.float_format = '{:,.3f}'.format
df= pd.DataFrame(columns=['n', 'c', 'h', 'w', 'f', 'r', 's', 'ph', 'pw', 'sh', 'sw', 'dh', 'dw', 'oph', 'opw', 'group']) df= pd.DataFrame(columns=['n', 'c', 'h', 'w', 'f', 'r', 's', 'ph', 'pw', 'sh', 'sw', 'dh', 'dw', 'oph', 'opw', 'group'])
@ -72,9 +74,11 @@ def conv_tranpsposed2d_togemm(name, n, c, h, w, f, r, s, ph, pw, sh, sw, dh, dw,
conv_transposed2d_rules=[conv_tranpsposed2d_original, conv_tranpsposed2d_togemm] conv_transposed2d_rules=[conv_tranpsposed2d_original, conv_tranpsposed2d_togemm]
def print_result(): def print_result(model_fn):
pd.set_option('display.max_rows', 500)
df['t_min'] = df.filter(regex=("^t_.*")).min(axis=1) df['t_min'] = df.filter(regex=("^t_.*")).min(axis=1)
print(df) print(df)
print(f'Origin: {df["t_original"].sum():.3f} ms') print(f'Origin: {df["t_original"].sum():.3f} ms')
print(f'Min: {df["t_min"].sum():.3f} ms') print(f'Min: {df["t_min"].sum():.3f} ms')
print(f'Speedup: {df["t_original"].sum()/df["t_min"].sum():.3f} x') print(f'Speedup: {df["t_original"].sum()/df["t_min"].sum():.3f} x')
df.to_pickle(f'optime_{model_fn.split("/")[-1]}_{datetime.now().strftime("%m_%d_%H_%M_%S")}.pkl')