Skip to content

Commit 9bf8a7c

Browse files
vinhngxnarendasan
vinhngx
authored and
narendasan
committed
fix benchmark
Signed-off-by: Vinh Nguyen <[email protected]>
1 parent 60ab93a commit 9bf8a7c

File tree

1 file changed

+116
-85
lines changed

1 file changed

+116
-85
lines changed

‎notebooks/LeNet-example.ipynb

Lines changed: 116 additions & 85 deletions
Original file line numberDiff line numberDiff line change
@@ -2,7 +2,7 @@
22
"cells": [
33
{
44
"cell_type": "code",
5-
"execution_count": 9,
5+
"execution_count": 1,
66
"metadata": {},
77
"outputs": [],
88
"source": [
@@ -80,7 +80,7 @@
8080
},
8181
{
8282
"cell_type": "code",
83-
"execution_count": 3,
83+
"execution_count": 2,
8484
"metadata": {},
8585
"outputs": [],
8686
"source": [
@@ -134,33 +134,44 @@
134134
},
135135
{
136136
"cell_type": "code",
137-
"execution_count": 25,
137+
"execution_count": 9,
138138
"metadata": {},
139139
"outputs": [],
140140
"source": [
141141
"import time\n",
142142
"import numpy as np\n",
143143
"\n",
144+
"import torch.backends.cudnn as cudnn\n",
145+
"cudnn.benchmark = True\n",
146+
"\n",
144147
"def benchmark(model, input_shape=(1024, 1, 32, 32), dtype='fp32', nwarmup=50, nruns=10000):\n",
145148
" input_data = torch.randn(input_shape)\n",
146149
" input_data = input_data.to(\"cuda\")\n",
147150
" if dtype=='fp16':\n",
148151
" input_data = input_data.half()\n",
149152
" \n",
150-
" for _ in range(nwarmup):\n",
151-
" results = model(input_data)\n",
153+
" print(\"Warm up ...\")\n",
154+
" with torch.no_grad():\n",
155+
" for _ in range(nwarmup):\n",
156+
" features = model(input_data)\n",
157+
" torch.cuda.synchronize()\n",
158+
" print(\"Start timing ...\")\n",
159+
" timings = []\n",
160+
" with torch.no_grad():\n",
161+
" for i in range(1, nruns+1):\n",
162+
" start_time = time.time()\n",
163+
" features = model(input_data)\n",
164+
" torch.cuda.synchronize()\n",
165+
" end_time = time.time()\n",
166+
" timings.append(end_time - start_time)\n",
167+
" if i%1000==0:\n",
168+
" print('Iteration %d/%d, ave batch time %.2f ms'%(i, nruns, np.mean(timings)*1000))\n",
169+
"\n",
170+
" print(\"Input shape:\", input_data.size())\n",
171+
" print(\"Output features size:\", features.size())\n",
152172
" \n",
153-
" start_time = time.time()\n",
154-
" time_arr = []\n",
155-
" for _ in range(1, nruns+1):\n",
156-
" start_time = time.time()\n",
157-
" results = model(input_data)\n",
158-
" time_arr.append(time.time() - start_time)\n",
159-
" \n",
160-
" if _%1000==0:\n",
161-
" print('Iteration %d, ave batch time %.2f ms'%(_, np.mean(time_arr)*1000))\n",
162-
" \n",
163-
" print('Average batch time: %.2f ms'%(np.mean(time_arr)*1000)) "
173+
" print('Average batch time: %.2f ms'%(np.mean(timings)*1000))\n",
174+
" "
164175
]
165176
},
166177
{
@@ -172,7 +183,7 @@
172183
},
173184
{
174185
"cell_type": "code",
175-
"execution_count": 26,
186+
"execution_count": 10,
176187
"metadata": {},
177188
"outputs": [
178189
{
@@ -191,7 +202,7 @@
191202
")"
192203
]
193204
},
194-
"execution_count": 26,
205+
"execution_count": 10,
195206
"metadata": {},
196207
"output_type": "execute_result"
197208
}
@@ -203,24 +214,28 @@
203214
},
204215
{
205216
"cell_type": "code",
206-
"execution_count": 27,
217+
"execution_count": 11,
207218
"metadata": {},
208219
"outputs": [
209220
{
210221
"name": "stdout",
211222
"output_type": "stream",
212223
"text": [
213-
"Iteration 1000, ave batch time 0.91 ms\n",
214-
"Iteration 2000, ave batch time 0.90 ms\n",
215-
"Iteration 3000, ave batch time 0.90 ms\n",
216-
"Iteration 4000, ave batch time 0.90 ms\n",
217-
"Iteration 5000, ave batch time 0.90 ms\n",
218-
"Iteration 6000, ave batch time 0.90 ms\n",
219-
"Iteration 7000, ave batch time 0.90 ms\n",
220-
"Iteration 8000, ave batch time 0.90 ms\n",
221-
"Iteration 9000, ave batch time 0.90 ms\n",
222-
"Iteration 10000, ave batch time 0.90 ms\n",
223-
"Average batch time: 0.90 ms\n"
224+
"Warm up ...\n",
225+
"Start timing ...\n",
226+
"Iteration 1000/10000, ave batch time 0.93 ms\n",
227+
"Iteration 2000/10000, ave batch time 0.93 ms\n",
228+
"Iteration 3000/10000, ave batch time 0.93 ms\n",
229+
"Iteration 4000/10000, ave batch time 0.93 ms\n",
230+
"Iteration 5000/10000, ave batch time 0.93 ms\n",
231+
"Iteration 6000/10000, ave batch time 0.93 ms\n",
232+
"Iteration 7000/10000, ave batch time 0.93 ms\n",
233+
"Iteration 8000/10000, ave batch time 0.93 ms\n",
234+
"Iteration 9000/10000, ave batch time 0.93 ms\n",
235+
"Iteration 10000/10000, ave batch time 0.93 ms\n",
236+
"Input shape: torch.Size([1024, 1, 32, 32])\n",
237+
"Output features size: torch.Size([1024, 10])\n",
238+
"Average batch time: 0.93 ms\n"
224239
]
225240
}
226241
],
@@ -241,7 +256,7 @@
241256
},
242257
{
243258
"cell_type": "code",
244-
"execution_count": 29,
259+
"execution_count": 12,
245260
"metadata": {},
246261
"outputs": [
247262
{
@@ -263,7 +278,7 @@
263278
")"
264279
]
265280
},
266-
"execution_count": 29,
281+
"execution_count": 12,
267282
"metadata": {},
268283
"output_type": "execute_result"
269284
}
@@ -275,24 +290,28 @@
275290
},
276291
{
277292
"cell_type": "code",
278-
"execution_count": 30,
293+
"execution_count": 13,
279294
"metadata": {},
280295
"outputs": [
281296
{
282297
"name": "stdout",
283298
"output_type": "stream",
284299
"text": [
285-
"Iteration 1000, ave batch time 0.66 ms\n",
286-
"Iteration 2000, ave batch time 0.66 ms\n",
287-
"Iteration 3000, ave batch time 0.66 ms\n",
288-
"Iteration 4000, ave batch time 0.65 ms\n",
289-
"Iteration 5000, ave batch time 0.65 ms\n",
290-
"Iteration 6000, ave batch time 0.65 ms\n",
291-
"Iteration 7000, ave batch time 0.65 ms\n",
292-
"Iteration 8000, ave batch time 0.65 ms\n",
293-
"Iteration 9000, ave batch time 0.65 ms\n",
294-
"Iteration 10000, ave batch time 0.65 ms\n",
295-
"Average batch time: 0.65 ms\n"
300+
"Warm up ...\n",
301+
"Start timing ...\n",
302+
"Iteration 1000/10000, ave batch time 0.68 ms\n",
303+
"Iteration 2000/10000, ave batch time 0.68 ms\n",
304+
"Iteration 3000/10000, ave batch time 0.68 ms\n",
305+
"Iteration 4000/10000, ave batch time 0.68 ms\n",
306+
"Iteration 5000/10000, ave batch time 0.68 ms\n",
307+
"Iteration 6000/10000, ave batch time 0.68 ms\n",
308+
"Iteration 7000/10000, ave batch time 0.68 ms\n",
309+
"Iteration 8000/10000, ave batch time 0.68 ms\n",
310+
"Iteration 9000/10000, ave batch time 0.68 ms\n",
311+
"Iteration 10000/10000, ave batch time 0.68 ms\n",
312+
"Input shape: torch.Size([1024, 1, 32, 32])\n",
313+
"Output features size: torch.Size([1024, 10])\n",
314+
"Average batch time: 0.68 ms\n"
296315
]
297316
}
298317
],
@@ -311,7 +330,7 @@
311330
},
312331
{
313332
"cell_type": "code",
314-
"execution_count": 32,
333+
"execution_count": 14,
315334
"metadata": {},
316335
"outputs": [],
317336
"source": [
@@ -321,7 +340,7 @@
321340
},
322341
{
323342
"cell_type": "code",
324-
"execution_count": 33,
343+
"execution_count": 15,
325344
"metadata": {},
326345
"outputs": [
327346
{
@@ -343,7 +362,7 @@
343362
")"
344363
]
345364
},
346-
"execution_count": 33,
365+
"execution_count": 15,
347366
"metadata": {},
348367
"output_type": "execute_result"
349368
}
@@ -354,24 +373,28 @@
354373
},
355374
{
356375
"cell_type": "code",
357-
"execution_count": 35,
376+
"execution_count": 16,
358377
"metadata": {},
359378
"outputs": [
360379
{
361380
"name": "stdout",
362381
"output_type": "stream",
363382
"text": [
364-
"Iteration 1000, ave batch time 0.64 ms\n",
365-
"Iteration 2000, ave batch time 0.65 ms\n",
366-
"Iteration 3000, ave batch time 0.65 ms\n",
367-
"Iteration 4000, ave batch time 0.64 ms\n",
368-
"Iteration 5000, ave batch time 0.64 ms\n",
369-
"Iteration 6000, ave batch time 0.64 ms\n",
370-
"Iteration 7000, ave batch time 0.64 ms\n",
371-
"Iteration 8000, ave batch time 0.64 ms\n",
372-
"Iteration 9000, ave batch time 0.64 ms\n",
373-
"Iteration 10000, ave batch time 0.64 ms\n",
374-
"Average batch time: 0.64 ms\n"
383+
"Warm up ...\n",
384+
"Start timing ...\n",
385+
"Iteration 1000/10000, ave batch time 0.68 ms\n",
386+
"Iteration 2000/10000, ave batch time 0.68 ms\n",
387+
"Iteration 3000/10000, ave batch time 0.68 ms\n",
388+
"Iteration 4000/10000, ave batch time 0.68 ms\n",
389+
"Iteration 5000/10000, ave batch time 0.68 ms\n",
390+
"Iteration 6000/10000, ave batch time 0.68 ms\n",
391+
"Iteration 7000/10000, ave batch time 0.68 ms\n",
392+
"Iteration 8000/10000, ave batch time 0.68 ms\n",
393+
"Iteration 9000/10000, ave batch time 0.68 ms\n",
394+
"Iteration 10000/10000, ave batch time 0.68 ms\n",
395+
"Input shape: torch.Size([1024, 1, 32, 32])\n",
396+
"Output features size: torch.Size([1024, 10])\n",
397+
"Average batch time: 0.68 ms\n"
375398
]
376399
}
377400
],
@@ -398,7 +421,7 @@
398421
},
399422
{
400423
"cell_type": "code",
401-
"execution_count": 8,
424+
"execution_count": 17,
402425
"metadata": {},
403426
"outputs": [],
404427
"source": [
@@ -427,24 +450,28 @@
427450
},
428451
{
429452
"cell_type": "code",
430-
"execution_count": 37,
453+
"execution_count": 18,
431454
"metadata": {},
432455
"outputs": [
433456
{
434457
"name": "stdout",
435458
"output_type": "stream",
436459
"text": [
437-
"Iteration 1000, ave batch time 0.19 ms\n",
438-
"Iteration 2000, ave batch time 0.18 ms\n",
439-
"Iteration 3000, ave batch time 0.19 ms\n",
440-
"Iteration 4000, ave batch time 0.19 ms\n",
441-
"Iteration 5000, ave batch time 0.19 ms\n",
442-
"Iteration 6000, ave batch time 0.19 ms\n",
443-
"Iteration 7000, ave batch time 0.19 ms\n",
444-
"Iteration 8000, ave batch time 0.19 ms\n",
445-
"Iteration 9000, ave batch time 0.19 ms\n",
446-
"Iteration 10000, ave batch time 0.19 ms\n",
447-
"Average batch time: 0.19 ms\n"
460+
"Warm up ...\n",
461+
"Start timing ...\n",
462+
"Iteration 1000/10000, ave batch time 0.24 ms\n",
463+
"Iteration 2000/10000, ave batch time 0.24 ms\n",
464+
"Iteration 3000/10000, ave batch time 0.24 ms\n",
465+
"Iteration 4000/10000, ave batch time 0.24 ms\n",
466+
"Iteration 5000/10000, ave batch time 0.24 ms\n",
467+
"Iteration 6000/10000, ave batch time 0.24 ms\n",
468+
"Iteration 7000/10000, ave batch time 0.23 ms\n",
469+
"Iteration 8000/10000, ave batch time 0.23 ms\n",
470+
"Iteration 9000/10000, ave batch time 0.23 ms\n",
471+
"Iteration 10000/10000, ave batch time 0.23 ms\n",
472+
"Input shape: torch.Size([1024, 1, 32, 32])\n",
473+
"Output features size: torch.Size([1, 10])\n",
474+
"Average batch time: 0.23 ms\n"
448475
]
449476
}
450477
],
@@ -463,7 +490,7 @@
463490
},
464491
{
465492
"cell_type": "code",
466-
"execution_count": 39,
493+
"execution_count": 19,
467494
"metadata": {},
468495
"outputs": [],
469496
"source": [
@@ -492,24 +519,28 @@
492519
},
493520
{
494521
"cell_type": "code",
495-
"execution_count": 40,
522+
"execution_count": 20,
496523
"metadata": {},
497524
"outputs": [
498525
{
499526
"name": "stdout",
500527
"output_type": "stream",
501528
"text": [
502-
"Iteration 1000, ave batch time 0.18 ms\n",
503-
"Iteration 2000, ave batch time 0.18 ms\n",
504-
"Iteration 3000, ave batch time 0.18 ms\n",
505-
"Iteration 4000, ave batch time 0.18 ms\n",
506-
"Iteration 5000, ave batch time 0.18 ms\n",
507-
"Iteration 6000, ave batch time 0.18 ms\n",
508-
"Iteration 7000, ave batch time 0.18 ms\n",
509-
"Iteration 8000, ave batch time 0.18 ms\n",
510-
"Iteration 9000, ave batch time 0.18 ms\n",
511-
"Iteration 10000, ave batch time 0.18 ms\n",
512-
"Average batch time: 0.18 ms\n"
529+
"Warm up ...\n",
530+
"Start timing ...\n",
531+
"Iteration 1000/10000, ave batch time 0.20 ms\n",
532+
"Iteration 2000/10000, ave batch time 0.20 ms\n",
533+
"Iteration 3000/10000, ave batch time 0.20 ms\n",
534+
"Iteration 4000/10000, ave batch time 0.21 ms\n",
535+
"Iteration 5000/10000, ave batch time 0.21 ms\n",
536+
"Iteration 6000/10000, ave batch time 0.21 ms\n",
537+
"Iteration 7000/10000, ave batch time 0.21 ms\n",
538+
"Iteration 8000/10000, ave batch time 0.21 ms\n",
539+
"Iteration 9000/10000, ave batch time 0.21 ms\n",
540+
"Iteration 10000/10000, ave batch time 0.21 ms\n",
541+
"Input shape: torch.Size([1024, 1, 32, 32])\n",
542+
"Output features size: torch.Size([1, 10])\n",
543+
"Average batch time: 0.21 ms\n"
513544
]
514545
}
515546
],

0 commit comments

Comments
 (0)