def runThroughput(modelName, net, params):
log.info('Running prediction...')
b = params[const.BATCH]
fns = getter.apiGetTestInputs(modelName, b)
tsum = 0
# Load the entire batch into system/accelerator memory for iterating
imgs = []
for i in range(b):
img = preprocessor.apiProcess(modelName, fns.pop())
imgs.append(img)
# TPU takes different format, unlike others
imgs = np.array(imgs)
imgs = imgs.flatten()
# Allow one warmup prediction outside the overall timing loop
times = []
t0 = time.time()
net.predict(imgs)
tn = time.time()
t = tn - t0
times.append(t)
# Overall timing loop for throughput
t_start = time.time()
for i in range(params[const.RUNITERATIONS]):
t0 = time.time()
net.predict(imgs)
tn = time.time()
t = tn - t0
times.append(t)
t_finish = time.time()
times.insert(0, t_finish - t_start)
return times
def runThroughput(modelName, net, params):
log.info('Running prediction...')
b = params[const.BATCH]
fns = getter.apiGetTestInputs(modelName, b)
tsum = 0
# Load the entire batch into system/accelerator memory for iterating
imgs = []
for i in range(b):
if params[const.PRECISION] == const.FP32:
img = preprocessor.apiProcess(modelName, fns.pop())
if params[const.PRECISION] == const.INT8:
img = preprocessor.apiProcess_int8(modelName, fns.pop())
imgs.append(img)
# Allow one warmup prediction outside the overall timing loop
times = []
t = net.predict_runtime(imgs, params)
times.append(t)
# Overall timing loop for throughput
t_start = time.time()
for i in range(params[const.RUNITERATIONS]):
# individual timing for latency; we don't support concurrency > 1
t = net.predict_runtime(imgs, params)
times.append(t)
t_finish = time.time()
times.insert(0, t_finish - t_start)
return times
def runLatency(modelName, net, params):
log.info('Running prediction...')
fns = getter.apiGetTestInputs(modelName, 1)
img = preprocessor.apiProcess(modelName, fns[0])
times = []
for j in range(params[const.RUNITERATIONS]):
t0 = time.time()
res = net.predict(img)
tn = time.time()
t = tn - t0
log.debug('%f sec' % t)
times.append(t)
return times
def runThroughput(modelName, net, params):
log.info('Running prediction...')
b = params[const.BATCH]
fns = getter.apiGetTestInputs(modelName, b)
tsum = 0
# Load the entire batch into system/accelerator memory for iterating
imgs_list = []
for i in range(b):
img = preprocessor.apiProcess(modelName, fns[i])
imgs_list.append(img)
imgs_np = numpy.asarray(imgs_list)
results = net.predict(fns[0], imgs_np, 5, 0, params[const.RUNITERATIONS])
#tsum=sum(results['seconds'])
# New schema is [ total_time_s, time0, time1, ..., timeN ]
return results['seconds']
def run(modelName, modelFileName, params):
# TODO careful this over-writes (not appends to) original
if (params[const.HARDWARE] == 'gpu'):
os.environ['LD_LIBRARY_PATH'] = commonDir + ':' + os.path.join(
commonDir, 'GpuAcc')
else:
os.environ['LD_LIBRARY_PATH'] = commonDir + ':' + os.path.join(
commonDir, 'CpuAcc')
tmpdir = tempfile.mkdtemp()
inputFn = os.path.join(tmpdir, 'inputs.json')
outputFn = os.path.join(tmpdir, 'outputs.json')
if params[const.MODE] == const.ACCURACY:
imageFileNames = getter.apiGetValidationInputs(modelName, cache=True)
else:
imageFileNames = getter.apiGetTestInputs(modelName,
params[const.BATCH],
cache=True)
cxxParams = {
'images':
imageFileNames, # Note: this can be up to 5,000 filenames
'model':
os.path.join(paths.MODELS, 'tensorflow', modelName, 'frozen_graph.pb'),
'params':
params
}
with open(inputFn, 'w') as fp:
json.dump(cxxParams, fp)
exeCmd = os.path.join(commonDir, '..', modelName, modelName + '.exe')
cmd = [exeCmd, inputFn, outputFn]
log.info('Running prediction...')
log.debug(cmd)
ret = subprocess.call(cmd)
if ret:
log.error('Inference failed')
return None
log.info('Loading results file %s' % outputFn)
with open(outputFn) as fp:
returnData = json.load(fp)
if params[const.MODE] == const.ACCURACY:
return returnData['predictions']
else:
return returnData['times']
def runThroughput(modelName, net, params):
log.info('Running throughput...')
fns = getter.apiGetTestInputs(modelName, params[const.BATCH], cache=True)
times = net.predict(fns, params[const.RUNITERATIONS])
return times
def runLatency(modelName, net, params):
log.info('Running latency...')
fns = getter.apiGetTestInputs(modelName, 1, cache=True)
times = net.predict(fns, params[const.RUNITERATIONS])
return times
