def run_benchmark(self):
# Warm up, ignore execution time value
_, _ = block_forward_backward_and_time(block=self.block, runs=self.warmup, x=self.data)
# Run Benchmarks
exe_time, _ = block_forward_backward_and_time(block=self.block, runs=self.runs, x=self.data)
self.results["MX_Gluon_Imperative_PRelu_Forward_Backward_Time"] = exe_time / self.runs
def run_benchmark(self):
# NOTE: Below we are calling as (self.data, block...) purposefully.
# Because, Lambda(..) block DO NOT take a named input and we can't call
# block_forward_backward_and_time(..) with unnamed parameter self.data at the end.
# Warm up, ignore execution time value
_, _ = block_forward_backward_and_time(self.data,
block=self.block,
runs=self.warmup)
# Run Benchmarks
exe_time, _ = block_forward_backward_and_time(self.data,
block=self.block,
runs=self.runs)
self.results[
"MX_Gluon_Imperative_Lambda_Batchify_Forward_Backward_Time"] = exe_time / self.runs
def run_benchmark(self):
# TODO: Conv2DTranspose performance on CPU is real bad. Not recommended to run for now.
if self.ctx == mx.cpu():
print(
"WARNING: Conv2DTranspose performance on CPU is real bad. Not recommended to run for now!!"
)
self.results[
"MX_Gluon_Imperative_Conv2DTranspose_Forward_Backward_Time"] = 0
return
# Warm up, ignore execution time value
_, _ = block_forward_backward_and_time(block=self.block,
runs=self.warmup,
x=self.data)
# Run Benchmarks
exe_time, _ = block_forward_backward_and_time(block=self.block,
runs=self.runs,
x=self.data)
self.results[
"MX_Gluon_Imperative_Conv2DTranspose_Forward_Backward_Time"] = exe_time / self.runs
