def benchmark(mod, dry_run=10, iterations=10):
if len(mod._context) == 1:
ctx = mod._context[0]
else:
ctx = mx.cpu()
data = [mx.random.uniform(-1.0, 1.0, shape=shape, ctx=ctx) for _, shape in mod.data_shapes]
label = [mx.nd.array(np.random.randint(1, 100, size=shape), ctx=ctx) for _, shape in mod.label_shapes]
batch = mx.io.DataBatch(data, label)
# dry run
for i in range(dry_run):
mod.forward(batch, is_train=True)
mod.backward()
for output in mod.get_outputs(merge_multi_context=False)[0]:
output.wait_to_read()
mod.update()
t0 = time.clock()
profiler.profiler_set_state('run')
# real run
for i in range(iterations):
mod.forward(batch, is_train=True)
mod.backward()
mod.update()
for output in mod.get_outputs(merge_multi_context=False)[0]:
output.wait_to_read()
profiler.profiler_set_state('stop')
t1 = time.clock()
return (t1 - t0)*1000.0 / iterations
def benchmark(mod, dry_run=10, iterations=10):
if len(mod._context) == 1:
ctx = mod._context[0]
else:
ctx = mx.cpu()
data = [mx.random.uniform(-1.0, 1.0, shape=shape, ctx=ctx) for _, shape in mod.data_shapes]
label = [mx.nd.array(np.random.randint(1, 100, size=shape), ctx=ctx) for _, shape in mod.label_shapes]
batch = mx.io.DataBatch(data, label)
# dry run
for i in range(dry_run):
mod.forward(batch, is_train=True)
mod.backward()
for output in mod.get_outputs(merge_multi_context=False)[0]:
output.wait_to_read()
mod.update()
t0 = time.clock()
profiler.profiler_set_state('run')
# real run
for i in range(iterations):
mod.forward(batch, is_train=True)
mod.backward()
mod.update()
for output in mod.get_outputs(merge_multi_context=False)[0]:
output.wait_to_read()
profiler.profiler_set_state('stop')
t1 = time.clock()
return (t1 - t0)*1000.0 / iterations
def test_profiler():
profile_filename = "test_profile.json"
iter_num = 100
begin_profiling_iter = 50
end_profiling_iter = 50
profiler.profiler_set_config(mode='symbolic', filename=profile_filename)
print('profile file save to {0}'.format(profile_filename))
A = mx.sym.Variable('A')
B = mx.sym.Variable('B')
C = mx.symbol.dot(A, B)
executor = C.simple_bind(mx.cpu(1),
'write',
A=(4096, 4096),
B=(4096, 4096))
a = mx.random.uniform(-1.0, 1.0, shape=(4096, 4096))
b = mx.random.uniform(-1.0, 1.0, shape=(4096, 4096))
a.copyto(executor.arg_dict['A'])
b.copyto(executor.arg_dict['B'])
flag = False
print("execution begin")
for i in range(iter_num):
if i == begin_profiling_iter:
t0 = time.clock()
profiler.profiler_set_state('run')
if i == end_profiling_iter:
t1 = time.clock()
profiler.profiler_set_state('stop')
executor.forward()
c = executor.outputs[0]
c.wait_to_read()
print("execution end")
duration = t1 - t0
print('duration: {0}s'.format(duration))
print(' {0}ms/operator'.format(duration * 1000 / iter_num))
def test_profiler():
profile_filename = "test_profile.json"
iter_num = 100
begin_profiling_iter = 50
end_profiling_iter = 50
profiler.profiler_set_config(mode='symbolic', filename=profile_filename)
print('profile file save to {0}'.format(profile_filename))
A = mx.sym.Variable('A')
B = mx.sym.Variable('B')
C = mx.symbol.dot(A, B)
executor = C.simple_bind(mx.cpu(1), 'write', A=(4096, 4096), B=(4096, 4096))
a = mx.random.uniform(-1.0, 1.0, shape=(4096, 4096))
b = mx.random.uniform(-1.0, 1.0, shape=(4096, 4096))
a.copyto(executor.arg_dict['A'])
b.copyto(executor.arg_dict['B'])
flag = False
print("execution begin")
for i in range(iter_num):
if i == begin_profiling_iter:
t0 = time.clock()
profiler.profiler_set_state('run')
if i == end_profiling_iter:
t1 = time.clock()
profiler.profiler_set_state('stop')
executor.forward()
c = executor.outputs[0]
c.wait_to_read()
print("execution end")
duration = t1 - t0
print('duration: {0}s'.format(duration))
print(' {0}ms/operator'.format(duration*1000/iter_num))
def switch_profiler(param):
if param.epoch == 0 and param.nbatch == 100:
profiler.profiler_set_state('run')
if param.epoch == 0 and param.nbatch == 110:
profiler.profiler_set_state('stop')
profiler.dump_profile()