def test_all_networks(self):
"""
Get all_networks.
"""
self.assertIn('alex_net', nns.all_networks())
self.assertIn('vgg_net', nns.all_networks())
self.assertGreater(len(nns.all_networks()), 5)
def do_scheduling_efficients():
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
# Resource.
arch_info, dataflow_info = extract_info(arch_file, dataflow_file)
resource = Resource.arch(arch_info)
# Unroll loop lower bound
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
# batch size = 4
batch_size.init(4)
for net in all_networks():
# Network.
network = import_network(net)
print("\n============================================")
print(network.net_name)
print("waiting...")
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network, resource, cost_model, loop_lower_bound, z_fusion=True, womincost=True)
schedule_info_list, _ = sg.schedule_search()
print("done!\n\n")
res_parse(schedule_info_list, resource,
cost_model, sg, network,
loop_lower_bound,
'./result/overall_experiment/efficients', arch_info, True)
def argparser():
"""
Argument parser.
"""
ap = argparse.ArgumentParser(description="AutoMorpher")
# ===================================================================================
# accelerator architecture
# ===================================================================================
ap.add_argument("arch", help="architecture specification")
# ===================================================================================
# model and batch size
# ===================================================================================
ap.add_argument('net',
help='network name, should be a .py file under "nns". '
'Choices: {}.'.format(', '.join(all_networks())))
ap.add_argument("-d",
"--dataflow",
help="restriction of the dataflow space")
ap.add_argument('--batch', type=int, default=1, help='batch size')
# ===================================================================================
# reuse optimization selection
# ===================================================================================
ap.add_argument('--path', default='.', help='path to store results')
# ===================================================================================
# verbose
# ===================================================================================
ap.add_argument('-v',
'--verbose',
action='store_true',
help='show progress and details.')
args = ap.parse_args()
return args
def do_scheduling_woFullSpace1():
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
# Resource.
arch_info, dataflow_info = extract_info(arch_file, dataflow_file)
resource = Resource.arch(arch_info)
# Unroll loop lower bound
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
# batch size = 4
access_list = []
batch_size.init(4)
for net in all_networks():
# Network.
network = import_network(net)
print("\n============================================")
print(network.net_name)
print("waiting...")
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network, resource, cost_model, loop_lower_bound, d_fusion=True)
schedule_info_list, _ = sg.schedule_search()
print("done!\n\n")
_, access = res_parse(schedule_info_list, resource,
cost_model, sg, network,
loop_lower_bound,
'./result/analysis/woFullSpace1', arch_info, True, is_access=True)
access_list.append(int(access))
return access_list
def main():
"""
Main function.
"""
print('\n')
print("*" * 60)
print('HaFS:')
access_hafs = do_scheduling_optimus()
print('\n')
print("*" * 60)
print('w/o MinCost:')
access_woMinCost = do_scheduling_woMinCost()
print('\n')
print("*" * 60)
print('w/o full space1:')
access_woFullSpace1 = do_scheduling_woFullSpace1()
print('\n')
print("*" * 60)
print('w/o full space2:')
access_woFullSpace2 = do_scheduling_woFullSpace2()
print('\n')
print("*" * 60)
print('w/o fusion:')
access_woFusion = do_scheduling_woFusion()
x = list(range(len(access_hafs)))
total_width, n = 1, 6
width = total_width / n
plt.figure(figsize=(10, 4))
plt.bar(x, list(np.array(access_hafs) / np.array(access_hafs)),
width=width, label='HaFS', color='r', edgecolor='black')
for i in range(len(x)):
x[i] = x[i] + width
plt.bar(x, list(np.array(access_woMinCost) / np.array(access_hafs)),
width=width, label='w/o MinCost', color='r', edgecolor='black')
for i in range(len(x)):
x[i] = x[i] + width
plt.bar(x, list(np.array(access_woFullSpace1) / np.array(access_hafs)), width=width,
label='w/o full space1', tick_label=list(all_networks()), color='r', edgecolor='black')
for i in range(len(x)):
x[i] = x[i] + width
plt.bar(x, list(np.array(access_woFullSpace2) / np.array(access_hafs)),
width=width, label='w/o full space2', color='r', edgecolor='black')
for i in range(len(x)):
x[i] = x[i] + width
plt.bar(x, list(np.array(access_woFusion) / np.array(access_hafs)),
width=width, label='w/o fusion', color='r', edgecolor='black')
plt.ylabel("Normalized Access")
plt.xlabel("models")
plt.ylim(0.0, 4.0)
plt.legend()
plt.savefig('./result/analysis/analysis.png')
plt.show()
return 0
def test_import_network(self):
"""
Get import_network.
"""
batch_size.init(3)
for name in nns.all_networks():
network = nns.import_network(name)
self.assertIsInstance(network, Network)
