def build_lookup_table(self, network_def_full, resource_type, lookup_table_path,
min_conv_feature_size=_MIN_CONV_FEATURE_SIZE,
min_fc_feature_size=_MIN_FC_FEATURE_SIZE,
measure_latency_batch_size=_MEASURE_LATENCY_BATCH_SIZE,
measure_latency_sample_times=_MEASURE_LATENCY_SAMPLE_TIMES,
verbose=True):
# Build lookup table for latency
'''
please refer to def build_latency_lookup_table(...) in functions.py
'''
return fns.build_latency_lookup_table(network_def_full, lookup_table_path,
min_conv_feature_size=min_conv_feature_size,
min_fc_feature_size=min_fc_feature_size,
measure_latency_batch_size=measure_latency_batch_size,
measure_latency_sample_times=measure_latency_sample_times,
verbose=verbose)
num_w = fns.compute_resource(network_def, 'WEIGHTS')
flops = fns.compute_resource(network_def, 'FLOPS')
num_param = fns.compute_resource(network_def, 'WEIGHTS')
print('Number of FLOPs: ', flops)
print('Number of weights: ', num_w)
print('Number of parameters: ', num_param)
print('-------------------------------------------')
model = model.cuda()
print('Building latency lookup table for', torch.cuda.get_device_name())
if build_lookup_table:
fns.build_latency_lookup_table(
network_def,
lookup_table_path=lookup_table_path,
min_fc_feature_size=MIN_FC_FEATRE_SIZE,
min_conv_feature_size=MIN_CONV_FEATURE_SIZE,
measure_latency_batch_size=MEASURE_LATENCY_BATCH_SIZE,
measure_latency_sample_times=MEASURE_LATENCY_SAMPLE_TIMES,
verbose=True)
print('-------------------------------------------')
print('Finish building latency lookup table.')
print(' Device:', torch.cuda.get_device_name())
print(' Model: ', model_arch)
print('-------------------------------------------')
latency = fns.compute_resource(network_def, 'LATENCY', lookup_table_path)
print('Computed latency: ', latency)
latency = fns.measure_latency(
model, [MEASURE_LATENCY_BATCH_SIZE, *INPUT_DATA_SHAPE])
print('Exact latency: ', latency)