def performance_summary(model, dummy_input, opt=None, prefix=""):
try:
df = distiller.model_performance_summary(model.module, dummy_input)
except AttributeError:
df = distiller.model_performance_summary(model, dummy_input)
new_entry = {
"Name": ["Total"],
"MACs": [df["MACs"].sum()],
}
MAC_total = df["MACs"].sum()
return MAC_total
def performance_summary(model, dummy_input, opt=None, prefix=""):
try:
df = distiller.model_performance_summary(model.module, dummy_input)
except AttributeError:
df = distiller.model_performance_summary(model, dummy_input)
new_entry = {
'Name':['Total'],
'MACs':[df['MACs'].sum()],
}
MAC_total = df['MACs'].sum()
return MAC_total
def get_experiment_performance_summary(chkpt_fname, dataset, arch, validate_fn):
model = create_model(False, dataset, arch)
model, compression_scheduler, start_epoch = apputils.load_checkpoint(model, chkpt_fname)
dummy_input = get_dummy_input(dataset)
perf_df = distiller.model_performance_summary(model, dummy_input, 1)
total_macs = perf_df['MACs'].sum()
top1, top5, vloss = validate_fn(model=model, epoch=-1)
return total_macs, distiller.model_numel(model), top1
def test_sg_macs():
'''Compare the MACs of different modules as computed by a SummaryGraph
and model summary.'''
import common
sg = create_graph('imagenet', 'mobilenet')
assert sg
model, _ = common.setup_test('mobilenet', 'imagenet', parallel=False)
df_compute = distiller.model_performance_summary(model, distiller.get_dummy_input('imagenet'))
modules_macs = df_compute.loc[:, ['Name', 'MACs']]
for name, mod in model.named_modules():
if isinstance(mod, (nn.Conv2d, nn.Linear)):
summary_macs = int(modules_macs.loc[modules_macs.Name == name].MACs)
sg_macs = sg.find_op(name)['attrs']['MACs']
assert summary_macs == sg_macs
def test_compute_summary():
dataset = "cifar10"
arch = "simplenet_cifar"
model, _ = common.setup_test(arch, dataset, parallel=True)
df_compute = distiller.model_performance_summary(
model, distiller.get_dummy_input(dataset))
module_macs = df_compute.loc[:, 'MACs'].to_list()
# [conv1, conv2, fc1, fc2, fc3]
assert module_macs == [352800, 240000, 48000, 10080, 840]
dataset = "imagenet"
arch = "mobilenet"
model, _ = common.setup_test(arch, dataset, parallel=True)
df_compute = distiller.model_performance_summary(
model, distiller.get_dummy_input(dataset))
module_macs = df_compute.loc[:, 'MACs'].to_list()
expected_macs = [
10838016, 3612672, 25690112, 1806336, 25690112, 3612672, 51380224,
903168, 25690112, 1806336, 51380224, 451584, 25690112, 903168,
51380224, 903168, 51380224, 903168, 51380224, 903168, 51380224, 903168,
51380224, 225792, 25690112, 451584, 51380224, 1024000
]
assert module_macs == expected_macs
def save_model_stats(model, save_path, name=None):
if not os.path.exists(save_path):
os.makedirs(save_path)
sparsity_df, _ = distiller.weights_sparsity_summary(model, True)
performance_df = distiller.model_performance_summary(
model, torch.rand([1, 3, 352, 352]), 1)
if name:
sparsity_df.to_csv('{}/{}_{}.sparsity.csv'.format(
save_path, model.name, name))
performance_df.to_csv('{}/{}_{}.performance.csv'.format(
save_path, model.name, name))
else:
sparsity_df.to_csv('{}/{}.sparsity.csv'.format(save_path, model.name))
performance_df.to_csv('{}/{}.performance.csv'.format(
save_path, model.name))
def macs_display(model, macs_file, dummy_input):
"""
分析网络的计算量MACs
Arguments:
model (class Net): 已训练好的模型 \n
macs_file (str): 存储计算量结果的文件名称和位置 \n
dummy_input (tensor): 和网络输入相同维度的dummy数据, for example, torch.FloatTensor(1, 3, 128, 128), [batch, channel, height, width]
Examples:
>>> from apputils.platform_summaries import *
>>> macs_display(model, 'macs_file.xlsx'， torch.FloatTensor(1, 3, 128, 128))
"""
df_macs = distiller.model_performance_summary(model, dummy_input, 1)
df_macs.to_csv(macs_file)
netG_MACs = df_macs['MACs'].sum()
msglogger.info("netG MACs: " + "{:,}".format(int(netG_MACs)))
