def evaluate(model, search, seq_dict, run_box_head, run_cls_head):
"""Compute FLOPs, Params, and Speed"""
custom_ops = {nn.MultiheadAttention: get_complexity_MHA}
# # backbone
macs1, params1 = profile(model,
inputs=(search, None, "backbone", False, False),
custom_ops=None,
verbose=False)
macs, params = clever_format([macs1, params1], "%.3f")
print('backbone (search) macs is ', macs)
print('backbone params is ', params)
# transformer and head
macs2, params2 = profile(model,
inputs=(None, seq_dict, "transformer", True,
True),
custom_ops=custom_ops,
verbose=False)
macs, params = clever_format([macs2, params2], "%.3f")
print('transformer and head macs is ', macs)
print('transformer and head params is ', params)
# the whole model
macs, params = clever_format([macs1 + macs2, params1 + params2], "%.3f")
print('overall macs is ', macs)
print('overall params is ', params)
'''Speed Test'''
T_w = 10
T_t = 100
print("testing speed ...")
with torch.no_grad():
# overall
for i in range(T_w):
_ = model(search, None, "backbone", run_box_head, run_cls_head)
_ = model(None, seq_dict, "transformer", run_box_head,
run_cls_head)
start = time.time()
for i in range(T_t):
_ = model(search, None, "backbone", run_box_head, run_cls_head)
_ = model(None, seq_dict, "transformer", run_box_head,
run_cls_head)
end = time.time()
avg_lat = (end - start) / T_t
print("The average overall latency is %.2f ms" % (avg_lat * 1000))
# backbone
for i in range(T_w):
_ = model(search, None, "backbone", run_box_head, run_cls_head)
start = time.time()
for i in range(T_t):
_ = model(search, None, "backbone", run_box_head, run_cls_head)
end = time.time()
avg_lat = (end - start) / T_t
print("The average backbone latency is %.2f ms" % (avg_lat * 1000))
def evaluate(model, img_x, att_x, q, k, v, key_padding_mask):
"""Compute FLOPs, Params, and Speed"""
# backbone
macs1, params1 = profile(model,
inputs=(img_x, att_x, None, None, None, None,
"backbone", "search"),
custom_ops=None,
verbose=False)
macs, params = clever_format([macs1, params1], "%.3f")
print('backbone (search) macs is ', macs)
print('backbone params is ', params)
# transformer and head
macs2, params2 = profile(model,
inputs=(None, None, q, k, v, key_padding_mask,
"transformer", "search"),
custom_ops=None,
verbose=False)
macs, params = clever_format([macs2, params2], "%.3f")
print('transformer and head macs is ', macs)
print('transformer and head params is ', params)
# the whole model
macs, params = clever_format([macs1 + macs2, params1 + params2], "%.3f")
print('overall macs is ', macs)
print('overall params is ', params)
def main():
cfg.merge_from_file('./models/config/config.yaml')
model = ModelBuilder()
x = torch.randn(1, 3, 255, 255)
zf = torch.randn(1, 3, 127, 127)
model.template(zf)
macs, params = profile(model, inputs=(x,), verbose = False)
macs, params = clever_format([macs, params], "%.3f")
print('overall macs is ', macs)
print('overall params is ', params)
'reg': torch.randn(1, 128, 16, 16)
}
oup = model(x, zf)
# custom_ops = {
# Conv2dDynamicSamePadding: count_convNd,
# Conv2dStaticSamePadding: count_convNd,
# MemoryEfficientSwish: zero_ops,
# }
# compute FLOPs and Params
# the whole model
macs, params = profile(model,
inputs=(x, zf),
custom_ops=None,
verbose=False)
macs, params = clever_format([macs, params], "%.3f")
print('overall macs is ', macs)
print('overall params is ', params)
# backbone
macs, params = profile(backbone,
inputs=(x, ),
custom_ops=None,
verbose=False)
macs, params = clever_format([macs, params], "%.3f")
print('backbone macs is ', macs)
print('backbone params is ', params)
# head
macs, params = profile(head, inputs=(inp, ), verbose=False)
macs, params = clever_format([macs, params], "%.3f")
print('head macs is ', macs)
print('head params is ', params)
def test_clever_format_returns_formatted_number(self):
nums = 1
format = "%.2f"
clever_nums = utils.clever_format(nums, format)
assert clever_nums == '1.00B'
def test_clever_format_returns_formatted_numbers(self):
nums = [1, 2]
format = "%.2f"
clever_nums = utils.clever_format(nums, format)
assert clever_nums == ('1.00B', '2.00B')