Inference time of each layer

[ghost]

Hello I'm trying to profile inference time of each layer of yolov5.
I have added the code below to yolo.py between Create model and Options:

......
# Create model
im = torch.rand(opt.batch_size, 3, 640, 640).to(device)
model = Model(opt.cfg).to(device)

x = torch.randn(1, 3, 640, 640).cuda()

with torch.autograd.profiler.profile(enabled=True, use_cuda=True, record_shapes=False,
                                     profile_memory=False) as prof:
    output = model(x)
print(prof.table())
prof.export_chrome_trace('./model.json')

summary(model, x)

# Options
if opt.line_profile:   # profile layer by layer
    _ = model(im, profile=True)
......

the result is as follow:

We can clearly see that the first convolutional layer takes more than 90% of the inference time, but when I search the reason from source code, I haven't find any difference between the first conv layer and the other conv layers.

From my experiment，other models like mask-rcnn，mobileNet，vgg，the inference time of their conv layer is not much different.

Question

So I want to know the reason of this phenomenon. Appreciate for any help!

[glenn-jocher]

@769430755 your profiler is incorrect.

We already have built-in line profiling of YOLOv5 models, they display even resource utilization throughout the model generally speaking.

python models/yolo.py --cfg yolov5s.yaml --line-profile