feature map

[limbo-zhj]

Hello, how can I output the feature map or hot map of each layer

[glenn-jocher]

@limbo-zhj this has been asked before, but typically users don't understand the number of feature maps they would be dealing with. The YOLOv5l model for example has over a thousand output feature maps, each with it's own 80x80, 40x40 or 20x20 gird that you could look at. We don't have a standard method for viewing these, but they are easily extractable in the Detect() layer of any YOLOv5 model:

yolov5/models/yolo.py

Lines 24 to 58 in 5f7d39f

	class Detect(nn.Module):
	stride = None # strides computed during build
	export = False # onnx export
	def __init__(self, nc=80, anchors=(), ch=()): # detection layer
	super(Detect, self).__init__()
	self.nc = nc # number of classes
	self.no = nc + 5 # number of outputs per anchor
	self.nl = len(anchors) # number of detection layers
	self.na = len(anchors[0]) // 2 # number of anchors
	self.grid = [torch.zeros(1)] * self.nl # init grid
	a = torch.tensor(anchors).float().view(self.nl, -1, 2)
	self.register_buffer('anchors', a) # shape(nl,na,2)
	self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2)
	self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv
	def forward(self, x):
	# x = x.copy() # for profiling
	z = [] # inference output
	self.training |= self.export
	for i in range(self.nl):
	x[i] = self.m[i](x[i]) # conv
	bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
	x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
	if not self.training: # inference
	if self.grid[i].shape[2:4] != x[i].shape[2:4]:
	self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
	y = x[i].sigmoid()
	y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
	y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
	z.append(y.view(bs, -1, self.no))
	return x if self.training else (torch.cat(z, 1), x)

[kimsung-git]

@glenn-jocher
I'm also in attempt to extract feature map from the Detect as you mentioned. After printing out the input x to forward(), I verified that the shape of x is different from what you mentioned above. The the dimension I got for x is the following:

torch.Size([1, 256, 48, 80]), torch.Size([1, 512, 24, 40]), torch.Size([1, 1024, 12, 20])

I understand that the shape changes to [1, 255, 48, 80] [1, 255, 24, 40] [1, 255, 12, 20] after conv in foward(), but how come the input x dimension is not 80x80, 40x40, 20x20, but 48x80, 24x40, 12x20?

Can you explain why it is so?

Thank you

[glenn-jocher]

@kimsung-git output shapes are a function of input shapes.

[kimsung-git]

@glenn-jocher when you said "output shape", you meant x in Detect's forward method and "input shape" an image size?

Isn't input shapes(img size) always fixed before feeding to the network i.e 640? I also tested image with different size but it's the same shapes: 48x80, 24x40, 12x20..

Thank you

[glenn-jocher]

@kimsung-git detect.py displays input shapes directly, you might want to start there:

python detect.py