預備知識，mxnet.ndarray的一些操作
shape_array
shape_array([[1,2,3,4], [5,6,7,8]]) = [2,4] # 獲取ndarry的shape
# Returns a 1D int64 array containing the shape of data.
split
Splits an array along a particular axis into multiple sub-arrays.
cc = [ 1 80 128 128]
N, C, H, W = cc.split(num_outputs=4, axis=0)
# N = [1], C = [80], H = [128], W = [128]
topk
topk(data=None, axis=_Null, k=_Null, ret_typ=_Null, is_ascend=_Null, dtype=_Null, out=None, name=None, **kwargs)
# data : The input array
# axis : 從哪個維度選取topk，默認-1，即最後一個維度
# ret_typ : return type，可以選擇返回值，或者index，或者both
# is_ascend : 是否逆序，默認爲0，返回 top k largest
cast
Casts all elements of the input to a new type.
# 將array中元素轉換爲指定的類型
cast([0.9, 1.3], dtype='int32') = [0, 1]
slice_like
# Slices a region of the array like the shape of another array.
x = [[ 1., 2., 3., 4.],
[ 5., 6., 7., 8.],
[ 9., 10., 11., 12.]]
y = [[ 0., 0., 0.],
[ 0., 0., 0.]]
slice_like(x, y) = [[ 1., 2., 3.]
[ 5., 6., 7.]]
slice_like(x, y, axes=(0, 1)) = [[ 1., 2., 3.]
[ 5., 6., 7.]]
slice_like(x, y, axes=(0)) = [[ 1., 2., 3., 4.]
[ 5., 6., 7., 8.]]
slice_like(x, y, axes=(-1)) = [[ 1., 2., 3.]
[ 5., 6., 7.]
[ 9., 10., 11.]]
expand_dims
# Inserts a new axis of size 1 into the array shape For example, given x with shape (2,3,4),
# then expand_dims(x, axis=1) will return a new array with shape (2,1,3,4).
mxnet.ndarray.expand_dims(data=None, axis=_Null, out=None, name=None, **kwargs)
# axis=-1, 則在末尾增加一個維度
tile
# 將array複製n次
x = [[1, 2],
[3, 4]]
tile(x, reps=(2,3)) = [[ 1., 2., 1., 2., 1., 2.],
[ 3., 4., 3., 4., 3., 4.],
[ 1., 2., 1., 2., 1., 2.],
[ 3., 4., 3., 4., 3., 4.]]
gather_nd
參考：https://discuss.gluon.ai/t/topic/2413/3
# 從array中，按照輸入的索引取出其中的元素
data = [[0, 1], [2, 3]]
indices = [[1, 1, 0], [0, 1, 0]]
gather_nd(data, indices) = [2, 3, 0] # 取出(1,0),(1,1),(0,0)的元素
centernet訓練與預測過程中，圖片的幾何變化
訓練階段：原圖 -> (512,512) -> (128,128)
預測階段：(128,128) -> (512,512) -> 原圖
在inference階段，模型的輸出是在128x128的feature shape上，在feature map上得到的結果，需要乘以scale=4.0，才能映射到512x512上，接着，還必須經過512x512到原圖的一個仿射變換，才能得到模型最終輸出值
核心操作及解碼
核心操作
# self.heatmap_nms = nn.MaxPool2D(pool_size=3, strides=1, padding=1)
heatmap = outs[0]
keep = F.broadcast_equal(self.heatmap_nms(heatmap), heatmap)
results = self.decoder(keep * heatmap, outs[1], outs[2])
outs爲模型輸出，在heatmap所在維度上做3x3的max pooling，把峯值點找出來
解碼
def hybrid_forward(self, F, x, wh, reg):
"""Forward of decoder"""
import pdb
pdb.set_trace()
# 這裏假設batch_size = 4, resize w和h爲512x512
_, _, out_h, out_w = x.shape_array().split(num_outputs=4, axis=0) # 獲取feature map的H，W
scores, indices = x.reshape((0, -1)).topk(k=self._topk, ret_typ='both') # 獲取top_100的scores和indices，x nx80x128x128 -> nx(80x128x128), 某類的128x128元素連在一起
indices = F.cast(indices, 'int64')
topk_classes = F.cast(
F.broadcast_div(
indices,
(out_h * out_w)),
'float32') # 根據indices，獲取top_100對應的類別，0-79
topk_indices = F.broadcast_mod(indices, (out_h * out_w)) # 獲取在128x128平面拉平後的索引
topk_ys = F.broadcast_div(topk_indices, out_w) # 1維索引恢復出2維平面上的y
topk_xs = F.broadcast_mod(topk_indices, out_w) # 1維索引恢復出2維平面上的x
center = reg.transpose((0, 2, 3, 1)).reshape((0, -1, 2)) # shape: (4, 16384, 2)
wh = wh.transpose((0, 2, 3, 1)).reshape((0, -1, 2)) # shape: (4, 16384, 2)
batch_indices = F.cast(F.arange(256).slice_like(center, axes=(
0)).expand_dims(-1).tile(reps=(1, self._topk)), 'int64') # shape: (4, 100), 依次爲全0，全1，全2，全3
reg_xs_indices = F.zeros_like(batch_indices, dtype='int64') # shape: (4, 100)，全0
reg_ys_indices = F.ones_like(batch_indices, dtype='int64') # shape: (4, 100)，全1
reg_xs = F.concat(
batch_indices, topk_indices, reg_xs_indices, dim=0).reshape(
(3, -1)) # shape： (3, 400)
‘’‘reg_xs =
[[ 0 0 0 ... 3 3 3]
[9664 8207 9425 ... 9639 5593 9044]
[ 0 0 0 ... 0 0 0]]
’‘’
reg_ys = F.concat(
batch_indices, topk_indices, reg_ys_indices, dim=0).reshape(
(3, -1))
‘’‘reg_ys =
[[ 0 0 0 ... 3 3 3]
[9664 8207 9425 ... 9639 5593 9044]
[ 1 1 1 ... 1 1 1]]
’‘’
xs = F.cast(
F.gather_nd(
center, reg_xs).reshape(
(-1, self._topk)), 'float32') # shape : (4, 100)
ys = F.cast(
F.gather_nd(
center, reg_ys).reshape(
(-1, self._topk)), 'float32') # shape : (4, 100)
topk_xs = F.cast(topk_xs, 'float32') + xs
topk_ys = F.cast(topk_ys, 'float32') + ys # feature map上座標+偏移 = 真實座標
w = F.cast(F.gather_nd(wh, reg_xs).reshape((-1, self._topk)), 'float32') # noqa
h = F.cast(F.gather_nd(wh, reg_ys).reshape((-1, self._topk)), 'float32') # noqa
half_w = w / 2
half_h = h / 2
results = [
topk_xs - half_w,
topk_ys - half_h,
topk_xs + half_w,
topk_ys + half_h]
results = F.concat(*[tmp.expand_dims(-1) for tmp in results], dim=-1) # shape: (4, 100, 4)
return topk_classes, scores, results * self._scale # (4, 100),(4, 100),(4, 100, 4)
由512x512映射到原圖
affine_mat = get_post_transform(orig_width, orig_height, 512, 512) # 獲取仿射變換矩陣
bbox[0:2] = affine_transform(bbox[0:2], affine_mat) # 對x1,y1做仿射變換
bbox[2:4] = affine_transform(bbox[2:4], affine_mat) # 對x2,y2做仿射變換