pytorch和tensorflow实现one-hot编码,label编码为target

在我们实现网络训练或者自定义loss函数式,要计算预测图score和GT图target之间的loss时，score和target的shape需要保持一致，而标签数据label通常只是索引信息，我们需要将label数据通过onehot编码转化为向量(矩阵)target,在target中对应label索引的位置为1.0,其余为0.0，通常label类型为int，转化后的target为float

首先定义batch_szie，num_classes,H,W

bs=16
num_classes=5
H=128
W=128

1.pytorch实现¶

1. 2D Tensor：

   • label：shape:[bs] dtype:torch.int64===>target: shape:[bs,num_classes] dtype:torch.float32
   • score: shape:[bs,num_classes] dtype:torch.float32

2. KD Tensor:

   • label：shape:[bs,H,W] dtype:torch.int64===>target: shape:[bs,num_classes,H,W] dtype:torch.float32
   • score: shape:[bs,num_classes,H,W] dtype:torch.float32

scatter_(dim, index, src/value) → Tensor

根据index(LongTensor)中的索引值将value(float)或src(Tensor)中的所有值写入self(Tensor)中，写入位置在dim维度的索引由index确定

• 对于2D tensor,slef更新规则:

  • dim=0时self[index[i][j]][j]=src[i][j]
  • dim=1时self[i][index[i][j]]=src[i][j]

• 对于4D tensor,self更新规则:

  • dim=0时self[index[i][j][h][w]][j][h][w]=src[i][j][h][w]
  • dim=1时self[i][index[i][j][h][w]][h][w]=src[i][j][h][w]
  • dim=2时self[i][j][index[i][j][h][w]][w]=src[i][j][h][w]
  • dim=3时self[i][j][h][index[i][j][h][w]]=src[i][j][h][w]

可见self,index和src必须有相同的维度

参数：

• dim(int):索引的轴
• index(Long Tensor):index为空时，返回原self
• src(Tensor)/value(float)

例子：

1. 2D Tensor¶

1. 定义1D label和2D score

score=torch.rand(bs,num_classes)
label=torch.randint(low=0,high=num_classes,size=(bs,))
print("score shape:{},score dtype:{}".format(score.size(),score.dtype))
print("label shape:{},label dtype:{}".format(label.size(),label.dtype))

 输出：

score shape:torch.Size([16, 5]),score dtype:torch.float32
label shape:torch.Size([16]),label dtype:torch.int64

2. 将1D label转化为one-hot矩阵2D target¶

1. 将label在num_classes对应的维度扩展到2D
2. one-hot编码

label=torch.unsqueeze(label,dim=1)
print("label shape:{},label dtype:{}".format(label.size(),label.dtype))
target=torch.zeros_like(score).scatter_(1, label, torch.ones_like(label,dtype=torch.float32))
print("target shape:{},target dtype:{}".format(target.size(),target.dtype))

输出：

label shape:torch.Size([16, 1]),label dtype:torch.int64
target shape:torch.Size([16, 5]),target dtype:torch.float32

3.测试：

print(label[0:5,0])
print(target[0:5,:])

输出：

tensor([4, 3, 4, 0, 2])
tensor([[0., 0., 0., 0., 1.],
        [0., 0., 0., 1., 0.],
        [0., 0., 0., 0., 1.],
        [1., 0., 0., 0., 0.],
        [0., 0., 1., 0., 0.]])

2. 4D Tensor¶

1. 定义3D label和4D score

score=torch.rand(bs,num_classes,H,W)
label=torch.randint(low=0,high=num_classes,size=(bs,H,W))
print("score shape:{},score dtype:{}".format(score.size(),score.dtype))
print("label shape:{},label dtype:{}".format(label.size(),label.dtype))

输出：

score shape:torch.Size([16, 5, 128, 128]),score dtype:torch.float32
label shape:torch.Size([16, 128, 128]),label dtype:torch.int64

2. 将3D label转化为one-hot矩阵4D target¶

1. 将label在num_classes对应的维度扩展到4D
2. one-hot编码

label=torch.unsqueeze(label,dim=1)
print("label shape:{},label dtype:{}".format(label.size(),label.dtype))
target=torch.zeros(bs,num_classes,H,W).scatter_(1, label, torch.ones_like(label,dtype=torch.float32))
print("target shape:{},target dtype:{}".format(target.size(),target.dtype))

输出：

label shape:torch.Size([16, 1, 128, 128]),label dtype:torch.int64
target shape:torch.Size([16, 5, 128, 128]),target dtype:torch.float32

3.测试

print(label[2:6,0,66,60])
print(target[2:6,:,66,60])

输出：

tensor([4, 1, 2, 3])
tensor([[0., 0., 0., 0., 1.],
        [0., 1., 0., 0., 0.],
        [0., 0., 1., 0., 0.],
        [0., 0., 0., 1., 0.]])

2.tensorflow实现

1. 2D Tensor：

   • label：shape:[bs] dtype:tf.int32===>target: shape:[bs,num_classes] dtype:tf.float32
   • score: shape:[bs,num_classes] dtype:tf.float32

2. KD Tensor:

   • label：shape:[bs,H,W] dtype:tf.int32===>target: shape:[bs,H,W,num_classes] dtype:tf.float32
   • score: shape:[bs,H,W,num_classes] dtype:tf.float32

tf.one_hot(indices,depth,on_value=None,off_value=None,axis=None,dtype=None,name=None)

• indices:若输入的indices维度为n,则输出的Tensor维度为n+1，扩展的维度位置由axis指定(默认新增维度在最后添加)
  • 若indices是标量,输出的shape为[depth,]
  • 若indices是向量shape:[bs,],输出的shape为[bs,depth](axis=-1)或者[depth,bs](axis=0)
  • ....

参数：

• indices:
• depth:一个标量定义one-hot维度的长度
• on_value:定义在 indices[j] = i 时填充输出的值的标量，默认为1
• off_value:定义在 indices[j] != i 时填充输出的值的标量，默认为0
• axis:要填充的轴，默认为-1，即一个新的最内层轴
• dtype:输出Tensor的数据类型
• name:

例子

1. 2D Tensor¶

1. 定义1D label和2D score

score=tf.random.uniform((bs,num_classes),0,1)
label_numpy=np.random.randint(low=0,high=num_classes,size=(bs,))
label=tf.convert_to_tensor(label_numpy)
print("score shape:{},score dtype:{}".format(score.shape,score.dtype))
print("label shape:{},label dtype:{}".format(label.shape,label.dtype))

输出：

score shape:(16, 5),score dtype:
label shape:(16,),label dtype:

2.将1D label转化为one-hot矩阵2D target

target=tf.one_hot(indices=label,depth=num_classes,on_value=1.0,off_value=0.0,axis=1,dtype=tf.float32)
print("target shape:{},target dtype:{}".format(target.shape,target.dtype))

输出：

target shape:(16, 5),target dtype:

3.测试

print(label[1:5,])
print(target[1:5,:])

输出：

tf.Tensor([3 2 4 4], shape=(4,), dtype=int32)
tf.Tensor(
[[0. 0. 0. 1. 0.]
 [0. 0. 1. 0. 0.]
 [0. 0. 0. 0. 1.]
 [0. 0. 0. 0. 1.]], shape=(4, 5), dtype=float32)

2. 4D Tensor¶

1. 定义3D label和4D score

score=tf.random.uniform((bs,H,W,num_classes),0,1)
label_numpy=np.random.randint(low=0,high=num_classes,size=(bs,H,W))
label=tf.convert_to_tensor(label_numpy)
print("score shape:{},score dtype:{}".format(score.shape,score.dtype))
print("label shape:{},label dtype:{}".format(label.shape,label.dtype))

输出：

score shape:(16, 128, 128, 5),score dtype:
label shape:(16, 128, 128),label dtype:

2.将3D label转化为one-hot矩阵4D target

target=tf.one_hot(indices=label,depth=num_classes,on_value=1.0,off_value=0.,axis=3,dtype=tf.float32)
print("target shape:{},target dtype:{}".format(target.shape,target.dtype))

输出：

target shape:(16, 128, 128, 5),target dtype:

3. 测试

print(label[2:6,66,66])
print(target[2:6,66,66,:])

输出：

tf.Tensor([0 4 4 3], shape=(4,), dtype=int32)
tf.Tensor(
[[1. 0. 0. 0. 0.]
 [0. 0. 0. 0. 1.]
 [0. 0. 0. 0. 1.]
 [0. 0. 0. 1. 0.]], shape=(4, 5), dtype=float32)