Update deephashloss.py

ppcls/loss/deephashloss.py

@@ -57,3 +57,35 @@ class DSHSDLoss(nn.Layer):
 
         return {"dshsdloss": Lc + Ld * self.alpha}
 
+class LCDSHLoss(paddle.nn.Layer):
+    """
+    # paper [Locality-Constrained Deep Supervised Hashing for Image Retrieval](https://www.ijcai.org/Proceedings/2017/0499.pdf)
+    # [LCDSH] epoch:145, bit:48, dataset:cifar10-1,  MAP:0.798, Best MAP: 0.798
+    # [LCDSH] epoch:183, bit:48, dataset:nuswide_21, MAP:0.833, Best MAP: 0.834
+    """
+    def __init__(self, n_class, _lambda):
+        super(LCDSHLoss, self).__init__()
+        self._lambda = _lambda
+        self.n_class = n_class
+
+    def forward(self, input, label):
+        feature = input["features"]
+        label = label.astype("float32")
+
+        # label to ont-hot
+        label = paddle.flatten(label)
+        label = paddle.nn.functional.one_hot(label,  self.n_class).astype("float32")
+        
+        s = 2 * (paddle.matmul(label, label, transpose_y=True) > 0).astype("float32") - 1
+        inner_product = paddle.matmul(feature, feature, transpose_y=True) * 0.5
+
+        inner_product = inner_product.clip(min=-50, max=50)
+        L1 = paddle.log(1 + paddle.exp(-s * inner_product)).mean()
+
+        b = feature.sign()
+        inner_product_ = paddle.matmul(b, b, transpose_y=True) * 0.5
+        sigmoid = paddle.nn.Sigmoid()
+        L2 = (sigmoid(inner_product) - sigmoid(inner_product_)).pow(2).mean()
+
+        return {"lcdshloss": L1 + self._lambda * L2}
+