mirrors
/
PaddleClas
mirror of https://github.com/PaddlePaddle/PaddleClas.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Update deephashloss.py

pull/1390/head
Bin Lu 2021-11-06 23:32:19 +08:00 committed by GitHub
parent f7ee2b98a6
commit 3b4a45f8ab
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 19 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
ppcls/loss/deephashloss.py
View File

@ -23,38 +23,42 @@ class DSHSDLoss(nn.Layer):
# [DSHSD] epoch:250, bit:48, dataset:nuswide_21, MAP:0.809, Best MAP: 0.815 # [DSHSD] epoch:250, bit:48, dataset:nuswide_21, MAP:0.809, Best MAP: 0.815
# [DSHSD] epoch:135, bit:48, dataset:imagenet, MAP:0.647, Best MAP: 0.647 # [DSHSD] epoch:135, bit:48, dataset:imagenet, MAP:0.647, Best MAP: 0.647
""" """
def __init__(self, n_class, bit, alpha, multi_label=False): def __init__(self, alpha, multi_label=False):
super(DSHSDLoss, self).__init__() super(DSHSDLoss, self).__init__()
self.alpha = alpha self.alpha = alpha
self.multi_label = multi_label self.multi_label = multi_label
def forward(self, input, label): def forward(self, input, label):
feature = input["features"] feature = input["features"]
logits = input["logits"] logits = input["logits"]
dist = paddle.sum( dist = paddle.sum(paddle.square(
paddle.square((paddle.unsqueeze(feature, 1) - paddle.unsqueeze(feature, 0))), (paddle.unsqueeze(feature, 1) - paddle.unsqueeze(feature, 0))),
axis=2) axis=2)
# label to ont-hot # label to ont-hot
label = paddle.flatten(label) label = paddle.flatten(label)
n_class = logits.shape[1] n_class = logits.shape[1]
label = paddle.nn.functional.one_hot(label, n_class).astype("float32") label = paddle.nn.functional.one_hot(label, n_class).astype("float32")
s = (paddle.matmul(label, label, transpose_y=True) == 0).astype("float32") s = (paddle.matmul(
label, label, transpose_y=True) == 0).astype("float32")
margin = 2 * feature.shape[1] margin = 2 * feature.shape[1]
Ld = (1 - s) / 2 * dist + s / 2 * (margin - dist).clip(min=0) Ld = (1 - s) / 2 * dist + s / 2 * (margin - dist).clip(min=0)
Ld = Ld.mean() Ld = Ld.mean()
if self.multi_label: if self.multi_label:
# multiple labels classification loss # multiple labels classification loss
Lc = (logits - label * logits + ((1 + (-logits).exp()).log())).sum(axis=1).mean() Lc = (logits - label * logits + (
(1 + (-logits).exp()).log())).sum(axis=1).mean()
else: else:
# single labels classification loss # single labels classification loss
Lc = (-paddle.nn.functional.softmax(logits).log() * label).sum(axis=1).mean() Lc = (-paddle.nn.functional.softmax(logits).log() * label).sum(
axis=1).mean()
return {"dshsdloss": Lc + Ld * self.alpha} return {"dshsdloss": Lc + Ld * self.alpha}
class LCDSHLoss(nn.Layer): class LCDSHLoss(nn.Layer):
""" """
# paper [Locality-Constrained Deep Supervised Hashing for Image Retrieval](https://www.ijcai.org/Proceedings/2017/0499.pdf) # paper [Locality-Constrained Deep Supervised Hashing for Image Retrieval](https://www.ijcai.org/Proceedings/2017/0499.pdf)