def forward(self, x):
x = self.backbone(x)
dis_gen2cen, dis_gen2ori, thresholds, amplified_thresholds, embed_gen = None, None, None, None, None
gap = (F.adaptive_avg_pool2d(x, 1)).view(x.size(0), -1)
if hasattr(self,'decorrelation'):
gap = self.decorrelation(gap)
embed_fea = self.embeddingLayer(gap) if hasattr(self, 'embeddingLayer') else gap
centroids = F.normalize(self.centroids, dim=1, p=2) if self.norm_centroid else self.centroids
SIMI = Similarity(scaled=self.scaled)
sim_fea2cen = getattr(SIMI, self.similarity)(embed_fea, centroids)
DIST = Distance(scaled=self.scaled)
dis_fea2cen = getattr(DIST, self.distance)(embed_fea, centroids)
dis_cen2cen = getattr(DIST, self.distance)(centroids, centroids)
dis_thr2thr = self.cal_thr2thr()
# if hasattr(self, 'estimator'):
# dis_gen2cen = getattr(DIST, self.distance)(embed_gen, centroids)
# dis_gen2ori = getattr(DIST, self.distance)(embed_gen, self.origin)
return {
"gap": gap,
"embed_fea": embed_fea,
"embed_gen": embed_gen,
"sim_fea2cen": sim_fea2cen,
"dis_fea2cen": dis_fea2cen,
"dis_cen2cen": dis_cen2cen,
"dis_thr2thr": dis_thr2thr,
"thresholds": self.thresholds
}
def forward(self, x):
x = self.backbone(x)
gap = (F.adaptive_avg_pool2d(x, 1)).view(x.size(0),
-1) # [n, backbone_c]
embed_fea = gap
# embed_fea = self.embeddingLayer(gap) # [n, embed_dim]
norm_fea = torch.norm(embed_fea, dim=1, p=2,
keepdim=True) # norm length for each image [n,1]
embed_fea_normed = F.normalize(embed_fea, dim=1, p=2) # [n, embed_dim]
centroids = self.centroids # [class, embed_dim]
centroids_normed = F.normalize(centroids, dim=1,
p=2) # [class, embed_dim]
SIMI = Similarity()
# dotproduct: X*W = ||X|| * ||W|| * cos(X,W) # [n,class]
dotproduct_fea2cen = getattr(SIMI, "dotproduct")(embed_fea, centroids)
# cosine: cos(X,W) # [n,class]
# cosine_fea2cen = getattr(SIMI, "dotproduct")(embed_fea_normed, centroids_normed)
# normweight: ||X|| * cos(X,W) # [n,class]
normweight_fea2cen = getattr(SIMI, "dotproduct")(embed_fea,
centroids_normed)
energy = torch.logsumexp(normweight_fea2cen, dim=1,
keepdim=False) # [n]
return {
# "gap": gap, # [n,self.feat_dim] gap extracted by backbone
# "embed_fea": embed_fea, # [n,embed_dim] embedded features
"norm_fea":
norm_fea, # [n,1] the norm value of one images features, keepdim for post-processing
# "dotproduct_fea2cen": dotproduct_fea2cen, # [n,num_classes]
# "cosine_fea2cen": cosine_fea2cen, # [n,num_classes]
"normweight_fea2cen": normweight_fea2cen,
"energy": energy
}
def forward(self, x):
x = self.backbone(x)
dis_gen2cen, dis_gen2ori, thresholds, amplified_thresholds, embed_gen = None, None, None, None, None
gap = (F.adaptive_avg_pool2d(x, 1)).view(x.size(0), -1)
if hasattr(self, 'thresholds'):
thresholds = self.thresholds
gen = whitennoise_generator(self.estimator, gap)
embed_gen = self.embeddingLayer(gen) if hasattr(self, 'embeddingLayer') else gen
amplified_thresholds = self.thresholds * self.amplifier
embed_fea = self.embeddingLayer(gap) if hasattr(self, 'embeddingLayer') else gap
centroids = F.normalize(self.centroids, dim=1, p=2) if self.norm_centroid else self.centroids
SIMI = Similarity(scaled=self.scaled)
sim_fea2cen = getattr(SIMI, self.similarity)(embed_fea, centroids)
DIST = Distance(scaled=self.scaled)
dis_fea2cen = getattr(DIST, self.distance)(embed_fea, centroids)
if hasattr(self, 'thresholds'):
dis_gen2cen = getattr(DIST, self.distance)(embed_gen, centroids)
dis_gen2ori = getattr(DIST, self.distance)(embed_gen, self.origin)
return {
"gap": gap,
"embed_fea": embed_fea,
"embed_gen": embed_gen,
"sim_fea2cen": sim_fea2cen,
"dis_fea2cen": dis_fea2cen,
"dis_gen2cen": dis_gen2cen,
"dis_gen2ori": dis_gen2ori,
"amplified_thresholds": amplified_thresholds,
"thresholds": thresholds
}
def forward(self, input):
x = self.backbone(input)
gap = (F.adaptive_avg_pool2d(x, 1)).view(x.size(0), -1)
embed_fea = self.embeddingLayer(gap) if hasattr(
self, 'embeddingLayer') else gap
gen = self.generator(input)
embed_gen = self.embeddingLayer(gen) if hasattr(
self, 'embeddingLayer') else gen
centroids = F.normalize(self.centroids, dim=1,
p=2) if self.norm_centroid else self.centroids
SIMI = Similarity(scaled=self.scaled)
sim_fea2cen = getattr(SIMI, self.similarity)(embed_fea, centroids)
DIST = Distance(scaled=self.scaled)
dis_fea2cen = getattr(DIST, self.distance)(embed_fea, centroids)
dis_gen2cen = getattr(DIST, self.distance)(embed_gen, centroids)
dis_gen2ori = getattr(DIST, self.distance)(embed_gen, self.origin)
thresholds = None
if hasattr(self, 'thresholds'):
thresholds = self.thresholds
return {
"gap": gap,
"embed_fea": embed_fea,
"embed_gen": embed_gen,
"sim_fea2cen": sim_fea2cen,
"dis_fea2cen": dis_fea2cen,
"dis_gen2cen": dis_gen2cen,
"dis_gen2ori": dis_gen2ori,
"thresholds": thresholds
}
def forward(self, x):
x = self.backbone(x)
gap = (F.adaptive_avg_pool2d(x, 1)).view(x.size(0), -1)
embed_fea = self.embeddingLayer(gap)
norm_fea = torch.norm(embed_fea, dim=1, p=2, keepdim=True)
embed_fea_normed = F.normalize(embed_fea, dim=1, p=2)
centroids = self.centroids
centroids_normed = F.normalize(centroids, dim=1, p=2)
SIMI = Similarity()
# dotproduct: X*W = ||X|| * ||W|| * cos(X,W) ignoring transposition operation
dotproduct_fea2cen = getattr(SIMI, "dotproduct")(embed_fea, centroids)
# cosine: cos(X,W) ignoring transposition operation
cosine_fea2cen = getattr(SIMI, "dotproduct")(embed_fea_normed,
centroids_normed)
# normweight: ||X|| * cos(X,W)
normweight_fea2cen = getattr(SIMI, "dotproduct")(embed_fea,
centroids_normed)
return {
"gap": gap, # [n,self.feat_dim] gap extracted by backbone
"embed_fea": embed_fea, # [n,embed_dim] embedded features
"norm_fea":
norm_fea, # [n,1] the norm value of one images features, keepdim for post-processing
"dotproduct_fea2cen": dotproduct_fea2cen, # [n,num_classes]
"cosine_fea2cen": cosine_fea2cen, # [n,num_classes]
"normweight_fea2cen": normweight_fea2cen
}
def forward(self, x):
x = self.backbone(x)
dis_gen2cen, dis_gen2ori, thresholds, amplified_thresholds, embed_gen = None, None, None, None, None
gap = x
if hasattr(self, 'estimator'):
thresholds = self.thresholds
gen = self.estimator.sampler(gap)
embed_gen = self.embeddingLayer(gen) if hasattr(self, 'embeddingLayer') else gen
embed_fea = self.embeddingLayer(gap) if hasattr(self, 'embeddingLayer') else gap
embed_fea_2d = (self.fuse*embed_fea).sum(dim=2, keepdim=False).squeeze(dim=-1)
centroids = F.normalize(self.centroids, dim=1, p=2) if self.norm_centroid else self.centroids
SIMI = Similarity(scaled=self.scaled)
sim_fea2cen = getattr(SIMI, self.similarity)(embed_fea_2d, centroids)
DIST = Distance(scaled=self.scaled)
dis_fea2cen = getattr(DIST, self.distance)(embed_fea_2d, centroids)
if hasattr(self, 'estimator'):
dis_gen2cen = getattr(DIST, self.distance)(embed_gen, centroids)
dis_gen2ori = getattr(DIST, self.distance)(embed_gen, self.origin)
return {
"gap": gap,
"embed_fea": embed_fea,
"embed_gen": embed_gen,
"sim_fea2cen": sim_fea2cen,
"dis_fea2cen": dis_fea2cen,
"dis_gen2cen": dis_gen2cen,
"dis_gen2ori": dis_gen2ori,
"thresholds": thresholds
}
def forward(self, x):
x = self.backbone(x)
gap = (F.adaptive_avg_pool2d(x, 1)).view(x.size(0), -1)
embed_fea = self.embeddingLayer(gap) if hasattr(
self, 'embeddingLayer') else gap
SIMI = Similarity(scaled=self.scaled)
centroids = F.normalize(self.centroids, dim=1,
p=2) if self.norm_centroid else self.centroids
sim_fea2cen = getattr(SIMI, self.distance)(embed_fea, centroids)
return {"gap": x, "embed_fea": embed_fea, "sim_fea2cen": sim_fea2cen}
def forward(self, x):
x = self.backbone(x)
gap = (F.adaptive_avg_pool2d(x, 1)).view(x.size(0), -1)
embed_fea = self.embeddingLayer(gap)
embed_fea_norm = F.normalize(embed_fea, dim=1, p=2)
centroids = self.centroids
centroids_norm = F.normalize(centroids, dim=1, p=2)
SIMI = Similarity()
dotproduct_fea2cen = getattr(SIMI, "dotproduct")(embed_fea, centroids)
# cosine distance equals l2-normalized dotproduct distance. For efficient computation.
cosine_fea2cen = getattr(SIMI, "dotproduct")(embed_fea_norm, centroids_norm)
# # # re-range to distance defination, [0,1], smaller indicates more similar.
# cosine_fea2cen = (1.0-cosine_fea2cen)/2.0
return {
"gap": gap, # [n,self.feat_dim]
"embed_fea": embed_fea, # [n,embed_dim]
"dotproduct_fea2cen": dotproduct_fea2cen, # [n,num_classes]
"cosine_fea2cen": cosine_fea2cen # [n,num_classes]
}
def forward(self, x):
x = self.backbone(x)
gap = (F.adaptive_avg_pool2d(x, 1)).view(x.size(0), -1)
embed_fea = self.embeddingLayer(gap)
embed_fea_norm = F.normalize(embed_fea, dim=1, p=2)
centroids = self.centroids
centroids_norm = F.normalize(centroids, dim=1, p=2)
SIMI = Similarity()
dotproduct_fea2cen = getattr(SIMI, "dotproduct")(embed_fea, centroids)
cosine_fea2cen = getattr(SIMI, "dotproduct")(embed_fea_norm,
centroids_norm)
normweight_fea2cen = getattr(SIMI, "dotproduct")(embed_fea,
centroids_norm)
return {
"gap": gap, # [n,self.feat_dim]
"embed_fea": embed_fea, # [n,embed_dim]
"dotproduct_fea2cen": dotproduct_fea2cen, # [n,num_classes]
"cosine_fea2cen": cosine_fea2cen, # [n,num_classes]
"normweight_fea2cen": normweight_fea2cen
}
# sim_dis = SimDistance(dim_data)
# sim_cor = SimCorrelation(dim_data)
# sim_mu = SimMutual(dim_data)
def calc_and_output(sim):
# labels_predict = k_means(sim, 6)
model = AgglomerativeClustering(n_clusters=8, affinity='euclidean')
labels_predict = model.fit(sim).labels_
# labels_predict = knn_model.predict(np.max(sim) - sim)
print('ARI:', ARI(labels, labels_predict))
return labels_predict, ARI(labels, labels_predict)
sim_data = Similarity(dim_data, alpha=0.6, beta=0.3)
pred, ari = calc_and_output(sim_data)
rel = RelevanceMatrix(pred)
print(len(rel[rel > 1e-10]))
sim_next = sim_data
l0 = len(sim_next[sim_next > 1e-10])
while l0 != 0:
print('l0:', l0)
# print(sim_next, rel)
sim_next = 0.01 * sim_next + 0.99 * rel
pred, ari = calc_and_output(sim_next)
rel = get_normalize(RelevanceMatrix(pred))
l0 = len(sim_next[sim_next > 1e-5])
# calc_and_output(sim_dis)
# calc_and_output(sim_cor)