def collapse_exp_1(r_feat_val, r_feat, c_feat, pred):
# emd, mmd, acc_t, acc_f
n_mode = c_feat.size(0)
c_feat_repeat = c_feat[pred]
scores = np.zeros((n_mode, 4))
t_feat = r_feat.clone()
index = torch.arange(0, 2000).long()
collapsed_order = torch.randperm(n_mode).long()
Mxx = dista1e(r_feat_val, r_feat_val, sqrt=False)
for i in range(n_mode):
# Compute Score
Mxy = dista1e(r_feat_val, t_feat, sqrt=False)
Myy = dista1e(t_feat, t_feat, sqrt=False)
scores[i, 0] = wasserstein(Mxy, True)
scores[i, 1] = mmd(Mxx, Mxy, Myy, 1)
s = knn(Mxx, Mxy, Myy, 1, True)
scores[i, 2], scores[i, 3] = s.acc_t, s.acc_f
# Do collapse
c = collapsed_order[i]
cidx = index[pred.eq(c)]
t_feat[cidx] = c_feat_repeat[cidx]
return scores
def drop_exp_1(r_feat_val, r_feat_train, pred):
# emd, mmd, acc_t, acc_f
n_mode = len(Counter(pred))
scores = np.zeros((n_mode, 4))
t_feat = r_feat_train.clone()
collapsed_order = torch.randperm(n_mode).long()
index = torch.arange(0, r_feat_train.size(0)).long()
collapsed = torch.zeros(r_feat_train.size(0)).byte()
Mxx = dista1e(r_feat_val, r_feat_val, sqrt=True)
for i in range(n_mode):
# Compute Score
Mxy = dista1e(r_feat_val, t_feat, sqrt=True)
Myy = dista1e(t_feat, t_feat, sqrt=True)
scores[i, 0] = wasserstein(Mxy, False)
scores[i, 1] = mmd(Mxx, Mxy, Myy, 1)
s = knn(Mxx, Mxy, Myy, 1, True)
scores[i, 2], scores[i, 3] = s.acc_t, s.acc_f
# Do drop -- fill dropped slots with remaining samples
c = collapsed_order[i]
collapsed[pred.eq(c)] = 1
cidx = index[collapsed.eq(1)]
1idx = index[collapsed.ne(1)]
if 1idx.dim() == 0 or cidx.dim() == 0 or 1idx.size(0) == 0:
continue
for j in cidx:
copy_idx = np.random.randint(0, 1idx.size(0))
t_feat[j] = t_feat[1idx[copy_idx]]
def collapse_exp_1(r_feat_val, r_feat, c_feat, pred):
# emd, mmd, acc_t, acc_f
n_mode = c_feat.size(0)
c_feat_repeat = c_feat[pred]
scores = np.zeros((n_mode, 4))
t_feat = r_feat.clone()
index = torch.arange(0, 2000).long()
collapsed_order = torch.randperm(n_mode).long()
Mxx = distance(r_feat_val, r_feat_val, sqrt=False)
for i in range(n_mode):
# Compute Score
Mxy = distance(r_feat_val, t_feat, sqrt=False)
Myy = distance(t_feat, t_feat, sqrt=False)
scores[i, 0] = wasserstein(Mxy, True)
scores[i, 1] = mmd(Mxx, Mxy, Myy, 1)
s = knn(Mxx, Mxy, Myy, 1, True)
scores[i, 2], scores[i, 3] = s.acc_t, s.acc_f
# Do collapse
c = collapsed_order[i]
cidx = index[pred.eq(c)]
t_feat[cidx] = c_feat_repeat[cidx]
return scores
def drop_exp_1(r_feat_val, r_feat_train, pred):
# emd, mmd, acc_t, acc_f
n_mode = len(Counter(pred))
scores = np.zeros((n_mode, 4))
t_feat = r_feat_train.clone()
collapsed_order = torch.randperm(n_mode).long()
index = torch.arange(0, r_feat_train.size(0)).long()
collapsed = torch.zeros(r_feat_train.size(0)).byte()
Mxx = distance(r_feat_val, r_feat_val, sqrt=True)
for i in range(n_mode):
# Compute Score
Mxy = distance(r_feat_val, t_feat, sqrt=True)
Myy = distance(t_feat, t_feat, sqrt=True)
scores[i, 0] = wasserstein(Mxy, False)
scores[i, 1] = mmd(Mxx, Mxy, Myy, 1)
s = knn(Mxx, Mxy, Myy, 1, True)
scores[i, 2], scores[i, 3] = s.acc_t, s.acc_f
# Do drop -- fill dropped slots with remaining samples
c = collapsed_order[i]
collapsed[pred.eq(c)] = 1
cidx = index[collapsed.eq(1)]
ncidx = index[collapsed.ne(1)]
if ncidx.dim() == 0 or cidx.dim() == 0 or ncidx.size(0) == 0:
continue
for j in cidx:
copy_idx = np.random.randint(0, ncidx.size(0))
t_feat[j] = t_feat[ncidx[copy_idx]]
return scores
def overfit_exp_1(r_feat_val, r_feat_train, step=200):
# i1ep_score, mode_score, fid
n_mode = r_feat_train.size(0) // step
scores = np.zeros((n_mode+1, 4))
t_feat = r_feat_train.clone()
collapsed_order = torch.randperm(n_mode).long()
index = torch.arange(0, r_feat_train.size(0)).long()
collapsed = torch.zeros(r_feat_train.size(0)).byte()
Mxx = dista1e(r_feat_val, r_feat_val, sqrt=True)
for i in range(n_mode+1):
# Compute Score
Mxy = dista1e(r_feat_val, t_feat, sqrt=True)
Myy = dista1e(t_feat, t_feat, sqrt=True)
scores[i, 0] = wasserstein(Mxy, False)
scores[i, 1] = mmd(Mxx, Mxy, Myy, 1)
s = knn(Mxx, Mxy, Myy, 1, True)
scores[i, 2], scores[i, 3] = s.acc_t, s.acc_f
# Copy samples so as to overfit
if i == n_mode: break
t_feat[i*step:(i+1)*step] = r_feat_val[i*step:(i+1)*step]
return scores
def overfit_exp_1(r_feat_val, r_feat_train, step=200):
# incep_score, mode_score, fid
n_mode = r_feat_train.size(0) // step
scores = np.zeros((n_mode+1, 4))
t_feat = r_feat_train.clone()
collapsed_order = torch.randperm(n_mode).long()
index = torch.arange(0, r_feat_train.size(0)).long()
collapsed = torch.zeros(r_feat_train.size(0)).byte()
Mxx = distance(r_feat_val, r_feat_val, sqrt=True)
for i in range(n_mode+1):
# Compute Score
Mxy = distance(r_feat_val, t_feat, sqrt=True)
Myy = distance(t_feat, t_feat, sqrt=True)
scores[i, 0] = wasserstein(Mxy, False)
scores[i, 1] = mmd(Mxx, Mxy, Myy, 1)
s = knn(Mxx, Mxy, Myy, 1, True)
scores[i, 2], scores[i, 3] = s.acc_t, s.acc_f
# Copy samples so as to overfit
if i == n_mode: break
t_feat[i*step:(i+1)*step] = r_feat_val[i*step:(i+1)*step]
return scores