bug fix

src/pytorch_metric_learning/losses/dynamic_soft_margin_loss.py

@@ -81,6 +81,11 @@ def compute_loss_without_labels(
  ):
  mat = self.distance(embeddings, ref_emb)
  r, c = mat.size()
+
+ d_pos = torch.zeros(max(r, c))
+ d_pos = c_f.to_device(d_pos, tensor=embeddings, dtype=embeddings.dtype)
+ d_pos[: min(r, c)] = mat.diag()
+ mat.fill_diagonal_(np.inf)
 
  min_a, min_p = torch.zeros(max(r, c)), torch.zeros(
  max(r, c)
@@ -90,9 +95,7 @@ def compute_loss_without_labels(
  min_a[:c], _ = torch.min(mat, dim=0)
  min_p[:r], _ = torch.min(mat, dim=1)
 
- d_pos = torch.zeros(max(r, c))
- d_pos = c_f.to_device(d_pos, tensor=embeddings, dtype=embeddings.dtype)
- d_pos[: min(r, c)], d_neg = mat.diag(), torch.min(min_a, min_p)
+ d_neg = torch.min(min_a, min_p)
  return d_pos - d_neg
 
  def compute_loss_with_labels(

src/pytorch_metric_learning/losses/ranked_list_loss.py

@@ -70,7 +70,6 @@ def compute_loss(self, embeddings, labels, indices_tuple, ref_emb, ref_labels):
  )
  w_n[N_star > 0] = torch.exp(self.Tn * (self.alpha - N_star[N_star > 0]))
 
- print("w_P: ", w_p)
  loss_P = torch.sum(
  w_p * (P_star - (self.alpha - self.margin)), dim=1
  ) / torch.sum(w_p + 1e-5, dim=1)

tests/losses/test_dynamic_soft_margin_loss.py

@@ -110,6 +110,7 @@ def _compute_l2_distances(self, x, labels=None):
  return find_hard_negatives(dmat, output_index=False, empirical_thresh=0.008)
  else:
  dmat = compute_distance_matrix_unit_l2(x, x)
+ dmat.fill_diagonal_(0) # Put distance to itself to 0
  anchor_idx, positive_idx, negative_idx = lmu.convert_to_triplets(
  None, labels, labels, t_per_anchor="all"
  )
@@ -200,8 +201,10 @@ def forward(self, x, labels=None):
  bin_idx = torch.floor((hist_var - self._min_val) / self.bin_width).long()
  weight = CDF[bin_idx]
 
- loss = -(neg_dist * weight).mean() + (pos_dist * weight).mean()
- return loss.to(dtype=x.dtype)
+ # Changed to an equivalent version for making same computation as in dynamic_soft_margin_loss.py
+ # loss = -(neg_dist * weight).mean() + (pos_dist * weight).mean()
+ loss = (hist_var*weight).mean()
+ return loss.to(device=x.device, dtype=x.dtype) # Added cast to avoid errors
 
 
 import unittest