This is an automated email from the ASF dual-hosted git repository.
haibin pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git
The following commit(s) were added to refs/heads/master by this push:
new 632f514 Sampling negative samples other classes only (#10980)
632f514 is described below
commit 632f5140a69b51ff87129e6742e5be46684dc58c
Author: Jon <jonbakerf...@gmail.com>
AuthorDate: Thu May 17 13:39:37 2018 +0930
Sampling negative samples other classes only (#10980)
The original code will have a chance to sample negative samples from the
same class of the anchors
---
example/gluon/embedding_learning/model.py | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/example/gluon/embedding_learning/model.py
b/example/gluon/embedding_learning/model.py
index 91f7735..f82240e 100644
--- a/example/gluon/embedding_learning/model.py
+++ b/example/gluon/embedding_learning/model.py
@@ -108,6 +108,7 @@ class DistanceWeightedSampling(HybridBlock):
mask = np.ones(weights.shape)
for i in range(0, n, k):
mask[i:i+k, i:i+k] = 0
+ mask_uniform_probs = mask * (1.0/(n-k))
weights = weights * F.array(mask) * (distance <
self.nonzero_loss_cutoff)
weights_sum = F.sum(weights, axis=1, keepdims=True)
@@ -125,7 +126,7 @@ class DistanceWeightedSampling(HybridBlock):
n_indices += np.random.choice(n, k-1, p=np_weights[i]).tolist()
else:
# all samples are above the cutoff so we sample uniformly
- n_indices += np.random.choice(n, k-1).tolist()
+ n_indices += np.random.choice(n, k-1,
p=mask_uniform_probs[i]).tolist()
for j in range(block_idx * k, (block_idx + 1) * k):
if j != i:
a_indices.append(i)
--
To stop receiving notification emails like this one, please contact
hai...@apache.org.
