up

1 files changed, 2 insertions, 38 deletions

diff --git a/cli/app/search/search_dense.py b/cli/app/search/search_dense.py
index 27be015..7c0c728 100644
--- a/cli/app/search/search_dense.py
+++ b/cli/app/search/search_dense.py
@@ -284,42 +284,6 @@ def find_dense_embedding_for_images(params):
     feat_loss = tf.constant(0.0)
     img_feat_err = tf.constant(0.0)
 
-  # --------------------------
-  # Regularization losses.
-  # --------------------------
-  # Loss on the norm of the encoding.
-  if params.norm_loss:
-    dim = 20
-    chi2_dist = tfp.distributions.Chi2(dim)
-    mode = dim - 2
-    mode_log_prob = chi2_dist.log_prob(mode)
-    norm_loss = 0.0
-    for i in range(int(Z_DIM / dim)):
-      squared_l2 = tf.reduce_sum(tf.square(latent[:,i*dim:(i+1)*dim]), axis=1)
-      over_mode = tf.nn.relu(squared_l2 - mode)
-      norm_loss -= tf.reduce_mean(chi2_dist.log_prob(mode + over_mode))
-      norm_loss += mode_log_prob
-  else:
-    norm_loss = tf.constant(0.0)
-
-  # Loss on the likelihood of the encoding.
-  if params.likeli_loss:
-    norm_dist = tfp.distributions.Normal(0.0, 1.0)
-    likeli_loss = - tf.reduce_mean(norm_dist.log_prob(latent))
-    mode_log_prob = norm_dist.log_prob(0.0)
-    likeli_loss += mode_log_prob
-  else:
-    likeli_loss = tf.constant(0.0)
-
-  # Regularization loss.
-  reg_loss = norm_loss + likeli_loss
-
-  # Loss on the l1 distance between gen_encoding and inverted encoding.
-  if params.dist_loss:
-    dist_loss = tf.reduce_mean(tf.abs(encoding - gen_encoding))
-  else:
-    dist_loss = tf.constant(0.0)
-
   # Per image reconstruction error.
   img_rec_err = params.lambda_mse * img_mse_err\
       + params.lambda_feat * img_feat_err
@@ -328,8 +292,7 @@ def find_dense_embedding_for_images(params):
   rec_loss = params.lambda_mse * mse_loss + params.lambda_feat * feat_loss
 
   # Total inversion loss.
-  inv_loss = rec_loss + params.lambda_reg * reg_loss\
-      + params.lambda_dist * dist_loss
+  inv_loss = rec_loss
 
   # --------------------------
   # Optimizer.
@@ -480,6 +443,7 @@ def find_dense_embedding_for_images(params):
       out_i = out_pos + i
       sample_fn, ext = os.path.splitext(sample_fns[out_i])
       image = Image.fromarray(images[i])
+      image = vs.grid_transform(image)
       fp = BytesIO()
       image.save(fp, format='png')
       data = upload_bytes_to_cortex(params.folder_id, sample_fn + "-inverse.png", fp, "image/png")