fix dense

1 files changed, 20 insertions, 44 deletions

diff --git a/cli/app/search/search_dense.py b/cli/app/search/search_dense.py
index 46183c7..362e0ce 100644
--- a/cli/app/search/search_dense.py
+++ b/cli/app/search/search_dense.py
@@ -21,49 +21,25 @@ from app.settings import app_cfg
 from app.utils.file_utils import write_pickle
 from app.utils.cortex_utils import upload_bytes_to_cortex
 
-# --------------------------
-# Hyper-parameters.
-# --------------------------
-# Expected parameters:
-#  generator_path: path to generator module.
-#  generator_fixed_inputs: dictionary of fixed generator's input parameters.
-#  dataset: name of the dataset (hdf5 file).
-#  dataset_out: name for the output inverted dataset (hdf5 file).
-# General parameters:
-#  batch_size: number of images inverted at the same time.
-#  inv_it: number of iterations to invert an image.
-#  inv_layer: 'latent' or name of the tensor of the custom layer to be inverted.
-#  lr: learning rate.
-#  decay_lr: exponential decay on the learning rate.
-#  decay_n: number of exponential decays on the learning rate.
-#  custom_grad_relu: replace relus with custom gradient.
-# Logging:
-#  sample_size: number of images included in sampled images.
-#  save_progress: whether to save intermediate images during optimization.
-#  log_z_norm: log the norm of different sections of z.
-#  log_activation_layer: log the percentage of active neurons in this layer.
-# Losses:
-#  mse: use the mean squared error on pixels for image comparison.
-#  features: use features extracted by a feature extractor for image comparison.
-#  feature_extractor_path: path to feature extractor module.
-#  feature_extractor_output: output name from feature extractor.
-#  likeli_loss: regularization loss on the log likelihood of encodings.
-#  norm_loss: regularization loss on the norm of encodings.
-#  dist_loss: whether to include a loss on the dist between g1(z) and enc.
-#  lambda_mse: coefficient for mse loss.
-#  lambda_feat: coefficient for features loss.
-#  lambda_reg: coefficient for regularization loss on latent.
-#  lambda_dist: coefficient for l1 regularization on delta.
-# Latent:
-#  clipping: whether to clip encoding values after every update.
-#  stochastic_clipping: whether to consider stochastic clipping.
-#  clip: clipping bound.
-#  pretrained_latent: load pre trained fixed latent.
-#  fixed_z: do not train the latent vector.
-# Initialization:
-#  init_gen_dist: initialize encodings from the generated distribution.
-#  init_lo: init min value.
-#  init_hi: init max value.
+feature_layer_names = {
+  '1a': "InceptionV3/Conv2d_1a_3x3",
+  '2a': "InceptionV3/Conv2d_2a_3x3",
+  '2b': "InceptionV3/Conv2d_2b_3x3",
+  '3a': "InceptionV3/Conv2d_3a_3x3",
+  '3b': "InceptionV3/Conv2d_3b_3x3",
+  '4a': "InceptionV3/Conv2d_4a_3x3",
+  '5b': "InceptionV3/Mixed_5b",
+  '5c': "InceptionV3/Mixed_5c",
+  '5d': "InceptionV3/Mixed_5d",
+  '6a': "InceptionV3/Mixed_6a",
+  '6b': "InceptionV3/Mixed_6b",
+  '6c': "InceptionV3/Mixed_6c",
+  '6d': "InceptionV3/Mixed_6d",
+  '6e': "InceptionV3/Mixed_6e",
+  '7a': "InceptionV3/Mixed_7a",
+  '7b': "InceptionV3/Mixed_7b",
+  '7c': "InceptionV3/Mixed_7c",
+}
 
 def find_dense_embedding_for_images(params):
   # --------------------------
@@ -444,7 +420,7 @@ def find_dense_embedding_for_images(params):
     for i in range(BATCH_SIZE):
       out_i = out_pos + i
       sample_fn, ext = os.path.splitext(sample_fns[out_i])
-      image = Image.fromarray(image)
+      image = Image.fromarray(images[i])
       fp = BytesIO()
       image.save(fp, format='png')
       data = upload_bytes_to_cortex(params.folder_id, sample_fn + "-inverse.png", fp, "image/png")