dense search with args

10 files changed, 44 insertions, 27 deletions

diff --git a/cli/app/commands/biggan/search_dense.py b/cli/app/commands/biggan/search_dense.py
index 0cf56a3..b743350 100644
--- a/cli/app/commands/biggan/search_dense.py
+++ b/cli/app/commands/biggan/search_dense.py
@@ -1,15 +1,22 @@
 import click
 
-from app.search.params import Params
+from app.search.params import Params, timestamp
 from app.search.search_dense import find_dense_embedding_for_images
 
 @click.command('')
 @click.option('-i', '--input', 'opt_fp_in', required=True, 
   help='Path to input image')
+@click.option('-t', '--tag', 'opt_tag', default="inverse_" + timestamp(), 
+  help='Tag this build')
+@click.option('-ll', '--feature_layers', 'opt_feature_layers', default="1a,2a,4a,7a",
+  help='Feature layers used for loss')
+@click.option('-s', '--save_progress', 'opt_save_progress', is_flag=True,
+  help='Save example images every 500 frames')
 @click.pass_context
-def cli(ctx, opt_fp_in):
+def cli(ctx, opt_fp_in, opt_tag, opt_feature_layers, opt_save_progress):
   """
   Search for an image (class vector) in BigGAN using gradient descent
   """
   params = Params(opt_fp_in)
-  find_dense_embedding_for_images(params)
+  opt_feature_layers = opt_feature_layers.split(',')
+  find_dense_embedding_for_images(params, opt_tag=opt_tag, opt_feature_layers=opt_feature_layers, opt_save_progress=opt_save_progress)
diff --git a/cli/app/search/search_dense.py b/cli/app/search/search_dense.py
index ac67d07..ac66a73 100644
--- a/cli/app/search/search_dense.py
+++ b/cli/app/search/search_dense.py
@@ -42,7 +42,7 @@ feature_layer_names = {
   '7c': "InceptionV3/Mixed_7c",
 }
 
-def find_dense_embedding_for_images(params):
+def find_dense_embedding_for_images(params, opt_tag="inverse", opt_feature_layers=["1a,2a,3a,4a,7a"], opt_save_progress=True):
   # --------------------------
   # Global directories.
   # --------------------------
@@ -227,24 +227,35 @@ def find_dense_embedding_for_images(params):
     # # feat_loss += tf.reduce_mean(feat_square_diff) * 0.17
     # # img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.17
 
+    feat_square_diff = tf.constant(0.0)
+    img_feat_err = tf.constant(0.0)
+
+    for layer in opt_feature_layers:
+      layer_name = feature_layer_names[layer]
+      gen_feat = gen_feat_ex[layer_name]
+      target_feat = target_feat_ex[layer_name]
+      feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [batch_size, -1])
+      feat_loss += tf.reduce_mean(feat_square_diff) / len(opt_feature_layers)
+      img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) / len(opt_feature_layers)
+
     # conv1 1, conv1 2, conv3 2 and conv4 2
-    gen_feat = gen_feat_ex["InceptionV3/Conv2d_1a_3x3"]
-    target_feat = target_feat_ex["InceptionV3/Conv2d_1a_3x3"]
-    feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1])
-    feat_loss = tf.reduce_mean(feat_square_diff) * 0.25
-    img_feat_err = tf.reduce_mean(feat_square_diff, axis=1) * 0.25
+    # gen_feat = gen_feat_ex["InceptionV3/Conv2d_1a_3x3"]
+    # target_feat = target_feat_ex["InceptionV3/Conv2d_1a_3x3"]
+    # feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1])
+    # feat_loss = tf.reduce_mean(feat_square_diff) * 0.25
+    # img_feat_err = tf.reduce_mean(feat_square_diff, axis=1) * 0.25
 
-    gen_feat = gen_feat_ex["InceptionV3/Conv2d_2a_3x3"]
-    target_feat = target_feat_ex["InceptionV3/Conv2d_2a_3x3"]
-    feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1])
-    feat_loss += tf.reduce_mean(feat_square_diff) * 0.25
-    img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.25
+    # gen_feat = gen_feat_ex["InceptionV3/Conv2d_2a_3x3"]
+    # target_feat = target_feat_ex["InceptionV3/Conv2d_2a_3x3"]
+    # feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1])
+    # feat_loss += tf.reduce_mean(feat_square_diff) * 0.25
+    # img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.25
 
-    gen_feat = gen_feat_ex["InceptionV3/Conv2d_3b_1x1"]
-    target_feat = target_feat_ex["InceptionV3/Conv2d_3b_1x1"]
-    feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1])
-    feat_loss += tf.reduce_mean(feat_square_diff) * 0.25
-    img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.25
+    # gen_feat = gen_feat_ex["InceptionV3/Conv2d_3b_1x1"]
+    # target_feat = target_feat_ex["InceptionV3/Conv2d_3b_1x1"]
+    # feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1])
+    # feat_loss += tf.reduce_mean(feat_square_diff) * 0.25
+    # img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.25
 
     # gen_feat = gen_feat_ex["InceptionV3/Mixed_6a"]
     # target_feat = target_feat_ex["InceptionV3/Mixed_6a"]
@@ -252,11 +263,11 @@ def find_dense_embedding_for_images(params):
     # feat_loss += tf.reduce_mean(feat_square_diff) * 0.25
     # img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.25
 
-    gen_feat = gen_feat_ex["InceptionV3/Mixed_7a"]
-    target_feat = target_feat_ex["InceptionV3/Mixed_7a"]
-    feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1])
-    feat_loss += tf.reduce_mean(feat_square_diff) * 0.25
-    img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.25
+    # gen_feat = gen_feat_ex["InceptionV3/Mixed_7a"]
+    # target_feat = target_feat_ex["InceptionV3/Mixed_7a"]
+    # feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1])
+    # feat_loss += tf.reduce_mean(feat_square_diff) * 0.25
+    # img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.25
 
   else:
     feat_loss = tf.constant(0.0)
@@ -393,7 +404,7 @@ def find_dense_embedding_for_images(params):
         sys.stdout.flush()
 
         # Save target images and reconstructions.
-        if params.save_progress:
+        if opt_save_progress:
           assert SAMPLE_SIZE <= BATCH_SIZE
           gen_time = time.time()
           gen_images  = sess.run(gen_img)
@@ -423,7 +434,7 @@ def find_dense_embedding_for_images(params):
       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")
+      data = upload_bytes_to_cortex(params.folder_id, "{}-{}.png".format(sample_fn, tag), fp, "image/png")
       print(json.dumps(data, indent=2))
       if data is not None and 'files' in data:
         file_id = data['files'][0]['id']
diff --git a/cli/categories b/cli/categories
deleted file mode 120000
index 5668d42..0000000
--- a/cli/categories
+++ /dev/null
@@ -1 +0,0 @@
-../inversion/categories
\ No newline at end of file
diff --git a/inversion/categories/biggan_all.json b/cli/categories/biggan_all.json
index 9981a91..9981a91 100644
--- a/inversion/categories/biggan_all.json
+++ b/cli/categories/biggan_all.json
diff --git a/inversion/categories/biggan_all_categories.json b/cli/categories/biggan_all_categories.json
index 5212e90..5212e90 100644
--- a/inversion/categories/biggan_all_categories.json
+++ b/cli/categories/biggan_all_categories.json
diff --git a/inversion/categories/biggan_all_categories.txt b/cli/categories/biggan_all_categories.txt
index a0e37eb..a0e37eb 100644
--- a/inversion/categories/biggan_all_categories.txt
+++ b/cli/categories/biggan_all_categories.txt
diff --git a/inversion/categories/biggan_animate.json b/cli/categories/biggan_animate.json
index 35a35b0..35a35b0 100644
--- a/inversion/categories/biggan_animate.json
+++ b/cli/categories/biggan_animate.json
diff --git a/inversion/categories/biggan_animate.txt b/cli/categories/biggan_animate.txt
index 69386c9..69386c9 100644
--- a/inversion/categories/biggan_animate.txt
+++ b/cli/categories/biggan_animate.txt
diff --git a/inversion/categories/biggan_inanimate.json b/cli/categories/biggan_inanimate.json
index 6a7498c..6a7498c 100644
--- a/inversion/categories/biggan_inanimate.json
+++ b/cli/categories/biggan_inanimate.json
diff --git a/inversion/categories/biggan_inanimate.txt b/cli/categories/biggan_inanimate.txt
index 05540d9..05540d9 100644
--- a/inversion/categories/biggan_inanimate.txt
+++ b/cli/categories/biggan_inanimate.txt