summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--cli/app/search/json.py4
-rw-r--r--cli/app/search/search_class.py29
-rw-r--r--cli/app/search/search_dense.py40
3 files changed, 22 insertions, 51 deletions
diff --git a/cli/app/search/json.py b/cli/app/search/json.py
index 8c3c2ef..bf6bcb9 100644
--- a/cli/app/search/json.py
+++ b/cli/app/search/json.py
@@ -67,7 +67,7 @@ def make_params_dense(tag):
"dataset": os.path.join(app_cfg.DIR_INVERSES, tag, "dataset.latent.hdf5"),
"inv_layer": "Generator_2/G_Z/Reshape:0",
"decay_lr": False,
- "inv_it": 6000,
+ "inv_it": 15000,
"generator_path": "https://tfhub.dev/deepmind/biggan-512/2",
"attention_map_layer": "Generator_2/attention/Softmax:0",
"pre_trained_latent": True,
@@ -93,7 +93,7 @@ def make_params_dense(tag):
"init_lo": -0.001,
"lambda_mse": 1.0,
"lambda_reg": 0.1,
- "dist_loss": True,
+ "dist_loss": False,
"sample_size": 4,
"out_dataset": os.path.join(app_cfg.DIR_INVERSES, tag, "dataset.dense.hdf5"),
"save_progress": True,
diff --git a/cli/app/search/search_class.py b/cli/app/search/search_class.py
index 691a0a8..7961a0c 100644
--- a/cli/app/search/search_class.py
+++ b/cli/app/search/search_class.py
@@ -65,13 +65,13 @@ def find_nearest_vector_for_images(paths, opt_dims, opt_steps, opt_video, opt_ta
if index == opt_limit:
break
out_fns[index] = os.path.basename(path)
- fp_frames = find_nearest_vector(sess, generator, path, opt_dims, out_images, out_labels, out_latent, opt_steps, index,
+ fp_frames = find_nearest_vector(sess, generator, path, opt_dims, out_images, out_labels, out_latent, opt_steps, index, opt_tag,
opt_stochastic_clipping, opt_label_clipping, opt_use_feature_detector, opt_feature_layers, opt_snapshot_interval, opt_clip_interval)
if opt_video:
export_video(fp_frames)
sess.close()
-def find_nearest_vector(sess, generator, opt_fp_in, opt_dims, out_images, out_labels, out_latent, opt_steps, index,
+def find_nearest_vector(sess, generator, opt_fp_in, opt_dims, out_images, out_labels, out_latent, opt_steps, index, opt_tag,
opt_stochastic_clipping, opt_label_clipping, opt_use_feature_detector, opt_feature_layers, opt_snapshot_interval, opt_clip_interval):
"""
Find the closest latent and class vectors for an image. Store the class vector in an HDF5.
@@ -87,9 +87,6 @@ def find_nearest_vector(sess, generator, opt_fp_in, opt_dims, out_images, out_la
z_initial = truncated_z_sample(batch_size, z_dim, truncation/2)
y_initial = create_labels_uniform(batch_size, vocab_size)
- z_lr = 0.001
- y_lr = 0.001
-
input_z = tf.compat.v1.Variable(z_initial, dtype=np.float32, constraint=lambda t: tf.clip_by_value(t, -2, 2))
input_y = tf.compat.v1.Variable(y_initial, dtype=np.float32, constraint=lambda t: tf.clip_by_value(t, 0, 1))
input_trunc = tf.compat.v1.constant(1.0)
@@ -101,7 +98,7 @@ def find_nearest_vector(sess, generator, opt_fp_in, opt_dims, out_images, out_la
## clip the Z encoding
- opt_clip = 1.0
+ opt_clip = 2.0
clipped_encoding = tf.where(tf.abs(input_z) >= opt_clip,
tf.random.uniform([batch_size, z_dim], minval=-opt_clip, maxval=opt_clip), input_z)
@@ -145,12 +142,17 @@ def find_nearest_vector(sess, generator, opt_fp_in, opt_dims, out_images, out_la
feat_loss += tf.reduce_mean(feat_square_diff) / len(opt_feature_layers)
loss = 1.0 * mse_loss + 1.0 * feat_loss
+
+ z_lr = 0.001
+ y_lr = 0.001
train_step_z = tf.train.AdamOptimizer(learning_rate=z_lr, beta1=0.9, beta2=0.999) \
.minimize(loss, var_list=[input_z])
train_step_y = tf.train.AdamOptimizer(learning_rate=y_lr, beta1=0.9, beta2=0.999) \
.minimize(loss, var_list=[input_y])
else:
+ z_lr = 0.001
+ y_lr = 0.001
loss = tf.compat.v1.losses.mean_squared_error(target, output)
train_step_z = tf.train.AdamOptimizer(z_lr).minimize(loss, var_list=[input_z], name='AdamOpterZ')
train_step_y = tf.train.AdamOptimizer(y_lr).minimize(loss, var_list=[input_y], name='AdamOpterY')
@@ -183,17 +185,22 @@ def find_nearest_vector(sess, generator, opt_fp_in, opt_dims, out_images, out_la
print("Iterating!")
if i % 20 == 0:
print('iter: {}, loss: {}'.format(i, curr_loss))
- if opt_stochastic_clipping != 0 and (i % opt_stochastic_clipping) == 0:
- sess.run(clip_latent)
- if opt_label_clipping != 0 and (i % opt_label_clipping) == 0:
- sess.run(clip_labels)
- if opt_snapshot_interval != 0 and (i % opt_snapshot_interval) == 0:
+ if i > 0:
+ if opt_stochastic_clipping != 0 and (i % opt_stochastic_clipping) == 0:
+ sess.run(clip_latent)
+ if opt_label_clipping != 0 and (i % opt_label_clipping) == 0:
+ sess.run(clip_labels)
+ if opt_video and opt_snapshot_interval != 0 and (i % opt_snapshot_interval) == 0:
phi_guess = sess.run(output)
guess_im = imgrid(imconvert_uint8(phi_guess), cols=1)
imwrite(join(app_cfg.DIR_OUTPUTS, fp_frames, 'frame_{:04d}.png'.format(int(i / opt_snapshot_interval))), guess_im)
except KeyboardInterrupt:
pass
+ phi_guess = sess.run(output)
+ guess_im = imgrid(imconvert_uint8(phi_guess), cols=1)
+ imwrite(join(app_cfg.DIR_OUTPUTS, 'frame_{}_final.png'.format(opt_tag)), guess_im)
+
z_guess, y_guess = sess.run([input_z, input_y])
out_images[index] = phi_target_for_inversion
out_labels[index] = y_guess
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")
generated by cgit v1.2.3-70-g09d2 (git 2.51.0) at 2025-12-10 20:56:25 +0000