diff options
Diffstat (limited to 'inversion')
| -rw-r--r-- | inversion/image_inversion_inception.py | 66 |
1 files changed, 47 insertions, 19 deletions
diff --git a/inversion/image_inversion_inception.py b/inversion/image_inversion_inception.py index f9d1c79..ae4793f 100644 --- a/inversion/image_inversion_inception.py +++ b/inversion/image_inversion_inception.py @@ -220,23 +220,23 @@ if params.features: gen_feat_ex = feature_extractor(dict(images=gen_img_1), as_dict=True, signature='image_feature_vector') target_feat_ex = feature_extractor(dict(images=target_img_1), as_dict=True, signature='image_feature_vector') - 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) * 0.5 - img_feat_err = tf.reduce_mean(feat_square_diff, axis=1) * 0.5 + # 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) * 0.5 + # img_feat_err = tf.reduce_mean(feat_square_diff, axis=1) * 0.5 - gen_feat = gen_feat_ex["InceptionV3/Mixed_6b"] - target_feat = target_feat_ex["InceptionV3/Mixed_6b"] - feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1]) - feat_loss += tf.reduce_mean(feat_square_diff) * 0.16 - img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.16 + # # gen_feat = gen_feat_ex["InceptionV3/Mixed_6b"] + # # target_feat = target_feat_ex["InceptionV3/Mixed_6b"] + # # feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1]) + # # feat_loss += tf.reduce_mean(feat_square_diff) * 0.16 + # # img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.16 - gen_feat = gen_feat_ex["InceptionV3/Mixed_5a"] - target_feat = target_feat_ex["InceptionV3/Mixed_5a"] - feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1]) - feat_loss += tf.reduce_mean(feat_square_diff) * 0.16 - img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.16 + # # gen_feat = gen_feat_ex["InceptionV3/Mixed_5a"] + # # target_feat = target_feat_ex["InceptionV3/Mixed_5a"] + # # feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1]) + # # feat_loss += tf.reduce_mean(feat_square_diff) * 0.16 + # # img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.16 # gen_feat = gen_feat_ex["InceptionV3/Mixed_7b"] # target_feat = target_feat_ex["InceptionV3/Mixed_7b"] @@ -244,11 +244,36 @@ if params.features: # feat_loss += tf.reduce_mean(feat_square_diff) * 0.33 # img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) - gen_feat = gen_feat_ex["InceptionV3/Mixed_7c"] - target_feat = target_feat_ex["InceptionV3/Mixed_7c"] + # # gen_feat = gen_feat_ex["InceptionV3/Mixed_7c"] + # # target_feat = target_feat_ex["InceptionV3/Mixed_7c"] + # # feat_square_diff = tf.reshape(tf.square(gen_feat - target_feat), [BATCH_SIZE, -1]) + # # feat_loss += tf.reduce_mean(feat_square_diff) * 0.17 + # # img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.17 + + # 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_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) * 0.17 - img_feat_err += tf.reduce_mean(feat_square_diff, axis=1) * 0.17 + 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_4a_3x3"] + target_feat = target_feat_ex["InceptionV3/Conv2d_4a_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 else: feat_loss = tf.constant(0.0) @@ -481,11 +506,14 @@ for image_batch, label_batch in image_gen: if params.save_progress: # Save linear interpolation between the actual and generated encodings. if params.dist_loss: + print("Generating linear interpolation...") enc_batch, gen_enc = sess.run([encoding, gen_encoding]) for j in range(10): + gen_time = time.time() custom_enc = gen_enc * (1-(j/10.0)) + enc_batch * (j/10.0) sess.run(encoding.assign(custom_enc)) gen_images = sess.run(gen_img) + print("Generation time: {:.1f}s".format(time.time() - gen_time)) inv_batch = vs.interleave(vs.data2img(image_batch[BATCH_SIZE - SAMPLE_SIZE:]), vs.data2img(gen_images[BATCH_SIZE - SAMPLE_SIZE:])) inv_batch = vs.grid_transform(inv_batch) |