Added Ability to Create GIFs

README.md

@@ -166,6 +166,10 @@ path or a full absolute path.
 * `-use_fft`: Whether to enable Fast Fourier transform (FFT) decorrelation.
 * `-fft_block`: The size of your FFT frequency filtering block. Default is `25`.
 
+**GIF Options:**
+* `-create_gif`: Whether to create a GIF from the output images after all iterations have been completed.
+* `-frame_duration`: The duration for each GIF frame in milliseconds. Default is `100`.
+
 **Help Options:**
 * `-print_layers`: Pass this flag to print the names of all usable layers for the selected model.
 * `-print_channels`: Pass this flag to print all the selected channels.
@@ -234,4 +238,4 @@ For example in a server with four GPUs, you can give the flag `-gpu 0,1,2,3` to
 We can achieve very high quality results at high resolution by combining multi-GPU processing with multiscale
 generation as described in the paper
 <a href="https://arxiv.org/abs/1611.07865">**Controlling Perceptual Factors in Neural Style Transfer**</a> by Leon A. Gatys,
-Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann and Eli Shechtman.
+Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann and Eli Shechtman.

neural_dream.py

@@ -1,6 +1,5 @@
 import os
 import copy
-import math
 import random
 import torch
 import torch.nn as nn
@@ -67,6 +66,10 @@
 parser.add_argument("-zoom", type=int, default=0)
 parser.add_argument("-zoom_mode", choices=['percent', 'pixel'], default='percent')
 
+# Gif options
+parser.add_argument("-create_gif", action='store_true')
+parser.add_argument("-frame_duration", type=int, default=100)
+
 # Other options
 parser.add_argument("-original_colors", type=int, choices=[0, 1], default=0)
 parser.add_argument("-pooling", choices=['avg', 'max'], default='max')
@@ -296,6 +299,9 @@ def save_output(t, save_img, content_image, iter_name, model_mean):
 
     disp.save(str(filename))
 
+    if t == params.num_iterations and params.create_gif:
+        dream_image.create_gif(output_filename, params.frame_duration)
+
 
 def maybe_save(t, save_img, content_image, input_mean):
     should_save = params.save_iter > 0 and t % params.save_iter == 0
@@ -579,4 +585,4 @@ def new_img(input_image, scale_factor, mode='bilinear'):
 
 
 if __name__ == "__main__":
-    main()
+    main()

neural_dream/dream_image.py

@@ -1,6 +1,8 @@
+import os
 import torch
 import torch.nn as nn
 import neural_dream.dream_utils as dream_utils
+from PIL import Image
 
 
 # Adjust tensor contrast
@@ -31,4 +33,32 @@ def zoom(input, crop_val, mode='percent'):
     h, w = input.size(2), input.size(3)
     input = center_crop(input.clone(), crop_val, mode=mode)
     input = resize_tensor(input, (h,w))
-    return input
+    return input
+
+
+# Get most common Pillow Image size from list
+def common_size(l, v):
+    l = [list(im.size)[v] for im in l]
+    return max(set(l), key = l.count)
+
+
+# Create GIF from images
+def create_gif(base_name, duration=100):
+    frames_dir, base_name = os.path.split(base_name)
+    if frames_dir == '':
+        frames_dir = '.'
+    base_name = base_name.rsplit('.', 1)[0]
+
+    ext = [".jpg", ".jpeg", ".png", ".tiff"]
+    image_list = [file for file in os.listdir(frames_dir) if os.path.splitext(file)[1].lower() in ext]
+    image_list = [im for im in image_list if base_name in im]
+    if "_" in image_list[0]:
+        fsorted = sorted(image_list,key=lambda x: int(os.path.splitext(x)[0].rsplit('_', 1)[1]))
+    else:
+        fsorted = sorted(image_list[1:],key=lambda x: int(os.path.splitext(x)[0].rsplit('_', 1)[1]))
+        fsorted.append(image_list[0])
+
+    frames = [Image.open(os.path.join(frames_dir, im)).convert('RGBA') for im in fsorted]
+    w, h = common_size(frames, 0), common_size(frames, 1)
+    frames = [im for im in frames if list(im.size)[0] == w and list(im.size)[1] == h]
+    frames[0].save(os.path.join(frames_dir, base_name+'.gif'), format='GIF', append_images=frames[1:], save_all=True, duration=duration, loop=0)