surabaya on various multi style transfer learning
| original image | starry night |
|---|---|
 |
 |
 |
 |
| original image | great wave off |
|---|---|
|
 |
|
 |
| original image | Luncheon of the Boating Party |
|---|---|
|
 |
|
 |
| original image | levitan |
|---|---|
|
 |
|
 |
| original image | montmartre |
|---|---|
|
 |
|
 |
"insert famous painters name painted insert painting with different style"
training
at initiation, uses a randomly init (white noise) image,
using the bash script, content and style pics will be trained 10 times where each layer were configured like the bash script below, gradually increasing on image_size, and making the init_image based on the output of previous result:
python3 neural_style.py \
-content_image examples/inputs/its.jpg\
-style_image wave_off.jpg\
-init image -init_image out5.png \
-style_scale 1.0 \
-print_iter 10 \
-style_weight 2500 \
-image_size 1024 \
-num_iterations 2000 \
-output_image out6.png \
-tv_weight 0 \
-gpu 0 \
-backend cudnn
tpu vs gpu
running on starry_bigger_time.sh, measured on time_check.py
- gpu@colab time: 3045.925956964493 s
- gpu@p5000 time: 971.1943960189819 s
neural-style-transfer bot
I've created a telegram bot @neuralstyletransferbot you can see the live demo here
pycon 2019 slide
you can view it here