Update README.md
Browse files
README.md
CHANGED
@@ -25,13 +25,7 @@ For the classifier models, the final output goes through `nn.Softmax`.
|
|
25 |
|
26 |
# Models
|
27 |
|
28 |
-
##
|
29 |
-
|
30 |
-
- Unified (by joining the datasets of the other classifiers)
|
31 |
-
- Compression (jpg/webp/gif/dithering/etc)
|
32 |
-
- Noise
|
33 |
-
|
34 |
-
## ChromaticAberration - Anime
|
35 |
|
36 |
### Design goals
|
37 |
|
@@ -74,3 +68,54 @@ Version history:
|
|
74 |
- v1.1 - Added 300 images tagged "chromatic_aberration" from gelbooru. Added first 1000 images from danbooru2021 as reg images
|
75 |
- v1.2 - Used the newly trained predictor to filter the existing datasets - found ~70 positives in the reg set and ~30 false positives in the target set.
|
76 |
- v1.3-v1.16 - Repeatedly ran predictor against various datasets, adding false positives/negatives back into the dataset, sometimes running against the training set to filter out misclassified images as the predictor got better. Added/removed images were manually checked (My eyes hurt).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
# Models
|
27 |
|
28 |
+
## Chromatic Aberration - Anime
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
### Design goals
|
31 |
|
|
|
68 |
- v1.1 - Added 300 images tagged "chromatic_aberration" from gelbooru. Added first 1000 images from danbooru2021 as reg images
|
69 |
- v1.2 - Used the newly trained predictor to filter the existing datasets - found ~70 positives in the reg set and ~30 false positives in the target set.
|
70 |
- v1.3-v1.16 - Repeatedly ran predictor against various datasets, adding false positives/negatives back into the dataset, sometimes running against the training set to filter out misclassified images as the predictor got better. Added/removed images were manually checked (My eyes hurt).
|
71 |
+
|
72 |
+
## Image Compression - Anime
|
73 |
+
|
74 |
+
### Design goals
|
75 |
+
|
76 |
+
The goal was to detect [compression artifacts](https://en.wikipedia.org/wiki/Compression_artifact?useskin=vector) in images.
|
77 |
+
|
78 |
+
This seems like the next logical step in dataset filtering. The flagged images can either be cleaned up or tagged correctly so the resulting network won't inherit the image artifacts.
|
79 |
+
|
80 |
+
### Issues
|
81 |
+
|
82 |
+
- Low accuracy on 3D/2.5D with possible false positives.
|
83 |
+
|
84 |
+
### Training
|
85 |
+
|
86 |
+
The training settings can be found in the `config/CCAnime-Compression-v1.yaml` file (2.7e-6 LR, cosine scheduler, 40K steps).
|
87 |
+
|
88 |
+
![loss](https://github.com/city96/CityClassifiers/assets/125218114/9d0294bf-81ee-4b30-89ae-3b1aca27788e)
|
89 |
+
|
90 |
+
The eval loss only uses a single image for each target class, hence the questionable nature of the graph.
|
91 |
+
|
92 |
+
![loss-eval](https://github.com/city96/CityClassifiers/assets/125218114/77c9882f-6263-4926-b3ee-a032ef7784ea)
|
93 |
+
|
94 |
+
|
95 |
+
Final dataset score distribution for v1.5:
|
96 |
+
```
|
97 |
+
22736 images in dataset.
|
98 |
+
0_fpl - 108
|
99 |
+
0_reg_aes - 142
|
100 |
+
0_reg_gel - 7445 |||||||||||||
|
101 |
+
1_aes_jpg - 103
|
102 |
+
1_fpl - 8
|
103 |
+
1_syn_gel - 7445 |||||||||||||
|
104 |
+
1_syn_jpg - 40
|
105 |
+
2_syn_gel - 7445 |||||||||||||
|
106 |
+
2_syn_webp - 0
|
107 |
+
|
108 |
+
Class ratios:
|
109 |
+
00 - 7695 |||||||||||||
|
110 |
+
01 - 7596 |||||||||||||
|
111 |
+
02 - 7445 |||||||||||||
|
112 |
+
```
|
113 |
+
|
114 |
+
Version history:
|
115 |
+
|
116 |
+
- v1.0 - Initial test model, dataset consists of 40 hand picked images and their jpeg compressed counterpart. Compression is done with ChaiNNer, compression rate is randomized.
|
117 |
+
- v1.1 - Added more images by re-filtering the input dataset using the v1 model, keeping only the top/bottom 10%.
|
118 |
+
- v1.2 - Used the newly trained predictor to filter the existing datasets - found ~70 positives in the reg set and ~30 false positives in the target set.
|
119 |
+
- v1.3 - Scraped ~7500 images from gelbooru, filtering for min. image size of at least 3000 and a file size larger than 8MB. Compressed using ChaiNNer as before.
|
120 |
+
- v1.4 - Added webm compression to the list, decided against adding GIF/dithering since it's rarely used nowadays.
|
121 |
+
- v1.5 - Changed LR/step count to better match larger dataset. Added false positives/negatives from v1.4.
|