Hugging Face's logo

SCU-VIP-Lab
/
FlexICM

PyTorch
compression
detection
Model card Files
xet
 Community
FlexICM
File size: 3,167 Bytes
4c4a6cb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
223873d
 
 
 
 
 
 
 
 
 
5c11913
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
# Checkpoints Layout (Placeholders)

Put pretrained / trained weights here before running **test / eval**.
Training still writes to `logs/` by default; after training, copy (or symlink) best
checkpoints into this tree so eval configs have a stable path.

> Files named `PLACEHOLDER` are not real weights. Replace each with the matching
> `.pth.tar` checkpoint, then update or keep the path expected by eval scripts.

## Directory map

```text
checkpoints/
β”œβ”€β”€ base_codec/                 # frozen TIC (TransTIC / AdaptiveICMH)
β”‚   β”œβ”€β”€ base_codec_1.pth.tar    # Ξ» = 0.0035
β”‚   β”œβ”€β”€ base_codec_2.pth.tar    # Ξ» = 0.0067
β”‚   β”œβ”€β”€ base_codec_3.pth.tar    # Ξ» = 0.0130
β”‚   └── base_codec_4.pth.tar    # Ξ» = 0.0250
β”‚
β”œβ”€β”€ taic/                       # five single-task TAIC codecs
β”‚   β”œβ”€β”€ detection/{1,2,3,4}/checkpoint_best_loss.pth.tar
β”‚   β”œβ”€β”€ semantic/{1,2,3,4}/checkpoint_best_loss.pth.tar
β”‚   β”œβ”€β”€ instance/{1,2,3,4}/checkpoint_best_loss.pth.tar
β”‚   β”œβ”€β”€ panoptic/{1,2,3,4}/checkpoint_best_loss.pth.tar
β”‚   └── pose/{1,2,3,4}/checkpoint_best_loss.pth.tar
β”‚
└── ctaic/                      # three multi-task scenarios
    β”œβ”€β”€ s1_det_instance/
    β”‚   β”œβ”€β”€ stage1/{1,2,3,4}/checkpoint_best_loss.pth.tar
    β”‚   └── stage2/{1,2,3,4}/checkpoint_best_loss.pth.tar
    β”œβ”€β”€ s2_sem_panoptic/
    β”‚   β”œβ”€β”€ stage1/{1,2,3,4}/checkpoint_best_loss.pth.tar
    β”‚   └── stage2/{1,2,3,4}/checkpoint_best_loss.pth.tar
    └── s3_det_pose/
        β”œβ”€β”€ stage1/{1,2,3,4}/checkpoint_best_loss.pth.tar
        └── stage2/{1,2,3,4}/checkpoint_best_loss.pth.tar
```

Quality folders `{1,2,3,4}` match paper Ξ» / TIC quality levels.

## Download base TIC codecs

```bash
bash scripts/download_base_codecs.sh
# downloads into checkpoints/base_codec/
```

## After training: copy into placeholders

```bash
# example: TAIC detection, quality 1
cp logs/taic_detection/1/checkpoint_best_loss.pth.tar \
   checkpoints/taic/detection/1/checkpoint_best_loss.pth.tar

# example: C-TAIC s1 stage2, quality 1
cp logs/ctaic_s1_stage2/1/checkpoint_best_loss.pth.tar \
   checkpoints/ctaic/s1_det_instance/stage2/1/checkpoint_best_loss.pth.tar
```

## Eval configs

See `configs/eval/` β€” they point to these placeholder paths.

Codec test (bpp / feature distortion):

```bash
python scripts/eval_taic.py -c configs/eval/taic_detection.yaml
python scripts/eval_ctaic.py -c configs/eval/ctaic_s1.yaml
```

Scripts refuse to run if a `PLACEHOLDER` file is still present or if the `.pth.tar` is missing.
Task rate–accuracy metrics are not implemented in these scripts yet.

## Task networks (for metric evaluation)

```text
checkpoints/task_networks/
β”œβ”€β”€ detection/model.pth
β”œβ”€β”€ instance/model.pth
β”œβ”€β”€ semantic/model.pth
β”œβ”€β”€ panoptic/model.pth
└── pose/model.pth
```

These are **official pretrained task networks** (not codec weights).
Required when running:

```bash
python scripts/eval_taic.py -c configs/eval/taic_detection.yaml --with-metrics
```

See `configs/task_networks/README.md` for config/checkpoint pairing.