File size: 1,377 Bytes
4102df7
 
 
 
 
 
 
 
 
 
 
eefa55d
90c3c77
eefa55d
669cb7f
 
 
 
 
 
 
 
 
 
 
 
eefa55d
4102df7
1dd2363
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4102df7
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
---
license: gpl-3.0
datasets:
- stanford-vision-lab/gpic
language:
- en
pipeline_tag: image-text-to-image
tags:
- VAE
- mmDiT
---
# Weights for mmDiT and VAE from [Diffusion Transformer](https://github.com/someonecantcode/diffusion-transformer).

## Weights structure:
```python
flow_match_checkpoint = {
        'model': raw_model.state_dict(),
        'optimizer': optimizer.state_dict(),
        'ema': ema.state_dict,
        'step': step
}

vae_checkpoint = vae.state_dict(),
```


## Hyperparameters used for model.
```python
@dataclass
class TrainConfig:
    lr: float = 1e-4
    batch_size: int = 8
    total_steps: int = 10000 # 50k steps is ideal.
    lr_cooldown_steps: int = 20000
    starting_step: int = 0
    logging_interval: int = 100
    grad_accum_steps: int = 2 # not really needed
    save_dir: str = "./saved_models/"
    ckpt: str = "flowmatch-ckpt.pt"

@dataclass
class VAEConfig:
    kl_beta: float = 5e-2
    fft_weight: float = 0.05
    lpips_weight: float = 0.5
    latent_channels: tuple[int, ...] = (128, 256, 512, 512)
    z_channels: int = 32
    n_heads: int = 8
    resnet_blocks: int  = 2

@dataclass
class mmDiTConfig:
    d_model: int = 768
    n_heads: int = 32
    n_blocks: int = 16
    patch_size: int = 2
    color_channels: int = 32 # z_channels
    height: int = 32 # latent
    width: int = 32 # latent
    t_channels: int = 64
```