Spaces:
Running
on
Zero
Running
on
Zero
File size: 1,595 Bytes
246c106 |
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 |
"""
Not in original Open-MAGVIT2.
"""
import json
from dataclasses import dataclass
@dataclass
class VQConfig:
# Model Arch
in_channels: int = 3
z_channels: int = 18
out_channels: int = 3
base_channels: int = 128 # Initial hidden width, all subsequent blocks are multiples (`ch_mult`) of this width
ch_mult: tuple[int] = (1, 1, 2, 2, 4)
num_res_blocks: int = 2
# Loss Config (uncertain about some of the types)
# Hardcoding for `magvit2.modules.losses.vqperceptual.VQLPIPSWithDiscriminator`
disc_conditional: bool = False
disc_in_channels: int = 3
disc_start: int = 0 # from 0 epoch
disc_loss: str = "hinge"
disc_ndf: int = 64
disc_num_layers: int = 3
use_actnorm: bool = False
disc_factor: float = 1.0
disc_weight: float = 0.8
gen_loss_weight: float = 0.1
lecam_loss_weight: float = 0.005
codebook_weight: float = 0.1
commit_weight: float = 0.25
pixelloss_weight: float = 1.0
perceptual_weight: float = 1.0
codebook_enlarge_ratio: float = 0
codebook_enlarge_steps: int = 2000
num_codebooks: int = 1
codebook_size: int = 262144
sample_minimization_weight: float = 1.0
batch_maximization_weight: float = 1.0
token_factorization: bool = False
# TODO: duplicated from GenieConfig
def save_pretrained(self, json_path):
with open(json_path, "w") as f:
json.dump(vars(self), f)
@classmethod
def from_pretrained(cls, json_path):
with open(json_path, "r") as f:
config = json.load(f)
return cls(**config)
|