Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,125 Bytes
a7dedf9 c628976 a7dedf9 a5dc50a 227d44e a7dedf9 c628976 a7dedf9 |
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 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 |
from torch import nn, Tensor
import open_clip
from ..utils import ConvRefine
from ..utils import ConvUpsample, _get_norm_layer, _get_activation
convnext_names_and_weights = {
"convnext_base": ["laion400m_s13b_b51k"], # 107.49M
"convnext_base_w": ["laion2b_s13b_b82k", "laion2b_s13b_b82k_augreg", "laion_aesthetic_s13b_b82k"], # 107.75M
"convnext_base_w_320": ["laion_aesthetic_s13b_b82k", "laion_aesthetic_s13b_b82k_augreg"], # 107.75M
"convnext_large_d": ["laion2b_s26b_b102k_augreg"], # 217.46M
"convnext_large_d_320": ["laion2b_s29b_b131k_ft", "laion2b_s29b_b131k_ft_soup"], # 217.46M
"convnext_xxlarge": ["laion2b_s34b_b82k_augreg", "laion2b_s34b_b82k_augreg_rewind", "laion2b_s34b_b82k_augreg_soup"] # 896.88M
}
refiner_channels = {
"convnext_base": 1024,
"convnext_base_w": 1024,
"convnext_base_w_320": 1024,
"convnext_large_d": 1536,
"convnext_large_d_320": 1536,
"convnext_xxlarge": 3072,
}
refiner_groups = {
"convnext_base": 1,
"convnext_base_w": 1,
"convnext_base_w_320": 1,
"convnext_large_d": refiner_channels["convnext_large_d"] // 512, # 3
"convnext_large_d_320": refiner_channels["convnext_large_d_320"] // 512, # 3
"convnext_xxlarge": refiner_channels["convnext_xxlarge"] // 512, # 6
}
class ConvNeXt(nn.Module):
def __init__(
self,
model_name: str,
weight_name: str,
block_size: int = 16,
norm: str = "none",
act: str = "none"
) -> None:
super(ConvNeXt, self).__init__()
assert model_name in convnext_names_and_weights, f"Model name should be one of {list(convnext_names_and_weights.keys())}, but got {model_name}."
assert weight_name in convnext_names_and_weights[model_name], f"Pretrained should be one of {convnext_names_and_weights[model_name]}, but got {weight_name}."
assert block_size in [32, 16, 8], f"block_size should be one of [32, 16, 8], got {block_size}"
self.model_name, self.weight_name = model_name, weight_name
self.block_size = block_size
# model = open_clip.create_model_from_pretrained(model_name, weight_name, return_transform=False).visual
model = open_clip.create_model(model_name=model_name, pretrained=False, load_weights=False).visual
self.stem = model.trunk.stem
self.depth = len(model.trunk.stages)
for idx, stage in enumerate(model.trunk.stages):
setattr(self, f"stage{idx}", stage)
if self.model_name in ["convnext_base", "convnext_base_w", "convnext_base_w_320", "convnext_xxlarge"]:
self.in_features, self.out_features = model.head.proj.in_features, model.head.proj.out_features
else: # "convnext_large_d", "convnext_large_d_320":
self.in_features, self.out_features = model.head.mlp.fc1.in_features, model.head.mlp.fc2.out_features
if norm == "bn":
norm_layer = nn.BatchNorm2d
elif norm == "ln":
norm_layer = nn.LayerNorm
else:
norm_layer = _get_norm_layer(model)
if act == "relu":
activation = nn.ReLU(inplace=True)
elif act == "gelu":
activation = nn.GELU()
else:
activation = _get_activation(model)
if block_size == 32:
self.refiner = ConvRefine(
in_channels=self.in_features,
out_channels=self.in_features,
norm_layer=norm_layer,
activation=activation,
groups=refiner_groups[self.model_name],
)
elif block_size == 16:
self.refiner = ConvUpsample(
in_channels=self.in_features,
out_channels=self.in_features,
norm_layer=norm_layer,
activation=activation,
groups=refiner_groups[self.model_name],
)
else: # block_size == 8
self.refiner = nn.Sequential(
ConvUpsample(
in_channels=self.in_features,
out_channels=self.in_features,
norm_layer=norm_layer,
activation=activation,
groups=refiner_groups[self.model_name],
),
ConvUpsample(
in_channels=self.in_features,
out_channels=self.in_features,
norm_layer=norm_layer,
activation=activation,
groups=refiner_groups[self.model_name],
),
)
def forward(self, x: Tensor) -> Tensor:
x = self.stem(x)
for idx in range(self.depth):
x = getattr(self, f"stage{idx}")(x)
x = self.refiner(x)
return x
def _convnext(
model_name: str,
weight_name: str,
block_size: int = 16,
norm: str = "none",
act: str = "none"
) -> ConvNeXt:
model = ConvNeXt(
model_name=model_name,
weight_name=weight_name,
block_size=block_size,
norm=norm,
act=act
)
return model |