Spaces:
Running
on
Zero
Running
on
Zero
File size: 4,900 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 |
from torch import nn, Tensor
import open_clip
from ..utils import ConvRefine, ConvUpsample
from ..utils import _get_norm_layer, _get_activation
resnet_names_and_weights = {
"RN50": ["openai", "yfcc15m", "cc12m"],
"RN101": ["openai", "yfcc15m", "cc12m"],
"RN50x4": ["openai", "yfcc15m", "cc12m"],
"RN50x16": ["openai", "yfcc15m", "cc12m"],
"RN50x64": ["openai", "yfcc15m", "cc12m"],
}
refiner_channels = {
"RN50": 2048,
"RN101": 2048,
"RN50x4": 2560,
"RN50x16": 3072,
"RN50x64": 4096,
}
refiner_groups = {
"RN50": refiner_channels["RN50"] // 512, # 4
"RN101": refiner_channels["RN101"] // 512, # 4
"RN50x4": refiner_channels["RN50x4"] // 512, # 5
"RN50x16": refiner_channels["RN50x16"] // 512, # 6
"RN50x64": refiner_channels["RN50x64"] // 512, # 8
}
class ResNet(nn.Module):
def __init__(
self,
model_name: str,
weight_name: str,
block_size: int = 16,
norm: str = "none",
act: str = "none"
) -> None:
super(ResNet, self).__init__()
assert model_name in resnet_names_and_weights, f"Model name should be one of {list(resnet_names_and_weights.keys())}, but got {model_name}."
assert weight_name in resnet_names_and_weights[model_name], f"Pretrained should be one of {resnet_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
# Stem
self.conv1 = model.conv1
self.bn1 = model.bn1
self.act1 = model.act1
self.conv2 = model.conv2
self.bn2 = model.bn2
self.act2 = model.act2
self.conv3 = model.conv3
self.bn3 = model.bn3
self.act3 = model.act3
self.avgpool = model.avgpool
# Stem: reduction = 4
# Layers
for idx in range(1, 5):
setattr(self, f"layer{idx}", getattr(model, f"layer{idx}"))
self.in_features = model.attnpool.c_proj.weight.shape[1]
self.out_features = model.attnpool.c_proj.weight.shape[0]
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 stem(self, x: Tensor) -> Tensor:
x = self.act1(self.bn1(self.conv1(x)))
x = self.act2(self.bn2(self.conv2(x)))
x = self.act3(self.bn3(self.conv3(x)))
x = self.avgpool(x)
return x
def forward(self, x: Tensor) -> Tensor:
x = self.stem(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.refiner(x)
return x
def _resnet(
model_name: str,
weight_name: str,
block_size: int = 16,
norm: str = "none",
act: str = "none"
) -> ResNet:
model = ResNet(
model_name=model_name,
weight_name=weight_name,
block_size=block_size,
norm=norm,
act=act
)
return model
|