Spaces:
Running
Running
File size: 4,738 Bytes
20e29fa |
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 |
import torch
import torch.nn as nn
import torchvision.models as models
from modelscope.msdatasets import MsDataset
from utils import MODEL_DIR
def get_backbone(ver, backbone_list):
for bb in backbone_list:
if ver == bb["ver"]:
return bb
print("Backbone name not found, using default option - alexnet.")
return backbone_list[0]
def model_info(m_ver):
backbone_list = MsDataset.load(
"monetjoe/cv_backbones", split="train"
)
backbone = get_backbone(m_ver, backbone_list)
m_type = str(backbone["type"])
input_size = int(backbone["input_size"])
return m_type, input_size
def Classifier(cls_num: int, output_size: int, linear_output: bool):
q = (1.0 * output_size / cls_num) ** 0.25
l1 = int(q * cls_num)
l2 = int(q * l1)
l3 = int(q * l2)
if linear_output:
return torch.nn.Sequential(
nn.Dropout(),
nn.Linear(output_size, l3),
nn.ReLU(inplace=True),
nn.Dropout(),
nn.Linear(l3, l2),
nn.ReLU(inplace=True),
nn.Dropout(),
nn.Linear(l2, l1),
nn.ReLU(inplace=True),
nn.Linear(l1, cls_num),
)
else:
return torch.nn.Sequential(
nn.Dropout(),
nn.Conv2d(output_size, l3, kernel_size=(1, 1), stride=(1, 1)),
nn.ReLU(inplace=True),
nn.AdaptiveAvgPool2d(output_size=(1, 1)),
nn.Flatten(),
nn.Linear(l3, l2),
nn.ReLU(inplace=True),
nn.Dropout(),
nn.Linear(l2, l1),
nn.ReLU(inplace=True),
nn.Linear(l1, cls_num),
)
class EvalNet:
model = None
m_type = "squeezenet"
input_size = 224
output_size = 512
def __init__(self, log_name: str, cls_num=4):
saved_model_path = f"{MODEL_DIR}/{log_name}/save.pt"
m_ver = "_".join(log_name.split("_")[:-1])
self.m_type, self.input_size = model_info(m_ver)
if not hasattr(models, m_ver):
print("Unsupported model.")
exit()
self.model = eval("models.%s()" % m_ver)
linear_output = self._set_outsize()
self._set_classifier(cls_num, linear_output)
checkpoint = torch.load(saved_model_path, map_location="cpu")
if torch.cuda.is_available():
checkpoint = torch.load(saved_model_path)
self.model.load_state_dict(checkpoint, False)
self.model.eval()
def _set_outsize(self, debug_mode=False):
for name, module in self.model.named_modules():
if (
str(name).__contains__("classifier")
or str(name).__eq__("fc")
or str(name).__contains__("head")
):
if isinstance(module, torch.nn.Linear):
self.output_size = module.in_features
if debug_mode:
print(
f"{name}(Linear): {self.output_size} -> {module.out_features}"
)
return True
if isinstance(module, torch.nn.Conv2d):
self.output_size = module.in_channels
if debug_mode:
print(
f"{name}(Conv2d): {self.output_size} -> {module.out_channels}"
)
return False
return False
def _set_classifier(self, cls_num, linear_output):
if self.m_type == "convnext":
del self.model.classifier[2]
self.model.classifier = nn.Sequential(
*list(self.model.classifier)
+ list(Classifier(cls_num, self.output_size, linear_output))
)
return
if hasattr(self.model, "classifier"):
self.model.classifier = Classifier(cls_num, self.output_size, linear_output)
return
elif hasattr(self.model, "fc"):
self.model.fc = Classifier(cls_num, self.output_size, linear_output)
return
elif hasattr(self.model, "head"):
self.model.head = Classifier(cls_num, self.output_size, linear_output)
return
self.model.heads.head = Classifier(cls_num, self.output_size, linear_output)
def forward(self, x):
if torch.cuda.is_available():
x = x.cuda()
self.model = self.model.cuda()
if self.m_type == "googlenet" and self.training:
return self.model(x)[0]
else:
return self.model(x)
|