Spaces:
Sleeping
Sleeping
File size: 4,922 Bytes
a9d56ef | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 | import torch
from torch import nn
from torchvision.models import (
efficientnet_v2_s,
mobilenet_v3_large,
resnet101,
swin_v2_b,
)
import math
NUM_GRADUAL_UNFREEZING_STAGES = 5
SEED = 123
ACT_FUNCS = {
"relu": nn.ReLU,
"tanh": nn.Tanh, # Tanh is not used
}
def classification_head(in_features: int, config: dict, flatten=False) -> nn.Sequential:
torch.manual_seed(SEED)
first_linear = nn.Linear(in_features, config["units"], bias=False)
nn.init.kaiming_uniform_(first_linear.weight, nonlinearity=config["activation"])
head = nn.Sequential(
first_linear,
nn.LayerNorm(config["units"]),
ACT_FUNCS[config["activation"]](),
nn.Dropout(config["dropout"]),
nn.Linear(config["units"], 7),
)
if flatten:
head.insert(0, nn.Flatten())
return head
class PretrainedModel(nn.Module):
def __init__(self, config):
super().__init__()
self.unfreezing_stage = 0
# The layers in forwarding order
self.layers_to_unfreeze: list[nn.Module] = []
self.model_type: str = config["model_type"]
self.grad_cam_layer: list[nn.Module] = []
def set_head_trainable(self):
"""
Requires overriding if the classification head is not called
"model.classifier"
"""
self.model.classifier.requires_grad_(True)
def inc_grad_unfreezing(self):
"""
Increments the gradual unfreezing process by unfreezing
the next 100% / NUM_GRADUAL_UNFREEZING_STAGES layers
"""
if self.unfreezing_stage <= NUM_GRADUAL_UNFREEZING_STAGES:
self.unfreezing_stage += 1
self.set_unfreezing_stage(self.unfreezing_stage)
def set_unfreezing_stage(self, unfreezing_stage: int):
self.unfreezing_stage = unfreezing_stage
if self.unfreezing_stage > NUM_GRADUAL_UNFREEZING_STAGES:
self.unfreezing_stage = NUM_GRADUAL_UNFREEZING_STAGES
self.requires_grad_(True)
return
else:
# Make sure all layers are untrainable before
# setting the trainable layers to be trainable
self.requires_grad_(False)
layer_index = math.ceil(
self.unfreezing_stage
* len(self.layers_to_unfreeze)
/ NUM_GRADUAL_UNFREEZING_STAGES
)
for module in self.layers_to_unfreeze[-layer_index:]:
module.requires_grad_(True)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.model(x)
class EfficientNet(PretrainedModel):
def __init__(self, config: dict):
super().__init__(config)
self.model = efficientnet_v2_s()
in_features = self.model.classifier[1].in_features
self.model.classifier = classification_head(in_features, config)
self.layers_to_unfreeze = [
self.model.features[i] for i in range(len(self.model.features))
]
self.grad_cam_layer = [self.model.features[-1][-1]]
class MobileNet(PretrainedModel):
"""
MobileNet V3 or V4, customized for our transfer learning
V4 paper:
https://arxiv.org/abs/2404.10518
"""
def __init__(self, config: dict, version: str = "v3"):
super().__init__(config)
# MBNetV4 is in a MBNetV3 object for some reason
if version == "v3":
self.model = mobilenet_v3_large()
in_features = self.model.classifier[0].in_features
self.layers_to_unfreeze = [
self.model.features[i] for i in range(len(self.model.features))
]
self.grad_cam_layer = [self.model.features[-1][-1]]
else:
raise NotImplementedError()
self.model.classifier = classification_head(in_features, config)
class ResNet(PretrainedModel):
def __init__(self, config: dict):
super().__init__(config)
self.model = resnet101()
in_features = self.model.fc.in_features
self.model.fc = classification_head(in_features, config)
self.layers_to_unfreeze = [
self.model.conv1,
self.model.bn1,
self.model.layer1,
self.model.layer2,
self.model.layer3,
self.model.layer4,
]
self.grad_cam_layer = [self.model.layer4[-1]]
def set_head_trainable(self):
self.model.fc.requires_grad_(True)
class Swin(PretrainedModel):
def __init__(self, config: dict):
super().__init__(config)
self.model = swin_v2_b()
in_features = self.model.head.in_features
self.model.head = classification_head(in_features, config)
self.layers_to_unfreeze = [
self.model.features[i] for i in range(len(self.model.features))
] + [self.model.norm]
self.grad_cam_layer = [self.model.permute]
def set_head_trainable(self):
self.model.head.requires_grad_(True)
|