File size: 11,524 Bytes
02ba886 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 |
import logging
import torch
import torch.nn as nn
from torchvision import transforms
from copy import deepcopy
from functools import wraps
logger = logging.getLogger(__name__)
class TTAMethod(nn.Module):
def __init__(self, cfg, model, num_classes):
super().__init__()
self.cfg = cfg
self.device = "cuda" if torch.cuda.is_available() else "cpu"
self.model = self.decorate_model(model)
self.num_classes = num_classes
self.episodic = cfg.MODEL.EPISODIC
self.dataset_name = cfg.CORRUPTION.DATASET
self.steps = cfg.OPTIM.STEPS
self.current_grad_norm = 0.0
assert self.steps > 0, "requires >= 1 step(s) to forward and update"
# variables for resetting the model after a certain amount of performed update steps
self.performed_updates = 0
self.reset_after_num_updates = cfg.MODEL.RESET_AFTER_NUM_UPDATES
# restore the image input size from the model pre-processing if it is defined
# this is required for methods relying on test-time augmentation
if "cifar" in self.dataset_name:
self.img_size = (32, 32)
if "imagenet" in self.dataset_name or "ccc" in self.dataset_name:
self.img_size = (224, 224)
if hasattr(self.model, "model_preprocess") and isinstance(self.model.model_preprocess, transforms.Compose):
for transf in self.model.model_preprocess.transforms[::-1]:
if hasattr(transf, "size"):
self.img_size = getattr(transf, "size")
if self.dataset_name in ["imagenet_c", "ccc"] and max(self.img_size) > 224:
raise ValueError(f"The specified model with pre-processing {model.model_preprocess} "
f"is not suited in combination with ImageNet-C and CCC! "
f"These datasets are already resized and center cropped to 224")
break
# configure model and optimizer
self.configure_model()
self.params, self.param_names = self.collect_params()
self.optimizer = self.setup_optimizer() if len(self.params) > 0 or len(self.param_names) > 0 else None
self.num_trainable_params, self.num_total_params = self.get_number_trainable_params()
# variables needed for single sample test-time adaptation (sstta) using a sliding window (buffer) approach
self.input_buffer = None
self.window_length = cfg.TEST.WINDOW_LENGTH
self.pointer = torch.tensor([0], dtype=torch.long).to(self.device)
# sstta: if the model has no batchnorm layers, we do not need to forward the whole buffer when not performing any updates
self.has_bn = any([isinstance(m, (nn.BatchNorm1d, nn.BatchNorm2d)) for m in model.modules()])
# note: if the self.model is never reset, like for continual adaptation,
# then skipping the state copy would save memory
self.models = [self.model]
self.model_states, self.optimizer_state = self.copy_model_and_optimizer()
# setup for mixed-precision or single precision
self.mixed_precision = cfg.MIXED_PRECISION
self.scaler = torch.cuda.amp.GradScaler() if cfg.MIXED_PRECISION else None
def decorate_model(self, model):
return model
def forward(self, x):
if self.episodic:
self.reset()
x = x if isinstance(x, list) else [x]
if x[0].shape[0] == 1: # single sample test-time adaptation
# create the sliding window input
if self.input_buffer is None:
self.input_buffer = [x_item for x_item in x]
# set bn1d layers into eval mode, since no statistics can be extracted from 1 sample
self.change_mode_of_batchnorm1d(self.models, to_train_mode=False)
elif self.input_buffer[0].shape[0] < self.window_length:
self.input_buffer = [torch.cat([self.input_buffer[i], x_item], dim=0) for i, x_item in enumerate(x)]
# set bn1d layers into train mode
self.change_mode_of_batchnorm1d(self.models, to_train_mode=True)
else:
for i, x_item in enumerate(x):
self.input_buffer[i][self.pointer] = x_item
if self.pointer == (self.window_length - 1):
# update the model, since the complete buffer has changed
for _ in range(self.steps):
outputs = self.forward_and_adapt(self.input_buffer)
# if specified, reset the model after a certain amount of update steps
self.performed_updates += 1
if self.reset_after_num_updates > 0 and self.performed_updates % self.reset_after_num_updates == 0:
self.reset()
outputs = outputs[self.pointer.long()]
else:
# create the prediction without updating the model
if self.has_bn:
# forward the whole buffer to get good batchnorm statistics
outputs = self.forward_sliding_window(self.input_buffer)
outputs = outputs[self.pointer.long()]
else:
# only forward the current test sample, since there are no batchnorm layers
outputs = self.forward_sliding_window(x)
# increase the pointer
self.pointer += 1
self.pointer %= self.window_length
else: # common batch adaptation setting
for _ in range(self.steps):
outputs = self.forward_and_adapt(x)
# if specified, reset the model after a certain amount of update steps
self.performed_updates += 1
if self.reset_after_num_updates > 0 and self.performed_updates % self.reset_after_num_updates == 0:
logger.info(f"Reset the model after {self.reset_after_num_updates} updates")
self.reset()
return outputs
def loss_calculation(self, x):
"""
Loss calculation.
"""
raise NotImplementedError
def forward_and_adapt(self, x):
"""
Forward and adapt the model on a batch of data.
"""
raise NotImplementedError
@torch.no_grad()
def forward_sliding_window(self, x):
"""
Create the prediction for single sample test-time adaptation with a sliding window
:param x: The buffered data created with a sliding window
:return: Model predictions
"""
imgs_test = x[0]
return self.model(imgs_test)
def configure_model(self):
raise NotImplementedError
def collect_params(self):
"""Collect all trainable parameters.
Walk the model's modules and collect all parameters.
Return the parameters and their names.
Note: other choices of parameterization are possible!
"""
params = []
names = []
for nm, m in self.model.named_modules():
for np, p in m.named_parameters():
if np in ['weight', 'bias', 'prompts'] and p.requires_grad:
params.append(p)
names.append(f"{nm}.{np}")
return params, names
def setup_optimizer(self):
if self.cfg.OPTIM.METHOD == 'Adam':
return torch.optim.Adam(self.params,
lr=self.cfg.OPTIM.LR,
betas=(self.cfg.OPTIM.BETA, 0.999),
weight_decay=self.cfg.OPTIM.WD)
elif self.cfg.OPTIM.METHOD == 'AdamW':
return torch.optim.AdamW(self.params,
lr=self.cfg.OPTIM.LR,
betas=(self.cfg.OPTIM.BETA, 0.999),
weight_decay=self.cfg.OPTIM.WD)
elif self.cfg.OPTIM.METHOD == 'SGD':
return torch.optim.SGD(self.params,
lr=self.cfg.OPTIM.LR,
momentum=self.cfg.OPTIM.MOMENTUM,
dampening=self.cfg.OPTIM.DAMPENING,
weight_decay=self.cfg.OPTIM.WD,
nesterov=self.cfg.OPTIM.NESTEROV)
else:
raise NotImplementedError
def get_number_trainable_params(self):
if isinstance(self.params, list):
trainable = sum(p.numel() for p in self.params) if len(self.params) > 0 else 0
elif isinstance(self.params, dict):
trainable = []
for _, param in self.params.items():
if len(param) > 0:
trainable.append(sum(p.numel() for p in param))
trainable = sum(trainable)
total = sum(p.numel() for p in self.model.parameters())
logger.info(f"#Trainable/total parameters: {trainable:,}/{total:,} \t Ratio: {trainable / total * 100:.3f}% ")
return trainable, total
def reset(self):
"""Reset the model and optimizer state to the initial source state"""
if self.model_states is None or self.optimizer_state is None:
raise Exception("cannot reset without saved model/optimizer state")
self.load_model_and_optimizer()
def copy_model_and_optimizer(self):
"""Copy the model and optimizer states for resetting after adaptation."""
model_states = [deepcopy(model.state_dict()) for model in self.models]
optimizer_state = deepcopy(self.optimizer.state_dict())
return model_states, optimizer_state
def load_model_and_optimizer(self):
"""Restore the model and optimizer states from copies."""
for model, model_state in zip(self.models, self.model_states):
model.load_state_dict(model_state, strict=True)
self.optimizer.load_state_dict(self.optimizer_state)
def save_model(self, save_path, r, errs_5, errs, current_domain_step, current_global_step, accuracy_buffer):
pass
def load_model(self, ckpt):
raise NotImplementedError
def average_grad_norm(self,):
raise NotImplementedError
@staticmethod
def copy_model(model):
is_parallel = isinstance(model, nn.DataParallel)
if is_parallel:
model = model.module
coppied_model = deepcopy(model)
for param in coppied_model.parameters():
param.detach_()
if is_parallel:
model = nn.DataParallel(model)
coppied_model = nn.DataParallel(coppied_model)
return coppied_model
@staticmethod
def change_mode_of_batchnorm1d(model_list, to_train_mode=True):
# batchnorm1d layers do not work with single sample inputs
for model in model_list:
for m in model.modules():
if isinstance(m, nn.BatchNorm1d):
if to_train_mode:
m.train()
else:
m.eval()
def forward_decorator(fn):
@wraps(fn)
def decorator(self, *args, **kwargs):
if self.mixed_precision:
with torch.cuda.amp.autocast():
outputs = fn(self, *args, **kwargs)
else:
outputs = fn(self, *args, **kwargs)
return outputs
return decorator
|