EdgeTA / methods /feat_align /main_partial.py
LINC-BIT's picture
Upload 1912 files
b84549f verified
from typing import Any, Dict, List
from schema import Schema
from data import Scenario, MergedDataset
from methods.base.alg import BaseAlg
from methods.base.model import BaseModel
from data import build_dataloader
import torch.optim
import tqdm
import os
import time
from abc import abstractmethod
import matplotlib.pyplot as plt
from copy import deepcopy
from torch import nn
import torch.optim
def tent_as_detector(online_model, x, num_iters=1, lr=1e-4, l1_wd=0., strategy='ours'):
model = online_model.models_dict['main']
before_model = deepcopy(model)
# from methods.tent import tent
optimizer = torch.optim.SGD(
model.parameters(), lr=lr, weight_decay=l1_wd)
# from .tent import configure_model, forward_and_adapt
# configure_model(model)
output = online_model.infer(x)
entropy = online_model.get_output_entropy(output).mean()
entropy.backward()
# for _ in range(num_iters):
# forward_and_adapt(x, model, optimizer)
# entropy_loss = model.
filters_sen_info = {}
last_conv_name = None
for (name, m1), m2 in zip(model.named_modules(), before_model.modules()):
if isinstance(m1, nn.Linear):
last_conv_name = name
if not isinstance(m1, nn.LayerNorm):
continue
with torch.no_grad():
features_weight_diff = ((m1.weight.data - m2.weight.data).abs())
features_bias_diff = ((m1.bias.data - m2.bias.data).abs())
features_diff = features_weight_diff + features_bias_diff
features_diff_order = features_diff.argsort(descending=False)
if strategy == 'ours':
untrained_filters_index = features_diff_order[: int(len(features_diff) * 0.8)]
elif strategy == 'random':
untrained_filters_index = torch.randperm(len(features_diff))[: int(len(features_diff) * 0.8)]
elif strategy == 'inversed_ours':
untrained_filters_index = features_diff_order.flip(0)[: int(len(features_diff) * 0.8)]
elif strategy == 'none':
untrained_filters_index = None
filters_sen_info[name] = dict(untrained_filters_index=untrained_filters_index, conv_name=last_conv_name)
return filters_sen_info
class SGDF(torch.optim.SGD):
@torch.no_grad()
def step(self, p_names, conv_filters_sen_info, filters_sen_info, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
for group in self.param_groups:
weight_decay = group['weight_decay']
momentum = group['momentum']
dampening = group['dampening']
nesterov = group['nesterov']
# assert len([i for i in model.named_parameters()]) == len([j for j in group['params']])
for name, p in zip(p_names, group['params']):
if p.grad is None:
continue
layer_name = '.'.join(name.split('.')[0:-1])
if layer_name in filters_sen_info.keys():
untrained_filters_index = filters_sen_info[layer_name]['untrained_filters_index']
elif layer_name in conv_filters_sen_info.keys():
untrained_filters_index = conv_filters_sen_info[layer_name]['untrained_filters_index']
else:
untrained_filters_index = []
d_p = p.grad
if weight_decay != 0:
d_p = d_p.add(p, alpha=weight_decay)
if momentum != 0:
param_state = self.state[p]
if 'momentum_buffer' not in param_state:
buf = param_state['momentum_buffer'] = torch.clone(d_p).detach()
else:
buf = param_state['momentum_buffer']
buf.mul_(momentum).add_(d_p, alpha=1 - dampening)
if nesterov:
d_p = d_p.add(buf, alpha=momentum)
else:
d_p = buf
try:
d_p[untrained_filters_index] = 0.
p.add_(d_p, alpha=-group['lr'])
except Exception as e:
print('SGDF error', name)
return loss
class OnlineFeatAlignModel(BaseModel):
def get_required_model_components(self) -> List[str]:
return ['main']
@abstractmethod
def get_feature_hook(self):
pass
@abstractmethod
def forward_to_get_task_loss(self, x, y):
pass
@abstractmethod
def get_trained_params(self):
pass
@abstractmethod
def get_mmd_loss(self, f1, f2):
pass
@abstractmethod
def get_output_entropy(self, output):
pass
class FeatAlignAlg(BaseAlg):
def get_required_models_schema(self) -> Schema:
return Schema({
'main': OnlineFeatAlignModel
})
def get_required_hyp_schema(self) -> Schema:
return Schema({
'train_batch_size': int,
'val_batch_size': int,
'num_workers': int,
'optimizer': str,
'optimizer_args': dict,
'scheduler': str,
'scheduler_args': dict,
'num_iters': int,
'val_freq': int,
'feat_align_loss_weight': float,
'trained_neuron_selection_strategy': str
})
def run(self, scenario: Scenario, hyps: Dict) -> Dict[str, Any]:
super().run(scenario, hyps)
assert isinstance(self.models['main'], OnlineFeatAlignModel) # for auto completion
cur_domain_name = scenario.target_domains_order[scenario.cur_domain_index]
datasets_for_training = scenario.get_online_cur_domain_datasets_for_training()
train_dataset = datasets_for_training[cur_domain_name]['train']
val_dataset = datasets_for_training[cur_domain_name]['val']
datasets_for_inference = scenario.get_online_cur_domain_datasets_for_inference()
test_dataset = datasets_for_inference
train_loader = iter(build_dataloader(train_dataset, hyps['train_batch_size'], hyps['num_workers'],
True, None))
test_loader = build_dataloader(test_dataset, hyps['val_batch_size'], hyps['num_workers'],
False, False)
source_datasets = [d['train'] for n, d in datasets_for_training.items() if n != cur_domain_name]
source_dataset = MergedDataset(source_datasets)
source_train_loader = iter(build_dataloader(source_dataset, hyps['train_batch_size'], hyps['num_workers'],
True, None))
# 1. generate surrogate DNN
# for n, m in self.models['main'].models_dict['md'].named_modules():
# if isinstance(m, nn.Linear):
# m.reset_parameters()
# from utils.dl.common.model import set_module
# for n, m in self.models['main'].models_dict['md'].named_modules():
# if m.__class__.__name__ == 'KTakesAll':
# set_module(self.models['main'].models_dict['md'], n, KTakesAll(0.5))
# self.models['main'].set_sd_sparsity(hyps['sd_sparsity'])
device = self.models['main'].device
# surrogate_dnn = self.models['main'].generate_sd_by_target_samples(next(train_loader)[0].to(device))
# self.models['sd'] = surrogate_dnn
# 2. train surrogate DNN
# TODO: train only a part of filters
trained_params, p_name = self.models['main'].get_trained_params()
# optimizer = torch.optim.__dict__[hyps['optimizer']](trained_params, **hyps['optimizer_args'])
optimizer = SGDF(trained_params, **hyps['optimizer_args'])
if hyps['scheduler'] != '':
scheduler = torch.optim.lr_scheduler.__dict__[hyps['scheduler']](optimizer, **hyps['scheduler_args'])
else:
scheduler = None
pbar = tqdm.tqdm(range(hyps['num_iters']), dynamic_ncols=True, desc='da...')
task_losses, mmd_losses = [], []
accs = []
x, _ = next(train_loader)
if isinstance(x, dict):
for k, v in x.items():
if isinstance(v, torch.Tensor):
x[k] = v.to(device)
else:
x = x.to(device)
filters_sen_info = tent_as_detector(self.models['main'], x, strategy=hyps['trained_neuron_selection_strategy'])
conv_filters_sen_info = {v['conv_name']: v for _, v in filters_sen_info.items()}
total_train_time = 0.
feature_hook = self.models['main'].get_feature_hook()
for iter_index in pbar:
if iter_index % hyps['val_freq'] == 0:
from data import split_dataset
cur_test_batch_dataset = split_dataset(test_dataset, hyps['val_batch_size'], iter_index)[0]
cur_test_batch_dataloader = build_dataloader(cur_test_batch_dataset, hyps['train_batch_size'], hyps['num_workers'], False, False)
cur_acc = self.models['main'].get_accuracy(cur_test_batch_dataloader)
accs += [{
'iter': iter_index,
'acc': cur_acc
}]
cur_start_time = time.time()
self.models['main'].to_train_mode()
x, _ = next(train_loader)
if isinstance(x, dict):
for k, v in x.items():
if isinstance(v, torch.Tensor):
x[k] = v.to(device)
else:
x = x.to(device)
source_x, source_y = next(source_train_loader)
if isinstance(source_x, dict):
for k, v in source_x.items():
if isinstance(v, torch.Tensor):
source_x[k] = v.to(device)
source_y = source_y.to(device)
else:
source_x, source_y = source_x.to(device), source_y.to(device)
task_loss = self.models['main'].forward_to_get_task_loss(source_x, source_y)
source_features = feature_hook.input
self.models['main'].infer(x)
target_features = feature_hook.input
mmd_loss = hyps['feat_align_loss_weight'] * self.models['main'].get_mmd_loss(source_features, target_features)
loss = task_loss + mmd_loss
optimizer.zero_grad()
loss.backward()
# optimizer.step()
optimizer.step(p_name, conv_filters_sen_info, filters_sen_info)
if scheduler is not None:
scheduler.step()
pbar.set_description(f'da... | cur_acc: {cur_acc:.4f}, task_loss: {task_loss:.6f}, mmd_loss: {mmd_loss:.6f}')
task_losses += [float(task_loss.cpu().item())]
mmd_losses += [float(mmd_loss.cpu().item())]
total_train_time += time.time() - cur_start_time
feature_hook.remove()
time_usage = total_train_time
plt.plot(task_losses, label='task')
plt.plot(mmd_losses, label='mmd')
plt.xlabel('iteration')
plt.ylabel('loss')
plt.savefig(os.path.join(self.res_save_dir, 'loss.png'))
plt.clf()
cur_test_batch_dataset = split_dataset(test_dataset, hyps['train_batch_size'], iter_index + 1)[0]
cur_test_batch_dataloader = build_dataloader(cur_test_batch_dataset, len(cur_test_batch_dataset), hyps['num_workers'], False, False)
cur_acc = self.models['main'].get_accuracy(cur_test_batch_dataloader)
accs += [{
'iter': iter_index + 1,
'acc': cur_acc
}]
return {
'accs': accs,
'time': time_usage
}, self.models