我正在增量学习中体验知识蒸馏。基本上,在每个阶段我都会初始化一个新模型并将其作为学生模型在当前数据上进行训练,并使用旧模型(在前一阶段训练过的)作为老师。到目前为止,代码中没有错误,但问题是在每个阶段之后损失永远不会减少。另外,PL 中有没有办法在 on_train_epoch_start 中初始化新的优化器?
def on_train_epoch_start(self) :
if self.new_phase:
self.old_backbone = copy.deepcopy(self.backbone)
self.old_head = copy.deepcopy(self.head)
self.backbone = None
self.head = None
for p in self.old_backbone.parameters():
p.requires_grad = False
for p in self.old_head.parameters():
p.requires_grad = False
self.backbone = create_backbone(model_name=self.params.backbone_name,
**self.params.backbone_params)
# create LINEAR head
self.params.head_params['in_features'] = self.backbone.num_features
self.head = HEADS.get(self.params.head_name)(**self.params.head_params)
self.backbone = self.backbone.to(self.device)
self.head = self.head.to(self.device)
self.old_backbone.eval()
self.old_head.eval()
self.backbone.train()
self.head.train()