Customize Model Template

Here, we present how to develop a new Model Template, and apply it into EduStudio. EduStudio provides the Model Protocol in EduStudio.edustudio.model.basemodeltpl.BaseModelTPL (BaseModelTPL).

Protocol

BaseModelTPL

The protocols in BaseModelTPL are listed as follows.

name	description	type	note
default_cfg	the default configuration	class variable
add_extra_data	add extra data in addition to the data loaders	function interface	implemented by subclass

BaseProxyModelTPL

This protocol is for some works proposed general frameworks which can replace the backbone model. We implement the BaseProxyModelTPL Protocol. Users only need to inherit this protocol and set the backbone_model_cls parameter in the default_cfg.

ModelTPL

EduStudio provides the ModelTPL in EduStudio.edustudio.model.gd_basemodeltpl.GDBaseModelTPL (GDBaseModelTPL). GDBaseModelTPL inherents BaseModelTPL, and the methods in GDBaseModelTPL are listed as follows.

name	description	type	note
default_cfg	the default configuration	class variable
build_cfg	construct model config	abstract method	implemented by subclass
build_model	construct model component	abstract method	implemented by subclass
predict	predict function	function interface	implemented by subclass
get_loss_dict	obtain loss	function interface	implemented by subclass
_init_params	initial parameters	function interface	implemented by subclass
_load_params_from_pretrained	load parameters as dict	function interface	implemented by subclass

Develop a New ModelTPL in EduStudio

When you develope a new model in EduStudio, then you can inherent GDBaseModelTPL and implement the abstract methods build_cfg() and build_model(). Then, you can revise the function predict() and get_loss_dict(). You can also define the configuration of the new model template in the dictionary default_cfg.

If you want to develope a new ModelTPL for a backbone-style model, then you can inherent BaseProxyModelTPL and set the backbone_model_cls parameter in default_cfg. Then, you can implement the some methods such as build_model(), get_loss_dict(), and predict().

Example 1: Develop a traditional ModelTPL

from ..gd_basemodeltpl import GDBaseModelTPL

class NewModelTPL(GDBaseModelTPL):
    default_cfg = {
    'dnn_units': [512, 256],
    'dropout_rate': 0.5,
    'activation': 'sigmoid',
    'disc_scale': 10
}

def __init__(self, cfg):
    super().__init__(cfg)

def build_cfg(self):
    self.n_user = self.datafmt_cfg['dt_info']['stu_count']
    ...

def build_model(self):
    ...
    self.pd_net = PosMLP(
        input_dim=self.n_cpt, output_dim=1, activation=self.model_cfg['activation'],
        dnn_units=self.model_cfg['dnn_units'], dropout_rate=self.model_cfg['dropout_rate']
    )

def forward(self, stu_id, exer_id, Q_mat, **kwargs):
    ...
    pd = self.pd_net(input_x).sigmoid()
    return pd

@torch.no_grad()
def predict(self, stu_id, exer_id, Q_mat, **kwargs):
    return {
        'y_pd': self(stu_id, exer_id, Q_mat).flatten(),
    }

def get_main_loss(self, **kwargs):
    ...
    pd = self(stu_id, exer_id, Q_mat).flatten()
    loss = F.binary_cross_entropy(input=pd, target=label)
    return {
        'loss_main': loss
    }

def get_loss_dict(self, **kwargs):
    return self.get_main_loss(**kwargs)

Example 2: Develop a backbone-style ModelTPL

from ..basemodeltpl import BaseProxyModelTPL

class NewBackboneModelTPL(BaseProxyModelTPL):
    default_cfg = {
    "backbone_model_cls": "IRT",
    }

    def __init__(self, cfg):
        super().__init__(cfg)

    def build_model(self):
        super().build_model()
        self.irr_pair_loss = PairSCELoss()

    def get_main_loss(self, **kwargs):
        pair_exer = kwargs['pair_exer']
        pair_pos_stu = kwargs['pair_pos_stu']
        pair_neg_stu = kwargs['pair_neg_stu']

        kwargs['exer_id'] = pair_exer
        kwargs['stu_id'] = pair_pos_stu
        pos_pd = self(**kwargs).flatten()
        kwargs['stu_id'] = pair_neg_stu
        neg_pd = self(**kwargs).flatten()

        return {
            'loss_main': self.irr_pair_loss(pos_pd, neg_pd)
        }

class PairSCELoss(nn.Module):
    ...