Dubhe/model_measuring/kamal/slim/distillation/data_free/zskt.py

"""
 Copyright 2020 Tianshu AI Platform. All Rights Reserved.

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
 You may obtain a copy of the License at

     http://www.apache.org/licenses/LICENSE-2.0

 Unless required by applicable law or agreed to in writing, software
 distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
 =============================================================
"""

import torch
import time
import torch.nn.functional as F
from kamal.slim.distillation.kd import KDDistiller
from kamal.utils import set_mode
from kamal.core.tasks.loss import kldiv

class ZSKTDistiller(KDDistiller):
    def __init__(   self, 
                    student,
                    teacher,
                    generator,
                    z_dim,
                    logger=None,
                    viz=None):
        super(ZSKTDistiller, self).__init__(logger, viz)
        self.teacher = teacher
        self.model = self.student = student
        self.generator = generator
        self.z_dim = z_dim

    def train(self, start_iter, max_iter, optim_s, optim_g, device=None):
        if device is None:
            device = torch.device( 'cuda' if torch.cuda.is_available() else 'cpu' )
        self.device = device
        self.optim_s, self.optim_g = optim_s, optim_g

        self.model.to(self.device)
        self.teacher.to(self.device)
        self.generator.to(self.device)
        self.train_loader = [0, ]
        
        with set_mode(self.student, training=True), \
             set_mode(self.teacher, training=False), \
             set_mode(self.generator, training=True):
            super( ZSKTDistiller, self ).train( start_iter, max_iter )
    
    def search_optimizer(self, evaluator, train_loader, hpo_space=None, mode='min', max_evals=20, max_iters=400):
        optimizer = hpo.search_optimizer(self, train_loader, evaluator=evaluator, hpo_space=hpo_space, mode=mode, max_evals=max_evals, max_iters=max_iters)
        return optimizer
    
    def step(self):
        start_time = time.perf_counter()
        
        # Adv
        z = torch.randn( self.z_dim ).to(self.device)
        fake = self.generator( z )
        self.optim_g.zero_grad()
        t_out = self.teacher( fake )
        s_out = self.student( fake )
        loss_g = -kldiv( s_out, t_out )
        loss_g.backward()
        self.optim_g.step()

        with torch.no_grad():
            fake = self.generator( z )
            t_out = self.teacher( fake.detach() )
        for _ in range(10):
            self.optim_s.zero_grad()
            s_out = self.student( fake.detach() )
            loss_s = kldiv( s_out, t_out )
            loss_s.backward()
            self.optim_s.step()
        
        loss_dict = {
            'loss_g': loss_g,
            'loss_s': loss_s,
        }

        step_time = time.perf_counter() - start_time

        # record training info
        info = loss_dict
        info['step_time'] = step_time
        info['lr_s'] = float( self.optim_s.param_groups[0]['lr'] )
        info['lr_g'] = float( self.optim_g.param_groups[0]['lr'] )
        self.history.put_scalars( **info )
    
    def reset(self):
        self.history = None
        self._train_loader_iter = iter(train_loader)
        self.iter = self.start_iter