Beispiel #1
0
class MoCoSolver(ClsSolver):

    def build_model(self):
        """
        Build encode_q and encoder_k.
        """
        if hasattr(self.config, 'lms'):
            if self.config.lms.enable:
                torch.cuda.set_enabled_lms(True)
                byte_limit = self.config.lms.kwargs.limit * (1 << 30)
                torch.cuda.set_limit_lms(byte_limit)
                self.logger.info('Enable large model support, limit of {}G!'.format(
                    self.config.lms.kwargs.limit))

        encoder_q = model_entry(self.config.model)
        encoder_k = model_entry(self.config.model)
        self.model = MoCo(encoder_q, encoder_k, **self.config.moco.kwargs)
        self.model.cuda()
        count_params(self.model.encoder_k)
        count_flops(self.model.encoder_k, input_shape=[1, 3, self.config.data.input_size, self.config.data.input_size])

        # handle fp16
        if self.config.optimizer.type in ['FP16SGD', 'FusedFP16SGD', 'FP16RMSprop']:
            self.fp16 = True
        else:
            self.fp16 = False

        if self.fp16:
            # if you have modules that must use fp32 parameters, and need fp32 input
            # try use link.fp16.register_float_module(your_module)
            # if you only need fp32 parameters set cast_args=False when call this
            # function, then call link.fp16.init() before call model.half()
            if self.config.optimizer.get('fp16_normal_bn', False):
                self.logger.info('using normal bn for fp16')
                link.fp16.register_float_module(link.nn.SyncBatchNorm2d, cast_args=False)
                link.fp16.register_float_module(torch.nn.BatchNorm2d, cast_args=False)
            if self.config.optimizer.get('fp16_normal_fc', False):
                self.logger.info('using normal fc for fp16')
                link.fp16.register_float_module(torch.nn.Linear, cast_args=True)
            link.fp16.init()
            self.model.half()

        self.model = DistModule(self.model, self.config.dist.sync)
        if 'model' in self.state:
            load_state_model(self.model, self.state['model'])

    def build_data(self):
        """
        Unsupervised training: only training data is needed.
        """
        self.config.data.max_iter = self.config.lr_scheduler.kwargs.max_iter
        self.config.data.last_iter = self.state['last_iter']
        if self.config.data.last_iter < self.config.data.max_iter:
            if self.config.data.type == 'imagenet':
                self.train_data = build_imagenet_train_dataloader(self.config.data)
            elif self.config.data.type == 'custom':
                self.train_data = build_custom_dataloader('train', self.config.data)
            else:
                raise RuntimeError("undefined data type!")

    def train(self):

        self.pre_train()
        total_step = len(self.train_data['loader'])
        start_step = self.state['last_iter'] + 1
        end = time.time()
        for i, batch in enumerate(self.train_data['loader']):
            input = batch['image']
            curr_step = start_step + i
            self.lr_scheduler.step(curr_step)
            # lr_scheduler.get_lr()[0] is the main lr
            current_lr = self.lr_scheduler.get_lr()[0]
            # measure data loading time
            self.meters.data_time.update(time.time() - end)
            # transfer input to gpu
            input = input.cuda().half() if self.fp16 else input.cuda()
            # forward
            logits, target = self.model(input)
            loss = self.criterion(logits, target) / self.dist.world_size
            reduced_loss = loss.clone()
            self.meters.losses.reduce_update(reduced_loss)
            self.optimizer.zero_grad()
            if FusedFP16SGD is not None and isinstance(self.optimizer, FusedFP16SGD):
                self.optimizer.backward(loss)
                self.model.sync_gradients()
                self.optimizer.step()
            elif isinstance(self.optimizer, FP16SGD) or isinstance(self.optimizer, FP16RMSprop):
                def closure():
                    self.optimizer.backward(loss, False)
                    self.model.sync_gradients()
                    # check overflow, convert to fp32 grads, downscale
                    self.optimizer.update_master_grads()
                    return loss
                self.optimizer.step(closure)
            else:
                loss.backward()
                self.model.sync_gradients()
                self.optimizer.step()

            # measure elapsed time
            self.meters.batch_time.update(time.time() - end)
            if curr_step % self.config.saver.print_freq == 0 and self.dist.rank == 0:
                self.tb_logger.add_scalar('loss_train', self.meters.losses.avg, curr_step)
                self.tb_logger.add_scalar('lr', current_lr, curr_step)
                remain_secs = (total_step - curr_step) * self.meters.batch_time.avg
                remain_time = datetime.timedelta(seconds=round(remain_secs))
                finish_time = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime(time.time()+remain_secs))
                log_msg = f'Iter: [{curr_step}/{total_step}]\t' \
                    f'Time {self.meters.batch_time.val:.3f} ({self.meters.batch_time.avg:.3f})\t' \
                    f'Data {self.meters.data_time.val:.3f} ({self.meters.data_time.avg:.3f})\t' \
                    f'Loss {self.meters.losses.val:.4f} ({self.meters.losses.avg:.4f})\t' \
                    f'LR {current_lr:.4f}\t' \
                    f'Remaining Time {remain_time} ({finish_time})'
                self.logger.info(log_msg)

            if curr_step > 0 and curr_step % self.config.saver.val_freq == 0:
                if self.dist.rank == 0:
                    if self.config.saver.save_many:
                        ckpt_name = f'{self.path.save_path}/ckpt_{curr_step}.pth.tar'
                    else:
                        ckpt_name = f'{self.path.save_path}/ckpt.pth.tar'
                    self.state['model'] = self.model.state_dict()
                    self.state['optimizer'] = self.optimizer.state_dict()
                    self.state['last_iter'] = curr_step
                    torch.save(self.state, ckpt_name)
            end = time.time()
Beispiel #2
0
class SimCLRSolver(ClsSolver):
    def build_model(self):
        encoder = model_entry(self.config.model)
        self.model = SimCLR(encoder)
        self.model.cuda()
        count_params(self.model.encoder)
        count_flops(self.model.encoder,
                    input_shape=[
                        1, 3, self.config.data.input_size,
                        self.config.data.input_size
                    ])

        # handle fp16
        if self.config.optimizer.type in [
                'FP16SGD', 'FusedFP16SGD', 'FP16RMSprop'
        ]:
            self.fp16 = True
        else:
            self.fp16 = False

        if self.fp16:
            # if you have modules that must use fp32 parameters, and need fp32 input
            # try use link.fp16.register_float_module(your_module)
            # if you only need fp32 parameters set cast_args=False when call this
            # function, then call link.fp16.init() before call model.half()
            if self.config.optimizer.get('fp16_normal_bn', False):
                self.logger.info('using normal bn for fp16')
                link.fp16.register_float_module(link.nn.SyncBatchNorm2d,
                                                cast_args=False)
                link.fp16.register_float_module(torch.nn.BatchNorm2d,
                                                cast_args=False)
            if self.config.optimizer.get('fp16_normal_fc', False):
                self.logger.info('using normal fc for fp16')
                link.fp16.register_float_module(torch.nn.Linear,
                                                cast_args=True)
            link.fp16.init()
            self.model.half()

        self.model = DistModule(self.model, self.config.dist.sync)

        if 'model' in self.state:
            load_state_model(self.model, self.state['model'])

    def pre_train(self):
        super().pre_train()
        self.criterion = NT_Xent(self.config.data.batch_size,
                                 self.config.temperature)

    def train(self):

        self.pre_train()
        total_step = len(self.train_data['loader'])
        start_step = self.state['last_iter'] + 1
        end = time.time()

        for i, batch in enumerate(self.train_data['loader']):
            input = batch['image']
            curr_step = start_step + i
            self.lr_scheduler.step(curr_step)
            # lr_scheduler.get_lr()[0] is the main lr
            current_lr = self.lr_scheduler.get_lr()[0]
            # measure data loading time
            self.meters.data_time.update(time.time() - end)
            # transfer input to gpu
            input = input.cuda().half() if self.fp16 else input.cuda()

            # forward
            z_i, z_j = self.model(input)
            # normalize projection feature vectors
            z_i = F.normalize(z_i, dim=1)
            z_j = F.normalize(z_j, dim=1)
            loss = self.criterion(z_i, z_j) / self.dist.world_size
            reduced_loss = loss.clone()
            self.meters.losses.reduce_update(reduced_loss)
            self.optimizer.zero_grad()

            if FusedFP16SGD is not None and isinstance(self.optimizer,
                                                       FusedFP16SGD):
                self.optimizer.backward(loss)
                self.model.sync_gradients()
                self.optimizer.step()
            elif isinstance(self.optimizer, FP16SGD) or isinstance(
                    self.optimizer, FP16RMSprop):

                def closure():
                    self.optimizer.backward(loss, False)
                    self.model.sync_gradients()
                    # check overflow, convert to fp32 grads, downscale
                    self.optimizer.update_master_grads()
                    return loss

                self.optimizer.step(closure)
            else:
                loss.backward()
                self.model.sync_gradients()
                self.optimizer.step()

            # measure elapsed time
            self.meters.batch_time.update(time.time() - end)
            if curr_step % self.config.saver.print_freq == 0 and self.dist.rank == 0:
                self.tb_logger.add_scalar('loss_train', self.meters.losses.avg,
                                          curr_step)
                self.tb_logger.add_scalar('lr', current_lr, curr_step)
                remain_secs = (total_step -
                               curr_step) * self.meters.batch_time.avg
                remain_time = datetime.timedelta(seconds=round(remain_secs))
                finish_time = time.strftime(
                    "%Y-%m-%d %H:%M:%S",
                    time.localtime(time.time() + remain_secs))
                log_msg = f'Iter: [{curr_step}/{total_step}]\t' \
                    f'Time {self.meters.batch_time.val:.3f} ({self.meters.batch_time.avg:.3f})\t' \
                    f'Data {self.meters.data_time.val:.3f} ({self.meters.data_time.avg:.3f})\t' \
                    f'Loss {self.meters.losses.val:.4f} ({self.meters.losses.avg:.4f})\t' \
                    f'LR {current_lr:.4f}\t' \
                    f'Remaining Time {remain_time} ({finish_time})'
                self.logger.info(log_msg)

            if curr_step > 0 and curr_step % self.config.saver.val_freq == 0:
                if self.dist.rank == 0:
                    if self.config.saver.save_many:
                        ckpt_name = f'{self.path.save_path}/ckpt_{curr_step}.pth.tar'
                    else:
                        ckpt_name = f'{self.path.save_path}/ckpt.pth.tar'
                    self.state['model'] = self.model.state_dict()
                    self.state['optimizer'] = self.optimizer.state_dict()
                    self.state['last_iter'] = curr_step
                    torch.save(self.state, ckpt_name)

            end = time.time()