update

游雁

2024-02-20 58b6154a73331a8807127d4579ed473432ce88de

 funasr/train_utils/trainer.py

@@ -69,6 +69,7 @@
        self.device = next(model.parameters()).device
        self.avg_nbest_model = kwargs.get("avg_nbest_model", 5)
        self.kwargs = kwargs
        self.log_interval = kwargs.get("log_interval", 50)
        
    
        try:
@@ -204,7 +205,25 @@
            my_context = self.model.no_sync if batch_idx % accum_grad != 0 else nullcontext
            with my_context():
                time2 = time.perf_counter()
                # print("before, GPU, memory: {:.3f} GB, "
                #       "{:.3f} GB, "
                #       "{:.3f} GB, "
                #       "{:.3f} GB".format(torch.cuda.memory_allocated()/1024/1024/1024,
                #                      torch.cuda.max_memory_allocated()/1024/1024/1024,
                #                      torch.cuda.memory_reserved()/1024/1024/1024,
                #                      torch.cuda.max_memory_reserved()/1024/1024/1024,
                #                      ))

                retval = self.model(**batch)
                torch.cuda.empty_cache()
                # print("after, GPU, memory: {:.3f} GB, "
                #       "{:.3f} GB, "
                #       "{:.3f} GB, "
                #       "{:.3f} GB".format(torch.cuda.memory_allocated()/1024/1024/1024,
                #                      torch.cuda.max_memory_allocated()/1024/1024/1024,
                #                      torch.cuda.memory_reserved()/1024/1024/1024,
                #                      torch.cuda.max_memory_reserved()/1024/1024/1024,
                #                      ))
                time3 = time.perf_counter()
                speed_stats["forward_time"] = f"{time3 - time2:0.3f}"
                loss, stats, weight = retval
@@ -255,24 +274,35 @@
                speed_stats["total_time"] = total_time


            pbar.update(1)
            if self.local_rank == 0:
            
            if batch_idx % self.log_interval == 0 or batch_idx == len(self.dataloader_train) - 1:
                pbar.update(self.log_interval)
                gpu_info = "GPU, memory: {:.3f} GB, " \
                           "{:.3f} GB, "\
                           "{:.3f} GB, "\
                           "{:.3f} GB".format(torch.cuda.memory_allocated()/1024/1024/1024,
                                             torch.cuda.max_memory_allocated()/1024/1024/1024,
                                             torch.cuda.memory_reserved()/1024/1024/1024,
                                             torch.cuda.max_memory_reserved()/1024/1024/1024,
                                             )
                description = (
                    f"rank: {self.local_rank}, "
                    f"Train epoch: {epoch}/{self.max_epoch}, "
                    f"step {batch_idx}/{len(self.dataloader_train)}, "
                    f"{speed_stats}, "
                    f"(loss: {loss.detach().cpu().item():.3f}), "
                    f"{[(k, round(v.cpu().item(), 3)) for k, v in stats.items()]}"
                    f"{gpu_info}"
                )
                pbar.set_description(description)
                if self.writer:
                    self.writer.add_scalar('Loss/train', loss.item(),
                    self.writer.add_scalar(f'rank{self.local_rank}, Loss/train', loss.item(),
                                           epoch*len(self.dataloader_train) + batch_idx)
                    for key, var in stats.items():
                        self.writer.add_scalar(f'{key}/train', var.item(),
                        self.writer.add_scalar(f'rank{self.local_rank}, {key}/train', var.item(),
                                               epoch * len(self.dataloader_train) + batch_idx)
                    for key, var in speed_stats.items():
                        self.writer.add_scalar(f'{key}/train', eval(var),
                        self.writer.add_scalar(f'rank{self.local_rank}, {key}/train', eval(var),
                                               epoch * len(self.dataloader_train) + batch_idx)
                    
            # if batch_idx == 2:
@@ -317,22 +347,25 @@
                loss = loss
                time4 = time.perf_counter()

                pbar.update(1)
                if self.local_rank == 0:
                
                if batch_idx % self.log_interval == 0 or batch_idx == len(self.dataloader_train) - 1:
                    pbar.update(self.log_interval)
                    description = (
                        f"rank: {self.local_rank}, "
                        f"validation epoch: {epoch}/{self.max_epoch}, "
                        f"step {batch_idx}/{len(self.dataloader_train)}, "
                        f"{speed_stats}, "
                        f"(loss: {loss.detach().cpu().item():.3f}), "
                        f"{[(k, round(v.cpu().item(), 3)) for k, v in stats.items()]}"
                        f"rank: {self.local_rank}"
                    )
                    pbar.set_description(description)
                    if self.writer:
                        self.writer.add_scalar('Loss/val', loss.item(),
                        self.writer.add_scalar(f"rank{self.local_rank}, Loss/val", loss.item(),
                                               epoch*len(self.dataloader_train) + batch_idx)
                        for key, var in stats.items():
                            self.writer.add_scalar(f'{key}/val', var.item(),
                            self.writer.add_scalar(f'rank{self.local_rank}, {key}/val', var.item(),
                                                   epoch * len(self.dataloader_train) + batch_idx)
                        for key, var in speed_stats.items():
                            self.writer.add_scalar(f'{key}/val', eval(var),
                            self.writer.add_scalar(f'rank{self.local_rank}, {key}/val', eval(var),
                                                   epoch * len(self.dataloader_train) + batch_idx)