diff --git a/PyTorch/build-in/Classification/XCiT/readme.md b/PyTorch/build-in/Classification/XCiT/readme.md new file mode 100644 index 000000000..e193b9aad --- /dev/null +++ b/PyTorch/build-in/Classification/XCiT/readme.md @@ -0,0 +1,65 @@ +```markdown +## 1. 模型链接 +- 原始仓库链接： +https://github.com/huggingface/pytorch-image-models?tab=readme-ov-file#models + +## 2. 快速开始 + +使用本模型执行训练的主要流程如下： + +1. **基础环境安装**：介绍训练前需要完成的基础环境检查和安装。 +2. **获取数据集**：介绍如何获取训练所需的数据集。 +3. **构建环境**：介绍如何构建模型运行所需要的环境。 +4. **启动训练**：介绍如何运行训练。 + +### 2.1 基础环境安装 + +请参考主仓库的基础环境安装章节，完成训练前的基础环境检查和安装（如驱动、固件等）。 + +### 2.2 准备数据集 + +#### 2.2.1 获取数据集 + +训练使用 **CIFAR-100** 数据集。该数据集为开源数据集，包含 100 个类别的 60000 张彩色图像。 + +#### 2.2.2 处理数据集 + +请确保数据集已下载并解压。根据训练脚本的默认配置，建议将数据集存放在模型目录的上级 `data` 目录中（即 `../data`），或者根据实际路径修改训练命令中的 `--datapath` 参数。 + +### 2.3 构建环境 + +所使用的环境下需包含 PyTorch 框架虚拟环境。 + +1. 执行以下命令，启动虚拟环境（根据实际环境名称修改）： + + ```bash + conda activate torch_env_py310 + +``` + +2. 安装 Python 依赖。确保已安装项目所需的依赖包： +```bash +pip install -r requirements_exact.txt + +``` + + + +### 2.4 启动训练 + +1. 在构建好的环境中，进入模型训练脚本所在目录。 + +2. 运行训练。该模型支持单机单卡训练。 +执行以下命令启动训练（使用 CIFAR-100 数据集，Batch Size 为 128）： +```bash +python weloTrainStep.py \ + --name train \ + --arch xcit \ + --print_freq 1 \ + --steps 100 \ + --dataset cifar100 \ + --datapath ../data \ + --batch_size 32 \ + --epochs 100 + +``` diff --git a/PyTorch/build-in/Classification/XCiT/requirements.txt b/PyTorch/build-in/Classification/XCiT/requirements.txt new file mode 100644 index 000000000..7394b3319 --- /dev/null +++ b/PyTorch/build-in/Classification/XCiT/requirements.txt @@ -0,0 +1,89 @@ +addict==2.4.0 +aliyun-python-sdk-core==2.16.0 +aliyun-python-sdk-kms==2.16.5 +anyio==4.11.0 +astunparse==1.6.3 +certifi==2024.12.14 +cffi==2.0.0 +charset-normalizer==3.4.1 +click==8.3.1 +colorama==0.4.6 +contourpy==1.3.2 +crcmod==1.7 +cryptography==46.0.3 +cycler==0.12.1 +einops==0.8.1 +exceptiongroup==1.3.1 +filelock==3.14.0 +fonttools==4.60.1 +fsspec==2024.12.0 +future @ file:///croot/future_1730902796226/work +git-filter-repo==2.47.0 +h11==0.16.0 +hf-xet==1.2.0 +httpcore==1.0.9 +httpx==0.28.1 +huggingface_hub==1.1.5 +idna==3.10 +inplace-abn @ git+https://github.com/mapillary/inplace_abn.git@b50bfe9c7cd7116a3ab091a352b48d6ba5ee701c +Jinja2==3.1.5 +jmespath==0.10.0 +joblib==1.5.2 +kiwisolver==1.4.9 +Markdown==3.10 +markdown-it-py==4.0.0 +MarkupSafe==3.0.2 +matplotlib==3.10.7 +mdurl==0.1.2 +mmdet==3.3.0 +mmengine==0.10.7 +model-index==0.1.11 +mpmath==1.3.0 +networkx==3.4.2 +numpy==1.23.5 +opencv-python==4.12.0.88 +opendatalab==0.0.10 +openmim==0.3.9 +openxlab==0.1.3 +ordered-set==4.1.0 +oss2==2.17.0 +packaging @ file:///croot/packaging_1734472117206/work +pandas==2.3.3 +pillow==11.1.0 +platformdirs==4.5.1 +pycocotools==2.0.11 +pycparser @ file:///tmp/build/80754af9/pycparser_1636541352034/work +pycryptodome==3.23.0 +Pygments==2.19.2 +pyparsing==3.2.5 +python-dateutil==2.9.0.post0 +pytz==2023.4 +PyYAML @ file:///croot/pyyaml_1728657952215/work +requests==2.28.2 +rich==13.4.2 +safetensors==0.7.0 +scikit-learn==1.7.2 +scipy==1.15.3 +shapely==2.1.2 +shellingham==1.5.4 +six @ file:///tmp/build/80754af9/six_1644875935023/work +sniffio==1.3.1 +sympy==1.13.3 +tabulate==0.9.0 +termcolor==3.2.0 +terminaltables==3.1.10 +threadpoolctl==3.6.0 +timm==1.0.22 +tomli==2.3.0 +torch @ file:///apps/torch-2.4.0a0%2Bgit4451b0e-cp310-cp310-linux_x86_64.whl#sha256=2e472c916044cac5a1a0e0d8b0e12bb943d8522b24ff826c8014dd444dccd378 +torch_sdaa @ file:///apps/torch_sdaa-2.0.0-cp310-cp310-linux_x86_64.whl#sha256=5aa57889b002e1231fbf806642e1353bfa016297bc25178396e89adc2b1f92e7 +torchaudio @ file:///apps/torchaudio-2.0.2%2Bda3eb8d-cp310-cp310-linux_x86_64.whl#sha256=46525c02fb7eaa8dafea860428de3d01e437ba8d6ff2cc228d7c71975ac4054b +torchdata @ file:///apps/torchdata-0.6.1%2Be1feeb2-py3-none-any.whl#sha256=aa2dc1a7732ea68adfad186978049bf68cc1afdbbdd1e17a8024227ab770e433 +torchtext @ file:///apps/torchtext-0.15.2a0%2B4571036-cp310-cp310-linux_x86_64.whl#sha256=7e42c684ba366f97b59ec37488bf95e416cce3892b6589200d2b3ad159ee5788 +torchvision @ file:///apps/torchvision-0.15.1a0%2B42759b1-cp310-cp310-linux_x86_64.whl#sha256=4b904db2d50102415536bc764bbc31c669b90b1b014f90964e9eccaadb2fd9eb +tqdm==4.65.2 +typer-slim==0.20.0 +typing_extensions==4.15.0 +tzdata==2025.2 +urllib3==1.26.20 +yapf==0.43.0 diff --git a/PyTorch/build-in/Classification/XCiT/weloTrainStep.py b/PyTorch/build-in/Classification/XCiT/weloTrainStep.py new file mode 100644 index 000000000..13297c11b --- /dev/null +++ b/PyTorch/build-in/Classification/XCiT/weloTrainStep.py @@ -0,0 +1,692 @@ +#!/usr/bin/env python3 +# coding: utf-8 + +import os +import random +import sys +import time +import json +import argparse +from collections import OrderedDict +from pathlib import Path +import numpy as np +import pandas as pd +from tqdm import tqdm +import importlib + +os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8" # 强烈推荐在 shell/最顶端设置 +os.environ["PYTHONHASHSEED"] = "12345" +os.environ["OMP_NUM_THREADS"] = "1" +os.environ["MKL_NUM_THREADS"] = "1" + +def ensure_cublas_workspace(config=":4096:8"): + """ + 尝试为 cuBLAS 设置可复现 workspace。强烈建议在主脚本入口处（import torch 之前） + 通过 export 设置该 env。此函数会在运行时设置，但如果 torch 已经被 import， + 则可能为时已晚——函数会打印提醒。 + """ + already = os.environ.get("CUBLAS_WORKSPACE_CONFIG") + if already: + print(f"[seed_utils] CUBLAS_WORKSPACE_CONFIG 已存在：{already}") + else: + os.environ["CUBLAS_WORKSPACE_CONFIG"] = config + print(f"[seed_utils] 已设置 CUBLAS_WORKSPACE_CONFIG={config} （注意：请在 import torch 前设置以保证生效）") + +def set_global_seed(seed: int = 42, set_threads: bool = True): + """ + 统一随机性设置。注意：若希望完全发挥效果，请在主脚本入口（import torch 之前） + 先调用 ensure_cublas_workspace(...) 或在 shell 中 export CUBLAS_WORKSPACE_CONFIG。 + """ + ensure_cublas_workspace() # 会设置 env 并提醒 + os.environ["PYTHONHASHSEED"] = str(seed) + + if set_threads: + os.environ["OMP_NUM_THREADS"] = "1" + os.environ["MKL_NUM_THREADS"] = "1" + + random.seed(seed) + np.random.seed(seed) + + # 现在导入 torch（晚导入以便前面 env 生效） + import torch + torch.manual_seed(seed) + if torch.cuda.is_available(): + torch.cuda.manual_seed(seed) + torch.cuda.manual_seed_all(seed) + # 强制确定性（如果存在不确定性算子，PyTorch 会报错并提示） + try: + torch.use_deterministic_algorithms(True) + except Exception as e: + print("[seed_utils] 设置 deterministic 模式时出错：", e) + print("[seed_utils] 请确认 CUBLAS_WORKSPACE_CONFIG 已在 import torch 之前设置。") + + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + + if set_threads: + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + + print(f"[seed_utils] 全局 seed 已设置为 {seed}") + +set_global_seed(2025) + +""" +通用训练模版（优先从本地导入 Model -> 支持 DDP / 单卡，AMP，resume，日志，checkpoint） +保存为 train_template_localmodel.py +""" +import torch +import torch.nn as nn +import torch.optim as optim +import torch.backends.cudnn as cudnn +import torchvision.transforms as transforms +import torchvision.datasets as datasets +import torchvision.models as tv_models + +import torch.distributed as dist +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.data import DataLoader +from torch.utils.data.distributed import DistributedSampler + +from torch.sdaa import amp +# from torch.cuda import amp + + +# ---------------------------- +# Helper utilities (self-contained) +# ---------------------------- +class AverageMeter(object): + def __init__(self, name='Meter', fmt=':.4f'): + self.name = name + self.fmt = fmt + self.reset() + def reset(self): + self.val = 0 + self.avg = 0 + self.sum = 0 + self.count = 0 + def update(self, val, n=1): + self.val = val + self.sum += val * n + self.count += n + self.avg = self.sum / max(1, self.count) + def __str__(self): + fmtstr = '{name} {val' + self.fmt + '} (avg {avg' + self.fmt + '})' + return fmtstr.format(name=self.name, val=self.val, avg=self.avg) + +def accuracy(output, target, topk=(1,)): + """Computes the precision@k for the specified values of k + 返回一个 list，每个元素是 tensor（百分比形式） + """ + with torch.no_grad(): + maxk = max(topk) + batch_size = target.size(0) + + # output: (N, C) -> pred: (maxk, N) + _, pred = output.topk(maxk, 1, True, True) + pred = pred.t() # (maxk, N) + correct = pred.eq(target.view(1, -1).expand_as(pred)) # (maxk, N) bool + + res = [] + for k in topk: + # 把前 k 行展平后求和（返回 0-dim tensor），随后换算为百分比 + correct_k = correct[:k].reshape(-1).float().sum() # 注意：不传 keepdim + # 乘以 100.0 / batch_size，保持返回 tensor（和之前代码兼容） + res.append(correct_k.mul_(100.0 / batch_size)) + return res + +def save_checkpoint(state, is_best, save_dir, filename='checkpoint.pth'): + save_path = os.path.join(save_dir, filename) + torch.save(state, save_path) + if is_best: + best_path = os.path.join(save_dir, 'model_best.pth') + torch.save(state, best_path) + +def set_seed(seed, deterministic=False): + random.seed(seed) + np.random.seed(seed) + torch.manual_seed(seed) + torch.cuda.manual_seed_all(seed) + if deterministic: + cudnn.deterministic = True + cudnn.benchmark = False + else: + cudnn.deterministic = False + cudnn.benchmark = True + +# ---------------------------- +# Argument parser +# ---------------------------- +def parse_args(): + parser = argparse.ArgumentParser(description='Generic PyTorch training template (DDP/AMP) with LocalModel priority') + parser.add_argument('--name', default='run', type=str, help='experiment name (log/checkpoints dir)') + parser.add_argument('--seed', default=42, type=int, help='random seed') + parser.add_argument('--arch', default='None', type=str, help='model name') + parser.add_argument('--deterministic', action='store_true', help='set cudnn deterministic (may be slower)') + parser.add_argument('--dataset', default='cifar10', choices=['cifar10','cifar100','imagenet','custom'], help='which dataset') + parser.add_argument('--datapath', default='./data', type=str, help='dataset root / imagenet root / custom root') + parser.add_argument('--imagenet_dir', default='./imagenet', type=str, help='if dataset=imagenet, path to imagenet root') + parser.add_argument('--custom_eval_dir', default=None, help='if dataset=custom, provide val dir') + parser.add_argument('--num_workers', default=4, type=int, help='dataloader workers per process') + parser.add_argument('--epochs', default=200, type=int) + parser.add_argument('--steps', default=0, type=int, help='max steps to run (if >0, training will stop when global_step reaches this).') + parser.add_argument('--batch_size', default=128, type=int) + parser.add_argument('--model_name', default='resnet18', help='torchvision model name or python path e.g. mypkg.mymodule.Model (used if no local Model)') + parser.add_argument('--num_classes', default=None, type=int, help='override num classes (auto-detect for common sets)') + parser.add_argument('--pretrained', action='store_true', help='use torchvision pretrained weights when available') + parser.add_argument('--optimizer', default='sgd', choices=['sgd','adam','adamw'], help='optimizer') + parser.add_argument('--lr', '--learning_rate', default=0.1, type=float) + parser.add_argument('--momentum', default=0.9, type=float) + parser.add_argument('--weight_decay', default=5e-4, type=float) + parser.add_argument('--nesterov', action='store_true') + parser.add_argument('--scheduler', default='multistep', choices=['multistep','step','cosine','none'], help='lr scheduler') + parser.add_argument('--milestones', default='100,150', type=str, help='milestones for multistep (comma sep)') + parser.add_argument('--step_size', default=30, type=int, help='step size for StepLR or cosine max epochs') + parser.add_argument('--gamma', default=0.1, type=float) + parser.add_argument('--scheduler_step_per_batch', action='store_true', help='call scheduler.step() per batch (for some schedulers)') + parser.add_argument('--resume', default='', type=str, help='path to checkpoint to resume from') + parser.add_argument('--start_epoch', default=0, type=int) + parser.add_argument('--print_freq', default=100, type=int) + parser.add_argument('--save_freq', default=10, type=int, help='save checkpoint every N epochs (rank0 only)') + parser.add_argument('--amp', action='store_true', default = True,help='use automatic mixed precision (AMP)') + parser.add_argument('--grad_accum_steps', default=1, type=int, help='gradient accumulation steps') + parser.add_argument('--local_rank', default=None, type=int, help='local rank passed by torchrun (if any). Use -1 or None for non-distributed') + parser.add_argument('--cutmix_prob', default=0.0, type=float) + parser.add_argument('--beta', default=1.0, type=float) + parser.add_argument('--seed_sampler', default=False, action='store_true', help='set sampler epoch seeds to make deterministic distributed shuffling') + args = parser.parse_args() + args.milestones = [int(x) for x in args.milestones.split(',')] if args.milestones else [] + return args + +# ---------------------------- +# build model (优先 LocalModel) +# ---------------------------- +def build_model_with_local_priority(args, device=None): + """ + 用参数 args.arch 作为模块名导入 Model() + 如果模块不存在或没有 Model 类，则报错停止。 + """ + try: + # 动态导入模块，比如 args.arch = "rexnet" + mod = importlib.import_module(args.arch) + Model = getattr(mod, "Model") # 从模块中获取 Model 类 + except Exception as e: + raise RuntimeError( + f"无法导入模型模块 '{args.arch}' 或未找到类 Model。" + f"\n错误信息：{e}" + ) + + # 解析数据集类别数 + if args.dataset == 'cifar10': + num_classes = 10 + elif args.dataset == 'cifar100': + num_classes = 100 + else: + print(f"[ERROR] 不支持的数据集类型：{args.dataset}，无法确定类别数。程序终止。") + sys.exit(1) + + + # 实例化 + try: + model = Model(num_classes) + except Exception as e: + raise RuntimeError( + f"Model() 实例化失败，请检查模型构造函数。\n错误信息：{e}" + ) + + return model + +# ---------------------------- +# Data loader factory +# ---------------------------- +def build_dataloaders(args, rank, world_size): + if args.dataset == 'cifar10' or args.dataset == 'cifar100': + mean = (0.4914, 0.4822, 0.4465) + std = (0.2470, 0.2435, 0.2616) if args.dataset == 'cifar10' else (0.2023, 0.1994, 0.2010) + # train_transform = transforms.Compose([ + # transforms.RandomCrop(32, padding=4), + # transforms.RandomHorizontalFlip(), + # transforms.ToTensor(), + # transforms.Normalize(mean, std), + # ]) + # test_transform = transforms.Compose([ + # transforms.ToTensor(), + # transforms.Normalize(mean, std), + # ]) + + train_transform = transforms.Compose([ # 2025/12/3 从visformer模型开始 + transforms.Resize(256), # 先放大到 256 + transforms.RandomCrop(224), # 再随机裁剪为 224（更符合 ImageNet 风格增强） + transforms.RandomHorizontalFlip(), + transforms.ToTensor(), + transforms.Normalize(mean, std), + ]) + test_transform = transforms.Compose([ + transforms.Resize(256), + transforms.CenterCrop(224), + transforms.ToTensor(), + transforms.Normalize(mean, std), + ]) + root = args.datapath + if args.dataset == 'cifar10': + train_set = datasets.CIFAR10(root=root, train=True, download=False, transform=train_transform) + val_set = datasets.CIFAR10(root=root, train=False, download=False, transform=test_transform) + num_classes = 10 + else: + train_set = datasets.CIFAR100(root=root, train=True, download=False, transform=train_transform) + val_set = datasets.CIFAR100(root=root, train=False, download=False, transform=test_transform) + num_classes = 100 + + elif args.dataset == 'imagenet': + train_dir = os.path.join(args.imagenet_dir, 'train') + val_dir = os.path.join(args.imagenet_dir, 'val') + train_transform = transforms.Compose([ + transforms.RandomResizedCrop(224), + transforms.RandomHorizontalFlip(), + transforms.ToTensor(), + transforms.Normalize((0.485,0.456,0.406), (0.229,0.224,0.225)), + ]) + test_transform = transforms.Compose([ + transforms.Resize(256), + transforms.CenterCrop(224), + transforms.ToTensor(), + transforms.Normalize((0.485,0.456,0.406), (0.229,0.224,0.225)), + ]) + train_set = datasets.ImageFolder(train_dir, train_transform) + val_set = datasets.ImageFolder(val_dir, test_transform) + num_classes = args.num_classes or 1000 + + elif args.dataset == 'custom': + train_dir = os.path.join(args.datapath, 'train') + val_dir = args.custom_eval_dir or os.path.join(args.datapath, 'val') + train_transform = transforms.Compose([ + transforms.RandomResizedCrop(224), + transforms.RandomHorizontalFlip(), + transforms.ToTensor(), + ]) + test_transform = transforms.Compose([ + transforms.Resize(256), + transforms.CenterCrop(224), + transforms.ToTensor(), + ]) + train_set = datasets.ImageFolder(train_dir, train_transform) + val_set = datasets.ImageFolder(val_dir, test_transform) + num_classes = len(train_set.classes) + else: + raise ValueError("Unknown dataset") + + if dist.is_initialized() and world_size > 1: + train_sampler = DistributedSampler(train_set, num_replicas=world_size, rank=rank, shuffle=True) + else: + train_sampler = None + + train_loader = DataLoader(train_set, + batch_size=args.batch_size, + shuffle=(train_sampler is None), + num_workers=args.num_workers, + pin_memory=True, + sampler=train_sampler, + drop_last=False) + val_loader = DataLoader(val_set, + batch_size=args.batch_size, + shuffle=False, + num_workers=args.num_workers, + pin_memory=True) + + return train_loader, val_loader, num_classes, train_sampler + +# ---------------------------- +# Train & validate +# ---------------------------- +def train_one_epoch(args, epoch, model, criterion, optimizer, train_loader, device, scaler, scheduler=None, train_sampler=None, global_step_start=0, max_global_steps=None): + """ + 现在支持：若 max_global_steps 非 None，则当 global_step 达到该值时提前退出 + 返回: epoch_summary_dict, step_logs_list, global_step_end + step_logs_list: list of dicts with per-step info (for logging to CSV if需要) + """ + batch_time = AverageMeter('Time') + data_time = AverageMeter('Data') + losses = AverageMeter('Loss') + top1 = AverageMeter('Acc@1') + top5 = AverageMeter('Acc@5') + + model.train() + end = time.time() + optimizer.zero_grad() + + iters = len(train_loader) + step_logs = [] + global_step = global_step_start + + for i, (images, targets) in enumerate(train_loader): + # check global steps limit + if (max_global_steps is not None) and (global_step >= max_global_steps): + break + + data_time.update(time.time() - end) + images = images.to(device, non_blocking=True) + targets = targets.to(device, non_blocking=True) + + if args.amp: + with amp.autocast(): + outputs = model(images) + loss = criterion(outputs, targets) / args.grad_accum_steps + else: + outputs = model(images) + loss = criterion(outputs, targets) / args.grad_accum_steps + + if args.amp: + scaler.scale(loss).backward() + else: + loss.backward() + + # 每当累积步满足 grad_accum_steps 就 step + if (i + 1) % args.grad_accum_steps == 0: + if args.amp: + scaler.step(optimizer) + scaler.update() + else: + optimizer.step() + optimizer.zero_grad() + if scheduler is not None and args.scheduler_step_per_batch: + scheduler.step() + + with torch.no_grad(): + acc1, acc5 = accuracy(outputs, targets, topk=(1,5)) + losses.update(loss.item() * args.grad_accum_steps, images.size(0)) + top1.update(acc1.item(), images.size(0)) + top5.update(acc5.item(), images.size(0)) + + batch_time.update(time.time() - end) + end = time.time() + + # increment global step AFTER processing this batch + global_step += 1 + + # per-step print (controlled by print_freq) + if ((global_step % args.print_freq == 0) or (i == iters - 1)) and ((dist.get_rank() if dist.is_initialized() else 0) == 0): + lr = optimizer.param_groups[0]['lr'] + print(f"Epoch[{epoch}]:step[{i+1}/{iters}] step_train_loss {losses.val:.4f} acc1 {top1.val:.2f} acc5 {top5.val:.2f}") + + # collect per-step log + step_logs.append({ + 'epoch': epoch, + 'batch_idx': i, + 'global_step': global_step, + 'lr': optimizer.param_groups[0]['lr'], + 'loss': losses.val, + 'loss_avg': losses.avg, + 'acc1': top1.val, + 'acc1_avg': top1.avg, + 'acc5': top5.val, + 'acc5_avg': top5.avg, + 'time': batch_time.val + }) + + # if reached max_global_steps inside epoch, break (handled at loop start next iter) + if (max_global_steps is not None) and (global_step >= max_global_steps): + if (dist.get_rank() if dist.is_initialized() else 0) == 0: + print(f"[Info] 达到 max_global_steps={max_global_steps}，将在 epoch 内提前停止。") + break + + # --- flush remaining grads if needed (handle gradient accumulation leftovers) --- + processed_batches = global_step - global_step_start # 实际处理的 batch 数 + if args.grad_accum_steps > 1 and (processed_batches % args.grad_accum_steps) != 0: + # only step if there are gradients + grads_present = any((p.grad is not None and p.requires_grad) for p in model.parameters()) + if grads_present: + if args.amp: + try: + scaler.step(optimizer) + scaler.update() + except Exception as e: + # 防御性：若 scaler.step 因某些原因失败，尝试普通 step（只在极端情况下） + print("[Warning] scaler.step 失败，尝试普通 optimizer.step():", e) + optimizer.step() + else: + optimizer.step() + optimizer.zero_grad() + if scheduler is not None and args.scheduler_step_per_batch: + scheduler.step() + if (dist.get_rank() if dist.is_initialized() else 0) == 0: + print(f"[Info] flushed remaining gradients after early stop (processed_batches={processed_batches}, grad_accum={args.grad_accum_steps}).") + + if scheduler is not None and not args.scheduler_step_per_batch: + scheduler.step() + + return OrderedDict([('loss', losses.avg), ('acc1', top1.avg), ('acc5', top5.avg)]), step_logs, global_step + +def validate(args, model, val_loader, criterion, device, max_batches=None): + """ + Validate on the val_loader. + If max_batches is not None, only process up to that many batches (useful for quick checks). + Returns an OrderedDict with loss/acc1/acc5 (averaged over processed samples). + """ + losses = AverageMeter('Loss') + top1 = AverageMeter('Acc@1') + top5 = AverageMeter('Acc@5') + + model.eval() + processed_batches = 0 + processed_samples = 0 + with torch.no_grad(): + for i, (images, targets) in enumerate(tqdm(val_loader)): + images = images.to(device, non_blocking=True) + targets = targets.to(device, non_blocking=True) + outputs = model(images) + loss = criterion(outputs, targets) + acc1, acc5 = accuracy(outputs, targets, topk=(1,5)) + batch_n = images.size(0) + losses.update(loss.item(), batch_n) + top1.update(acc1.item(), batch_n) + top5.update(acc5.item(), batch_n) + + processed_batches += 1 + processed_samples += batch_n + + if (max_batches is not None) and (processed_batches >= max_batches): + break + + # 如果没处理任何样本，避免除0（不太可能，但防御性） + if processed_samples == 0: + return OrderedDict([('loss', 0.0), ('acc1', 0.0), ('acc5', 0.0)]) + return OrderedDict([('loss', losses.avg), ('acc1', top1.avg), ('acc5', top5.avg)]) + +# ---------------------------- +# Main +# ---------------------------- +def main(): + args = parse_args() + + # handle local_rank from env if not provided + local_rank_env = os.environ.get('LOCAL_RANK', None) + if args.local_rank is None and local_rank_env is not None: + args.local_rank = int(local_rank_env) + + distributed = (args.local_rank is not None and args.local_rank != -1) + if distributed: + dist.init_process_group(backend='nccl', init_method='env://') + rank = dist.get_rank() + world_size = dist.get_world_size() + else: + rank = 0 + world_size = 1 + + if distributed: + torch.cuda.set_device(args.local_rank) + device = torch.device('cuda', args.local_rank) + else: + device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') + + set_seed(args.seed + (rank if distributed else 0), deterministic=args.deterministic) + + save_dir = os.path.join('models', args.name) + if rank == 0: + os.makedirs(save_dir, exist_ok=True) + with open(os.path.join(save_dir, 'args.json'), 'w') as f: + json.dump(vars(args), f, indent=2) + if distributed: + dist.barrier() + + train_loader, val_loader, auto_num_classes, train_sampler = build_dataloaders(args, rank, world_size) + if args.num_classes is None: + args.num_classes = auto_num_classes + + # 使用本地 Model 优先（LocalModel 已在文件顶部尝试导入） + model = build_model_with_local_priority(args, device) + model.to(device) + + if distributed: + model = DDP(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True) + + criterion = nn.CrossEntropyLoss().to(device) + params = [p for p in model.parameters() if p.requires_grad] + if args.optimizer == 'sgd': + optimizer = optim.SGD(params, lr=args.lr, momentum=args.momentum, + weight_decay=args.weight_decay, nesterov=args.nesterov) + elif args.optimizer == 'adam': + optimizer = optim.Adam(params, lr=args.lr, weight_decay=args.weight_decay) + elif args.optimizer == 'adamw': + optimizer = optim.AdamW(params, lr=args.lr, weight_decay=args.weight_decay) + else: + raise ValueError('Unknown optimizer') + + scheduler = None + if args.scheduler == 'multistep': + scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.milestones, gamma=args.gamma) + elif args.scheduler == 'step': + scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=args.step_size, gamma=args.gamma) + elif args.scheduler == 'cosine': + scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=args.epochs) + elif args.scheduler == 'none': + scheduler = None + + scaler = amp.GradScaler() if args.amp else None + + start_epoch = args.start_epoch + best_acc = 0.0 + if args.resume: + if os.path.isfile(args.resume): + ckpt = torch.load(args.resume, map_location='cpu') + model_state = ckpt.get('state_dict', ckpt) + if isinstance(model, DDP): + model.module.load_state_dict(model_state) + else: + model.load_state_dict(model_state) + if 'optimizer' in ckpt: + optimizer.load_state_dict(ckpt['optimizer']) + start_epoch = ckpt.get('epoch', start_epoch) + best_acc = ckpt.get('best_acc', best_acc) + print(f"=> resumed from {args.resume}, start_epoch={start_epoch}") + else: + print(f"=> resume path {args.resume} not found") + + log_columns = ['epoch', 'lr', 'loss', 'acc1', 'acc5', 'val_loss', 'val_acc1', 'val_acc5'] + log_df = pd.DataFrame(columns=log_columns) + # step-level log + step_log_columns = ['epoch', 'batch_idx', 'global_step', 'lr', 'loss', 'loss_avg', 'acc1', 'acc1_avg', 'acc5', 'acc5_avg', 'time'] + step_log_df = pd.DataFrame(columns=step_log_columns) + + total_epochs = args.epochs + # global_step计数器（训练过程中跨epoch持续） + global_step = 0 + + epoch = start_epoch + # loop until either epoch criteria or step criteria met + while True: + if train_sampler is not None: + if args.seed_sampler: + train_sampler.set_epoch(epoch + args.seed) + else: + train_sampler.set_epoch(epoch) + + if rank == 0: + print(f"==== Epoch {epoch}/{total_epochs - 1} ====") + + # 如果传入了 args.steps (>0)，则把剩余允许的 step 数传给 train_one_epoch， + # 否则 max_global_steps=None（按整 epoch 执行完） + if args.steps and args.steps > 0: + max_global_steps = args.steps + else: + max_global_steps = None + + train_log, step_logs, global_step = train_one_epoch( + args, epoch, model, criterion, optimizer, train_loader, device, scaler, + scheduler, train_sampler, global_step_start=global_step, max_global_steps=max_global_steps + ) + + # 如果启用了按 steps 的模式且已经达到上限，标记需要在做一次验证后退出 + if max_global_steps is not None and global_step >= max_global_steps: + if rank == 0: + print(f"[Main] 达到 max_global_steps={max_global_steps}（global_step={global_step}），将在完成验证后退出训练。") + # 我们不 return 立刻退出；后面的 validate / 保存逻辑会执行一次，然后 main 返回/结束 + end_due_to_steps = True + else: + end_due_to_steps = False + + # 验证并记录 epoch 级别日志（如果在 step 模式下很可能在中间某个 epoch 提前结束，但我们仍做一次 validate） + val_log = validate(args, model, val_loader, criterion, device, args.batch_size) + current_lr = optimizer.param_groups[0]['lr'] + + if rank == 0: + # epoch summary print, 格式与示例对齐 + print(f"Epoch[{epoch}]: epoch_train_loss {train_log['loss']:.4f} acc1 {train_log['acc1']:.2f} acc5 {train_log['acc5']:.2f} | " + f"val_loss {val_log['loss']:.4f} acc1 {val_log['acc1']:.2f} acc5 {val_log['acc5']:.2f} lr {current_lr:.6f}") + row = { + 'epoch': epoch, + 'lr': current_lr, + 'loss': train_log['loss'], + 'acc1': train_log['acc1'], + 'acc5': train_log['acc5'], + 'val_loss': val_log['loss'], + 'val_acc1': val_log['acc1'], + 'val_acc5': val_log['acc5'], + } + new_row_df = pd.DataFrame([row]) + log_df = pd.concat([log_df, new_row_df], ignore_index=True) + log_df.to_csv(os.path.join(save_dir, 'log.csv'), index=False) + + is_best = val_log['acc1'] > best_acc + if is_best: + best_acc = val_log['acc1'] + if (epoch % args.save_freq == 0) or is_best or ( (max_global_steps is None) and (epoch == total_epochs - 1) ) : + state = { + 'epoch': epoch, + 'state_dict': model.module.state_dict() if isinstance(model, DDP) else model.state_dict(), + 'best_acc': best_acc, + 'optimizer': optimizer.state_dict(), + 'args': vars(args) + } + save_checkpoint(state, is_best, save_dir, filename=f'checkpoint_epoch_{epoch}.pth') + + # 如果是因为 steps 模式达到上限，则在完成 validation / 保存后退出训练 + if end_due_to_steps: + if rank == 0: + print(f"[Main] 已在 steps 模式下完成最后一次验证并保存，训练结束（global_step={global_step}）。") + break + + # increment epoch + epoch += 1 + + # stopping conditions: + # 1) if steps mode enabled and reached steps -> stop + if args.steps and args.steps > 0: + if global_step >= args.steps: + if rank == 0: + print(f"[Main] 已达到指定 steps={args.steps}（global_step={global_step}），训练结束。") + break + + # 2) if steps not used, stop when epoch >= epochs + else: + if epoch >= total_epochs: + if rank == 0: + print(f"[Main] 已达到指定 epochs={total_epochs}（epoch={epoch}），训练结束。") + break + + if dist.is_initialized(): + dist.barrier() + if rank == 0: + print("Training finished. Best val acc1: {:.2f}".format(best_acc)) + +if __name__ == '__main__': + main() \ No newline at end of file diff --git a/PyTorch/build-in/Classification/XCiT/xcit.py b/PyTorch/build-in/Classification/XCiT/xcit.py new file mode 100644 index 000000000..8098b5e99 --- /dev/null +++ b/PyTorch/build-in/Classification/XCiT/xcit.py @@ -0,0 +1,641 @@ +# Copyright (c) 2015-present, Facebook, Inc. +# All rights reserved. +""" +Implementation of Cross-Covariance Image Transformer (XCiT) +Based on timm and DeiT code bases +https://github.com/rwightman/pytorch-image-models/tree/master/timm +https://github.com/facebookresearch/deit/ +""" +import math + +import torch +import torch.nn as nn +from functools import partial + +# 屏蔽 warning +import warnings +warnings.filterwarnings("ignore", category=FutureWarning) + +# 注册模型 +from timm.models.registry import register_model +from timm.models.vision_transformer import _cfg, Mlp +from timm.models.layers import DropPath, trunc_normal_, to_2tuple + + +class PositionalEncodingFourier(nn.Module): + """ + Positional encoding relying on a fourier kernel matching the one used in the + "Attention is all of Need" paper. The implementation builds on DeTR code + https://github.com/facebookresearch/detr/blob/master/models/position_encoding.py + """ + + def __init__(self, hidden_dim=32, dim=768, temperature=10000): + super().__init__() + self.token_projection = nn.Conv2d(hidden_dim * 2, dim, kernel_size=1) + self.scale = 2 * math.pi + self.temperature = temperature + self.hidden_dim = hidden_dim + self.dim = dim + + def forward(self, B, H, W): + mask = torch.zeros(B, H, W).bool().to(self.token_projection.weight.device) + not_mask = ~mask + # y_embed = not_mask.cumsum(1, dtype=torch.float32) + # x_embed = not_mask.cumsum(2, dtype=torch.float32) + + device = not_mask.device # 保存原始设备 + + # 在 CPU 上计算累积和，保证 deterministic + y_embed = not_mask.to('cpu').cumsum(1, dtype=torch.float32).to(device) + x_embed = not_mask.to('cpu').cumsum(2, dtype=torch.float32).to(device) + + eps = 1e-6 + y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale + x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale + + dim_t = torch.arange(self.hidden_dim, dtype=torch.float32, device=mask.device) + dim_t = self.temperature ** (2 * (dim_t // 2) / self.hidden_dim) + + pos_x = x_embed[:, :, :, None] / dim_t + pos_y = y_embed[:, :, :, None] / dim_t + pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), + pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3) + pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), + pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3) + pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) + pos = self.token_projection(pos) + return pos + + +def conv3x3(in_planes, out_planes, stride=1): + """3x3 convolution with padding""" + return torch.nn.Sequential( + nn.Conv2d( + in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False + ), + # nn.SyncBatchNorm(out_planes) + nn.BatchNorm2d(out_planes) # ← 改成这个！ + ) + + +class ConvPatchEmbed(nn.Module): + """ Image to Patch Embedding using multiple convolutional layers + """ + + def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768): + super().__init__() + img_size = to_2tuple(img_size) + patch_size = to_2tuple(patch_size) + num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) + self.img_size = img_size + self.patch_size = patch_size + self.num_patches = num_patches + + if patch_size[0] == 16: + self.proj = torch.nn.Sequential( + conv3x3(3, embed_dim // 8, 2), + nn.GELU(), + conv3x3(embed_dim // 8, embed_dim // 4, 2), + nn.GELU(), + conv3x3(embed_dim // 4, embed_dim // 2, 2), + nn.GELU(), + conv3x3(embed_dim // 2, embed_dim, 2), + ) + elif patch_size[0] == 8: + self.proj = torch.nn.Sequential( + conv3x3(3, embed_dim // 4, 2), + nn.GELU(), + conv3x3(embed_dim // 4, embed_dim // 2, 2), + nn.GELU(), + conv3x3(embed_dim // 2, embed_dim, 2), + ) + elif patch_size[0] == 4: + # For CIFAR-32: produce 8x8 patches from 32x32 input using two stride-2 stages: + # 32 -> 16 -> 8 spatially. Use a compact two-stage conv stack. + self.proj = torch.nn.Sequential( + conv3x3(3, embed_dim // 4, 2), # 32 -> 16 + nn.GELU(), + conv3x3(embed_dim // 4, embed_dim, 2), # 16 -> 8, output embed_dim channels + ) + else: + raise("For convolutional projection, patch size has to be in [8, 16]") + + def forward(self, x, padding_size=None): + B, C, H, W = x.shape + x = self.proj(x) + Hp, Wp = x.shape[2], x.shape[3] + x = x.flatten(2).transpose(1, 2) + + return x, (Hp, Wp) + + +class LPI(nn.Module): + """ + Local Patch Interaction module that allows explicit communication between tokens in 3x3 windows + to augment the implicit communcation performed by the block diagonal scatter attention. + Implemented using 2 layers of separable 3x3 convolutions with GeLU and BatchNorm2d + """ + + def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, + drop=0., kernel_size=3): + super().__init__() + out_features = out_features or in_features + + padding = kernel_size // 2 + + self.conv1 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size, + padding=padding, groups=out_features) + self.act = act_layer() + # self.bn = nn.SyncBatchNorm(in_features) + self.bn = nn.BatchNorm2d(in_features) + self.conv2 = torch.nn.Conv2d(in_features, out_features, kernel_size=kernel_size, + padding=padding, groups=out_features) + + def forward(self, x, H, W): + B, N, C = x.shape + x = x.permute(0, 2, 1).contiguous().reshape(B, C, H, W) + x = self.conv1(x) + x = self.act(x) + x = self.bn(x) + x = self.conv2(x) + x = x.reshape(B, C, N).contiguous().permute(0, 2, 1) + + return x + + +class ClassAttention(nn.Module): + """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239 + """ + + def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + def forward(self, x): + B, N, C = x.shape + qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) + qkv = qkv.permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) + + qc = q[:, :, 0:1] # CLS token + attn_cls = (qc * k).sum(dim=-1) * self.scale + attn_cls = attn_cls.softmax(dim=-1) + attn_cls = self.attn_drop(attn_cls) + + cls_tkn = (attn_cls.unsqueeze(2) @ v).transpose(1, 2).reshape(B, 1, C) + cls_tkn = self.proj(cls_tkn) + x = torch.cat([self.proj_drop(cls_tkn), x[:, 1:]], dim=1) + return x + + +class ClassAttentionBlock(nn.Module): + """Class Attention Layer as in CaiT https://arxiv.org/abs/2103.17239 + """ + + def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., + attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, eta=None, + tokens_norm=False): + super().__init__() + self.norm1 = norm_layer(dim) + + self.attn = ClassAttention( + dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, + proj_drop=drop + ) + + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, + drop=drop) + + if eta is not None: # LayerScale Initialization (no layerscale when None) + self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) + self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) + else: + self.gamma1, self.gamma2 = 1.0, 1.0 + + # FIXME: A hack for models pre-trained with layernorm over all the tokens not just the CLS + self.tokens_norm = tokens_norm + + def forward(self, x, H, W, mask=None): + x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x))) + if self.tokens_norm: + x = self.norm2(x) + else: + # x[:, 0:1] = self.norm2(x[:, 0:1]) #welo + cls_norm = self.norm2(x[:, 0:1]) + x = torch.cat([cls_norm, x[:, 1:]], dim=1).contiguous() + + x_res = x + cls_token = x[:, 0:1] + cls_token = self.gamma2 * self.mlp(cls_token) + x = torch.cat([cls_token, x[:, 1:]], dim=1) + x = x_res + self.drop_path(x) + return x + + +class XCA(nn.Module): + """ Cross-Covariance Attention (XCA) operation where the channels are updated using a weighted + sum. The weights are obtained from the (softmax normalized) Cross-covariance + matrix (Q^T K \\in d_h \\times d_h) + """ + + def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): + super().__init__() + self.num_heads = num_heads + self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1)) + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + def forward(self, x): + B, N, C = x.shape + qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads) + qkv = qkv.permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) + + q = q.transpose(-2, -1) + k = k.transpose(-2, -1) + v = v.transpose(-2, -1) + + q = torch.nn.functional.normalize(q, dim=-1) + k = torch.nn.functional.normalize(k, dim=-1) + + attn = (q @ k.transpose(-2, -1)) * self.temperature + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + + x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + @torch.jit.ignore + def no_weight_decay(self): + return {'temperature'} + + +class XCABlock(nn.Module): + def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., + attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, + num_tokens=196, eta=None): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = XCA( + dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, + proj_drop=drop + ) + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, + drop=drop) + + self.norm3 = norm_layer(dim) + self.local_mp = LPI(in_features=dim, act_layer=act_layer) + + self.gamma1 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) + self.gamma2 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) + self.gamma3 = nn.Parameter(eta * torch.ones(dim), requires_grad=True) + + def forward(self, x, H, W): + x = x + self.drop_path(self.gamma1 * self.attn(self.norm1(x))) + x = x + self.drop_path(self.gamma3 * self.local_mp(self.norm3(x), H, W)) + x = x + self.drop_path(self.gamma2 * self.mlp(self.norm2(x))) + return x + + +class XCiT(nn.Module): + """ + Based on timm and DeiT code bases + https://github.com/rwightman/pytorch-image-models/tree/master/timm + https://github.com/facebookresearch/deit/ + """ + + def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, + depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, + drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None, + cls_attn_layers=2, use_pos=True, patch_proj='linear', eta=None, tokens_norm=False): + """ + Args: + img_size (int, tuple): input image size + patch_size (int, tuple): patch size + in_chans (int): number of input channels + num_classes (int): number of classes for classification head + embed_dim (int): embedding dimension + depth (int): depth of transformer + num_heads (int): number of attention heads + mlp_ratio (int): ratio of mlp hidden dim to embedding dim + qkv_bias (bool): enable bias for qkv if True + qk_scale (float): override default qk scale of head_dim ** -0.5 if set + drop_rate (float): dropout rate + attn_drop_rate (float): attention dropout rate + drop_path_rate (float): stochastic depth rate + norm_layer: (nn.Module): normalization layer + cls_attn_layers: (int) Depth of Class attention layers + use_pos: (bool) whether to use positional encoding + eta: (float) layerscale initialization value + tokens_norm: (bool) Whether to normalize all tokens or just the cls_token in the CA + """ + super().__init__() + self.num_classes = num_classes + self.num_features = self.embed_dim = embed_dim + norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6) + + self.patch_embed = ConvPatchEmbed(img_size=img_size, embed_dim=embed_dim, + patch_size=patch_size) + + num_patches = self.patch_embed.num_patches + + self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) + self.pos_drop = nn.Dropout(p=drop_rate) + + dpr = [drop_path_rate for i in range(depth)] + self.blocks = nn.ModuleList([ + XCABlock( + dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, + qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], + norm_layer=norm_layer, num_tokens=num_patches, eta=eta) + for i in range(depth)]) + + self.cls_attn_blocks = nn.ModuleList([ + ClassAttentionBlock( + dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, + qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, norm_layer=norm_layer, + eta=eta, tokens_norm=tokens_norm) + for i in range(cls_attn_layers)]) + self.norm = norm_layer(embed_dim) + self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity() + + self.pos_embeder = PositionalEncodingFourier(dim=embed_dim) + self.use_pos = use_pos + + # Classifier head + trunc_normal_(self.cls_token, std=.02) + self.apply(self._init_weights) + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + @torch.jit.ignore + def no_weight_decay(self): + return {'pos_embed', 'cls_token', 'dist_token'} + + def forward_features(self, x): + B, C, H, W = x.shape + + x, (Hp, Wp) = self.patch_embed(x) + + if self.use_pos: + pos_encoding = self.pos_embeder(B, Hp, Wp).reshape(B, -1, x.shape[1]).permute(0, 2, 1) + x = x + pos_encoding + + x = self.pos_drop(x) + + for blk in self.blocks: + x = blk(x, Hp, Wp) + + cls_tokens = self.cls_token.expand(B, -1, -1) + x = torch.cat((cls_tokens, x), dim=1) + + for blk in self.cls_attn_blocks: + x = blk(x, Hp, Wp) + + x = self.norm(x)[:, 0] + return x + + def forward(self, x): + x = self.forward_features(x) + x = self.head(x) + + if self.training: + return x + else: + return x + + +# Patch size 16x16 models +@register_model +def xcit_nano_12_p16(pretrained=False, **kwargs): + model = XCiT( + patch_size=16, embed_dim=128, depth=12, num_heads=4, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=False, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_tiny_12_p16(pretrained=False, **kwargs): + model = XCiT( + patch_size=16, embed_dim=192, depth=12, num_heads=4, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_small_12_p16(pretrained=False, **kwargs): + model = XCiT( + patch_size=16, embed_dim=384, depth=12, num_heads=8, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_tiny_24_p16(pretrained=False, **kwargs): + model = XCiT( + patch_size=16, embed_dim=192, depth=24, num_heads=4, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_small_24_p16(pretrained=False, **kwargs): + model = XCiT( + patch_size=16, embed_dim=384, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_medium_24_p16(pretrained=False, **kwargs): + model = XCiT( + patch_size=16, embed_dim=512, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_large_24_p16(pretrained=False, **kwargs): + model = XCiT( + patch_size=16, embed_dim=768, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +# Patch size 8x8 models +@register_model +def xcit_nano_12_p8(pretrained=False, **kwargs): + model = XCiT( + patch_size=8, embed_dim=128, depth=12, num_heads=4, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=False, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_tiny_12_p8(pretrained=False, **kwargs): + model = XCiT( + patch_size=8, embed_dim=192, depth=12, num_heads=4, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_small_12_p8(pretrained=False, **kwargs): + model = XCiT( + patch_size=8, embed_dim=384, depth=12, num_heads=8, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1.0, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_tiny_24_p8(pretrained=False, **kwargs): + model = XCiT( + patch_size=8, embed_dim=192, depth=24, num_heads=4, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_small_24_p8(pretrained=False, **kwargs): + model = XCiT( + patch_size=8, embed_dim=384, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_medium_24_p8(pretrained=False, **kwargs): + model = XCiT( + patch_size=8, embed_dim=512, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +@register_model +def xcit_large_24_p8(pretrained=False, **kwargs): + model = XCiT( + patch_size=8, embed_dim=768, depth=24, num_heads=16, mlp_ratio=4, qkv_bias=True, + norm_layer=partial(nn.LayerNorm, eps=1e-6), eta=1e-5, tokens_norm=True, **kwargs) + model.default_cfg = _cfg() + return model + + +# welo add +# --- helper to ensure divisible dims (add this near the top of xcit.py, after imports) --- +def ensure_divisible(value, divisor): + """Return the smallest integer >= value that is divisible by divisor.""" + if divisor <= 0: + raise ValueError("divisor must be > 0") + return int((int(value) + divisor - 1) // divisor) * divisor + +# --- 注册用于 CIFAR-32 的 XCiT 变体 --- +@register_model +def xcit_cifar32_patch4(pretrained=False, num_classes=10, *, # 默认 CIFAR-10 的 num_classes=10，可传 100 用于 CIFAR-100 + img_size=32, patch_size=4, + embed_dim=128, depth=6, num_heads=4, + mlp_ratio=2., qkv_bias=True, qk_scale=None, + drop_rate=0., attn_drop_rate=0., drop_path_rate=0.0, + cls_attn_layers=1, use_pos=True, eta=1e-5, tokens_norm=False, + **kwargs): + """ + XCiT variant for CIFAR-32. + - Default: patch_size=4 -> (32/4)^2 = 8x8 = 64 patches + - Lightweight: smaller embed_dim and depth suitable for CIFAR datasets. + - If you want CIFAR-100, call with num_classes=100. + """ + # basic sanity checks and adjustments + if img_size % patch_size != 0: + raise ValueError(f"img_size ({img_size}) must be divisible by patch_size ({patch_size}).") + if patch_size not in (4, 8, 16): + raise ValueError("xcit_cifar32_patch4 supports patch_size 4, 8 or 16 (for CIFAR/ImageNet use).") + + # ensure embed_dim divisible by num_heads + embed_dim = ensure_divisible(embed_dim, num_heads) + + # Create model with lightweight defaults + model = XCiT( + img_size=img_size, + patch_size=patch_size, + in_chans=3, + num_classes=num_classes, + embed_dim=embed_dim, + depth=depth, + num_heads=num_heads, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop_rate=drop_rate, + attn_drop_rate=attn_drop_rate, + drop_path_rate=drop_path_rate, + norm_layer=partial(nn.LayerNorm, eps=1e-6), + cls_attn_layers=cls_attn_layers, + use_pos=use_pos, + eta=eta, + tokens_norm=tokens_norm, + patch_proj='conv', # informational + **kwargs + ) + + # Use the (updated) ConvPatchEmbed implementation (supports patch_size 4/8/16) + model.patch_embed = ConvPatchEmbed(img_size=img_size, patch_size=patch_size, in_chans=3, embed_dim=embed_dim) + + # default cfg suitable for CIFAR datasets + model.default_cfg = _cfg( + url='', + input_size=(3, img_size, img_size), + num_classes=num_classes, + mean=(0.4914, 0.4822, 0.4465), + std=(0.2470, 0.2435, 0.2616), + crop_pct=1.0, + interpolation='bicubic' + ) + + if pretrained: + raise NotImplementedError("Pretrained weights not available for CIFAR XCiT variant") + + return model + + +# convenience factory expected by your external code: Model(100) -> CIFAR-100 XCiT +def Model(num_classes): + r"""Return XCiT model customized for CIFAR. Usage: Model(100) for CIFAR-100.""" + return xcit_cifar32_patch4(pretrained=False, num_classes=num_classes) \ No newline at end of file