MAE代码阅读（一）
参数
parser = argparse.ArgumentParser('MAE pre-training', add_help=False)
parser.add_argument('--batch_size', default=64, type=int,
help='Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus')
parser.add_argument('--epochs', default=400, type=int)
parser.add_argument('--accum_iter', default=1, type=int,
help='Accumulate gradient iterations (for increasing the effective batch size under memory constraints)')
# Model parameters
parser.add_argument('--model', default='mae_vit_large_patch16', type=str, metavar='MODEL',
help='Name of model to train')
parser.add_argument('--input_size', default=224, type=int,
help='images input size')
parser.add_argument('--mask_ratio', default=0.75, type=float,
help='Masking ratio (percentage of removed patches).')
parser.add_argument('--norm_pix_loss', action='store_true',
help='Use (per-patch) normalized pixels as targets for computing loss')
parser.set_defaults(norm_pix_loss=False)
# Optimizer parameters
parser.add_argument('--weight_decay', type=float, default=0.05,
help='weight decay (default: 0.05)')
parser.add_argument('--lr', type=float, default=None, metavar='LR',
help='learning rate (absolute lr)')
parser.add_argument('--blr', type=float, default=1e-3, metavar='LR',
help='base learning rate: absolute_lr = base_lr * total_batch_size / 256')
parser.add_argument('--min_lr', type=float, default=0., metavar='LR',
help='lower lr bound for cyclic schedulers that hit 0')
parser.add_argument('--warmup_epochs', type=int, default=40, metavar='N',
help='epochs to warmup LR')
# Dataset parameters
parser.add_argument('--data_path', default='/datasets01/imagenet_full_size/061417/', type=str,
help='dataset path')
parser.add_argument('--output_dir', default='./output_dir',
help='path where to save, empty for no saving')
parser.add_argument('--log_dir', default='./output_dir',
help='path where to tensorboard log')
parser.add_argument('--device', default='cuda',
help='device to use for training / testing')
parser.add_argument('--seed', default=0, type=int)
parser.add_argument('--resume', default='',
help='resume from checkpoint')
parser.add_argument('--start_epoch', default=0, type=int, metavar='N',
help='start epoch')
parser.add_argument('--num_workers', default=10, type=int)
parser.add_argument('--pin_mem', action='store_true',
help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')
parser.add_argument('--no_pin_mem', action='store_false', dest='pin_mem')
parser.set_defaults(pin_mem=True)
# distributed training parameters
parser.add_argument('--world_size', default=1, type=int,
help='number of distributed processes')
parser.add_argument('--local_rank', default=-1, type=int)
parser.add_argument('--dist_on_itp', action='store_true')
parser.add_argument('--dist_url', default='env://',
help='url used to set up distributed training')
模型（以mae_vit_large_patch16为例）
def mae_vit_large_patch16_dec512d8b(**kwargs):
model = MaskedAutoencoderViT(
patch_size=16, embed_dim=1024, depth=24, num_heads=16,
decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
mlp_ratio=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
return model
main_pretrain
param_groups = optim_factory.add_weight_decay(model_without_ddp, args.weight_decay)
optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))
#作者使用AdamW作为优化器。常用的adam 优化器虽然收敛速度快，但是一定程度上存在参数过拟合的问题。AdamW的出现就为了解决这个问题，根据一些实验，作者表示该算法是目前训练神经网络最快的方式（AdamW and Super-convergence is now the fastest way to train neural nets）。
print(optimizer)
loss_scaler = NativeScaler()#定义了混合精度训练
misc.load_model(args=args, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler)
print(f"Start training for {args.epochs} epochs")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
data_loader_train.sampler.set_epoch(epoch)
train_stats = train_one_epoch(
model, data_loader_train,
optimizer, device, epoch, loss_scaler,
log_writer=log_writer,
args=args
)
if args.output_dir and (epoch % 20 == 0 or epoch + 1 == args.epochs):
misc.save_model(
args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
loss_scaler=loss_scaler, epoch=epoch)
log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
'epoch': epoch,}
if args.output_dir and misc.is_main_process():
if log_writer is not None:
log_writer.flush()
with open(os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8") as f:
f.write(json.dumps(log_stats) + "\n")
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
train_one_epoch
这里是逐个epoch训练的代码
def train_one_epoch(model: torch.nn.Module,
data_loader: Iterable, optimizer: torch.optim.Optimizer,
device: torch.device, epoch: int, loss_scaler,
log_writer=None,
args=None):
model.train(True)
metric_logger = misc.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', misc.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
print_freq = 20
accum_iter = args.accum_iter
optimizer.zero_grad()
if log_writer is not None:
print('log_dir: {}'.format(log_writer.log_dir))
for data_iter_step, (samples, _) in enumerate(metric_logger.log_every(data_loader, print_freq, header)):
# we use a per iteration (instead of per epoch) lr scheduler
if data_iter_step % accum_iter == 0:
lr_sched.adjust_learning_rate(optimizer, data_iter_step / len(data_loader) + epoch, args)
samples = samples.to(device, non_blocking=True)
with torch.cuda.amp.autocast():#混合精度训练
loss, _, _ = model(samples, mask_ratio=args.mask_ratio)
loss_value = loss.item()
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
sys.exit(1)
loss /= accum_iter
loss_scaler(loss, optimizer, parameters=model.parameters(),
update_grad=(data_iter_step + 1) % accum_iter == 0)
if (data_iter_step + 1) % accum_iter == 0:
optimizer.zero_grad()
torch.cuda.synchronize()
#等待当前设备上所有流中的所有核心完成。
#因为在pytorch里面，程序的执行都是异步的。代码会同步cu的操作，等待gpu上的操作都完成了再继续成形。
metric_logger.update(loss=loss_value)
lr = optimizer.param_groups[0]["lr"]
metric_logger.update(lr=lr)
loss_value_reduce = misc.all_reduce_mean(loss_value)
if log_writer is not None and (data_iter_step + 1) % accum_iter == 0:
""" We use epoch_1000x as the x-axis in tensorboard.
This calibrates different curves when batch size changes.
"""
epoch_1000x = int((data_iter_step / len(data_loader) + epoch) * 1000)
log_writer.add_scalar('train_loss', loss_value_reduce, epoch_1000x)
log_writer.add_scalar('lr', lr, epoch_1000x)
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
half-cycle cosine after warmup
def adjust_learning_rate(optimizer, epoch, args):
"""Decay the learning rate with half-cycle cosine after warmup"""
if epoch < args.warmup_epochs:
lr = args.lr * epoch / args.warmup_epochs
else:
lr = args.min_lr + (args.lr - args.min_lr) * 0.5 * \
(1. + math.cos(math.pi * (epoch - args.warmup_epochs) / (args.epochs - args.warmup_epochs)))
for param_group in optimizer.param_groups:
if "lr_scale" in param_group:
param_group["lr"] = lr * param_group["lr_scale"]
else:
param_group["lr"] = lr
return lr
MaskedAutoencoderViT
class MaskedAutoencoderViT(nn.Module):
""" Masked Autoencoder with VisionTransformer backbone
"""
def __init__(self, img_size=224, patch_size=16, in_chans=3,
embed_dim=1024, depth=24, num_heads=16,
decoder_embed_dim=512, decoder_depth=8, decoder_num_heads=16,
mlp_ratio=4., norm_layer=nn.LayerNorm, norm_pix_loss=False):
super().__init__()
# --------------------------------------------------------------------------
# MAE encoder specifics
self.patch_embed = PatchEmbed(img_size, patch_size, in_chans, embed_dim)
num_patches = self.patch_embed.num_patches
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim), requires_grad=False) # fixed sin-cos embedding
self.blocks = nn.ModuleList([
Block(embed_dim, num_heads, mlp_ratio, qkv_bias=True, qk_scale=None, norm_layer=norm_layer)
for i in range(depth)])
#self.blocks属于encoder
self.norm = norm_layer(embed_dim)
# --------------------------------------------------------------------------
# --------------------------------------------------------------------------
# MAE decoder specifics
self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True)
self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim))
self.decoder_pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, decoder_embed_dim), requires_grad=False) # fixed sin-cos embedding
self.decoder_blocks = nn.ModuleList([
Block(decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True, qk_scale=None, norm_layer=norm_layer)
for i in range(decoder_depth)])
self.decoder_norm = norm_layer(decoder_embed_dim)
self.decoder_pred = nn.Linear(decoder_embed_dim, patch_size**2 * in_chans, bias=True) # decoder to patch
# --------------------------------------------------------------------------
self.norm_pix_loss = norm_pix_loss
self.initialize_weights()
def initialize_weights(self):
# initialization
# initialize (and freeze) pos_embed by sin-cos embedding
pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.patch_embed.num_patches**.5), cls_token=True)
self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
decoder_pos_embed = get_2d_sincos_pos_embed(self.decoder_pos_embed.shape[-1], int(self.patch_embed.num_patches**.5), cls_token=True)
self.decoder_pos_embed.data.copy_(torch.from_numpy(decoder_pos_embed).float().unsqueeze(0))
# initialize patch_embed like nn.Linear (instead of nn.Conv2d)
w = self.patch_embed.proj.weight.data
torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
# timm's trunc_normal_(std=.02) is effectively normal_(std=0.02) as cutoff is too big (2.)
torch.nn.init.normal_(self.cls_token, std=.02)
torch.nn.init.normal_(self.mask_token, std=.02)
# initialize nn.Linear and nn.LayerNorm
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
# we use xavier_uniform following official JAX ViT:
torch.nn.init.xavier_uniform_(m.weight)
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)
def patchify(self, imgs):
"""
imgs: (N, 3, H, W)
x: (N, L, patch_size**2 *3)
"""
p = self.patch_embed.patch_size[0]
assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0
h = w = imgs.shape[2] // p
x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))
x = torch.einsum('nchpwq->nhwpqc', x)
x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))
return x
def unpatchify(self, x):
"""
x: (N, L, patch_size**2 *3)
imgs: (N, 3, H, W)
"""
p = self.patch_embed.patch_size[0]
h = w = int(x.shape[1]**.5)
assert h * w == x.shape[1]
x = x.reshape(shape=(x.shape[0], h, w, p, p, 3))
x = torch.einsum('nhwpqc->nchpwq', x)
imgs = x.reshape(shape=(x.shape[0], 3, h * p, h * p))
return imgs
def random_masking(self, x, mask_ratio):
"""
Perform per-sample random masking by per-sample shuffling.
Per-sample shuffling is done by argsort random noise.
x: [N, L, D], sequence
"""
N, L, D = x.shape # batch, length, dim
len_keep = int(L * (1 - mask_ratio))
noise = torch.rand(N, L, device=x.device) # noise in [0, 1]
# sort noise for each sample
ids_shuffle = torch.argsort(noise, dim=1) # ascend: small is keep, large is remove
ids_restore = torch.argsort(ids_shuffle, dim=1)
# keep the first subset
ids_keep = ids_shuffle[:, :len_keep]
x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
# generate the binary mask: 0 is keep, 1 is remove
mask = torch.ones([N, L], device=x.device)
mask[:, :len_keep] = 0
# unshuffle to get the binary mask
mask = torch.gather(mask, dim=1, index=ids_restore)
return x_masked, mask, ids_restore
def forward_encoder(self, x, mask_ratio):
# embed patches
x = self.patch_embed(x)
# add pos embed w/o cls token
x = x + self.pos_embed[:, 1:, :]
# masking: length -> length * mask_ratio
x, mask, ids_restore = self.random_masking(x, mask_ratio)
# append cls token
cls_token = self.cls_token + self.pos_embed[:, :1, :]
cls_tokens = cls_token.expand(x.shape[0], -1, -1)
x = torch.cat((cls_tokens, x), dim=1)
# apply Transformer blocks
for blk in self.blocks:
x = blk(x)
x = self.norm(x)
return x, mask, ids_restore
def forward_decoder(self, x, ids_restore):
# embed tokens
x = self.decoder_embed(x)
# append mask tokens to sequence
mask_tokens = self.mask_token.repeat(x.shape[0], ids_restore.shape[1] + 1 - x.shape[1], 1)
x_ = torch.cat([x[:, 1:, :], mask_tokens], dim=1) # no cls token
x_ = torch.gather(x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2])) # unshuffle
x = torch.cat([x[:, :1, :], x_], dim=1) # append cls token
# add pos embed
x = x + self.decoder_pos_embed
# apply Transformer blocks
for blk in self.decoder_blocks:
x = blk(x)
x = self.decoder_norm(x)
# predictor projection
x = self.decoder_pred(x)
# remove cls token
x = x[:, 1:, :]
return x
def forward_loss(self, imgs, pred, mask):
"""
imgs: [N, 3, H, W]
pred: [N, L, p*p*3]
mask: [N, L], 0 is keep, 1 is remove,
"""
target = self.patchify(imgs)
if self.norm_pix_loss:
mean = target.mean(dim=-1, keepdim=True)
var = target.var(dim=-1, keepdim=True)
target = (target - mean) / (var + 1.e-6)**.5
loss = (pred - target) ** 2
loss = loss.mean(dim=-1) # [N, L], mean loss per patch
loss = (loss * mask).sum() / mask.sum() # mean loss on removed patches
return loss
def forward(self, imgs, mask_ratio=0.75):
latent, mask, ids_restore = self.forward_encoder(imgs, mask_ratio)
pred = self.forward_decoder(latent, ids_restore) # [N, L, p*p*3]
loss = self.forward_loss(imgs, pred, mask)
return loss, pred, mask
PatchEmbed
PatchEmbed将传进来的B3224224切成B7681414
class PatchEmbed(nn.Module):
""" 2D Image to Patch Embedding
"""
def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
self.img_size = img_size
self.patch_size = patch_size
self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
self.num_patches = self.grid_size[0] * self.grid_size[1]
self.flatten = flatten
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x):
B, C, H, W = x.shape
assert H == self.img_size[0] and W == self.img_size[1], \
f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
x = self.proj(x)#B*3*224*224->B*768*14*14
if self.flatten:
x = x.flatten(2).transpose(1, 2) # BCHW -> BNC B*768*14*14->B*768*196->B*196*768
x = self.norm(x)
return x
Block
class Attention(nn.Module):
def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = 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).permute(2, 0, 3, 1, 4)
#B*N*C - >B*N*(3*C) - > B*N*3*num_heads*（C//num_heads) - > 3 * B * num_heads * N * (C//num_heads)
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
#B * num_heads * N * (C//num_heads) || B * num_heads * (C//num_heads) * N 表示常规的数学上定义的矩阵相乘
attn = (q @ k.transpose(-2, -1)) * self.scale #B * num_heads * N * N
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
# B * num_heads * N * N || B * num_heads * N * (C//num_heads) - > B*num_heads*N*(C//num_heads) - > B*N*num_heads*(C//num_heads)->B*N*C
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class Block(nn.Module):
def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0.,
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
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)
def forward(self, x):
x = x + self.drop_path(self.attn(self.norm1(x)))
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x