vit for bmtrain

[qyc-98]

`import torch import torch.nn as nn import torch.nn.functional as F from itertools import repeat import collections.abc import bmtrain as bmt from model_center.layer import LayerNorm from functools import partial from timm.models.layers import truncnormal, DropPath try: from torch import _assert except ImportError: def _assert(condition:bool, message:str): assert condition, message

From PyTorch internals

def _ntuple(n): def parse(x): if isinstance(x, collections.abc.Iterable): return x return tuple(repeat(x, n)) return parse

to_2tuple = _ntuple(2)

class Identity(bmt.DistributedModule): def init(self, *args, **kwargs): super(Identity, self).init()

def forward(self, input):
    return input

class Conv2d(bmt.DistributedModule): def init(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, dtype=torch.float, int8: bool=False, init_mean : float=0.0, init_std : float = 1, bias : bool=True, padding_mode='zeros', ): super().init() self.in_channels = in_channels self.out_channels = out_channels self.kernel_size = kernel_size self.transposed = None self.output_padding = None

    self.stride = stride
    self.dilation = dilation
    self.groups = groups
    self.padding = padding
    self.padding_mode = padding_mode

    kernel = to_2tuple(kernel_size)
    self.weight = bmt.DistributedParameter(
        torch.empty((out_channels, int(in_channels/groups), kernel[0], kernel[1]), dtype=dtype),
        init_method=bmt.ParameterInitializer(torch.nn.init.normal_, mean=init_mean, std=init_std)
    )
    self.bias = bmt.DistributedParameter(
        torch.empty((out_channels,), dtype=dtype),
        init_method=bmt.ParameterInitializer(torch.nn.init.zeros_)
    ) if bias else None
    self.int8=int8

def forward(self, x : torch.Tensor):
    x = F.conv2d(x,
                weight=self.weight, 
                bias=self.bias, 
                stride=self.stride,
                padding=self.padding,
                dilation=self.dilation,
                groups=self.groups,
                )

    return x

class Linear(bmt.DistributedModule): r"""A fully connected layer, which performs :math:\pmb{y} = \mathbf{W} \pmb{x} + \pmb{b} Args: dim_in (int): input dimension of :math:\pmb{x} dim_out (int): output dimension of :math:\pmb{y} dtype (optional): Defaults to torch.half. init_mean (float, optional): mean of :math:\mathbf{W}\sim\mathcal{N}(\text{mean}, \text{std}^2). Defaults to 0. init_std (float, optional): std of :math:\mathbf{W}\sim\mathcal{N}(\text{mean}, \text{std}^2). Defaults to 1. bias (bool, optional): whether to add bias term :math:\pmb{b}. Defaults to False. """ def init(self, in_features : int, out_features : int, length_scale : bool = False, length_scale_before : bool = False, dtype = torch.float, int8 : bool = False, init_mean : float = 0.0, init_std : float = 1, bias : bool = True, ): super().init() self.in_features = in_features self.weight = bmt.DistributedParameter( torch.empty((out_features, in_features), dtype=dtype), initmethod=bmt.ParameterInitializer(torch.nn.init.normal, mean=init_mean, std=init_std) ) self.bias = bmt.DistributedParameter( torch.empty((out_features,), dtype=dtype), initmethod=bmt.ParameterInitializer(torch.nn.init.zeros) ) if bias else None self.length_scale = length_scale self.length_scale_before = length_scale_before self.int8 = int8

def forward(self, x : torch.Tensor):
    """ 
    Args:
        x (:obj:`torch.Tensor` of shape ``(batch, seq_len, dim_in)``): The input of linear layer
    Returns:
        :obj:`torch.Tensor` of shape ``(batch, seq_len, dim_out)``: The output of the linear transform y.
    """
    if self.length_scale and self.length_scale_before:
        x = x / math.sqrt(self.in_features)
    x = F.linear(x, self.weight)
    if self.length_scale and not self.length_scale_before:
        x = x / math.sqrt(self.in_features)
    if self.bias is not None:
        x = x + self.bias
    return x

class DropPath(nn.Module): """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). """ def init(self, drop_prob: float = 0., scale_by_keep: bool = True): super(DropPath, self).init() self.drop_prob = drop_prob self.scale_by_keep = scale_by_keep

def forward(self, x):
    return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)

class PatchEmbed(bmt.DistributedModule): """ 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, dtype=torch.half ): 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 = Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, dtype=dtype) self.norm = norm_layer(embed_dim) if norm_layer else Identity

def forward(self, x):
    B,C,H,W = x.shape
    _assert(H == self.img_size[0], f"Input image height ({H}) doesn't match model ({self.img_size[0]}).")
    _assert(W == self.img_size[1], f"Input image width ({W}) doesn't match model ({self.img_size[1]}).")
    x = self.proj(x)
    if self.flatten:
        x = x.flatten(2).transpose(1, 2)
    # x = self.norm(x)
    return x

class Mlp(bmt.DistributedModule): """ MLP as used in Vision Transformer, MLP-Mixer and related networks """ def init(self, in_features, hidden_features=None, out_features=None, act_layer=torch.nn.functional.gelu, drop=0.0, dtype=torch.half ): super().init() out_features = out_features or in_features hidden_features = hidden_features or in_features self.fc1 = Linear(in_features, hidden_features, dtype=dtype) self.act = act_layer self.fc2 = Linear(hidden_features, out_features, dtype=dtype) self.drop = nn.Dropout(drop)

def forward(self, x):
    x = self.fc1(x)
    x = self.act(x)
    x = self.drop(x)
    x = self.fc2(x)
    x = self.drop(x)
    return x

class Attention(bmt.DistributedModule): def init(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., length_scale=False, dtype=torch.float, int8=False, init_mean : float = 0.0, init_std : float = 1, bias : bool = True, ): super().init() self.num_heads = num_heads head_dim = dim // num_heads

NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights

    self.scale = qk_scale or head_dim ** -0.5
    self.attn_drop = nn.Dropout(attn_drop)
    self.proj_drop = nn.Dropout(proj_drop)
    self.attn_gradients = None
    self.attention_map = None

    self.qkv = Linear(
        in_features = dim,
        out_features = dim*3,
        bias=qkv_bias,
        length_scale=length_scale,
        length_scale_before=False,
        dtype=dtype,
        int8=int8,
        init_mean=init_mean,
        init_std=init_std,
    )

    self.proj = Linear(
        in_features = dim,
        out_features = dim,
        bias=qkv_bias,
        length_scale=length_scale,
        length_scale_before=False,
        dtype=dtype,
        int8=int8,
        init_mean=init_mean,
        init_std=init_std,
    )

def save_attn_gradients(self, attn_gradients):
    self.attn_gradients = attn_gradients

def get_attn_gradients(self):
    return self.attn_gradients

def save_attention_map(self, attention_map):
    self.attention_map = attention_map

def get_attention_map(self):
    return self.attention_map

def forward(self, x, register_hook=False):
    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)
    q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)

    attn = (q @ k.transpose(-2, -1)) * self.scale
    attn = attn.softmax(dim=-1)
    attn = self.attn_drop(attn)

    if register_hook:
        self.save_attention_map(attn)
        attn.register_hook(self.save_attn_gradients)        

    x = (attn @ v).transpose(1, 2).reshape(B, N, C)
    x = self.proj(x)
    x = self.proj_drop(x)
    return x

class Block(bmt.DistributedModule):

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=torch.nn.functional.gelu, norm_layer=LayerNorm,dtype=torch.float):
    super().__init__()
    self.norm1 = norm_layer(dim)
    self.attn = Attention(
        dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop,dtype=dtype)
    # 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 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,dtype=dtype)

def forward(self, x, register_hook=False):
    x = x + self.drop_path(self.attn(self.norm1(x), register_hook=register_hook))
    x = x + self.drop_path(self.mlp(self.norm2(x)))
    return x

class VisionTransformer(bmt.DistributedModule): """ Vision Transformer A PyTorch impl of : An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale - https://arxiv.org/abs/2010.11929 """ 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, representation_size=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None, dtype=torch.float): """ 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 representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set drop_rate (float): dropout rate attn_drop_rate (float): attention dropout rate drop_path_rate (float): stochastic depth rate norm_layer: (bmt.DistributedModule): normalization layer dtype: Defaults to torch.float.

    """
    super().__init__()
    self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
    norm_layer = norm_layer or LayerNorm
    self.patch_embed = PatchEmbed(
        img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, dtype=dtype)
    num_patches = self.patch_embed.num_patches

    self.cls_token = bmt.DistributedParameter(torch.empty((1,1,embed_dim), dtype=dtype))
    self.pos_embed = bmt.DistributedParameter(torch.empty((1,num_patches+1,embed_dim), dtype=dtype))
    self.pos_drop = nn.Dropout(p=drop_rate)

    dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule

    self.blocks =  bmt.TransformerBlockList(
                    [
                        bmt.CheckpointBlock(
                            Block(
                                  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, 
                                  dtype=dtype
                                )
                            ) for i in range(depth)
                        ]
                )
    self.norm = norm_layer(embed_dim)
    self.head = Linear(embed_dim, num_classes, dtype=dtype)
@torch.jit.ignore
def no_weight_decay(self):
    return {'pos_embed', 'cls_token'}

def forward(self, x, register_blk=-1):
    B = x.shape[0]
    x = self.patch_embed(x)

    cls_tokens = self.cls_token.expand(B, 1, -1)  # stole cls_tokens impl from Phil Wang, thanks
    x = torch.cat((cls_tokens, x), dim=1)
    x = x + self.pos_embed[:,:x.size(1),:]
    x = self.pos_drop(x)
    x = self.blocks(x)
    x = self.norm(x)
    x = self.head(x[:,0])
    return x

def interpolate_pos_embed(pos_embed_checkpoint, visual_encoder):

interpolate position embedding

embedding_size = pos_embed_checkpoint.shape[-1]
num_patches = visual_encoder.patch_embed.num_patches
num_extra_tokens = visual_encoder.pos_embed.shape[-2] - num_patches
# height (== width) for the checkpoint position embedding
orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
# height (== width) for the new position embedding
new_size = int(num_patches ** 0.5)
if orig_size!=new_size:
    # class_token and dist_token are kept unchanged
    extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
    # only the position tokens are interpolated
    pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
    pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
    pos_tokens = torch.nn.functional.interpolate(
        pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
    pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
    new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
    print('reshape position embedding from %d to %d'%(orig_size ** 2,new_size ** 2))

    return new_pos_embed    
else:
    return pos_embed_checkpoint

if name == 'main': import bmtrain as bmt from functools import partial from model_center.layer import LayerNorm bmt.init_distributed(seed=0) vit = VisionTransformer(img_size=256, patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True, norm_layer=partial(LayerNorm,dtype=torch.float,eps=1e-6), dtype=torch.float ) checkpoint = torch.hub.load_state_dict_from_url( url="https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth", map_location="cpu", check_hash=True) state_dict = checkpoint["model"] pos_embed_reshaped = interpolate_pos_embed(state_dict['pos_embed'], vit) state_dict['pos_embed'] = pos_embed_reshaped state_dict = bmt.store.DistributedStateDictWrapper(state_dict) msg = vit.load_state_dict(state_dict,strict=False)

`