##################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021 #
##################################################
from __future__ import division
from __future__ import print_function
import math
from functools import partial
from typing import Optional, Text
import torch
import torch.nn as nn
import torch.nn.functional as F
import layers as xlayers
DEFAULT_NET_CONFIG = dict(
d_feat=6,
embed_dim=64,
depth=5,
num_heads=4,
mlp_ratio=4.0,
qkv_bias=True,
pos_drop=0.0,
mlp_drop_rate=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.0,
)
# Real Model
class Attention(nn.Module):
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.0, proj_drop=0.0):
super(Attention, self).__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or math.sqrt(head_dim)
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)
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)
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=4.0,
qkv_bias=False,
qk_scale=None,
attn_drop=0.0,
mlp_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
):
super(Block, self).__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=mlp_drop
)
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
self.drop_path = xlayers.DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = xlayers.MLP(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=mlp_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
class SimpleEmbed(nn.Module):
def __init__(self, d_feat, embed_dim):
super(SimpleEmbed, self).__init__()
self.d_feat = d_feat
self.embed_dim = embed_dim
self.proj = nn.Linear(d_feat, embed_dim)
def forward(self, x):
x = x.reshape(len(x), self.d_feat, -1) # [N, F*T] -> [N, F, T]
x = x.permute(0, 2, 1) # [N, F, T] -> [N, T, F]
out = self.proj(x) * math.sqrt(self.embed_dim)
return out
class TransformerModel(nn.Module):
def __init__(
self,
d_feat: int = 6,
embed_dim: int = 64,
depth: int = 4,
num_heads: int = 4,
mlp_ratio: float = 4.0,
qkv_bias: bool = True,
qk_scale: Optional[float] = None,
pos_drop: float = 0.0,
mlp_drop_rate: float = 0.0,
attn_drop_rate: float = 0.0,
drop_path_rate: float = 0.0,
norm_layer: Optional[nn.Module] = None,
max_seq_len: int = 65,
):
"""
Args:
d_feat (int, tuple): input image size
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
pos_drop (float): dropout rate for the positional embedding
mlp_drop_rate (float): the dropout rate for MLP layers in a block
attn_drop_rate (float): attention dropout rate
drop_path_rate (float): stochastic depth rate
norm_layer: (nn.Module): normalization layer
"""
super(TransformerModel, self).__init__()
self.embed_dim = embed_dim
self.num_features = embed_dim
norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
self.input_embed = SimpleEmbed(d_feat, embed_dim=embed_dim)
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.pos_embed = xlayers.PositionalEncoder(d_model=embed_dim, max_seq_len=max_seq_len, dropout=pos_drop)
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule
self.blocks = nn.ModuleList(
[
Block(
dim=embed_dim,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop=attn_drop_rate,
mlp_drop=mlp_drop_rate,
drop_path=dpr[i],
norm_layer=norm_layer,
)
for i in range(depth)
]
)
self.norm = norm_layer(embed_dim)
# regression head
self.head = nn.Linear(self.num_features, 1)
xlayers.trunc_normal_(self.cls_token, std=0.02)
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
xlayers.trunc_normal_(m.weight, std=0.02)
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)
def forward_features(self, x):
batch, flatten_size = x.shape
feats = self.input_embed(x) # batch * 60 * 64
cls_tokens = self.cls_token.expand(batch, -1, -1) # stole cls_tokens impl from Phil Wang, thanks
feats_w_ct = torch.cat((cls_tokens, feats), dim=1)
feats_w_tp = self.pos_embed(feats_w_ct)
xfeats = feats_w_tp
for block in self.blocks:
xfeats = block(xfeats)
xfeats = self.norm(xfeats)[:, 0]
return xfeats
def forward(self, x):
feats = self.forward_features(x)
predicts = self.head(feats).squeeze(-1)
return predicts
def get_transformer(config):
if not isinstance(config, dict):
raise ValueError("Invalid Configuration: {:}".format(config))
name = config.get("name", "basic")
if name == "basic":
model = TransformerModel(
d_feat=config.get("d_feat"),
embed_dim=config.get("embed_dim"),
depth=config.get("depth"),
num_heads=config.get("num_heads"),
mlp_ratio=config.get("mlp_ratio"),
qkv_bias=config.get("qkv_bias"),
qk_scale=config.get("qkv_scale"),
pos_drop=config.get("pos_drop"),
mlp_drop_rate=config.get("mlp_drop_rate"),
attn_drop_rate=config.get("attn_drop_rate"),
drop_path_rate=config.get("drop_path_rate"),
norm_layer=config.get("norm_layer", None),
)
else:
raise ValueError("Unknown model name: {:}".format(name))
return model