Anitalker / model /blocks.py
Delik's picture
Upload 32 files
f1ea451 verified
raw
history blame
18.7 kB
import math
from abc import abstractmethod
from dataclasses import dataclass
from numbers import Number
import torch as th
import torch.nn.functional as F
from choices import *
from config_base import BaseConfig
from torch import nn
from .nn import (avg_pool_nd, conv_nd, linear, normalization,
timestep_embedding, torch_checkpoint, zero_module)
class ScaleAt(Enum):
after_norm = 'afternorm'
class TimestepBlock(nn.Module):
"""
Any module where forward() takes timestep embeddings as a second argument.
"""
@abstractmethod
def forward(self, x, emb=None, cond=None, lateral=None):
"""
Apply the module to `x` given `emb` timestep embeddings.
"""
class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
"""
A sequential module that passes timestep embeddings to the children that
support it as an extra input.
"""
def forward(self, x, emb=None, cond=None, lateral=None):
for layer in self:
if isinstance(layer, TimestepBlock):
x = layer(x, emb=emb, cond=cond, lateral=lateral)
else:
x = layer(x)
return x
@dataclass
class ResBlockConfig(BaseConfig):
channels: int
emb_channels: int
dropout: float
out_channels: int = None
# condition the resblock with time (and encoder's output)
use_condition: bool = True
# whether to use 3x3 conv for skip path when the channels aren't matched
use_conv: bool = False
# dimension of conv (always 2 = 2d)
dims: int = 2
# gradient checkpoint
use_checkpoint: bool = False
up: bool = False
down: bool = False
# whether to condition with both time & encoder's output
two_cond: bool = False
# number of encoders' output channels
cond_emb_channels: int = None
# suggest: False
has_lateral: bool = False
lateral_channels: int = None
# whether to init the convolution with zero weights
# this is default from BeatGANs and seems to help learning
use_zero_module: bool = True
def __post_init__(self):
self.out_channels = self.out_channels or self.channels
self.cond_emb_channels = self.cond_emb_channels or self.emb_channels
def make_model(self):
return ResBlock(self)
class ResBlock(TimestepBlock):
"""
A residual block that can optionally change the number of channels.
total layers:
in_layers
- norm
- act
- conv
out_layers
- norm
- (modulation)
- act
- conv
"""
def __init__(self, conf: ResBlockConfig):
super().__init__()
self.conf = conf
#############################
# IN LAYERS
#############################
assert conf.lateral_channels is None
layers = [
normalization(conf.channels),
nn.SiLU(),
conv_nd(conf.dims, conf.channels, conf.out_channels, 3, padding=1)
]
self.in_layers = nn.Sequential(*layers)
self.updown = conf.up or conf.down
if conf.up:
self.h_upd = Upsample(conf.channels, False, conf.dims)
self.x_upd = Upsample(conf.channels, False, conf.dims)
elif conf.down:
self.h_upd = Downsample(conf.channels, False, conf.dims)
self.x_upd = Downsample(conf.channels, False, conf.dims)
else:
self.h_upd = self.x_upd = nn.Identity()
#############################
# OUT LAYERS CONDITIONS
#############################
if conf.use_condition:
# condition layers for the out_layers
self.emb_layers = nn.Sequential(
nn.SiLU(),
linear(conf.emb_channels, 2 * conf.out_channels),
)
if conf.two_cond:
self.cond_emb_layers = nn.Sequential(
nn.SiLU(),
linear(conf.cond_emb_channels, conf.out_channels),
)
#############################
# OUT LAYERS (ignored when there is no condition)
#############################
# original version
conv = conv_nd(conf.dims,
conf.out_channels,
conf.out_channels,
3,
padding=1)
if conf.use_zero_module:
# zere out the weights
# it seems to help training
conv = zero_module(conv)
# construct the layers
# - norm
# - (modulation)
# - act
# - dropout
# - conv
layers = []
layers += [
normalization(conf.out_channels),
nn.SiLU(),
nn.Dropout(p=conf.dropout),
conv,
]
self.out_layers = nn.Sequential(*layers)
#############################
# SKIP LAYERS
#############################
if conf.out_channels == conf.channels:
# cannot be used with gatedconv, also gatedconv is alsways used as the first block
self.skip_connection = nn.Identity()
else:
if conf.use_conv:
kernel_size = 3
padding = 1
else:
kernel_size = 1
padding = 0
self.skip_connection = conv_nd(conf.dims,
conf.channels,
conf.out_channels,
kernel_size,
padding=padding)
def forward(self, x, emb=None, cond=None, lateral=None):
"""
Apply the block to a Tensor, conditioned on a timestep embedding.
Args:
x: input
lateral: lateral connection from the encoder
"""
return torch_checkpoint(self._forward, (x, emb, cond, lateral),
self.conf.use_checkpoint)
def _forward(
self,
x,
emb=None,
cond=None,
lateral=None,
):
"""
Args:
lateral: required if "has_lateral" and non-gated, with gated, it can be supplied optionally
"""
if self.conf.has_lateral:
# lateral may be supplied even if it doesn't require
# the model will take the lateral only if "has_lateral"
assert lateral is not None
x = th.cat([x, lateral], dim=1)
if self.updown:
in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
h = in_rest(x)
h = self.h_upd(h)
x = self.x_upd(x)
h = in_conv(h)
else:
h = self.in_layers(x)
if self.conf.use_condition:
# it's possible that the network may not receieve the time emb
# this happens with autoenc and setting the time_at
if emb is not None:
emb_out = self.emb_layers(emb).type(h.dtype)
else:
emb_out = None
if self.conf.two_cond:
# it's possible that the network is two_cond
# but it doesn't get the second condition
# in which case, we ignore the second condition
# and treat as if the network has one condition
if cond is None:
cond_out = None
else:
cond_out = self.cond_emb_layers(cond).type(h.dtype)
if cond_out is not None:
while len(cond_out.shape) < len(h.shape):
cond_out = cond_out[..., None]
else:
cond_out = None
# this is the new refactored code
h = apply_conditions(
h=h,
emb=emb_out,
cond=cond_out,
layers=self.out_layers,
scale_bias=1,
in_channels=self.conf.out_channels,
up_down_layer=None,
)
return self.skip_connection(x) + h
def apply_conditions(
h,
emb=None,
cond=None,
layers: nn.Sequential = None,
scale_bias: float = 1,
in_channels: int = 512,
up_down_layer: nn.Module = None,
):
"""
apply conditions on the feature maps
Args:
emb: time conditional (ready to scale + shift)
cond: encoder's conditional (read to scale + shift)
"""
two_cond = emb is not None and cond is not None
if emb is not None:
# adjusting shapes
while len(emb.shape) < len(h.shape):
emb = emb[..., None]
if two_cond:
# adjusting shapes
while len(cond.shape) < len(h.shape):
cond = cond[..., None]
# time first
scale_shifts = [emb, cond]
else:
# "cond" is not used with single cond mode
scale_shifts = [emb]
# support scale, shift or shift only
for i, each in enumerate(scale_shifts):
if each is None:
# special case: the condition is not provided
a = None
b = None
else:
if each.shape[1] == in_channels * 2:
a, b = th.chunk(each, 2, dim=1)
else:
a = each
b = None
scale_shifts[i] = (a, b)
# condition scale bias could be a list
if isinstance(scale_bias, Number):
biases = [scale_bias] * len(scale_shifts)
else:
# a list
biases = scale_bias
# default, the scale & shift are applied after the group norm but BEFORE SiLU
pre_layers, post_layers = layers[0], layers[1:]
# spilt the post layer to be able to scale up or down before conv
# post layers will contain only the conv
mid_layers, post_layers = post_layers[:-2], post_layers[-2:]
h = pre_layers(h)
# scale and shift for each condition
for i, (scale, shift) in enumerate(scale_shifts):
# if scale is None, it indicates that the condition is not provided
if scale is not None:
h = h * (biases[i] + scale)
if shift is not None:
h = h + shift
h = mid_layers(h)
# upscale or downscale if any just before the last conv
if up_down_layer is not None:
h = up_down_layer(h)
h = post_layers(h)
return h
class Upsample(nn.Module):
"""
An upsampling layer with an optional convolution.
:param channels: channels in the inputs and outputs.
:param use_conv: a bool determining if a convolution is applied.
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
upsampling occurs in the inner-two dimensions.
"""
def __init__(self, channels, use_conv, dims=2, out_channels=None):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.dims = dims
if use_conv:
self.conv = conv_nd(dims,
self.channels,
self.out_channels,
3,
padding=1)
def forward(self, x):
assert x.shape[1] == self.channels
if self.dims == 3:
x = F.interpolate(x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2),
mode="nearest")
else:
x = F.interpolate(x, scale_factor=2, mode="nearest")
if self.use_conv:
x = self.conv(x)
return x
class Downsample(nn.Module):
"""
A downsampling layer with an optional convolution.
:param channels: channels in the inputs and outputs.
:param use_conv: a bool determining if a convolution is applied.
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
downsampling occurs in the inner-two dimensions.
"""
def __init__(self, channels, use_conv, dims=2, out_channels=None):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.dims = dims
stride = 2 if dims != 3 else (1, 2, 2)
if use_conv:
self.op = conv_nd(dims,
self.channels,
self.out_channels,
3,
stride=stride,
padding=1)
else:
assert self.channels == self.out_channels
self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
def forward(self, x):
assert x.shape[1] == self.channels
return self.op(x)
class AttentionBlock(nn.Module):
"""
An attention block that allows spatial positions to attend to each other.
Originally ported from here, but adapted to the N-d case.
https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
"""
def __init__(
self,
channels,
num_heads=1,
num_head_channels=-1,
use_checkpoint=False,
use_new_attention_order=False,
):
super().__init__()
self.channels = channels
if num_head_channels == -1:
self.num_heads = num_heads
else:
assert (
channels % num_head_channels == 0
), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
self.num_heads = channels // num_head_channels
self.use_checkpoint = use_checkpoint
self.norm = normalization(channels)
self.qkv = conv_nd(1, channels, channels * 3, 1)
if use_new_attention_order:
# split qkv before split heads
self.attention = QKVAttention(self.num_heads)
else:
# split heads before split qkv
self.attention = QKVAttentionLegacy(self.num_heads)
self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
def forward(self, x):
return torch_checkpoint(self._forward, (x, ), self.use_checkpoint)
def _forward(self, x):
b, c, *spatial = x.shape
x = x.reshape(b, c, -1)
qkv = self.qkv(self.norm(x))
h = self.attention(qkv)
h = self.proj_out(h)
return (x + h).reshape(b, c, *spatial)
def count_flops_attn(model, _x, y):
"""
A counter for the `thop` package to count the operations in an
attention operation.
Meant to be used like:
macs, params = thop.profile(
model,
inputs=(inputs, timestamps),
custom_ops={QKVAttention: QKVAttention.count_flops},
)
"""
b, c, *spatial = y[0].shape
num_spatial = int(np.prod(spatial))
# We perform two matmuls with the same number of ops.
# The first computes the weight matrix, the second computes
# the combination of the value vectors.
matmul_ops = 2 * b * (num_spatial**2) * c
model.total_ops += th.DoubleTensor([matmul_ops])
class QKVAttentionLegacy(nn.Module):
"""
A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
"""
def __init__(self, n_heads):
super().__init__()
self.n_heads = n_heads
def forward(self, qkv):
"""
Apply QKV attention.
:param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
:return: an [N x (H * C) x T] tensor after attention.
"""
bs, width, length = qkv.shape
assert width % (3 * self.n_heads) == 0
ch = width // (3 * self.n_heads)
q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch,
dim=1)
scale = 1 / math.sqrt(math.sqrt(ch))
weight = th.einsum(
"bct,bcs->bts", q * scale,
k * scale) # More stable with f16 than dividing afterwards
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
a = th.einsum("bts,bcs->bct", weight, v)
return a.reshape(bs, -1, length)
@staticmethod
def count_flops(model, _x, y):
return count_flops_attn(model, _x, y)
class QKVAttention(nn.Module):
"""
A module which performs QKV attention and splits in a different order.
"""
def __init__(self, n_heads):
super().__init__()
self.n_heads = n_heads
def forward(self, qkv):
"""
Apply QKV attention.
:param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
:return: an [N x (H * C) x T] tensor after attention.
"""
bs, width, length = qkv.shape
assert width % (3 * self.n_heads) == 0
ch = width // (3 * self.n_heads)
q, k, v = qkv.chunk(3, dim=1)
scale = 1 / math.sqrt(math.sqrt(ch))
weight = th.einsum(
"bct,bcs->bts",
(q * scale).view(bs * self.n_heads, ch, length),
(k * scale).view(bs * self.n_heads, ch, length),
) # More stable with f16 than dividing afterwards
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
a = th.einsum("bts,bcs->bct", weight,
v.reshape(bs * self.n_heads, ch, length))
return a.reshape(bs, -1, length)
@staticmethod
def count_flops(model, _x, y):
return count_flops_attn(model, _x, y)
class AttentionPool2d(nn.Module):
"""
Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
"""
def __init__(
self,
spacial_dim: int,
embed_dim: int,
num_heads_channels: int,
output_dim: int = None,
):
super().__init__()
self.positional_embedding = nn.Parameter(
th.randn(embed_dim, spacial_dim**2 + 1) / embed_dim**0.5)
self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
self.num_heads = embed_dim // num_heads_channels
self.attention = QKVAttention(self.num_heads)
def forward(self, x):
b, c, *_spatial = x.shape
x = x.reshape(b, c, -1) # NC(HW)
x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1)
x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1)
x = self.qkv_proj(x)
x = self.attention(x)
x = self.c_proj(x)
return x[:, :, 0]