Adityadn's picture
Upload 524 files
617d388 verified
raw
history blame
No virus
26.8 kB
"""
Modified from https://github.com/sczhou/CodeFormer
VQGAN code, adapted from the original created by the Unleashing Transformers authors:
https://github.com/samb-t/unleashing-transformers/blob/master/models/vqgan.py
This version of the arch specifically was gathered from an old version of GFPGAN. If this is a problem, please contact me.
"""
import math
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
import logging as logger
from torch import Tensor
class VectorQuantizer(nn.Module):
def __init__(self, codebook_size, emb_dim, beta):
super(VectorQuantizer, self).__init__()
self.codebook_size = codebook_size # number of embeddings
self.emb_dim = emb_dim # dimension of embedding
self.beta = beta # commitment cost used in loss term, beta * ||z_e(x)-sg[e]||^2
self.embedding = nn.Embedding(self.codebook_size, self.emb_dim)
self.embedding.weight.data.uniform_(
-1.0 / self.codebook_size, 1.0 / self.codebook_size
)
def forward(self, z):
# reshape z -> (batch, height, width, channel) and flatten
z = z.permute(0, 2, 3, 1).contiguous()
z_flattened = z.view(-1, self.emb_dim)
# distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
d = (
(z_flattened**2).sum(dim=1, keepdim=True)
+ (self.embedding.weight**2).sum(1)
- 2 * torch.matmul(z_flattened, self.embedding.weight.t())
)
mean_distance = torch.mean(d)
# find closest encodings
# min_encoding_indices = torch.argmin(d, dim=1).unsqueeze(1)
min_encoding_scores, min_encoding_indices = torch.topk(
d, 1, dim=1, largest=False
)
# [0-1], higher score, higher confidence
min_encoding_scores = torch.exp(-min_encoding_scores / 10)
min_encodings = torch.zeros(
min_encoding_indices.shape[0], self.codebook_size
).to(z)
min_encodings.scatter_(1, min_encoding_indices, 1)
# get quantized latent vectors
z_q = torch.matmul(min_encodings, self.embedding.weight).view(z.shape)
# compute loss for embedding
loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean(
(z_q - z.detach()) ** 2
)
# preserve gradients
z_q = z + (z_q - z).detach()
# perplexity
e_mean = torch.mean(min_encodings, dim=0)
perplexity = torch.exp(-torch.sum(e_mean * torch.log(e_mean + 1e-10)))
# reshape back to match original input shape
z_q = z_q.permute(0, 3, 1, 2).contiguous()
return (
z_q,
loss,
{
"perplexity": perplexity,
"min_encodings": min_encodings,
"min_encoding_indices": min_encoding_indices,
"min_encoding_scores": min_encoding_scores,
"mean_distance": mean_distance,
},
)
def get_codebook_feat(self, indices, shape):
# input indices: batch*token_num -> (batch*token_num)*1
# shape: batch, height, width, channel
indices = indices.view(-1, 1)
min_encodings = torch.zeros(indices.shape[0], self.codebook_size).to(indices)
min_encodings.scatter_(1, indices, 1)
# get quantized latent vectors
z_q = torch.matmul(min_encodings.float(), self.embedding.weight)
if shape is not None: # reshape back to match original input shape
z_q = z_q.view(shape).permute(0, 3, 1, 2).contiguous()
return z_q
class GumbelQuantizer(nn.Module):
def __init__(
self,
codebook_size,
emb_dim,
num_hiddens,
straight_through=False,
kl_weight=5e-4,
temp_init=1.0,
):
super().__init__()
self.codebook_size = codebook_size # number of embeddings
self.emb_dim = emb_dim # dimension of embedding
self.straight_through = straight_through
self.temperature = temp_init
self.kl_weight = kl_weight
self.proj = nn.Conv2d(
num_hiddens, codebook_size, 1
) # projects last encoder layer to quantized logits
self.embed = nn.Embedding(codebook_size, emb_dim)
def forward(self, z):
hard = self.straight_through if self.training else True
logits = self.proj(z)
soft_one_hot = F.gumbel_softmax(logits, tau=self.temperature, dim=1, hard=hard)
z_q = torch.einsum("b n h w, n d -> b d h w", soft_one_hot, self.embed.weight)
# + kl divergence to the prior loss
qy = F.softmax(logits, dim=1)
diff = (
self.kl_weight
* torch.sum(qy * torch.log(qy * self.codebook_size + 1e-10), dim=1).mean()
)
min_encoding_indices = soft_one_hot.argmax(dim=1)
return z_q, diff, {"min_encoding_indices": min_encoding_indices}
class Downsample(nn.Module):
def __init__(self, in_channels):
super().__init__()
self.conv = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=3, stride=2, padding=0
)
def forward(self, x):
pad = (0, 1, 0, 1)
x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
x = self.conv(x)
return x
class Upsample(nn.Module):
def __init__(self, in_channels):
super().__init__()
self.conv = nn.Conv2d(
in_channels, in_channels, kernel_size=3, stride=1, padding=1
)
def forward(self, x):
x = F.interpolate(x, scale_factor=2.0, mode="nearest")
x = self.conv(x)
return x
class AttnBlock(nn.Module):
def __init__(self, in_channels):
super().__init__()
self.in_channels = in_channels
self.norm = normalize(in_channels)
self.q = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.k = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.v = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.proj_out = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
def forward(self, x):
h_ = x
h_ = self.norm(h_)
q = self.q(h_)
k = self.k(h_)
v = self.v(h_)
# compute attention
b, c, h, w = q.shape
q = q.reshape(b, c, h * w)
q = q.permute(0, 2, 1)
k = k.reshape(b, c, h * w)
w_ = torch.bmm(q, k)
w_ = w_ * (int(c) ** (-0.5))
w_ = F.softmax(w_, dim=2)
# attend to values
v = v.reshape(b, c, h * w)
w_ = w_.permute(0, 2, 1)
h_ = torch.bmm(v, w_)
h_ = h_.reshape(b, c, h, w)
h_ = self.proj_out(h_)
return x + h_
class Encoder(nn.Module):
def __init__(
self,
in_channels,
nf,
out_channels,
ch_mult,
num_res_blocks,
resolution,
attn_resolutions,
):
super().__init__()
self.nf = nf
self.num_resolutions = len(ch_mult)
self.num_res_blocks = num_res_blocks
self.resolution = resolution
self.attn_resolutions = attn_resolutions
curr_res = self.resolution
in_ch_mult = (1,) + tuple(ch_mult)
blocks = []
# initial convultion
blocks.append(nn.Conv2d(in_channels, nf, kernel_size=3, stride=1, padding=1))
# residual and downsampling blocks, with attention on smaller res (16x16)
for i in range(self.num_resolutions):
block_in_ch = nf * in_ch_mult[i]
block_out_ch = nf * ch_mult[i]
for _ in range(self.num_res_blocks):
blocks.append(ResBlock(block_in_ch, block_out_ch))
block_in_ch = block_out_ch
if curr_res in attn_resolutions:
blocks.append(AttnBlock(block_in_ch))
if i != self.num_resolutions - 1:
blocks.append(Downsample(block_in_ch))
curr_res = curr_res // 2
# non-local attention block
blocks.append(ResBlock(block_in_ch, block_in_ch)) # type: ignore
blocks.append(AttnBlock(block_in_ch)) # type: ignore
blocks.append(ResBlock(block_in_ch, block_in_ch)) # type: ignore
# normalise and convert to latent size
blocks.append(normalize(block_in_ch)) # type: ignore
blocks.append(
nn.Conv2d(block_in_ch, out_channels, kernel_size=3, stride=1, padding=1) # type: ignore
)
self.blocks = nn.ModuleList(blocks)
def forward(self, x):
for block in self.blocks:
x = block(x)
return x
class Generator(nn.Module):
def __init__(self, nf, ch_mult, res_blocks, img_size, attn_resolutions, emb_dim):
super().__init__()
self.nf = nf
self.ch_mult = ch_mult
self.num_resolutions = len(self.ch_mult)
self.num_res_blocks = res_blocks
self.resolution = img_size
self.attn_resolutions = attn_resolutions
self.in_channels = emb_dim
self.out_channels = 3
block_in_ch = self.nf * self.ch_mult[-1]
curr_res = self.resolution // 2 ** (self.num_resolutions - 1)
blocks = []
# initial conv
blocks.append(
nn.Conv2d(self.in_channels, block_in_ch, kernel_size=3, stride=1, padding=1)
)
# non-local attention block
blocks.append(ResBlock(block_in_ch, block_in_ch))
blocks.append(AttnBlock(block_in_ch))
blocks.append(ResBlock(block_in_ch, block_in_ch))
for i in reversed(range(self.num_resolutions)):
block_out_ch = self.nf * self.ch_mult[i]
for _ in range(self.num_res_blocks):
blocks.append(ResBlock(block_in_ch, block_out_ch))
block_in_ch = block_out_ch
if curr_res in self.attn_resolutions:
blocks.append(AttnBlock(block_in_ch))
if i != 0:
blocks.append(Upsample(block_in_ch))
curr_res = curr_res * 2
blocks.append(normalize(block_in_ch))
blocks.append(
nn.Conv2d(
block_in_ch, self.out_channels, kernel_size=3, stride=1, padding=1
)
)
self.blocks = nn.ModuleList(blocks)
def forward(self, x):
for block in self.blocks:
x = block(x)
return x
class VQAutoEncoder(nn.Module):
def __init__(
self,
img_size,
nf,
ch_mult,
quantizer="nearest",
res_blocks=2,
attn_resolutions=[16],
codebook_size=1024,
emb_dim=256,
beta=0.25,
gumbel_straight_through=False,
gumbel_kl_weight=1e-8,
model_path=None,
):
super().__init__()
self.in_channels = 3
self.nf = nf
self.n_blocks = res_blocks
self.codebook_size = codebook_size
self.embed_dim = emb_dim
self.ch_mult = ch_mult
self.resolution = img_size
self.attn_resolutions = attn_resolutions
self.quantizer_type = quantizer
self.encoder = Encoder(
self.in_channels,
self.nf,
self.embed_dim,
self.ch_mult,
self.n_blocks,
self.resolution,
self.attn_resolutions,
)
if self.quantizer_type == "nearest":
self.beta = beta # 0.25
self.quantize = VectorQuantizer(
self.codebook_size, self.embed_dim, self.beta
)
elif self.quantizer_type == "gumbel":
self.gumbel_num_hiddens = emb_dim
self.straight_through = gumbel_straight_through
self.kl_weight = gumbel_kl_weight
self.quantize = GumbelQuantizer(
self.codebook_size,
self.embed_dim,
self.gumbel_num_hiddens,
self.straight_through,
self.kl_weight,
)
self.generator = Generator(
nf, ch_mult, res_blocks, img_size, attn_resolutions, emb_dim
)
if model_path is not None:
chkpt = torch.load(model_path, map_location="cpu")
if "params_ema" in chkpt:
self.load_state_dict(
torch.load(model_path, map_location="cpu")["params_ema"]
)
logger.info(f"vqgan is loaded from: {model_path} [params_ema]")
elif "params" in chkpt:
self.load_state_dict(
torch.load(model_path, map_location="cpu")["params"]
)
logger.info(f"vqgan is loaded from: {model_path} [params]")
else:
raise ValueError("Wrong params!")
def forward(self, x):
x = self.encoder(x)
quant, codebook_loss, quant_stats = self.quantize(x)
x = self.generator(quant)
return x, codebook_loss, quant_stats
def calc_mean_std(feat, eps=1e-5):
"""Calculate mean and std for adaptive_instance_normalization.
Args:
feat (Tensor): 4D tensor.
eps (float): A small value added to the variance to avoid
divide-by-zero. Default: 1e-5.
"""
size = feat.size()
assert len(size) == 4, "The input feature should be 4D tensor."
b, c = size[:2]
feat_var = feat.view(b, c, -1).var(dim=2) + eps
feat_std = feat_var.sqrt().view(b, c, 1, 1)
feat_mean = feat.view(b, c, -1).mean(dim=2).view(b, c, 1, 1)
return feat_mean, feat_std
def adaptive_instance_normalization(content_feat, style_feat):
"""Adaptive instance normalization.
Adjust the reference features to have the similar color and illuminations
as those in the degradate features.
Args:
content_feat (Tensor): The reference feature.
style_feat (Tensor): The degradate features.
"""
size = content_feat.size()
style_mean, style_std = calc_mean_std(style_feat)
content_mean, content_std = calc_mean_std(content_feat)
normalized_feat = (content_feat - content_mean.expand(size)) / content_std.expand(
size
)
return normalized_feat * style_std.expand(size) + style_mean.expand(size)
class PositionEmbeddingSine(nn.Module):
"""
This is a more standard version of the position embedding, very similar to the one
used by the Attention is all you need paper, generalized to work on images.
"""
def __init__(
self, num_pos_feats=64, temperature=10000, normalize=False, scale=None
):
super().__init__()
self.num_pos_feats = num_pos_feats
self.temperature = temperature
self.normalize = normalize
if scale is not None and normalize is False:
raise ValueError("normalize should be True if scale is passed")
if scale is None:
scale = 2 * math.pi
self.scale = scale
def forward(self, x, mask=None):
if mask is None:
mask = torch.zeros(
(x.size(0), x.size(2), x.size(3)), device=x.device, dtype=torch.bool
)
not_mask = ~mask # pylint: disable=invalid-unary-operand-type
y_embed = not_mask.cumsum(1, dtype=torch.float32)
x_embed = not_mask.cumsum(2, dtype=torch.float32)
if self.normalize:
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.num_pos_feats, dtype=torch.float32, device=x.device)
dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
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)
return pos
def _get_activation_fn(activation):
"""Return an activation function given a string"""
if activation == "relu":
return F.relu
if activation == "gelu":
return F.gelu
if activation == "glu":
return F.glu
raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
class TransformerSALayer(nn.Module):
def __init__(
self, embed_dim, nhead=8, dim_mlp=2048, dropout=0.0, activation="gelu"
):
super().__init__()
self.self_attn = nn.MultiheadAttention(embed_dim, nhead, dropout=dropout)
# Implementation of Feedforward model - MLP
self.linear1 = nn.Linear(embed_dim, dim_mlp)
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_mlp, embed_dim)
self.norm1 = nn.LayerNorm(embed_dim)
self.norm2 = nn.LayerNorm(embed_dim)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.activation = _get_activation_fn(activation)
def with_pos_embed(self, tensor, pos: Optional[Tensor]):
return tensor if pos is None else tensor + pos
def forward(
self,
tgt,
tgt_mask: Optional[Tensor] = None,
tgt_key_padding_mask: Optional[Tensor] = None,
query_pos: Optional[Tensor] = None,
):
# self attention
tgt2 = self.norm1(tgt)
q = k = self.with_pos_embed(tgt2, query_pos)
tgt2 = self.self_attn(
q, k, value=tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask
)[0]
tgt = tgt + self.dropout1(tgt2)
# ffn
tgt2 = self.norm2(tgt)
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
tgt = tgt + self.dropout2(tgt2)
return tgt
def normalize(in_channels):
return torch.nn.GroupNorm(
num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
)
@torch.jit.script # type: ignore
def swish(x):
return x * torch.sigmoid(x)
class ResBlock(nn.Module):
def __init__(self, in_channels, out_channels=None):
super(ResBlock, self).__init__()
self.in_channels = in_channels
self.out_channels = in_channels if out_channels is None else out_channels
self.norm1 = normalize(in_channels)
self.conv1 = nn.Conv2d(
in_channels, out_channels, kernel_size=3, stride=1, padding=1 # type: ignore
)
self.norm2 = normalize(out_channels)
self.conv2 = nn.Conv2d(
out_channels, out_channels, kernel_size=3, stride=1, padding=1 # type: ignore
)
if self.in_channels != self.out_channels:
self.conv_out = nn.Conv2d(
in_channels, out_channels, kernel_size=1, stride=1, padding=0 # type: ignore
)
def forward(self, x_in):
x = x_in
x = self.norm1(x)
x = swish(x)
x = self.conv1(x)
x = self.norm2(x)
x = swish(x)
x = self.conv2(x)
if self.in_channels != self.out_channels:
x_in = self.conv_out(x_in)
return x + x_in
class Fuse_sft_block(nn.Module):
def __init__(self, in_ch, out_ch):
super().__init__()
self.encode_enc = ResBlock(2 * in_ch, out_ch)
self.scale = nn.Sequential(
nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1),
nn.LeakyReLU(0.2, True),
nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1),
)
self.shift = nn.Sequential(
nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1),
nn.LeakyReLU(0.2, True),
nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1),
)
def forward(self, enc_feat, dec_feat, w=1):
enc_feat = self.encode_enc(torch.cat([enc_feat, dec_feat], dim=1))
scale = self.scale(enc_feat)
shift = self.shift(enc_feat)
residual = w * (dec_feat * scale + shift)
out = dec_feat + residual
return out
class CodeFormer(VQAutoEncoder):
def __init__(self, state_dict):
dim_embd = 512
n_head = 8
n_layers = 9
codebook_size = 1024
latent_size = 256
connect_list = ["32", "64", "128", "256"]
fix_modules = ["quantize", "generator"]
# This is just a guess as I only have one model to look at
position_emb = state_dict["position_emb"]
dim_embd = position_emb.shape[1]
latent_size = position_emb.shape[0]
try:
n_layers = len(
set([x.split(".")[1] for x in state_dict.keys() if "ft_layers" in x])
)
except:
pass
codebook_size = state_dict["quantize.embedding.weight"].shape[0]
# This is also just another guess
n_head_exp = (
state_dict["ft_layers.0.self_attn.in_proj_weight"].shape[0] // dim_embd
)
n_head = 2**n_head_exp
in_nc = state_dict["encoder.blocks.0.weight"].shape[1]
self.model_arch = "CodeFormer"
self.sub_type = "Face SR"
self.scale = 8
self.in_nc = in_nc
self.out_nc = in_nc
self.state = state_dict
self.supports_fp16 = False
self.supports_bf16 = True
self.min_size_restriction = 16
super(CodeFormer, self).__init__(
512, 64, [1, 2, 2, 4, 4, 8], "nearest", 2, [16], codebook_size
)
if fix_modules is not None:
for module in fix_modules:
for param in getattr(self, module).parameters():
param.requires_grad = False
self.connect_list = connect_list
self.n_layers = n_layers
self.dim_embd = dim_embd
self.dim_mlp = dim_embd * 2
self.position_emb = nn.Parameter(torch.zeros(latent_size, self.dim_embd)) # type: ignore
self.feat_emb = nn.Linear(256, self.dim_embd)
# transformer
self.ft_layers = nn.Sequential(
*[
TransformerSALayer(
embed_dim=dim_embd, nhead=n_head, dim_mlp=self.dim_mlp, dropout=0.0
)
for _ in range(self.n_layers)
]
)
# logits_predict head
self.idx_pred_layer = nn.Sequential(
nn.LayerNorm(dim_embd), nn.Linear(dim_embd, codebook_size, bias=False)
)
self.channels = {
"16": 512,
"32": 256,
"64": 256,
"128": 128,
"256": 128,
"512": 64,
}
# after second residual block for > 16, before attn layer for ==16
self.fuse_encoder_block = {
"512": 2,
"256": 5,
"128": 8,
"64": 11,
"32": 14,
"16": 18,
}
# after first residual block for > 16, before attn layer for ==16
self.fuse_generator_block = {
"16": 6,
"32": 9,
"64": 12,
"128": 15,
"256": 18,
"512": 21,
}
# fuse_convs_dict
self.fuse_convs_dict = nn.ModuleDict()
for f_size in self.connect_list:
in_ch = self.channels[f_size]
self.fuse_convs_dict[f_size] = Fuse_sft_block(in_ch, in_ch)
self.load_state_dict(state_dict)
def _init_weights(self, module):
if isinstance(module, (nn.Linear, nn.Embedding)):
module.weight.data.normal_(mean=0.0, std=0.02)
if isinstance(module, nn.Linear) and module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
def forward(self, x, weight=0.5, **kwargs):
detach_16 = True
code_only = False
adain = True
# ################### Encoder #####################
enc_feat_dict = {}
out_list = [self.fuse_encoder_block[f_size] for f_size in self.connect_list]
for i, block in enumerate(self.encoder.blocks):
x = block(x)
if i in out_list:
enc_feat_dict[str(x.shape[-1])] = x.clone()
lq_feat = x
# ################# Transformer ###################
# quant_feat, codebook_loss, quant_stats = self.quantize(lq_feat)
pos_emb = self.position_emb.unsqueeze(1).repeat(1, x.shape[0], 1)
# BCHW -> BC(HW) -> (HW)BC
feat_emb = self.feat_emb(lq_feat.flatten(2).permute(2, 0, 1))
query_emb = feat_emb
# Transformer encoder
for layer in self.ft_layers:
query_emb = layer(query_emb, query_pos=pos_emb)
# output logits
logits = self.idx_pred_layer(query_emb) # (hw)bn
logits = logits.permute(1, 0, 2) # (hw)bn -> b(hw)n
if code_only: # for training stage II
# logits doesn't need softmax before cross_entropy loss
return logits, lq_feat
# ################# Quantization ###################
# if self.training:
# quant_feat = torch.einsum('btn,nc->btc', [soft_one_hot, self.quantize.embedding.weight])
# # b(hw)c -> bc(hw) -> bchw
# quant_feat = quant_feat.permute(0,2,1).view(lq_feat.shape)
# ------------
soft_one_hot = F.softmax(logits, dim=2)
_, top_idx = torch.topk(soft_one_hot, 1, dim=2)
quant_feat = self.quantize.get_codebook_feat(
top_idx, shape=[x.shape[0], 16, 16, 256] # type: ignore
)
# preserve gradients
# quant_feat = lq_feat + (quant_feat - lq_feat).detach()
if detach_16:
quant_feat = quant_feat.detach() # for training stage III
if adain:
quant_feat = adaptive_instance_normalization(quant_feat, lq_feat)
# ################## Generator ####################
x = quant_feat
fuse_list = [self.fuse_generator_block[f_size] for f_size in self.connect_list]
for i, block in enumerate(self.generator.blocks):
x = block(x)
if i in fuse_list: # fuse after i-th block
f_size = str(x.shape[-1])
if weight > 0:
x = self.fuse_convs_dict[f_size](
enc_feat_dict[f_size].detach(), x, weight
)
out = x
# logits doesn't need softmax before cross_entropy loss
# return out, logits, lq_feat
return out, logits