chore: import upstream snapshot with attribution
This commit is contained in:
wehub-resource-sync
@@ -0,0 +1,333 @@
|
||||
# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
#
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
"""
|
||||
GLUMBConv with 1x1 Conv replaced by Linear layers for better efficiency.
|
||||
The 1x1 Conv2d is mathematically equivalent to Linear layer.
|
||||
"""
|
||||
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from timm.models.vision_transformer import Mlp
|
||||
|
||||
from diffusion.model.act import build_act, get_act_name
|
||||
from diffusion.model.norms import build_norm, get_norm_name
|
||||
from diffusion.model.utils import get_same_padding, val2tuple
|
||||
|
||||
|
||||
class LinearLayer(nn.Module):
|
||||
"""Linear layer with optional normalization and activation."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
in_dim: int,
|
||||
out_dim: int,
|
||||
use_bias=True,
|
||||
norm=None,
|
||||
act=None,
|
||||
):
|
||||
super().__init__()
|
||||
self.in_dim = in_dim
|
||||
self.out_dim = out_dim
|
||||
self.use_bias = use_bias
|
||||
|
||||
self.linear = nn.Linear(in_dim, out_dim, bias=use_bias)
|
||||
self.norm = build_norm(norm, num_features=out_dim)
|
||||
self.act = build_act(act)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
x = self.linear(x)
|
||||
if self.norm:
|
||||
x = self.norm(x)
|
||||
if self.act:
|
||||
x = self.act(x)
|
||||
return x
|
||||
|
||||
|
||||
class ConvLayer(nn.Module):
|
||||
"""Conv layer for depthwise convolution (kernel_size > 1)."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
in_dim: int,
|
||||
out_dim: int,
|
||||
kernel_size=3,
|
||||
stride=1,
|
||||
dilation=1,
|
||||
groups=1,
|
||||
padding: int or None = None,
|
||||
use_bias=False,
|
||||
dropout=0.0,
|
||||
conv_type="2d",
|
||||
norm="bn2d",
|
||||
act="relu",
|
||||
):
|
||||
super().__init__()
|
||||
if padding is None:
|
||||
padding = get_same_padding(kernel_size)
|
||||
padding *= dilation
|
||||
|
||||
self.in_dim = in_dim
|
||||
self.out_dim = out_dim
|
||||
self.kernel_size = kernel_size
|
||||
self.stride = stride
|
||||
self.dilation = dilation
|
||||
self.groups = groups
|
||||
self.padding = padding
|
||||
self.use_bias = use_bias
|
||||
|
||||
self.dropout = nn.Dropout2d(dropout, inplace=False) if dropout > 0 else None
|
||||
if conv_type == "2d":
|
||||
self.conv = nn.Conv2d(
|
||||
in_dim,
|
||||
out_dim,
|
||||
kernel_size=(kernel_size, kernel_size),
|
||||
stride=(stride, stride),
|
||||
padding=padding,
|
||||
dilation=(dilation, dilation),
|
||||
groups=groups,
|
||||
bias=use_bias,
|
||||
)
|
||||
elif conv_type == "3d":
|
||||
self.conv = nn.Conv3d(
|
||||
in_dim,
|
||||
out_dim,
|
||||
kernel_size=(kernel_size, kernel_size, kernel_size),
|
||||
stride=(stride, stride, stride),
|
||||
padding=padding,
|
||||
dilation=(dilation, dilation, dilation),
|
||||
groups=groups,
|
||||
bias=use_bias,
|
||||
)
|
||||
else:
|
||||
self.conv = None
|
||||
|
||||
self.norm = build_norm(norm, num_features=out_dim)
|
||||
self.act = build_act(act)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
if self.dropout is not None:
|
||||
x = self.dropout(x)
|
||||
x = self.conv(x)
|
||||
if self.norm:
|
||||
x = self.norm(x)
|
||||
if self.act:
|
||||
x = self.act(x)
|
||||
return x
|
||||
|
||||
|
||||
class GLUMBConvLinear(nn.Module):
|
||||
"""
|
||||
GLUMBConv with 1x1 Conv replaced by Linear layers.
|
||||
|
||||
Original GLUMBConv structure:
|
||||
- inverted_conv: Conv2d(in_features, hidden_features*2, kernel=1x1) -> replaced with Linear
|
||||
- depth_conv: Conv2d(hidden_features*2, hidden_features*2, kernel=3x3, groups=hidden_features*2) -> keep Conv
|
||||
- point_conv: Conv2d(hidden_features, out_features, kernel=1x1) -> replaced with Linear
|
||||
|
||||
Weight mapping (for checkpoint conversion):
|
||||
- inverted_conv.conv.weight: [out, in, 1, 1] -> inverted_conv.linear.weight: [out, in]
|
||||
- inverted_conv.conv.bias: [out] -> inverted_conv.linear.bias: [out]
|
||||
- point_conv.conv.weight: [out, in, 1, 1] -> point_conv.linear.weight: [out, in]
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
in_features: int,
|
||||
hidden_features: int,
|
||||
out_feature=None,
|
||||
kernel_size=3,
|
||||
stride=1,
|
||||
padding: int or None = None,
|
||||
use_bias=False,
|
||||
norm=(None, None, None),
|
||||
act=("silu", "silu", None),
|
||||
dilation=1,
|
||||
):
|
||||
out_feature = out_feature or in_features
|
||||
super().__init__()
|
||||
use_bias = val2tuple(use_bias, 3)
|
||||
norm = val2tuple(norm, 3)
|
||||
act = val2tuple(act, 3)
|
||||
|
||||
self.hidden_features = hidden_features
|
||||
self.out_feature = out_feature
|
||||
|
||||
self.glu_act = build_act(act[1], inplace=False)
|
||||
|
||||
# Replace 1x1 conv with Linear
|
||||
self.inverted_conv = LinearLayer(
|
||||
in_features,
|
||||
hidden_features * 2,
|
||||
use_bias=use_bias[0],
|
||||
norm=norm[0],
|
||||
act=act[0],
|
||||
)
|
||||
|
||||
# Keep depthwise conv (kernel_size=3)
|
||||
self.depth_conv = ConvLayer(
|
||||
hidden_features * 2,
|
||||
hidden_features * 2,
|
||||
kernel_size,
|
||||
stride=stride,
|
||||
groups=hidden_features * 2,
|
||||
padding=padding,
|
||||
use_bias=use_bias[1],
|
||||
norm=norm[1],
|
||||
act=None,
|
||||
dilation=dilation,
|
||||
)
|
||||
|
||||
# Replace 1x1 conv with Linear
|
||||
self.point_conv = LinearLayer(
|
||||
hidden_features,
|
||||
out_feature,
|
||||
use_bias=use_bias[2],
|
||||
norm=norm[2],
|
||||
act=act[2],
|
||||
)
|
||||
|
||||
def forward(self, x: torch.Tensor, HW=None) -> torch.Tensor:
|
||||
B, N, C = x.shape
|
||||
if HW is None:
|
||||
H = W = int(N**0.5)
|
||||
elif len(HW) == 2:
|
||||
H, W = HW
|
||||
elif len(HW) == 3:
|
||||
T, H, W = HW
|
||||
|
||||
# inverted_conv: Linear on sequence dimension (B, N, C) -> (B, N, hidden*2)
|
||||
x = self.inverted_conv(x)
|
||||
|
||||
# Reshape for depthwise conv: (B, N, hidden*2) -> (B*T or B, hidden*2, H, W)
|
||||
if HW is not None and len(HW) == 3:
|
||||
x = x.reshape(B * T, H, W, -1).permute(0, 3, 1, 2)
|
||||
else:
|
||||
x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2)
|
||||
|
||||
# depth_conv: keep as Conv2d (depthwise 3x3)
|
||||
x = self.depth_conv(x)
|
||||
|
||||
# GLU activation
|
||||
x, gate = torch.chunk(x, 2, dim=1)
|
||||
gate = self.glu_act(gate)
|
||||
x = x * gate
|
||||
|
||||
# Reshape back to sequence for point_conv: (B*T or B, hidden, H, W) -> (B, N, hidden)
|
||||
if HW is not None and len(HW) == 3:
|
||||
x = x.reshape(B * T, self.hidden_features, H * W).permute(0, 2, 1)
|
||||
x = x.reshape(B, N, self.hidden_features)
|
||||
else:
|
||||
x = x.reshape(B, self.hidden_features, N).permute(0, 2, 1)
|
||||
|
||||
# point_conv: Linear on sequence dimension (B, N, hidden) -> (B, N, out)
|
||||
x = self.point_conv(x)
|
||||
|
||||
return x
|
||||
|
||||
|
||||
class GLUMBConvLinearTemp(GLUMBConvLinear):
|
||||
"""GLUMBConvLinear with temporal convolution support."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
in_features: int,
|
||||
hidden_features: int,
|
||||
out_feature=None,
|
||||
kernel_size=3,
|
||||
stride=1,
|
||||
padding: int or None = None,
|
||||
use_bias=False,
|
||||
norm=(None, None, None),
|
||||
act=("silu", "silu", None),
|
||||
t_kernel_size=3,
|
||||
):
|
||||
super().__init__(
|
||||
in_features=in_features,
|
||||
hidden_features=hidden_features,
|
||||
out_feature=out_feature,
|
||||
kernel_size=kernel_size,
|
||||
stride=stride,
|
||||
padding=padding,
|
||||
use_bias=use_bias,
|
||||
norm=norm,
|
||||
act=act,
|
||||
)
|
||||
|
||||
out_feature = out_feature or in_features
|
||||
t_padding = t_kernel_size // 2
|
||||
self.t_conv = nn.Conv2d(
|
||||
out_feature,
|
||||
out_feature,
|
||||
kernel_size=(t_kernel_size, 1),
|
||||
stride=1,
|
||||
padding=(t_padding, 0),
|
||||
bias=False,
|
||||
)
|
||||
|
||||
nn.init.zeros_(self.t_conv.weight)
|
||||
|
||||
def forward(self, x: torch.Tensor, HW=None, **kwargs) -> torch.Tensor:
|
||||
B, N, C = x.shape
|
||||
|
||||
assert len(HW) == 3, "HW must be a tuple of (T, H, W)"
|
||||
T, H, W = HW
|
||||
|
||||
# inverted_conv: Linear (B, N, C) -> (B, N, hidden*2)
|
||||
x = self.inverted_conv(x)
|
||||
|
||||
# Reshape for depth_conv: (B, T*H*W, hidden*2) -> (B*T, hidden*2, H, W)
|
||||
x = x.reshape(B * T, H, W, -1).permute(0, 3, 1, 2)
|
||||
x = self.depth_conv(x)
|
||||
|
||||
# Space aggregation (GLU)
|
||||
x, gate = torch.chunk(x, 2, dim=1)
|
||||
gate = self.glu_act(gate)
|
||||
x = x * gate
|
||||
|
||||
# Reshape for point_conv: (B*T, hidden, H, W) -> (B*T, H*W, hidden)
|
||||
x = x.reshape(B * T, self.hidden_features, H * W).permute(0, 2, 1)
|
||||
x = self.point_conv(x) # (B*T, H*W, out_feature)
|
||||
|
||||
# Temporal aggregation: reshape for t_conv
|
||||
x = x.reshape(B, T, H * W, self.out_feature).permute(0, 3, 1, 2) # B, out, T, H*W
|
||||
x_out = x + self.t_conv(x)
|
||||
|
||||
x_out = x_out.permute(0, 2, 3, 1).reshape(B, N, self.out_feature)
|
||||
|
||||
return x_out
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
# Test GLUMBConvLinear
|
||||
model = GLUMBConvLinear(
|
||||
2240,
|
||||
2240 * 5 // 2, # hidden_features = 5600
|
||||
2240,
|
||||
use_bias=(True, True, False),
|
||||
norm=(None, None, None),
|
||||
act=("silu", "silu", None),
|
||||
).cuda()
|
||||
|
||||
input_tensor = torch.randn(4, 1024, 2240).cuda() # (B, H*W, C)
|
||||
output = model(input_tensor, HW=(32, 32))
|
||||
print(f"Input shape: {input_tensor.shape}")
|
||||
print(f"Output shape: {output.shape}")
|
||||
|
||||
# Count parameters
|
||||
total_params = sum(p.numel() for p in model.parameters())
|
||||
print(f"Total parameters: {total_params:,}")
|
||||
Reference in New Issue
Block a user