File size: 4,709 Bytes
c7eddf0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 |
# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto. Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.
import math
from typing import Dict
import torch
from torch import nn
from einops import rearrange
from timm.models.vision_transformer import Block
class MLP(nn.Module):
def __init__(self, input_size: int, hidden_size: int, output_size: int,
num_inner: int = 0, device: torch.device = None, **kwargs):
super(MLP, self).__init__()
self.fc1 = nn.Linear(input_size, hidden_size, device=device)
self.norm = nn.LayerNorm(hidden_size, device=device)
self.relu = nn.ReLU()
inner = []
for _ in range(num_inner):
inner.extend([
nn.Linear(hidden_size, hidden_size, device=device),
nn.LayerNorm(hidden_size, device=device),
nn.ReLU(),
])
if inner:
self.inner = nn.Sequential(*inner)
else:
self.inner = nn.Identity()
self.fc2 = nn.Linear(hidden_size, output_size, device=device)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.fc1(x)
x = self.norm(x)
x = self.relu(x)
x = self.inner(x)
x = self.fc2(x)
return x
class MLP2(nn.Module):
def __init__(self, input_size: int, hidden_size: int, output_size: int,
num_inner: int = 0,
pre_norm: bool = False, device: torch.device = None,
upsample_factor: int = 1,
**kwargs):
super().__init__()
self.pre_norm = nn.Sequential(
nn.LayerNorm(input_size),
nn.GELU(),
) if pre_norm else nn.Identity()
self.upsample_factor = upsample_factor
self._real_output_dim = output_size
hidden_size *= upsample_factor
output_size *= (upsample_factor ** 2)
self.fc1 = nn.Linear(input_size, hidden_size, device=device)
blocks = []
for _ in range(num_inner):
blocks.append(nn.Sequential(
nn.LayerNorm(hidden_size, device=device),
nn.GELU(),
nn.Linear(hidden_size, hidden_size, device=device),
))
self.blocks = nn.ModuleList(blocks)
self.final = nn.Sequential(
nn.LayerNorm(hidden_size, device=device),
nn.GELU(),
nn.Linear(hidden_size, output_size, device=device),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.pre_norm(x)
x = self.fc1(x)
for block in self.blocks:
x = x + block(x)
x = self.final(x)
if self.upsample_factor > 1:
h = w = int(math.sqrt(x.shape[1]))
x = rearrange(x, 'b (h w) (u1 u2 c) -> b (u1 h u2 w) c',
h=h, w=w, u1=self.upsample_factor, u2=self.upsample_factor,
c=self._real_output_dim)
return x
MLP_FACTORY = {
'v1': MLP,
'v2': MLP2,
}
def strip_prefix(state: Dict[str, torch.Tensor], prefix: str):
state = {
k[len(prefix):]: v
for k, v in state.items()
if k.startswith(prefix)
}
return state
def get_mlp_info_from_state(version: str, state: Dict[str, torch.Tensor], prefix: str = ''):
state = strip_prefix(state, prefix)
if version == 'v1':
hidden_dim, input_dim = state['fc1.weight'].shape
output_dim = state['fc2.weight'].shape[0]
for num_inner in range(1000):
k = f'inner.{num_inner}.0.weight'
if k not in state:
break
elif version == 'v2':
hidden_dim, input_dim = state['fc1.weight'].shape
output_dim = state['final.2.weight'].shape[0]
for num_inner in range(1000):
k = f'blocks.{num_inner}.0.weight'
if k not in state:
break
else:
raise ValueError(f'Unsupported MLP version: {version}')
return input_dim, hidden_dim, output_dim, num_inner
def create_mlp_from_state(version: str, state: Dict[str, torch.Tensor], prefix: str = ''):
state = strip_prefix(state, prefix)
input_dim, hidden_dim, output_dim, num_inner = get_mlp_info_from_state(version, state)
ret: nn.Module = MLP_FACTORY[version](input_dim, hidden_dim, output_dim, num_inner)
ret.load_state_dict(state)
return ret
|