File size: 7,363 Bytes
e8f664f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 | from segment_anything import build_sam, SamPredictor
from segment_anything import sam_model_registry
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from torch.nn.parameter import Parameter
from segment_anything.modeling import Sam
from safetensors import safe_open
from safetensors.torch import save_file
from icecream import ic
class _LoRA_qkv(nn.Module):
"""In Sam it is implemented as
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
q, k, v = qkv.unbind(0)
"""
def __init__(
self,
qkv: nn.Module,
linear_a_q: nn.Module,
linear_b_q: nn.Module,
linear_a_v: nn.Module,
linear_b_v: nn.Module,
):
super().__init__()
self.qkv = qkv
self.linear_a_q = linear_a_q
self.linear_b_q = linear_b_q
self.linear_a_v = linear_a_v
self.linear_b_v = linear_b_v
self.dim = qkv.in_features
self.w_identity = torch.eye(qkv.in_features)
def forward(self, x):
qkv = self.qkv(x) # B,N,N,3*org_C
new_q = self.linear_b_q(self.linear_a_q(x))
new_v = self.linear_b_v(self.linear_a_v(x))
qkv[:, :, :, : self.dim] += new_q
qkv[:, :, :, -self.dim:] += new_v
return qkv
class LoRA_Sam(nn.Module):
"""Applies low-rank adaptation to a Sam model's image encoder.
Args:
sam_model: a vision transformer model, see base_vit.py
r: rank of LoRA
num_classes: how many classes the model output, default to the vit model
lora_layer: which layer we apply LoRA.
Examples::
>>> model = ViT('B_16_imagenet1k')
>>> lora_model = LoRA_ViT(model, r=4)
>>> preds = lora_model(img)
>>> print(preds.shape)
torch.Size([1, 1000])
"""
def __init__(self, sam_model: Sam, r: int, lora_layer=None):
super(LoRA_Sam, self).__init__()
assert r > 0
# base_vit_dim = sam_model.image_encoder.patch_embed.proj.out_channels
# dim = base_vit_dim
if lora_layer:
self.lora_layer = lora_layer
else:
self.lora_layer = list(
range(len(sam_model.image_encoder.blocks))) # Only apply lora to the image encoder by default
# create for storage, then we can init them or load weights
self.w_As = [] # These are linear layers
self.w_Bs = []
# lets freeze first
for param in sam_model.image_encoder.parameters():
param.requires_grad = False
# Here, we do the surgery
for t_layer_i, blk in enumerate(sam_model.image_encoder.blocks):
# If we only want few lora layer instead of all
if t_layer_i not in self.lora_layer:
continue
w_qkv_linear = blk.attn.qkv
self.dim = w_qkv_linear.in_features
w_a_linear_q = nn.Linear(self.dim, r, bias=False)
w_b_linear_q = nn.Linear(r, self.dim, bias=False)
w_a_linear_v = nn.Linear(self.dim, r, bias=False)
w_b_linear_v = nn.Linear(r, self.dim, bias=False)
self.w_As.append(w_a_linear_q)
self.w_Bs.append(w_b_linear_q)
self.w_As.append(w_a_linear_v)
self.w_Bs.append(w_b_linear_v)
blk.attn.qkv = _LoRA_qkv(
w_qkv_linear,
w_a_linear_q,
w_b_linear_q,
w_a_linear_v,
w_b_linear_v,
)
self.reset_parameters()
self.sam = sam_model
def save_lora_parameters(self, filename: str) -> None:
r"""Only safetensors is supported now.
pip install safetensor if you do not have one installed yet.
save both lora and fc parameters.
"""
assert filename.endswith(".pt") or filename.endswith('.pth')
num_layer = len(self.w_As) # actually, it is half
a_tensors = {f"w_a_{i:03d}": self.w_As[i].weight for i in range(num_layer)}
b_tensors = {f"w_b_{i:03d}": self.w_Bs[i].weight for i in range(num_layer)}
prompt_encoder_tensors = {}
mask_decoder_tensors = {}
# save prompt encoder, only `state_dict`, the `named_parameter` is not permitted
if isinstance(self.sam, torch.nn.DataParallel) or isinstance(self.sam, torch.nn.parallel.DistributedDataParallel):
state_dict = self.sam.module.state_dict()
else:
state_dict = self.sam.state_dict()
for key, value in state_dict.items():
if 'prompt_encoder' in key:
prompt_encoder_tensors[key] = value
if 'mask_decoder' in key:
mask_decoder_tensors[key] = value
merged_dict = {**a_tensors, **b_tensors, **prompt_encoder_tensors, **mask_decoder_tensors}
torch.save(merged_dict, filename)
def load_lora_parameters(self, filename: str) -> None:
r"""Only safetensors is supported now.
pip install safetensor if you do not have one installed yet.\
load both lora and fc parameters.
"""
assert filename.endswith(".pt") or filename.endswith('.pth')
state_dict = torch.load(filename)
for i, w_A_linear in enumerate(self.w_As):
saved_key = f"w_a_{i:03d}"
saved_tensor = state_dict[saved_key]
w_A_linear.weight = Parameter(saved_tensor)
for i, w_B_linear in enumerate(self.w_Bs):
saved_key = f"w_b_{i:03d}"
saved_tensor = state_dict[saved_key]
w_B_linear.weight = Parameter(saved_tensor)
sam_dict = self.sam.state_dict()
sam_keys = sam_dict.keys()
# load prompt encoder
prompt_encoder_keys = [k for k in sam_keys if 'prompt_encoder' in k]
prompt_encoder_values = [state_dict[k] for k in prompt_encoder_keys]
prompt_encoder_new_state_dict = {k: v for k, v in zip(prompt_encoder_keys, prompt_encoder_values)}
sam_dict.update(prompt_encoder_new_state_dict)
# load mask decoder
mask_decoder_keys = [k for k in sam_keys if 'mask_decoder' in k]
mask_decoder_values = [state_dict[k] for k in mask_decoder_keys]
mask_decoder_new_state_dict = {k: v for k, v in zip(mask_decoder_keys, mask_decoder_values)}
sam_dict.update(mask_decoder_new_state_dict)
self.sam.load_state_dict(sam_dict)
def reset_parameters(self) -> None:
for w_A in self.w_As:
nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))
for w_B in self.w_Bs:
nn.init.zeros_(w_B.weight)
def forward(self, batched_input, multimask_output, image_size):
return self.sam(batched_input, multimask_output, image_size)
# def forward(self, x: Tensor) -> Tensor:
# return self.lora_vit(x)
if __name__ == "__main__":
sam = sam_model_registry["vit_b"](checkpoint="sam_vit_b_01ec64.pth")
lora_sam = LoRA_Sam(sam, 4)
lora_sam.sam.image_encoder(torch.rand(size=(1, 3, 1024, 1024)))
|