Spaces:
Running
Running
File size: 9,502 Bytes
ae3b630 532251c ae3b630 532251c ae3b630 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 |
from extractor import BasicEncoder
from position_encoding import build_position_encoding
import argparse
import numpy as np
import torch
from torch import nn, Tensor
import torch.nn.functional as F
import copy
from typing import Optional
class attnLayer(nn.Module):
def __init__(self, d_model, nhead=8, dim_feedforward=2048, dropout=0.1,
activation="relu", normalize_before=False):
super().__init__()
self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout)
self.multihead_attn_list = nn.ModuleList([copy.deepcopy(nn.MultiheadAttention(d_model, nhead, dropout=dropout)) for i in range(2)])
# Implementation of Feedforward model
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.norm1 = nn.LayerNorm(d_model)
self.norm2_list = nn.ModuleList([copy.deepcopy(nn.LayerNorm(d_model)) for i in range(2)])
self.norm3 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2_list = nn.ModuleList([copy.deepcopy(nn.Dropout(dropout)) for i in range(2)])
self.dropout3 = nn.Dropout(dropout)
self.activation = _get_activation_fn(activation)
self.normalize_before = normalize_before
def with_pos_embed(self, tensor, pos: Optional[Tensor]):
return tensor if pos is None else tensor + pos
def forward_post(self, tgt, memory_list, tgt_mask=None, memory_mask=None,
tgt_key_padding_mask=None, memory_key_padding_mask=None,
pos=None, memory_pos=None):
q = k = self.with_pos_embed(tgt, pos)
tgt2 = self.self_attn(q, k, value=tgt, attn_mask=tgt_mask,
key_padding_mask=tgt_key_padding_mask)[0]
tgt = tgt + self.dropout1(tgt2)
tgt = self.norm1(tgt)
for memory, multihead_attn, norm2, dropout2, m_pos in zip(memory_list, self.multihead_attn_list, self.norm2_list, self.dropout2_list, memory_pos):
tgt2 = multihead_attn(query=self.with_pos_embed(tgt, pos),
key=self.with_pos_embed(memory, m_pos),
value=memory, attn_mask=memory_mask,
key_padding_mask=memory_key_padding_mask)[0]
tgt = tgt + dropout2(tgt2)
tgt = norm2(tgt)
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt))))
tgt = tgt + self.dropout3(tgt2)
tgt = self.norm3(tgt)
return tgt
def forward_pre(self, tgt, memory, tgt_mask=None, memory_mask=None,
tgt_key_padding_mask=None, memory_key_padding_mask=None,
pos=None, memory_pos=None):
tgt2 = self.norm1(tgt)
q = k = self.with_pos_embed(tgt2, 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)
tgt2 = self.norm2(tgt)
tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt2, pos),
key=self.with_pos_embed(memory, memory_pos),
value=memory, attn_mask=memory_mask,
key_padding_mask=memory_key_padding_mask)[0]
tgt = tgt + self.dropout2(tgt2)
tgt2 = self.norm3(tgt)
tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
tgt = tgt + self.dropout3(tgt2)
return tgt
def forward(self, tgt, memory_list, tgt_mask=None, memory_mask=None,
tgt_key_padding_mask=None, memory_key_padding_mask=None,
pos=None, memory_pos=None):
if self.normalize_before:
return self.forward_pre(tgt, memory_list, tgt_mask, memory_mask,
tgt_key_padding_mask, memory_key_padding_mask, pos, memory_pos)
return self.forward_post(tgt, memory_list, tgt_mask, memory_mask,
tgt_key_padding_mask, memory_key_padding_mask, pos, memory_pos)
def _get_clones(module, N):
return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
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 TransDecoder(nn.Module):
def __init__(self, num_attn_layers, hidden_dim=128):
super(TransDecoder, self).__init__()
attn_layer = attnLayer(hidden_dim)
self.layers = _get_clones(attn_layer, num_attn_layers)
self.position_embedding = build_position_encoding(hidden_dim)
def forward(self, imgf, query_embed):
pos = self.position_embedding(torch.ones(imgf.shape[0], imgf.shape[2], imgf.shape[3]).bool()) #.cuda()) # torch.Size([1, 128, 36, 36])
bs, c, h, w = imgf.shape
imgf = imgf.flatten(2).permute(2, 0, 1)
query_embed = query_embed.unsqueeze(1).repeat(1, bs, 1)
pos = pos.flatten(2).permute(2, 0, 1)
for layer in self.layers:
query_embed = layer(query_embed, [imgf], pos=pos, memory_pos=[pos, pos])
query_embed = query_embed.permute(1, 2, 0).reshape(bs, c, h, w)
return query_embed
class TransEncoder(nn.Module):
def __init__(self, num_attn_layers, hidden_dim=128):
super(TransEncoder, self).__init__()
attn_layer = attnLayer(hidden_dim)
self.layers = _get_clones(attn_layer, num_attn_layers)
self.position_embedding = build_position_encoding(hidden_dim)
def forward(self, imgf):
pos = self.position_embedding(torch.ones(imgf.shape[0], imgf.shape[2], imgf.shape[3]).bool()) #.cuda()) # torch.Size([1, 128, 36, 36])
bs, c, h, w = imgf.shape
imgf = imgf.flatten(2).permute(2, 0, 1)
pos = pos.flatten(2).permute(2, 0, 1)
for layer in self.layers:
imgf = layer(imgf, [imgf], pos=pos, memory_pos=[pos, pos])
imgf = imgf.permute(1, 2, 0).reshape(bs, c, h, w)
return imgf
class FlowHead(nn.Module):
def __init__(self, input_dim=128, hidden_dim=256):
super(FlowHead, self).__init__()
self.conv1 = nn.Conv2d(input_dim, hidden_dim, 3, padding=1)
self.conv2 = nn.Conv2d(hidden_dim, 2, 3, padding=1)
self.relu = nn.ReLU(inplace=True)
def forward(self, x):
return self.conv2(self.relu(self.conv1(x)))
class UpdateBlock(nn.Module):
def __init__(self, hidden_dim=128):
super(UpdateBlock, self).__init__()
self.flow_head = FlowHead(hidden_dim, hidden_dim=256)
self.mask = nn.Sequential(
nn.Conv2d(hidden_dim, 256, 3, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(256, 64*9, 1, padding=0))
def forward(self, imgf, coords1):
mask = .25 * self.mask(imgf) # scale mask to balence gradients
dflow = self.flow_head(imgf)
coords1 = coords1 + dflow
return mask, coords1
def coords_grid(batch, ht, wd):
coords = torch.meshgrid(torch.arange(ht), torch.arange(wd))
coords = torch.stack(coords[::-1], dim=0).float()
return coords[None].repeat(batch, 1, 1, 1)
def upflow8(flow, mode='bilinear'):
new_size = (8 * flow.shape[2], 8 * flow.shape[3])
return 8 * F.interpolate(flow, size=new_size, mode=mode, align_corners=True)
class GeoTr(nn.Module):
def __init__(self, num_attn_layers):
super(GeoTr, self).__init__()
self.num_attn_layers = num_attn_layers
self.hidden_dim = hdim = 256
self.fnet = BasicEncoder(output_dim=hdim, norm_fn='instance')
self.TransEncoder = TransEncoder(self.num_attn_layers, hidden_dim=hdim)
self.TransDecoder = TransDecoder(self.num_attn_layers, hidden_dim=hdim)
self.query_embed = nn.Embedding(1296, self.hidden_dim)
self.update_block = UpdateBlock(self.hidden_dim)
def initialize_flow(self, img):
N, C, H, W = img.shape
coodslar = coords_grid(N, H, W).to(img.device)
coords0 = coords_grid(N, H // 8, W // 8).to(img.device)
coords1 = coords_grid(N, H // 8, W // 8).to(img.device)
return coodslar, coords0, coords1
def upsample_flow(self, flow, mask):
N, _, H, W = flow.shape
mask = mask.view(N, 1, 9, 8, 8, H, W)
mask = torch.softmax(mask, dim=2)
up_flow = F.unfold(8 * flow, [3, 3], padding=1)
up_flow = up_flow.view(N, 2, 9, 1, 1, H, W)
up_flow = torch.sum(mask * up_flow, dim=2)
up_flow = up_flow.permute(0, 1, 4, 2, 5, 3)
return up_flow.reshape(N, 2, 8 * H, 8 * W)
def forward(self, image1):
fmap = self.fnet(image1)
fmap = torch.relu(fmap)
fmap = self.TransEncoder(fmap)
fmap = self.TransDecoder(fmap, self.query_embed.weight)
# convex upsample baesd on fmap
coodslar, coords0, coords1 = self.initialize_flow(image1)
coords1 = coords1.detach()
mask, coords1 = self.update_block(fmap, coords1)
flow_up = self.upsample_flow(coords1 - coords0, mask)
bm_up = coodslar + flow_up
return bm_up
|