Spaces:
No application file
No application file
File size: 9,386 Bytes
6755a2d |
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 |
import torch
import torch.nn.functional as F
def pairLoss(fea1, fea2, mask):
# fea1_size (bs, max_len, dim)
# fea2_size (bs, max_len, dim)
# mask_size (bs, max_len)
# '-Inf' for padded item, '0' for others
fea1 = F.normalize(fea1, p=2, dim=-1)
fea2 = F.normalize(fea2, p=2, dim=-1)
fea_sim = (fea1 * fea2).sum(dim=-1) # (bs, max_len)
fea_sim = torch.masked_select(fea_sim, mask == 0)
loss = 1.0 - torch.mean(fea_sim)
return loss
def SimpTripLoss(fea1, fea2):
# img fea1_size (bs, max_len1, dim) mask1_size (bs, max_len1)
# text fea2_size (bs, max_len2, dim) mask2_size (bs, max_len2)
# '-Inf' for padded item, '0' for others
# fea1 = fea1.mean(dim=1) #(bs, dim)
# mask2 = torch.where(mask2==0, torch.tensor([1.0],device=mask2.device), torch.tensor([0.0],device=mask2.device))
# fea2 = (fea2 * mask2.unsqueeze(-1)).sum(dim=1) / mask2.sum(dim=1).unsqueeze(-1) #(bs, dim)
fea1 = F.normalize(fea1, p=2, dim=-1)
fea2 = F.normalize(fea2, p=2, dim=-1)
# match fea1 to fea2
sim_pos1 = (fea1 * fea2).sum(dim=1) # (bs)
# (bs, 1, dim) (1, bs, dim)
sim_neg1_all = (fea1.unsqueeze(1) * fea2.unsqueeze(0)).sum(dim=-1) # (bs,bs)
unmask = torch.eye(sim_pos1.size(0), dtype=torch.float32, device=sim_pos1.device)
unmask = torch.where(unmask == 1, torch.tensor([float('-Inf')], device=unmask.device), unmask)
sim_neg1, _ = torch.max(sim_neg1_all + unmask, 1)
loss1 = -sim_pos1 + sim_neg1 + 0.2
loss1 = torch.maximum(loss1, torch.zeros_like(loss1)).mean()
# match fea2 to fea1
sim_pos2 = (fea2 * fea1).sum(
dim=1) # (bs) sim_neg2_all = (fea2.unsqueeze(1) * fea1.unsqueeze(0)).sum(dim=-1) #(bs,bs)
sim_neg2_all = (fea2.unsqueeze(1) * fea1.unsqueeze(0)).sum(dim=-1) # (bs,bs)
sim_neg2, _ = torch.max(sim_neg2_all + unmask, 1)
loss2 = -sim_pos2 + sim_neg2 + 0.2
loss2 = torch.maximum(loss2, torch.zeros_like(loss2)).mean()
loss = loss1 + loss2
return loss
def NCELoss(fea1, fea2):
# img fea1_size (bs, max_len1, dim) mask1_size (bs, max_len1)
# text fea2_size (bs, max_len2, dim) mask2_size (bs, max_len2)
# '-Inf' for padded item, '0' for others
# fea1 = fea1.mean(dim=1) #(bs, dim)
# mask2 = torch.where(mask2==0, torch.tensor([1.0],device=mask2.device), torch.tensor([0.0],device=mask2.device))
# fea2 = (fea2 * mask2.unsqueeze(-1)).sum(dim=1) / mask2.sum(dim=1).unsqueeze(-1) #(bs, dim)
fea1 = F.normalize(fea1, p=2, dim=-1)
fea2 = F.normalize(fea2, p=2, dim=-1)
# match fea1 to fea2
sim_pos1 = (fea1 * fea2).sum(dim=1).unsqueeze(-1) # (bs,1)
BS = sim_pos1.size(0)
# (bs, 1, dim) (1, bs, dim)
sim_neg1_all = (fea1.unsqueeze(1) * fea2.unsqueeze(0)).sum(dim=-1) # (bs,bs)
unmask = torch.eye(sim_pos1.size(0), dtype=torch.float32, device=sim_pos1.device)
sim_neg1_all = torch.masked_select(sim_neg1_all, unmask == 0).view(BS, BS - 1) # (bs, bs-1)
sim1_pos_neg = torch.cat((sim_pos1, sim_neg1_all), dim=1) / 0.07 # (bs, bs)
loss1 = -F.log_softmax(sim1_pos_neg, dim=1)[:, 0].mean()
# match fea2 to fea1
sim_pos2 = (fea2 * fea1).sum(dim=1).unsqueeze(-1) # (bs,1)
sim_neg2_all = (fea2.unsqueeze(1) * fea1.unsqueeze(0)).sum(dim=-1) # (bs,bs)
sim_neg2_all = torch.masked_select(sim_neg2_all, unmask == 0).view(BS, BS - 1) # (bs, bs-1)
sim2_pos_neg = torch.cat((sim_pos2, sim_neg2_all), dim=1) / 0.07 # (bs, bs)
loss2 = -F.log_softmax(sim2_pos_neg, dim=1)[:, 0].mean()
loss = (loss1 + loss2) / 2.0
return loss
def AlignTripLoss(fea1, fea2, mask1, mask2):
# fea1_size (bs, max_len1, dim) mask1_size (bs, max_len1)
# fea2_size (bs, max_len2, dim) mask2_size (bs, max_len2)
# '-Inf' for padded item, '0' for others
fea1 = F.normalize(fea1, p=2, dim=-1)
fea2 = F.normalize(fea2, p=2, dim=-1)
# match fea1 to fea2
sim_pos1 = cal_sim(fea1, fea2, mask1, mask2) # (bs)
# (bs, 1, max_len1, dim) (1, bs, max_len2, dim)
sim_neg1_all = cal_sim_all(fea1.unsqueeze(1), fea2.unsqueeze(0), mask1, mask2) # (bs,bs)
unmask = torch.eye(sim_pos1.size(0), dtype=torch.float32, device=sim_pos1.device)
unmask = torch.where(unmask == 1, torch.tensor([float('-Inf')], device=unmask.device), unmask)
sim_neg1, _ = torch.max(sim_neg1_all + unmask, 1)
loss1 = -sim_pos1 + sim_neg1 + 0.2
loss1 = torch.maximum(loss1, torch.zeros_like(loss1)).mean()
# match fea2 to fea1
sim_pos2 = cal_sim(fea2, fea1, mask2, mask1) # (bs)
# (bs, 1, max_len1, dim) (1, bs, max_len2, dim)
sim_neg2_all = cal_sim_all(fea2.unsqueeze(1), fea1.unsqueeze(0), mask2, mask1) # (bs,bs)
sim_neg2, _ = torch.max(sim_neg2_all + unmask, 1)
loss2 = -sim_pos2 + sim_neg2 + 0.2
loss2 = torch.maximum(loss2, torch.zeros_like(loss2)).mean()
loss = loss1 + loss2
return loss
def cal_sim_all(fea1, fea2, mask1, mask2):
# fea1_size (bs, 1, max_len1, dim) mask1_size (bs, max_len1)
# fea2_size (1, bs, max_len2, dim) mask2_size (bs, max_len2)
# '-Inf' for padded item, '0' for others
max_len1 = fea1.size(2)
max_len2 = fea2.size(2)
bs = fea1.size(0)
fea1_tmp = fea1.unsqueeze(3) # (bs, 1, max_len1, 1, dim)
fea2_tmp = fea2.unsqueeze(2) # (1, bs, 1, max_len2, dim)
fea_sim = (fea1_tmp * fea2_tmp).sum(dim=-1) # (bs, bs, max_len1, max_len2)
fea_sim = fea_sim + mask2.unsqueeze(dim=1) # (bs, bs, max_len1, max_len2)
idxs = torch.argmax(fea_sim, dim=-1).view(-1).unsqueeze(-1) # (bs*bs*max_len1, 1)
fea_sim = fea_sim.view(-1, max_len2) # (bs*bs*max_len1, max_len2)
select_sim = torch.gather(fea_sim, 1, idxs).view(bs, bs, max_len1) # (bs, bs, max_len1)
mask1_mult = torch.where(mask1 == 0, torch.tensor([1.0], device=mask1.device),
torch.tensor([0.0], device=mask1.device)).unsqueeze(1) # (bs, 1, max_len1)
select_sim = (select_sim * mask1_mult).sum(dim=-1) / mask1_mult.sum(dim=-1) # (bs, bs)
return select_sim
def cal_sim(fea1, fea2, mask1, mask2):
# fea1_size (bs, max_len1, dim) mask1_size (bs, max_len1)
# fea2_size (bs, max_len2, dim) mask2_size (bs, max_len2)
# '-Inf' for padded item, '0' for others
max_len1 = fea1.size(1)
max_len2 = fea2.size(1)
fea1_tmp = fea1.unsqueeze(2) # (bs, max_len1, 1, dim)
fea2_tmp = fea2.unsqueeze(1) # (bs, 1, max_len2, dim)
fea_sim = (fea1_tmp * fea2_tmp).sum(dim=-1) # (bs, max_len1, max_len2)
fea_sim = fea_sim + mask2.unsqueeze(dim=1) # (bs, max_len1, max_len2)
idxs = torch.argmax(fea_sim, dim=-1).view(-1).unsqueeze(-1) # (bs*max_len1, 1)
fea_sim = fea_sim.view(-1, max_len2) # (bs*max_len1, max_len2)
select_sim = torch.gather(fea_sim, 1, idxs).view(-1, max_len1) # (bs, max_len1)
mask1_mult = torch.where(mask1 == 0, 1, 0)
select_sim = (select_sim * mask1_mult).sum(dim=-1) / mask1_mult.sum(dim=-1) # (bs)
return select_sim
def alignmentLoss(fea1, fea2, mask1, mask2):
# fea1_size (bs, max_len1, dim) mask1_size (bs, max_len1)
# fea2_size (bs, max_len2, dim) mask2_size (bs, max_len2)
# '-Inf' for padded item, '0' for others
fea1 = F.normalize(fea1, p=2, dim=-1)
fea2 = F.normalize(fea2, p=2, dim=-1)
loss1 = alignSingleLoss(fea1, fea2, mask1, mask2)
loss2 = alignSingleLoss(fea2, fea1, mask2, mask1)
loss = (loss1 + loss2) / 2.0
return loss
def attAlignmentLoss(fea1, fea2, mask1, mask2, attFc):
# fea1_size (bs, max_len1, dim) mask1_size (bs, max_len1)
# fea2_size (bs, max_len2, dim) mask2_size (bs, max_len2)
# '-Inf' for padded item, '0' for others
fea1 = F.normalize(fea1, p=2, dim=-1)
fea2 = F.normalize(fea2, p=2, dim=-1)
fea1_tmp = fea1.unsqueeze(2) # (bs, max_len1, 1, dim)
fea2_tmp = fea2.unsqueeze(1) # (bs, 1, max_len2, dim)
fea_sim = fea1_tmp * fea2_tmp
att_sim = attFc(fea_sim).squeeze(-1) # (bs, max_len1, max_len2)
fea_sim = fea_sim.sum(dim=-1) # (bs, max_len1, max_len2)
fea_sim = fea_sim * att_sim # (bs, max_len1, max_len2)
###Simple as max_len1=49
loss = torch.masked_select(fea_sim, (mask2 == 0).unsqueeze(1))
loss = 1.0 - loss.mean()
return loss
def alignSingleLoss(fea1, fea2, mask1, mask2):
# fea1_size (bs, max_len1, dim) mask1_size (bs, max_len1)
# fea2_size (bs, max_len2, dim) mask2_size (bs, max_len2)
# '-Inf' for padded item, '0' for others
fea1_tmp = fea1.unsqueeze(2) # (bs, max_len1, 1, dim)
fea2_tmp = fea2.unsqueeze(1) # (bs, 1, max_len2, dim)
fea_sim = (fea1_tmp * fea2_tmp).sum(dim=-1) # (bs, max_len1, max_len2)
fea_sim = fea_sim + mask2.unsqueeze(dim=1) # (bs, max_len1, max_len2)
idxs = torch.argmax(fea_sim, dim=-1).view(-1).unsqueeze(-1) # (bs*max_len1, 1)
fea_sim = fea_sim.view(-1, fea_sim.size(-1)) # (bs*max_len1, max_len2)
select_sim = torch.gather(fea_sim, 1, idxs).view(-1) # (bs*max_len1)
select_sim = torch.masked_select(select_sim, (mask1 == 0).view(-1))
loss = 1.0 - torch.mean(select_sim)
return loss
def getLanMask(seq_lens, max_len):
# seq_lens (bs)
mask = torch.ones((seq_lens.size(0), max_len)) # (bs, max_len)
idxs = torch.arange(max_len).unsqueeze(dim=0) # (1, max_len)
seq_lens = seq_lens.unsqueeze(-1) # (bs, 1)
mask = torch.where(idxs < seq_lens, mask, torch.Tensor([0.0]))
return mask
|