File size: 9,281 Bytes
ea1014e |
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 |
import torch
import torch.nn as nn
from .layers import Decoder
from .layers_v2 import Decoder_v2
import torch.nn.functional as F
from bert.modeling_bert import BertModel
def dice_loss(inputs, targets):
"""
Compute the DICE loss, similar to generalized IOU for masks
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs
(0 for the negative class and 1 for the positive class).
"""
inputs = inputs.sigmoid()
inputs = inputs.flatten(1)
targets = targets.flatten(1)
numerator = 2 * (inputs * targets).sum(1)
denominator = inputs.sum(-1) + targets.sum(-1)
loss = 1 - (numerator + 1) / (denominator + 1)
return loss.mean()
def sigmoid_focal_loss(inputs, targets, alpha: float = 0.25, gamma: float = 2):
"""
Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs
(0 for the negative class and 1 for the positive class).
alpha: (optional) Weighting factor in range (0,1) to balance
positive vs negative examples. Default = -1 (no weighting).
gamma: Exponent of the modulating factor (1 - p_t) to
balance easy vs hard examples.
Returns:
Loss tensor
"""
prob = inputs.sigmoid()
ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
p_t = prob * targets + (1 - prob) * (1 - targets)
loss = ce_loss * ((1 - p_t) ** gamma)
if alpha >= 0:
alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
loss = alpha_t * loss
return loss.mean()
class CGFormer_RCC_sbert(nn.Module):
def __init__(self, backbone, args):
super(CGFormer_RCC_sbert, self).__init__()
self.backbone = backbone
self.mixup_lasttwo = args.mixup_lasttwo
if self.mixup_lasttwo :
self.decoder = Decoder_v2(args)
else :
self.decoder = Decoder(args)
self.text_encoder = BertModel.from_pretrained(args.bert)
self.text_encoder.pooler = None
self.args = args
if self.args.use_projections :
self.projection_1 = nn.Linear(1536, 1024, bias=True)
else :
self.projection_1 = None
self.use_projections = args.use_projections
self.filter_th = args.filter_threshold
# image, text, l_mask, target, hardpos, hp_emb
def forward(self, x, text, l_mask, mask=None, hp_bert_embs=None):
rows_to_filter, cols_to_filter = None, None
if self.training:
# In fact, we don't need verb_masks, cl_masks for ACE in RCC(+)
# Only calculate embedding similarity for ACE
norms = torch.norm(hp_bert_embs, dim=-1, keepdim=True)
normed_embs = hp_bert_embs / norms
cosime_sim = torch.mm(normed_embs, normed_embs.T)
rows_to_filter, cols_to_filter = torch.where(cosime_sim > self.filter_th)
# print(rows_to_filter, cols_to_filter)
input_shape = x.shape[-2:]
l_feats = self.text_encoder(text, attention_mask=l_mask)[0] # (6, 10, 768)
l_feats = l_feats.permute(0, 2, 1) # (B, 768, N_l) to make Conv1d happy
l_mask = l_mask.unsqueeze(dim=-1) # (batch, N_l, 1)
##########################
features = self.backbone(x, l_feats, l_mask)
x_c1, x_c2, x_c3, x_c4 = features
if self.mixup_lasttwo :
pred, maps, fq_fuse = self.decoder([x_c4, x_c3, x_c2, x_c1], l_feats, l_mask)
metric_tensor = F.adaptive_avg_pool2d(fq_fuse, (1, 1)).view(fq_fuse.shape[0], fq_fuse.shape[1])
# print(fq_fuse.shape, metric_tensor.shape)
else :
# not mixup_lasttwo and just using projection, or else use only the fq4 output
pred, maps = self.decoder([x_c4, x_c3, x_c2, x_c1], l_feats, l_mask)
if self.training :
if self.use_projections :
x_c3_proj = F.adaptive_avg_pool2d(x_c3, (1, 1)).view(x_c3.size(0), -1)
x_c4_proj = F.adaptive_avg_pool2d(x_c4, (1, 1)).view(x_c4.size(0), -1)
metric_tensor = torch.cat((x_c3_proj, x_c4_proj), dim=1)
metric_tensor = self.projection_1(metric_tensor)
else :
metric_tensor = F.adaptive_avg_pool2d(x_c4, (1, 1)).view(x_c4.size(0), -1)
pred = F.interpolate(pred, input_shape, mode='bilinear', align_corners=True)
# loss
if self.training:
loss = 0.
mask = mask.unsqueeze(1).float()
for m, lam in zip(maps, [0.001,0.01,0.1]):
m = m[:,1].unsqueeze(1)
if m.shape[-2:] != mask.shape[-2:]:
mask_ = F.interpolate(mask, m.shape[-2:], mode='nearest').detach()
loss += dice_loss(m, mask_) * lam
loss += dice_loss(pred, mask) + sigmoid_focal_loss(pred, mask, alpha=-1, gamma=0)
metric_loss = 0.
if hp_bert_embs.numel() > 0 :
metric_loss = self.compute_metric_loss(metric_tensor, rows_to_filter, cols_to_filter, self.args)
loss += metric_loss * self.args.metric_loss_weight
return pred.detach(), mask, loss
else:
return pred.detach(), maps
def compute_metric_loss(self, metric_tensor, rows_to_filter, cols_to_filter, args) :
if args.loss_option == "ACL_verbonly" :
raise ValueError("ACL_verbonly is not supported in CGFormer")
elif args.loss_option == "ACE_verbonly" :
metric_loss = self.UniAngularLogitContrastLoss(metric_tensor, rows_to_filter, cols_to_filter, m=args.margin_value, tau=args.temperature, verbonly=True, args=args)
return metric_loss
def return_mask(self, emb_distance, rows_to_filter=None, cols_to_filter=None):
B_, B_ = emb_distance.shape
positive_mask = torch.zeros_like(emb_distance)
positive_mask.fill_diagonal_(1) # Set diagonal elements to 1 for all cases
negative_mask = torch.ones_like(emb_distance) - positive_mask
negative_mask = negative_mask.clone()
if rows_to_filter is not None and cols_to_filter is not None :
for row, col in zip(rows_to_filter, cols_to_filter):
negative_mask[row , col] = 0
return positive_mask, negative_mask
def UniAngularLogitContrastLoss(self, total_fq, rows_to_filter, cols_to_filter, alpha=0.5, verbonly=True, m=0.5, tau=0.05, args=None):
_, HW = total_fq.shape
# _, C, H, W = total_fq.shape
emb = torch.mean(total_fq, dim=1, keepdim=True) # (B_, 1)
B_ = emb.shape[0]
emb_i = emb.unsqueeze(1).repeat(1, B_, 1) # (B_, B_, C)
emb_j = emb.unsqueeze(0).repeat(B_, 1, 1) # (B_, B_, C)
sim = nn.CosineSimilarity(dim=-1, eps=1e-6)
sim_matrix = sim(emb_i, emb_j).reshape(B_, B_) # (B_, B_)
sim_matrix = torch.clamp(sim_matrix, min=-0.999, max=0.999)
# print("sim matrix : ", sim_matrix)
margin_in_radians = m / 57.2958 # Convert degrees to radians
# print("sim_matrix : ", sim_matrix)
theta_matrix = (torch.pi / 2) - torch.acos(sim_matrix)
# print("theta_matrix : ", theta_matrix)
positive_mask, negative_mask = self.return_mask(sim_matrix, rows_to_filter, cols_to_filter)
# print("? `positive_mask` requires_grad:", positive_mask.requires_grad, positive_mask.device)
# print("? `negative_mask` requires_grad:", negative_mask.requires_grad, negative_mask.device)
# print("positive_mask : ", positive_mask)
# print("negative_mask : ", negative_mask)
# print("? `positive_mask` requires_grad:", positive_mask.requires_grad,
# "device:", positive_mask.device, "dtype:", positive_mask.dtype)
# print("? `negative_mask` requires_grad:", negative_mask.requires_grad,
# "device:", negative_mask.device, "dtype:", negative_mask.dtype)
theta_with_margin = theta_matrix.clone()
theta_with_margin[positive_mask.bool()] -= margin_in_radians
logits = theta_with_margin / tau # Scale with temperature
exp_logits = torch.exp(logits)
pos_exp_logits = exp_logits * positive_mask
pos_exp_logits = pos_exp_logits.sum(dim=-1)
neg_exp_logits = exp_logits * negative_mask
neg_exp_logits = neg_exp_logits.sum(dim=-1)
total_exp_logits = pos_exp_logits + neg_exp_logits
positive_loss = -torch.log(pos_exp_logits/ total_exp_logits)
angular_loss = positive_loss.mean()
# print("angular_loss : ", angular_loss)
return angular_loss |