About loss function

[wudongming97]

Hi, I found that the loss used in this repo is a cross-entropy loss between prediction and mask.

loss = F.binary_cross_entropy_with_logits(pred, mask)

But the loss mentioned in the paper is a contrastive loss between visual and textual features.

[Deepayan137]

I have the same query. Can the authors please clarify?

[Deepayan137]

Hello! I wrote the contrastive learning part by following the instructions in the paper. However, when training the model only with the contrastive loss, the training IOU doesn't seem to improve. Below, I am attaching the code snippet and the training IOU and precision curves. The training is done only for 1 epoch. The brown plots are for cross-entropy loss while the blue plots are for contrastive loss. I would be grateful if you could let me know what I am doing wrong and also if the contrastive loss is supposed to be used in addition to cross-entropy loss.
Thanks

def forward(self, x, word, mask):
      x = self.vis(x)
      B, C, H, W = x.size()
      word = self.txt(word)
      x = x.permute(0, 2, 3, 1)
      out = torch.einsum('nhwc,nc->nhw', x, word).unsqueeze(1)
      out = torch.sigmoid(out) #sigmoid of zt dot zv
      loss = torch.zeros((H, W)).cuda()
      pos_count, neg_count = 0, 0
      for i in range(word.size(0)):
          zt = word[i]
          zt = zt.unsqueeze(0)
          for j in range(x.size(0)):
              zv = x[j]
              zv = zv.reshape(self.in_dim, -1)
              prod = torch.mm(zt, zv).squeeze()
              prod = prod.reshape(H, W)
              if i == j:
                  pos = - torch.log(F.sigmoid(prod))
                  loss += pos
                  pos_count += 1
              else:
                  neg = - torch.log(1 - F.sigmoid(prod))
                  loss += neg
                  neg_count += 1
      total = pos_count + neg_count
      loss = torch.mean(loss)
      if out.shape[-2:] != mask.shape[-2:]:
          mask = F.interpolate(mask, out.shape[-2:],
              mode='nearest').detach()
      return out, loss/total, mask

[DerrickWang005]

please follow our implementation.

CRIS.pytorch/model/layers.py

Lines 47 to 84 in 0df39f0

	class Projector(nn.Module):
	def __init__(self, word_dim=1024, in_dim=256, kernel_size=3):
	super().__init__()
	self.in_dim = in_dim
	self.kernel_size = kernel_size
	# visual projector
	self.vis = nn.Sequential( # os16 -> os4
	nn.Upsample(scale_factor=2, mode='bilinear'),
	conv_layer(in_dim * 2, in_dim * 2, 3, padding=1),
	nn.Upsample(scale_factor=2, mode='bilinear'),
	conv_layer(in_dim * 2, in_dim, 3, padding=1),
	nn.Conv2d(in_dim, in_dim, 1))
	# textual projector
	out_dim = 1 * in_dim * kernel_size * kernel_size + 1
	self.txt = nn.Linear(word_dim, out_dim)
	def forward(self, x, word):
	'''
	x: b, 512, 26, 26
	word: b, 512
	'''
	x = self.vis(x)
	B, C, H, W = x.size()
	# 1, b*256, 104, 104
	x = x.reshape(1, B * C, H, W)
	# txt: b, (256*3*3 + 1) -> b, 256, 3, 3 / b
	word = self.txt(word)
	weight, bias = word[:, :-1], word[:, -1]
	weight = weight.reshape(B, C, self.kernel_size, self.kernel_size)
	# Conv2d - 1, b*256, 104, 104 -> 1, b, 104, 104
	out = F.conv2d(x,
	weight,
	padding=self.kernel_size // 2,
	groups=weight.size(0),
	bias=bias)
	out = out.transpose(0, 1)
	# b, 1, 104, 104
	return out

[Deepayan137]

Hello Derrick,

I had seen this implementation. In your paper, you have mentioned equations 9 and 10 as the contrastive loss between pixel embeddings and the text features. I am not able to understand, how it is taken care of in your above code snippet?

[tiger990111]

I have the same query. Can the authors please clarify?

[FabianRitter]

No follow up? looks like supervised learning on the code. I assume something is missing in the code.

[Starboy-at-earth]

@DerrickWang005 Could you please realse the code snippet of contrastive learning loss?

[clownrat6]

Actually, the implementation is in line with the description of the paper. However, this is actually not the standard contrastive learning.

[Fake10086]

you may take a deeper look at codes mentioned by the author above, and you'll find that conv2d actually acts like element wise product between text and image which can be considered as equation 9&10.

[lyu-yx]

I have the same question, could the authors release the latest version of code? @DerrickWang005

[DerrickWang005]

I think this article can answer your question to some extent. @lyu-yx
https://arxiv.org/pdf/2303.15343.pdf

[ccccchenllll]

I have the same question. I couldn't find the code about contrastive loss.

[Shaosifan]

I have the same question. I couldn't find the code about contrastive loss.

Me too...

[Kimsure]

I prefer that Eq. 9 and Eq. 10 actually represent a BCE loss, and agree that this is actually not the standard contrastive learning. @clownrat6