# Loss dropping to zero in two epochs **URL:** \ **Category:** vision\ **Created:** [February 22, 2020, 12:13am UTC](https://discuss.pytorch.org/t/loss-dropping-to-zero-in-two-epochs/70636 "2020-02-22T00:13:59Z")\ **Posts on this page:** 3\ **Page:** 1
**Author:** ![PCPJ](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/p/47e85d/32.png) [@PCPJ](https://discuss.pytorch.org/u/PCPJ)\ **Post date:** [February 22, 2020, 12:13am UTC](https://discuss.pytorch.org/t/loss-dropping-to-zero-in-two-epochs/70636/1 "2020-02-22T00:13:59Z")
Hello, everyone. I am new to pytorch, so sorry if something sounds stupid. My model consists of a vgg16 that outputs a 64x64 grid of boolean predictions. I have built my own custom loss functions, but I think that I am doing something really wrong. In two epochs the loss is dropping to zero, but when I try to test the model, with the exact same imagens used on training, I got results really weird. For all images the result was a bunch of zeros and a one in the first position: [1, 0, 0, 0, …0] [0, 0, 0, 0, …0] … … … [0, 0, 0, 0, …0] I will post here the loss function that I built. If someone could give me a hint on what I am doing wrong, I would be really grateful. If my question/explanation was confused/incomplete, please tell me, so I can try to explain better. ```auto true_scores = torch.zeros((batch_size, 1, self.grid_size, self.grid_size)).to(device) # (N, 1, 64, 64) # For each image for i in range(batch_size): for row in targets[i]: cells = torch.floor(row/self.cell_size).type(torch.long) # (n_points, 2) cells = cells[:, 1]*self.grid_size + cells[:, 0] # (n_points) true_scores[i][0].view(self.grid_size**2)[cells] = 1.0 score_loss = self.logist_bce(pred, true_scores) return score_loss ``` ---
**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\ **Post date:** [February 22, 2020, 7:26am UTC](https://discuss.pytorch.org/t/loss-dropping-to-zero-in-two-epochs/70636/2 "2020-02-22T07:26:57Z")
How did you implement `logist_bce`? Could you post a small code snippet with a random input and target tensor, so that we could have a better look at the use methods? 🙂 ---
**Author:** ![PCPJ](https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/p/47e85d/32.png) [@PCPJ](https://discuss.pytorch.org/u/PCPJ)\ **Post date:** [February 22, 2020, 3:17pm UTC](https://discuss.pytorch.org/t/loss-dropping-to-zero-in-two-epochs/70636/3 "2020-02-22T15:17:31Z")
Hello, @ptrblck. Thanks for the reply! Logist\_bce is the nn.BCEWithLogitsLoss from PyTorch. I will post the code from my custom loss function here, it is just a few lines. I wrote a code simulation of my loss with random input. For better visualization, I used a grid size of 4. So the input of the loss should be a tensor of size (batch\_size, 1, 4, 4), and a list of list of tensors with (x,y) coordinates. ```auto x = torch.randn(1,1,4,4) loss = MyLoss(4) y = list() rows = list() r1 = torch.Tensor([[0.1, 0.1], [0.3, 0.3]]) # tensor with two coordinates normalized to 0-1, (0.1, 0.1) and (0.3, 0.3) r2 = torch.Tensor([[0.6, 0.6], [0.9, 0.9]]) # tensor with two coordinates normalized to 0-1, (0.6, 0.6) and (0.9, 0.9) rows.append(r1) rows.append(r2) y.append(rows) x tensor([[[[0.1691, -1.2786, 1.3548, -0.3432], [0.8771, -0.8598, -1.0828, -0.7924], [0.1644, 1.0497, -0.4348, -1.5413], [0.7028, -0.6409, 0.5056, -1.9122]]]]) >>> y [[tensor([[0.1000, 0.1000], [0.3000, 0.3000]]), tensor([[0.6000, 0.6000], [0.9000, 0.9000]])]] >>> loss(x,y) true_scores = tensor([[[[1., 0., 0., 0.], [0., 1., 0., 0.], [0., 0., 1., 0.], [0., 0., 0., 1.]]]]) tensor(2.3711) ``` ```auto class MyLoss(nn.Module): def __init__ (self, grid_size=64): super(CropRowLoss, self). __init__ () # grid_size => The size of the grid, in cells self.grid_size = grid_size self.cell_size = 1.0/self.grid_size self.logist_bce = nn.BCEWithLogitsLoss(reduction='mean', pos_weight=torch.FloatTensor([6.0]).to(device)) def forward(self, pred, targets): """ Forward propagation. :param pred: a tensor of dimensions (N, 1, grid_size, grid_size) :param targets: list of list of tensor (n_points, 2) """ batch_size = pred.size(0) true_scores = torch.zeros((batch_size, 1, self.grid_size, self.grid_size)).to(device) # (N, 1, 64, 64) # For each image for i in range(batch_size): for row in targets[i]: cells = torch.floor(row/self.cell_size).type(torch.long) # (n_points, 2) cells = cells[:, 1]*self.grid_size + cells[:, 0] # (n_points) true_scores[i][0].view(self.grid_size**2)[cells] = 1.0 print("true_scores = \n", true_scores) score_loss = self.logist_bce(pred, true_scores) # Scalar return score_loss ```