SILog loss (pixel-wise)
| 40 | |
| 41 | # Main loss function used for ZoeDepth. Copy/paste from AdaBins repo (https://github.com/shariqfarooq123/AdaBins/blob/0952d91e9e762be310bb4cd055cbfe2448c0ce20/loss.py#L7) |
| 42 | class SILogLoss(nn.Module): |
| 43 | """SILog loss (pixel-wise)""" |
| 44 | def __init__(self, beta=0.15): |
| 45 | super(SILogLoss, self).__init__() |
| 46 | self.name = 'SILog' |
| 47 | self.beta = beta |
| 48 | |
| 49 | def forward(self, input, target, mask=None, interpolate=True, return_interpolated=False): |
| 50 | input = extract_key(input, KEY_OUTPUT) |
| 51 | if input.shape[-1] != target.shape[-1] and interpolate: |
| 52 | input = nn.functional.interpolate( |
| 53 | input, target.shape[-2:], mode='bilinear', align_corners=True) |
| 54 | intr_input = input |
| 55 | else: |
| 56 | intr_input = input |
| 57 | |
| 58 | if target.ndim == 3: |
| 59 | target = target.unsqueeze(1) |
| 60 | |
| 61 | if mask is not None: |
| 62 | if mask.ndim == 3: |
| 63 | mask = mask.unsqueeze(1) |
| 64 | |
| 65 | input = input[mask] |
| 66 | target = target[mask] |
| 67 | |
| 68 | with amp.autocast(enabled=False): # amp causes NaNs in this loss function |
| 69 | alpha = 1e-7 |
| 70 | g = torch.log(input + alpha) - torch.log(target + alpha) |
| 71 | |
| 72 | # n, c, h, w = g.shape |
| 73 | # norm = 1/(h*w) |
| 74 | # Dg = norm * torch.sum(g**2) - (0.85/(norm**2)) * (torch.sum(g))**2 |
| 75 | |
| 76 | Dg = torch.var(g) + self.beta * torch.pow(torch.mean(g), 2) |
| 77 | |
| 78 | loss = 10 * torch.sqrt(Dg) |
| 79 | |
| 80 | if torch.isnan(loss): |
| 81 | print("Nan SILog loss") |
| 82 | print("input:", input.shape) |
| 83 | print("target:", target.shape) |
| 84 | print("G", torch.sum(torch.isnan(g))) |
| 85 | print("Input min max", torch.min(input), torch.max(input)) |
| 86 | print("Target min max", torch.min(target), torch.max(target)) |
| 87 | print("Dg", torch.isnan(Dg)) |
| 88 | print("loss", torch.isnan(loss)) |
| 89 | |
| 90 | if not return_interpolated: |
| 91 | return loss |
| 92 | |
| 93 | return loss, intr_input |
| 94 | |
| 95 | |
| 96 | def grad(x): |