| import torch |
| import torch.nn as nn |
|
|
| from torchmetrics import CatMetric |
| from torchmetrics import SumMetric |
| from torchmetrics.image.fid import FrechetInceptionDistance |
| from torchmetrics.multimodal.clip_score import CLIPScore |
|
|
| from jutils.nn import DINOv2, preprocess_for_dinov2 |
| from jutils.nn.metric_kid import kid_features_to_metric |
|
|
|
|
| def un_normalize_ims(ims): |
| """Convert from [-1, 1] to [0, 255]""" |
| ims = ((ims * 127.5) + 127.5).clip(0, 255).to(torch.uint8) |
| return ims |
|
|
|
|
| class ImageMetricTracker(nn.Module): |
| def __init__(self): |
| super().__init__() |
| self.total_samples = SumMetric() |
|
|
| self.fid = FrechetInceptionDistance( |
| feature=2048, reset_real_features=True, normalize=False, sync_on_compute=True |
| ) |
|
|
| def __call__(self, target, pred): |
| """Assumes target and pred in [-1, 1] range""" |
| bs = target.shape[0] |
| real_ims = un_normalize_ims(target) |
| fake_ims = un_normalize_ims(pred) |
|
|
| self.fid.update(real_ims, real=True) |
| self.fid.update(fake_ims, real=False) |
|
|
| self.total_samples.update(bs) |
|
|
| def reset(self): |
| self.fid.reset() |
| self.total_samples.reset() |
|
|
| def aggregate(self): |
| """Compute the final metrics (automatically synced across devices)""" |
| n_total_samples = int(self.total_samples.compute()) |
| return { |
| f"fid-{n_total_samples}": self.fid.compute(), |
| "n_metric_samples": n_total_samples, |
| } |
|
|
|
|
| class Text2ImageMetricTracker(nn.Module): |
| def __init__(self, kid_subsets: int = 100, kid_subset_size: int = 200): |
| super().__init__() |
| self.total_samples = SumMetric() |
| self.fid = FrechetInceptionDistance( |
| feature=2048, reset_real_features=True, normalize=False, sync_on_compute=True |
| ) |
|
|
| self.clip = CLIPScore(model_name_or_path="openai/clip-vit-base-patch16") |
|
|
| self.dino = DINOv2(pretrained=True).eval() |
| self.dino_features_real = CatMetric() |
| self.dino_features_fake = CatMetric() |
| self.kid_subsets = kid_subsets |
| self.kid_subset_size = kid_subset_size |
|
|
| for p in self.parameters(): |
| p.requires_grad = False |
|
|
| @torch.no_grad() |
| def __call__(self, target, pred, txt): |
| """Assumes target and pred in [-1, 1] range""" |
| bs = target.shape[0] |
| real_ims = un_normalize_ims(target) |
| fake_ims = un_normalize_ims(pred) |
|
|
| self.fid.update(real_ims, real=True) |
| self.fid.update(fake_ims, real=False) |
|
|
| txt = [t.decode() if isinstance(t, bytes) else t for t in txt] |
| self.clip.update(fake_ims, list(txt)) |
|
|
| real_fts = self.dino(preprocess_for_dinov2(target, safe_mode=False)) |
| fake_fts = self.dino(preprocess_for_dinov2(pred, safe_mode=False)) |
| self.dino_features_real.update(real_fts) |
| self.dino_features_fake.update(fake_fts) |
|
|
| self.total_samples.update(bs) |
|
|
| def reset(self): |
| self.fid.reset() |
| self.clip.reset() |
| self.dino_features_real.reset() |
| self.dino_features_fake.reset() |
| self.total_samples.reset() |
|
|
| def aggregate(self): |
| """Compute the final metrics (automatically synced across devices)""" |
| n_total_samples = int(self.total_samples.compute()) |
|
|
| |
| real_fts = self.dino_features_real.compute() |
| fake_fts = self.dino_features_fake.compute() |
| kdd = kid_features_to_metric( |
| real_fts, |
| fake_fts, |
| kid_subsets=self.kid_subsets, |
| kid_subset_size=self.kid_subset_size, |
| verbose=False, |
| ) |
|
|
| return { |
| f"fid-{n_total_samples}": self.fid.compute(), |
| f"clip": self.clip.compute(), |
| **kdd, |
| "n_metric_samples": n_total_samples, |
| } |
|
|
