Get segformer to a trainable state

codes/data/image_folder_dataset.py

@@ -37,6 +37,8 @@ class ImageFolderDataset:
         if 'normalize' in opt.keys():
             if opt['normalize'] == 'stylegan2_norm':
                 self.normalize = Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
+            elif opt['normalize'] == 'imagenet':
+                self.normalize = Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225), inplace=True)
             else:
                 raise Exception('Unsupported normalize')
         else:

codes/models/byol/byol_for_semantic_chaining.py

@@ -51,6 +51,54 @@ def set_requires_grad(model, val):
         p.requires_grad = val
 
 
+# Specialized augmentor class that applies a set of image transformations on points as well, allowing one to track
+# where a point in the src image is located in the dest image. Restricts transformation such that this is possible.
+class PointwiseAugmentor(nn.Module):
+    def __init__(self, img_size=224):
+        super().__init__()
+        self.jitter = RandomApply(augs.ColorJitter(0.8, 0.8, 0.8, 0.2), p=0.8)
+        self.gray = augs.RandomGrayscale(p=0.2)
+        self.blur = RandomApply(filters.GaussianBlur2d((3, 3), (1.5, 1.5)), p=0.1)
+        self.rrc = augs.RandomResizedCrop((img_size, img_size), same_on_batch=True)
+
+    # Given a point in the source image, returns the same point in the source image, given the kornia RRC params.
+    def rrc_on_point(self, src_point, params):
+        dh, dw = params['dst'][:,2,1]-params['dst'][:,0,1], params['dst'][:,2,0] - params['dst'][:,0,0]
+        sh, sw = params['src'][:,2,1]-params['src'][:,0,1], params['src'][:,2,0] - params['src'][:,0,0]
+        scale_h, scale_w = sh.float() / dh.float(), sw.float() / dw.float()
+        t, l = src_point[0] - params['src'][0,0,1], src_point[1] - params['src'][0,0,0]
+        t = (t.float() / scale_h[0]).long()
+        l = (l.float() / scale_w[0]).long()
+        return torch.stack([t,l])
+
+    def flip_on_point(self, pt, input):
+        t, l = pt[0], pt[1]
+        center = input.shape[-1] // 2
+        return t, 2 * center - l
+
+    def forward(self, x, point):
+        d = self.jitter(x)
+        d = self.gray(d)
+        will_flip = random.random() > .5
+        if will_flip:
+            d = apply_hflip(d)
+            point = self.flip_on_point(point, x)
+        d = self.blur(d)
+
+        invalid = True
+        while invalid:
+            params = self.rrc.generate_parameters(d.shape)
+            potential = self.rrc_on_point(point, params)
+            # '10' is an arbitrary number: we want to provide some margin. Making predictions at the very edge of an image is not very useful.
+            if potential[0] <= 10 or potential[1] <= 10 or potential[0] > x.shape[-2]-10 or potential[1] > x.shape[-1]-10:
+                continue
+            d = self.rrc(d, params=params)
+            point = potential
+            invalid = False
+
+        return d, point
+
+
 # loss fn
 def loss_fn(x, y):
     x = F.normalize(x, dim=-1, p=2)
@@ -160,21 +208,21 @@ class NetWrapper(nn.Module):
             projector = MLP(dim, self.projection_size, self.projection_hidden_size)
         return projector.to(hidden)
 
-    def get_representation(self, x):
+    def get_representation(self, x, pt):
         if self.layer == -1:
-            return self.net(x)
+            return self.net(x, pt)
 
         if not self.hook_registered:
             self._register_hook()
 
-        unused = self.net(x)
+        unused = self.net(x, pt)
         hidden = self.hidden
         self.hidden = None
         assert hidden is not None, f'hidden layer {self.layer} never emitted an output'
         return hidden
 
-    def forward(self, x):
-        representation = self.get_representation(x)
+    def forward(self, x, pt):
+        representation = self.get_representation(x, pt)
         projector = self._get_projector(representation)
         projection = checkpoint(projector, representation)
         return projection
@@ -191,24 +239,13 @@ class BYOL(nn.Module):
             moving_average_decay=0.99,
             use_momentum=True,
             structural_mlp=False,
-            do_augmentation=False  # In DLAS this was intended to be done at the dataset level. For massive batch sizes
-                                   # this can overwhelm the CPU though, and it becomes desirable to do the augmentations
-                                   # on the GPU again.
     ):
         super().__init__()
 
         self.online_encoder = NetWrapper(net, projection_size, projection_hidden_size, layer=hidden_layer,
                                          use_structural_mlp=structural_mlp)
 
-        self.do_aug = do_augmentation
-        if self.do_aug:
-            augmentations = [ \
-                RandomApply(augs.ColorJitter(0.8, 0.8, 0.8, 0.2), p=0.8),
-                augs.RandomGrayscale(p=0.2),
-                augs.RandomHorizontalFlip(),
-                RandomApply(filters.GaussianBlur2d((3, 3), (1.5, 1.5)), p=0.1),
-                augs.RandomResizedCrop((self.cropped_img_size, self.cropped_img_size))]
-            self.aug = nn.Sequential(*augmentations)
+        self.aug = PointwiseAugmentor(image_size)
         self.use_momentum = use_momentum
         self.target_encoder = None
         self.target_ema_updater = EMA(moving_average_decay)
@@ -220,8 +257,7 @@ class BYOL(nn.Module):
         self.to(device)
 
         # send a mock image tensor to instantiate singleton parameters
-        self.forward(torch.randn(2, 3, image_size, image_size, device=device),
-                     torch.randn(2, 3, image_size, image_size, device=device))
+        self.forward(torch.randn(2, 3, image_size, image_size, device=device))
 
     @singleton('target_encoder')
     def _get_target_encoder(self):
@@ -245,28 +281,32 @@ class BYOL(nn.Module):
         return {'target_ema_beta': self.target_ema_updater.beta}
 
     def visual_dbg(self, step, path):
-        if self.do_aug:
-            torchvision.utils.save_image(self.im1.cpu().float(), os.path.join(path, "%i_image1.png" % (step,)))
-            torchvision.utils.save_image(self.im2.cpu().float(), os.path.join(path, "%i_image2.png" % (step,)))
+        torchvision.utils.save_image(self.im1.cpu().float(), os.path.join(path, "%i_image1.png" % (step,)))
+        torchvision.utils.save_image(self.im2.cpu().float(), os.path.join(path, "%i_image2.png" % (step,)))
 
-    def forward(self, image_one, image_two):
-        if self.do_aug:
-            image_one = self.aug(image_one)
-            image_two = self.aug(image_two)
-            # Keep copies on hand for visual_dbg.
-            self.im1 = image_one.detach().copy()
-            self.im2 = image_two.detach().copy()
+    def forward(self, image):
+        _, _, h, w = image.shape
+        point = torch.randint(h//8, 7*h//8, (2,)).long().to(image.device)
 
-        online_proj_one = self.online_encoder(image_one)
-        online_proj_two = self.online_encoder(image_two)
+        image_one, pt_one = self.aug(image, point)
+        image_two, pt_two = self.aug(image, point)
+
+        # Keep copies on hand for visual_dbg.
+        self.im1 = image_one.detach().clone()
+        self.im1[:,:,pt_one[0]-3:pt_one[0]+3,pt_one[1]-3:pt_one[1]+3] = 1
+        self.im2 = image_two.detach().clone()
+        self.im2[:,:,pt_two[0]-3:pt_two[0]+3,pt_two[1]-3:pt_two[1]+3] = 1
+
+        online_proj_one = self.online_encoder(image_one, pt_one)
+        online_proj_two = self.online_encoder(image_two, pt_two)
 
         online_pred_one = self.online_predictor(online_proj_one)
         online_pred_two = self.online_predictor(online_proj_two)
 
         with torch.no_grad():
             target_encoder = self._get_target_encoder() if self.use_momentum else self.online_encoder
-            target_proj_one = target_encoder(image_one).detach()
-            target_proj_two = target_encoder(image_two).detach()
+            target_proj_one = target_encoder(image_one, pt_one).detach()
+            target_proj_two = target_encoder(image_two, pt_two).detach()
 
         loss_one = loss_fn(online_pred_one, target_proj_two.detach())
         loss_two = loss_fn(online_pred_two, target_proj_one.detach())
@@ -275,53 +315,20 @@ class BYOL(nn.Module):
         return loss.mean()
 
 
-class PointwiseAugmentor(nn.Module):
-    def __init__(self, img_size=224):
-        super().__init__()
-        self.jitter = RandomApply(augs.ColorJitter(0.8, 0.8, 0.8, 0.2), p=0.8)
-        self.gray = augs.RandomGrayscale(p=0.2)
-        self.blur = RandomApply(filters.GaussianBlur2d((3, 3), (1.5, 1.5)), p=0.1)
-        self.rrc = augs.RandomResizedCrop((img_size, img_size))
-
-    # Given a point in the *destination* image, returns the same point in the source image, given the kornia RRC params.
-    def reverse_rrc(self, dest_point, params):
-        dh, dw = params['dst'][:,2,1]-params['dst'][:,0,1], params['dst'][:,2,0] - params['dst'][:,0,0]
-        sh, sw = params['src'][:,2,1]-params['src'][:,0,1], params['src'][:,2,0] - params['src'][:,0,0]
-        scale_h, scale_w = sh.float() / dh.float(), sw.float() / dw.float()
-        t, l = dest_point
-        t = (t.float() * scale_h).int()
-        l = (l.float() * scale_w).int()
-        return t + params['src'][:,0,1], l + params['src'][:,0,0]
-
-    def reverse_horizontal_flip(self, pt, input):
-        t, l = pt
-        center = input.shape[-1] // 2
-        return t, 2 * center - l
-
-    def forward(self, x, points):
-        d = self.jitter(x)
-        d = self.gray(d)
-        will_flip = random.random() > .5
-        if will_flip:
-            d = apply_hflip(d)
-        d = self.blur(d)
-        params = self.rrc.generate_parameters(d.shape)
-        d = self.rrc(d, params=params)
-
-        rev = self.reverse_rrc(points, params)
-        if will_flip:
-            rev = self.reverse_horizontal_flip(rev, x)
-
 if __name__ == '__main__':
-    p = PointwiseAugmentor(256)
-    t = ToTensor()(Image.open('E:\\4k6k\\datasets\\ns_images\\imagesets\\000001_152761.jpg')).unsqueeze(0).repeat(8,1,1,1)
-    points = (torch.randint(0,224,(t.shape[0],)),torch.randint(0,224,(t.shape[0],)))
-    p(t, points)
+    pa = PointwiseAugmentor(256)
+    for j in range(100):
+        t = ToTensor()(Image.open('E:\\4k6k\\datasets\\ns_images\\imagesets\\000001_152761.jpg')).unsqueeze(0).repeat(8,1,1,1)
+        p = torch.randint(50,180,(2,))
+        augmented, dp = pa(t, p)
+        t, p = pa(t, p)
+        t[:,:,p[0]-3:p[0]+3,p[1]-3:p[1]+3] = 0
+        torchvision.utils.save_image(t, f"{j}_src.png")
+        augmented[:,:,dp[0]-3:dp[0]+3,dp[1]-3:dp[1]+3] = 0
+        torchvision.utils.save_image(augmented, f"{j}_dst.png")
 
 
 @register_model
-def register_byol(opt_net, opt):
+def register_pixel_local_byol(opt_net, opt):
     subnet = create_model(opt, opt_net['subnet'])
-    return BYOL(subnet, opt_net['image_size'], opt_net['hidden_layer'],
-                structural_mlp=opt_get(opt_net, ['use_structural_mlp'], False),
-                do_augmentation=opt_get(opt_net, ['gpu_augmentation'], False))
+    return BYOL(subnet, opt_net['image_size'], opt_net['hidden_layer'])

codes/models/segformer/segformer.py

@@ -2,11 +2,30 @@ import math
 
 import torch
 import torch.nn as nn
+import torchvision
 from tqdm import tqdm
 
 from models.segformer.backbone import backbone50
 
 
+# torch.gather() which operates as it always fucking should have: pulling indexes from the input.
+from trainer.networks import register_model
+
+
+def gather_2d(input, index):
+    b, c, h, w = input.shape
+    nodim = input.view(b, c, h * w)
+    ind_nd = index[:, 0]*w + index[:, 1]
+    ind_nd = ind_nd.unsqueeze(1)
+    ind_nd = ind_nd.repeat((1, c))
+    ind_nd = ind_nd.unsqueeze(2)
+    result = torch.gather(nodim, dim=2, index=ind_nd)
+    result = result.squeeze()
+    if b == 1:
+        result = result.unsqueeze(0)
+    return result
+
+
 class DilatorModule(nn.Module):
     def __init__(self, input_channels, output_channels, max_dilation):
         super().__init__()
@@ -15,7 +34,7 @@ class DilatorModule(nn.Module):
         if max_dilation > 1:
             self.bn = nn.BatchNorm2d(input_channels)
             self.relu = nn.ReLU()
-            self.conv2 = nn.Conv2d(input_channels, input_channels, kernel_size=3, padding=1, dilation=max_dilation, bias=True)
+            self.conv2 = nn.Conv2d(input_channels, input_channels, kernel_size=3, padding=max_dilation, dilation=max_dilation, bias=True)
         self.dense = nn.Linear(input_channels, output_channels, bias=True)
 
     def forward(self, inp, loc):
@@ -24,9 +43,8 @@ class DilatorModule(nn.Module):
             x = self.bn(self.relu(x))
             x = self.conv2(x)
 
-        # This can be made (possibly substantially) more efficient by only computing these convolutions across a subset of the image. Possibly.
-        i, j = loc
-        x = x[:,:,i,j]
+        # This can be made more efficient by only computing these convolutions across a subset of the image. Possibly.
+        x = gather_2d(x, loc).contiguous()
         return self.dense(x)
 
 
@@ -48,13 +66,22 @@ class PositionalEncoding(nn.Module):
         return x
 
 
-class Segformer(nn.Module):
+# Simple mean() layer encoded into a class so that BYOL can grab it.
+class Tail(nn.Module):
     def __init__(self):
         super().__init__()
+
+    def forward(self, x):
+        return x.mean(dim=0)
+
+
+class Segformer(nn.Module):
+    def __init__(self, latent_channels=1024, layers=8):
+        super().__init__()
         self.backbone = backbone50()
         backbone_channels = [256, 512, 1024, 2048]
         dilations = [[1,2,3,4],[1,2,3],[1,2],[1]]
-        final_latent_channels = 2048
+        final_latent_channels = latent_channels
         dilators = []
         for ic, dis in zip(backbone_channels, dilations):
             layer_dilators = []
@@ -64,26 +91,37 @@ class Segformer(nn.Module):
         self.dilators = nn.ModuleList(dilators)
 
         self.token_position_encoder = PositionalEncoding(final_latent_channels, max_len=10)
-        self.transformer_layers = nn.Sequential(*[nn.TransformerEncoderLayer(final_latent_channels, nhead=4) for _ in range(16)])
+        self.transformer_layers = nn.Sequential(*[nn.TransformerEncoderLayer(final_latent_channels, nhead=4) for _ in range(layers)])
+        self.tail = Tail()
 
     def forward(self, x, pos):
         layers = self.backbone(x)
         set = []
-        i, j = pos[0] // 4, pos[1] // 4
+
+        # A single position can be optionally given, in which case we need to expand it to represent the entire input.
+        if pos.shape == (2,):
+            pos = pos.unsqueeze(0).repeat(x.shape[0],1)
+
+        pos = pos // 4
         for layer_out, dilator in zip(layers, self.dilators):
             for subdilator in dilator:
-                set.append(subdilator(layer_out, (i, j)))
-            i, j = i // 2, j // 2
+                set.append(subdilator(layer_out, pos))
+            pos = pos // 2
 
         # The torch transformer expects the set dimension to be 0.
         set = torch.stack(set, dim=0)
         set = self.token_position_encoder(set)
         set = self.transformer_layers(set)
-        return set
+        return self.tail(set)
+
+
+@register_model
+def register_segformer(opt_net, opt):
+    return Segformer()
 
 
 if __name__ == '__main__':
     model = Segformer().to('cuda')
     for j in tqdm(range(1000)):
         test_tensor = torch.randn(64,3,224,224).cuda()
-        model(test_tensor, (43, 73))
+        print(model(test_tensor, torch.randint(0,224,(64,2)).cuda()).shape)

codes/train.py

@@ -295,7 +295,7 @@ class Trainer:
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_cats_stylegan2_rosinality.yml')
+    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_byol_segformer_xx.yml')
     parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher')
     parser.add_argument('--local_rank', type=int, default=0)
     args = parser.parse_args()

codes/trainer/eval/single_point_pair_contrastive_eval.py

@@ -11,6 +11,7 @@ import trainer.eval.evaluator as evaluator
 from pytorch_fid import fid_score
 
 from data.image_pair_with_corresponding_points_dataset import ImagePairWithCorrespondingPointsDataset
+from models.segformer.segformer import Segformer
 from utils.util import opt_get
 
 # Uses two datasets: a "similar" and "dissimilar" dataset, each of which contains pairs of images and similar/dissimilar
@@ -23,15 +24,17 @@ class SinglePointPairContrastiveEval(evaluator.Evaluator):
         self.batch_sz = opt_eval['batch_size']
         self.eval_qty = opt_eval['quantity']
         assert self.eval_qty % self.batch_sz == 0
-        self.similar_set = DataLoader(ImagePairWithCorrespondingPointsDataset(**opt_eval['similar_set_args']), shuffle=False, batch_size=self.batch_sz)
-        self.dissimilar_set = DataLoader(ImagePairWithCorrespondingPointsDataset(**opt_eval['dissimilar_set_args']), shuffle=False, batch_size=self.batch_sz)
+        self.similar_set = DataLoader(ImagePairWithCorrespondingPointsDataset(opt_eval['similar_set_args']), shuffle=False, batch_size=self.batch_sz)
+        self.dissimilar_set = DataLoader(ImagePairWithCorrespondingPointsDataset(opt_eval['dissimilar_set_args']), shuffle=False, batch_size=self.batch_sz)
+        # Hack to make this work with the BYOL generator. TODO: fix
+        self.model = self.model.online_encoder.net
 
-    def get_l2_score(self, dl):
+    def get_l2_score(self, dl, dev):
         distances = []
         l2 = MSELoss()
         for i, data in tqdm(enumerate(dl)):
-            latent1 = self.model(data['img1'], data['coords1'])
-            latent2 = self.model(data['img2'], data['coords2'])
+            latent1 = self.model(data['img1'].to(dev), torch.stack(data['coords1'], dim=1).to(dev))
+            latent2 = self.model(data['img2'].to(dev), torch.stack(data['coords2'], dim=1).to(dev))
             distances.append(l2(latent1, latent2))
             if i * self.batch_sz >= self.eval_qty:
                 break
@@ -40,9 +43,30 @@ class SinglePointPairContrastiveEval(evaluator.Evaluator):
 
     def perform_eval(self):
         self.model.eval()
-        print("Computing contrastive eval on similar set")
-        similars = self.get_l2_score(self.similar_set)
-        print("Computing contrastive eval on dissimilar set")
-        dissimilars = self.get_l2_score(self.dissimilar_set)
-        print(f"Eval done. val_similar_lq: {similars.item()}; val_dissimilar_l2: {dissimilars.item()}")
-        return {"val_similar_l2": similars.item(), "val_dissimilar_l2": dissimilars.item()}
+        with torch.no_grad():
+            dev = next(self.model.parameters()).device
+            print("Computing contrastive eval on similar set")
+            similars = self.get_l2_score(self.similar_set, dev)
+            print("Computing contrastive eval on dissimilar set")
+            dissimilars = self.get_l2_score(self.dissimilar_set, dev)
+            diff = dissimilars.item() - similars.item()
+            print(f"Eval done. val_similar_lq: {similars.item()}; val_dissimilar_l2: {dissimilars.item()}; val_diff: {diff}")
+        self.model.train()
+        return {"val_similar_l2": similars.item(), "val_dissimilar_l2": dissimilars.item(), "val_diff": diff.item()}
+
+
+if __name__ == '__main__':
+    model = Segformer(1024, 4).cuda()
+    eval = SinglePointPairContrastiveEval(model, {
+        'batch_size': 8,
+        'quantity': 32,
+        'similar_set_args': {
+            'path': 'E:\\4k6k\\datasets\\ns_images\\segformer_validation\\similar',
+            'size': 256
+        },
+        'dissimilar_set_args': {
+            'path': 'E:\\4k6k\\datasets\\ns_images\\segformer_validation\\dissimilar',
+            'size': 256
+        },
+    }, {})
+    eval.perform_eval()