Working on CutMix impl as per #8, integrating with Mixup, currently experimenting...

timm/data/__init__.py

@@ -4,6 +4,6 @@ from .dataset import Dataset, DatasetTar, AugMixDataset
 from .transforms import *
 from .loader import create_loader
 from .transforms_factory import create_transform
-from .mixup import mixup_batch, FastCollateMixup
+from .mixup import mix_batch, FastCollateMixup, FastCollateMixupBatchwise, FastCollateMixupElementwise
 from .auto_augment import RandAugment, AutoAugment, rand_augment_ops, auto_augment_policy,\
     rand_augment_transform, auto_augment_transform

timm/data/dataset.py

@@ -89,7 +89,7 @@ class Dataset(data.Dataset):
         return img, target
 
     def __len__(self):
-        return len(self.imgs)
+        return len(self.samples)
 
     def filenames(self, indices=[], basename=False):
         if indices:

timm/data/mixup.py

@@ -1,5 +1,30 @@
+""" Mixup and Cutmix
+
+Papers:
+mixup: Beyond Empirical Risk Minimization (https://arxiv.org/abs/1710.09412)
+
+CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features (https://arxiv.org/abs/1905.04899)
+
+Code Reference:
+CutMix: https://github.com/clovaai/CutMix-PyTorch
+
+Hacked together by Ross Wightman
+"""
+
 import numpy as np
 import torch
+import math
+from enum import IntEnum
+
+
+class MixupMode(IntEnum):
+    MIXUP = 0
+    CUTMIX = 1
+    RANDOM = 2
+
+    @classmethod
+    def from_str(cls, value):
+        return cls[value.upper()]
 
 
 def one_hot(x, num_classes, on_value=1., off_value=0., device='cuda'):
@@ -12,7 +37,7 @@ def mixup_target(target, num_classes, lam=1., smoothing=0.0, device='cuda'):
     on_value = 1. - smoothing + off_value
     y1 = one_hot(target, num_classes, on_value=on_value, off_value=off_value, device=device)
     y2 = one_hot(target.flip(0), num_classes, on_value=on_value, off_value=off_value, device=device)
-    return lam*y1 + (1. - lam)*y2
+    return y1 * lam + y2 * (1. - lam)
 
 
 def mixup_batch(input, target, alpha=0.2, num_classes=1000, smoothing=0.1, disable=False):
@@ -24,28 +49,167 @@ def mixup_batch(input, target, alpha=0.2, num_classes=1000, smoothing=0.1, disab
     return input, target
 
 
+def rand_bbox(size, ratio):
+    H, W = size[-2:]
+    ratio = max(min(ratio, 0.8), 0.2)
+    cut_h, cut_w = int(H * ratio), int(W * ratio)
+    cy, cx = np.random.randint(H), np.random.randint(W)
+    yl, yh = np.clip(cy - cut_h // 2, 0, H), np.clip(cy + cut_h // 2, 0, H)
+    xl, xh = np.clip(cx - cut_w // 2, 0, W), np.clip(cx + cut_w // 2, 0, W)
+    return yl, yh, xl, xh
+
+
+def cutmix_batch(input, target, alpha=0.2, num_classes=1000, smoothing=0.1, disable=False):
+    lam = 1.
+    if not disable:
+        lam = np.random.beta(alpha, alpha)
+    if lam != 1:
+        ratio = math.sqrt(1. - lam)
+        yl, yh, xl, xh = rand_bbox(input.size(), ratio)
+        input[:, :, yl:yh, xl:xh] = input.flip(0)[:, :, yl:yh, xl:xh]
+    target = mixup_target(target, num_classes, lam, smoothing)
+    return input, target
+
+
+def _resolve_mode(mode):
+    mode = MixupMode.from_str(mode) if isinstance(mode, str) else mode
+    if mode == MixupMode.RANDOM:
+        mode = MixupMode(np.random.rand() > 0.5)
+    return mode # will be one of cutmix or mixup
+
+
+def mix_batch(
+        input, target, alpha=0.2, num_classes=1000, smoothing=0.1, disable=False, mode=MixupMode.MIXUP):
+    mode = _resolve_mode(mode)
+    if mode == MixupMode.CUTMIX:
+        return mixup_batch(input, target, alpha, num_classes, smoothing, disable)
+    else:
+        return cutmix_batch(input, target, alpha, num_classes, smoothing, disable)
+
+
 class FastCollateMixup:
+    """Fast Collate Mixup that applies different params to each element + flipped pair
 
-    def __init__(self, mixup_alpha=1., label_smoothing=0.1, num_classes=1000):
+    NOTE once experiments are done, one of the three variants will remain with this class name
+    """
+    def __init__(self, mixup_alpha=1., label_smoothing=0.1, num_classes=1000, mode=MixupMode.MIXUP):
         self.mixup_alpha = mixup_alpha
         self.label_smoothing = label_smoothing
         self.num_classes = num_classes
+        self.mode = MixupMode.from_str(mode) if isinstance(mode, str) else mode
         self.mixup_enabled = True
+        self.correct_lam = False  # correct lambda based on clipped area for cutmix
 
-    def __call__(self, batch):
+    def _do_mix(self, tensor, batch):
         batch_size = len(batch)
+        lam_out = torch.ones(batch_size)
+        for i in range(batch_size//2):
+            j = batch_size - i - 1
             lam = 1.
             if self.mixup_enabled:
                 lam = np.random.beta(self.mixup_alpha, self.mixup_alpha)
 
-        target = torch.tensor([b[1] for b in batch], dtype=torch.int64)
-        target = mixup_target(target, self.num_classes, lam, self.label_smoothing, device='cpu')
+            if _resolve_mode(self.mode) == MixupMode.CUTMIX:
+                mixed_i, mixed_j = batch[i][0].astype(np.float32), batch[j][0].astype(np.float32)
+                ratio = math.sqrt(1. - lam)
+                if lam != 1:
+                    yl, yh, xl, xh = rand_bbox(tensor.size(), ratio)
+                    mixed_i[:, yl:yh, xl:xh] = batch[j][0][:, yl:yh, xl:xh].astype(np.float32)
+                    mixed_j[:, yl:yh, xl:xh] = batch[i][0][:, yl:yh, xl:xh].astype(np.float32)
+                    if self.correct_lam:
+                        lam_corrected = (yh - yl) * (xh - xl) / (tensor.shape[-2] * tensor.shape[-1])
+                        lam_out[i] -= lam_corrected
+                        lam_out[j] -= lam_corrected
+                    else:
+                        lam_out[i] = lam
+                        lam_out[j] = lam
+            else:
+                mixed_i = batch[i][0].astype(np.float32) * lam + batch[j][0].astype(np.float32) * (1 - lam)
+                mixed_j = batch[j][0].astype(np.float32) * lam + batch[i][0].astype(np.float32) * (1 - lam)
+                lam_out[i] = lam
+                lam_out[j] = lam
+            np.round(mixed_i, out=mixed_i)
+            np.round(mixed_j, out=mixed_j)
+            tensor[i] += torch.from_numpy(mixed_i.astype(np.uint8))
+            tensor[j] += torch.from_numpy(mixed_j.astype(np.uint8))
+        return lam_out
 
+    def __call__(self, batch):
+        batch_size = len(batch)
+        assert batch_size % 2 == 0, 'Batch size should be even when using this'
         tensor = torch.zeros((batch_size, *batch[0][0].shape), dtype=torch.uint8)
+        lam = self._do_mix(tensor, batch)
+        target = torch.tensor([b[1] for b in batch], dtype=torch.int64)
+        target = mixup_target(target, self.num_classes, lam.unsqueeze(1), self.label_smoothing, device='cpu')
+
+        return tensor, target
+
+
+class FastCollateMixupElementwise(FastCollateMixup):
+    """Fast Collate Mixup that applies different params to each batch element
+
+    NOTE this is for experimentation, may remove at some point
+    """
+    def __init__(self, mixup_alpha=1., label_smoothing=0.1, num_classes=1000, mode=MixupMode.MIXUP):
+        super(FastCollateMixupElementwise, self).__init__(mixup_alpha, label_smoothing, num_classes, mode)
+
+    def _do_mix(self, tensor, batch):
+        batch_size = len(batch)
+        lam_out = torch.ones(batch_size)
         for i in range(batch_size):
+            lam = 1.
+            if self.mixup_enabled:
+                lam = np.random.beta(self.mixup_alpha, self.mixup_alpha)
+
+            if _resolve_mode(self.mode) == MixupMode.CUTMIX:
+                mixed = batch[i][0].astype(np.float32)
+                ratio = math.sqrt(1. - lam)
+                if lam != 1:
+                    yl, yh, xl, xh = rand_bbox(tensor.size(), ratio)
+                    mixed[:, yl:yh, xl:xh] = batch[batch_size - i - 1][0][:, yl:yh, xl:xh].astype(np.float32)
+                    if self.correct_lam:
+                        lam_out[i] -= (yh - yl) * (xh - xl) / (tensor.shape[-2] * tensor.shape[-1])
+                    else:
+                        lam_out[i] = lam
+            else:
                 mixed = batch[i][0].astype(np.float32) * lam + \
                         batch[batch_size - i - 1][0].astype(np.float32) * (1 - lam)
+                lam_out[i] = lam
             np.round(mixed, out=mixed)
             tensor[i] += torch.from_numpy(mixed.astype(np.uint8))
+        return lam_out
 
-        return tensor, target
+
+class FastCollateMixupBatchwise(FastCollateMixup):
+    """Fast Collate Mixup that applies same params to whole batch
+
+    NOTE this is for experimentation, may remove at some point
+    """
+
+    def __init__(self, mixup_alpha=1., label_smoothing=0.1, num_classes=1000, mode=MixupMode.MIXUP):
+        super(FastCollateMixupBatchwise, self).__init__(mixup_alpha, label_smoothing, num_classes, mode)
+
+    def _do_mix(self, tensor, batch):
+        batch_size = len(batch)
+        lam_out = torch.ones(batch_size)
+        lam = 1.
+        cutmix = _resolve_mode(self.mode) == MixupMode.CUTMIX
+        if self.mixup_enabled:
+            lam = np.random.beta(self.mixup_alpha, self.mixup_alpha)
+            if cutmix and self.correct_lam:
+                ratio = math.sqrt(1. - lam)
+                yl, yh, xl, xh = rand_bbox(batch[0][0].shape, ratio)
+                lam = 1 - (yh - yl) * (xh - xl) / (tensor.shape[-2] * tensor.shape[-1])
+
+        for i in range(batch_size):
+            if cutmix:
+                mixed = batch[i][0].astype(np.float32)
+                if lam != 1:
+                    mixed[:, yl:yh, xl:xh] = batch[batch_size - i - 1][0][:, yl:yh, xl:xh].astype(np.float32)
+                    lam_out[i] -= (yh - yl) * (xh - xl) / (tensor.shape[-2] * tensor.shape[-1])
+            else:
+                mixed = batch[i][0].astype(np.float32) * lam + \
+                        batch[batch_size - i - 1][0].astype(np.float32) * (1 - lam)
+            np.round(mixed, out=mixed)
+            tensor[i] += torch.from_numpy(mixed.astype(np.uint8))
+        return lam

train.py

@@ -28,7 +28,8 @@ except ImportError:
     from torch.nn.parallel import DistributedDataParallel as DDP
     has_apex = False
 
-from timm.data import Dataset, create_loader, resolve_data_config, FastCollateMixup, mixup_batch, AugMixDataset
+from timm.data import Dataset, create_loader, resolve_data_config, FastCollateMixup, mix_batch, AugMixDataset,\
+    FastCollateMixupElementwise, FastCollateMixupBatchwise
 from timm.models import create_model, resume_checkpoint, convert_splitbn_model
 from timm.utils import *
 from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy, JsdCrossEntropy
@@ -134,6 +135,8 @@ parser.add_argument('--resplit', action='store_true', default=False,
                     help='Do not random erase first (clean) augmentation split')
 parser.add_argument('--mixup', type=float, default=0.0,
                     help='mixup alpha, mixup enabled if > 0. (default: 0.)')
+parser.add_argument('--mixup-mode', type=str, default='mixup',
+                    help='Mixup mode ("mixup", "cutmix", "random", default: "mixup")')
 parser.add_argument('--mixup-off-epoch', default=0, type=int, metavar='N',
                     help='turn off mixup after this epoch, disabled if 0 (default: 0)')
 parser.add_argument('--smoothing', type=float, default=0.1,
@@ -352,7 +355,7 @@ def main():
     collate_fn = None
     if args.prefetcher and args.mixup > 0:
         assert not num_aug_splits  # collate conflict (need to support deinterleaving in collate mixup)
-        collate_fn = FastCollateMixup(args.mixup, args.smoothing, args.num_classes)
+        collate_fn = FastCollateMixup(args.mixup, args.smoothing, args.num_classes, args.mixup_mode)
 
     if num_aug_splits > 1:
         dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
@@ -504,10 +507,10 @@ def train_epoch(
         if not args.prefetcher:
             input, target = input.cuda(), target.cuda()
             if args.mixup > 0.:
-                input, target = mixup_batch(
+                input, target = mix_batch(
                     input, target,
                     alpha=args.mixup, num_classes=args.num_classes, smoothing=args.smoothing,
-                    disable=args.mixup_off_epoch and epoch >= args.mixup_off_epoch)
+                    disable=args.mixup_off_epoch and epoch >= args.mixup_off_epoch, mode=args.mixup_mode)
 
         output = model(input)