Improve creation of data pipeline with prefetch enabled vs disabled, fixup inception_res_v2 and dpn models

data/__init__.py

@@ -1,4 +1,4 @@
 from data.dataset import Dataset
-from data.transforms import transforms_imagenet_eval, transforms_imagenet_train
+from data.transforms import transforms_imagenet_eval, transforms_imagenet_train, get_model_meanstd
-from data.utils import fast_collate, PrefetchLoader
+from data.utils import create_loader
 from data.random_erasing import RandomErasingTorch, RandomErasingNumpy

data/dataset.py

@@ -54,7 +54,7 @@ class Dataset(data.Dataset):
     def __init__(
             self,
             root,
-            transform):
+            transform=None):
 
         imgs, _, _ = find_images_and_targets(root)
         if len(imgs) == 0:
@@ -67,7 +67,8 @@ class Dataset(data.Dataset):
     def __getitem__(self, index):
         path, target = self.imgs[index]
         img = Image.open(path).convert('RGB')
-        img = self.transform(img)
+        if self.transform is not None:
+            img = self.transform(img)
         if target is None:
             target = torch.zeros(1).long()
         return img, target

data/transforms.py

@@ -15,7 +15,38 @@ IMAGENET_DEFAULT_MEAN = [0.485, 0.456, 0.406]
 IMAGENET_DEFAULT_STD = [0.229, 0.224, 0.225]
 
 
-class AsNumpy:
+# FIXME replace these mean/std fn with model factory based values from config dict
+def get_model_meanstd(model_name):
+    model_name = model_name.lower()
+    if 'dpn' in model_name:
+        return IMAGENET_DPN_MEAN, IMAGENET_DPN_STD
+    elif 'ception' in model_name:
+        return IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD
+    else:
+        return IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+
+
+def get_model_mean(model_name):
+    model_name = model_name.lower()
+    if 'dpn' in model_name:
+        return IMAGENET_DPN_STD
+    elif 'ception' in model_name:
+        return IMAGENET_INCEPTION_MEAN
+    else:
+        return IMAGENET_DEFAULT_MEAN
+
+
+def get_model_std(model_name):
+    model_name = model_name.lower()
+    if 'dpn' in model_name:
+        return IMAGENET_DEFAULT_STD
+    elif 'ception' in model_name:
+        return IMAGENET_INCEPTION_STD
+    else:
+        return IMAGENET_DEFAULT_STD
+
+
+class ToNumpy:
 
     def __call__(self, pil_img):
         np_img = np.array(pil_img, dtype=np.uint8)
@@ -25,29 +56,79 @@ class AsNumpy:
         return np_img
 
 
+class ToTensor:
+
+    def __init__(self, dtype=torch.float32):
+        self.dtype = dtype
+
+    def __call__(self, pil_img):
+        np_img = np.array(pil_img, dtype=np.uint8)
+        if np_img.ndim < 3:
+            np_img = np.expand_dims(np_img, axis=-1)
+        np_img = np.rollaxis(np_img, 2)  # HWC to CHW
+        return torch.from_numpy(np_img).to(dtype=self.dtype)
+
+
 def transforms_imagenet_train(
         img_size=224,
         scale=(0.1, 1.0),
         color_jitter=(0.4, 0.4, 0.4),
-        random_erasing=0.4):
+        random_erasing=0.4,
+        use_prefetcher=False,
+        mean=IMAGENET_DEFAULT_MEAN,
+        std=IMAGENET_DEFAULT_STD
+):
 
     tfl = [
-        transforms.RandomResizedCrop(img_size, scale=scale),
+        transforms.RandomResizedCrop(
+            img_size, scale=scale, interpolation=Image.BICUBIC),
         transforms.RandomHorizontalFlip(),
         transforms.ColorJitter(*color_jitter),
-        AsNumpy(),
     ]
-    #if random_erasing > 0.:
+
-    #    tfl.append(RandomErasingNumpy(random_erasing, per_pixel=True))
+    if use_prefetcher:
+        # prefetcher and collate will handle tensor conversion and norm
+        tfl += [ToNumpy()]
+    else:
+        tfl += [
+            ToTensor(),
+            transforms.Normalize(
+                mean=torch.tensor(mean) * 255,
+                std=torch.tensor(std) * 255)
+        ]
+        if random_erasing > 0.:
+            tfl.append(RandomErasingNumpy(random_erasing, per_pixel=True))
     return transforms.Compose(tfl)
 
 
-def transforms_imagenet_eval(img_size=224, crop_pct=None):
+def transforms_imagenet_eval(
+        img_size=224,
+        crop_pct=None,
+        use_prefetcher=False,
+        mean=IMAGENET_DEFAULT_MEAN,
+        std=IMAGENET_DEFAULT_STD):
     crop_pct = crop_pct or DEFAULT_CROP_PCT
     scale_size = int(math.floor(img_size / crop_pct))
 
-    return transforms.Compose([
+    tfl = [
         transforms.Resize(scale_size, Image.BICUBIC),
         transforms.CenterCrop(img_size),
-        AsNumpy(),
+    ]
-    ])
+    if use_prefetcher:
+        # prefetcher and collate will handle tensor conversion and norm
+        tfl += [ToNumpy()]
+    else:
+        tfl += [
+            transforms.ToTensor(),
+            transforms.Normalize(
+                     mean=torch.tensor(mean),
+                     std=torch.tensor(std))
+        ]
+        # tfl += [
+        #     ToTensor(),
+        #     transforms.Normalize(
+        #         mean=torch.tensor(mean) * 255,
+        #         std=torch.tensor(std) * 255)
+        # ]
+
+    return transforms.Compose(tfl)

data/utils.py

@@ -1,5 +1,7 @@
 import torch
+import torch.utils.data as tdata
 from data.random_erasing import RandomErasingTorch
+from data.transforms import *
 
 
 def fast_collate(batch):
@@ -17,16 +19,17 @@ class PrefetchLoader:
     def __init__(self,
             loader,
             fp16=False,
-            random_erasing=True,
+            random_erasing=0.,
-            mean=[0.485, 0.456, 0.406],
+            mean=IMAGENET_DEFAULT_MEAN,
-            std=[0.229, 0.224, 0.225]):
+            std=IMAGENET_DEFAULT_STD):
         self.loader = loader
         self.fp16 = fp16
         self.random_erasing = random_erasing
         self.mean = torch.tensor([x * 255 for x in mean]).cuda().view(1, 3, 1, 1)
         self.std = torch.tensor([x * 255 for x in std]).cuda().view(1, 3, 1, 1)
         if random_erasing:
-            self.random_erasing = RandomErasingTorch(per_pixel=True)
+            self.random_erasing = RandomErasingTorch(
+                probability=random_erasing, per_pixel=True)
         else:
             self.random_erasing = None
 
@@ -63,3 +66,47 @@ class PrefetchLoader:
 
     def __len__(self):
         return len(self.loader)
+
+
+def create_loader(
+        dataset,
+        img_size,
+        batch_size,
+        is_training=False,
+        use_prefetcher=True,
+        random_erasing=0.,
+        mean=IMAGENET_DEFAULT_MEAN,
+        std=IMAGENET_DEFAULT_STD,
+        num_workers=1,
+):
+
+    if is_training:
+        transform = transforms_imagenet_train(
+            img_size,
+            use_prefetcher=use_prefetcher,
+            mean=mean,
+            std=std)
+    else:
+        transform = transforms_imagenet_eval(
+            img_size,
+            use_prefetcher=use_prefetcher,
+            mean=mean,
+            std=std)
+
+    dataset.transform = transform
+
+    loader = tdata.DataLoader(
+        dataset,
+        batch_size=batch_size,
+        shuffle=is_training,
+        num_workers=num_workers,
+        collate_fn=fast_collate if use_prefetcher else tdata.dataloader.default_collate,
+    )
+    if use_prefetcher:
+        loader = PrefetchLoader(
+            loader,
+            random_erasing=random_erasing if is_training else 0.,
+            mean=mean,
+            std=std)
+
+    return loader

models/dpn.py

@@ -21,15 +21,23 @@ from .adaptive_avgmax_pool import adaptive_avgmax_pool2d
 __all__ = ['DPN', 'dpn68', 'dpn92', 'dpn98', 'dpn131', 'dpn107']
 
 
-# If anyone able to provide direct link hosting, more than happy to fill these out.. -rwightman
 model_urls = {
-    'dpn68': '',
+    'dpn68':
-    'dpn68b_extra': 'dpn68_extra-87733ef7.pth',
+        'http://data.lip6.fr/cadene/pretrainedmodels/dpn68-66bebafa7.pth',
+    'dpn68b_extra':
+        'http://data.lip6.fr/cadene/pretrainedmodels/'
+        'dpn68b_extra-84854c156.pth',
     'dpn92': '',
-    'dpn92_extra': '',
+    'dpn92_extra':
-    'dpn98': '',
+        'http://data.lip6.fr/cadene/pretrainedmodels/'
-    'dpn131': 'dpn131-89380fa2.pth',
+        'dpn92_extra-b040e4a9b.pth',
-    'dpn107_extra': 'dpn107_extra-fc014e8ec.pth'
+    'dpn98':
+        'http://data.lip6.fr/cadene/pretrainedmodels/dpn98-5b90dec4d.pth',
+    'dpn131':
+        'http://data.lip6.fr/cadene/pretrainedmodels/dpn131-71dfe43e0.pth',
+    'dpn107_extra':
+        'http://data.lip6.fr/cadene/pretrainedmodels/'
+        'dpn107_extra-1ac7121e2.pth'
 }
 
 

models/inception_resnet_v2.py

@@ -10,7 +10,7 @@ import numpy as np
 from .adaptive_avgmax_pool import *
 
 model_urls = {
-    'imagenet': 'http://webia.lip6.fr/~cadene/Downloads/inceptionresnetv2-d579a627.pth'
+    'imagenet': 'http://data.lip6.fr/cadene/pretrainedmodels/inceptionresnetv2-520b38e4.pth'
 }
 
 
@@ -267,7 +267,7 @@ class InceptionResnetV2(nn.Module):
         self.block8 = Block8(noReLU=True)
         self.conv2d_7b = BasicConv2d(2080, 1536, kernel_size=1, stride=1)
         self.num_features = 1536
-        self.classif = nn.Linear(1536, num_classes)
+        self.last_linear = nn.Linear(1536, num_classes)
 
     def get_classifier(self):
         return self.classif
@@ -277,9 +277,16 @@ class InceptionResnetV2(nn.Module):
         self.num_classes = num_classes
         del self.classif
         if num_classes:
-            self.classif = torch.nn.Linear(1536, num_classes)
+            self.last_linear = torch.nn.Linear(1536, num_classes)
         else:
-            self.classif = None
+            self.last_linear = None
+
+    def trim_classifier(self, trim=1):
+        self.num_classes -= trim
+        new_last_linear = nn.Linear(1536, self.num_classes)
+        new_last_linear.weight.data = self.last_linear.weight.data[trim:]
+        new_last_linear.bias.data = self.last_linear.bias.data[trim:]
+        self.last_linear = new_last_linear
 
     def forward_features(self, x, pool=True):
         x = self.conv2d_1a(x)
@@ -298,7 +305,8 @@ class InceptionResnetV2(nn.Module):
         x = self.block8(x)
         x = self.conv2d_7b(x)
         if pool:
-            x = adaptive_avgmax_pool2d(x, self.global_pool, count_include_pad=False)
+            x = adaptive_avgmax_pool2d(x, self.global_pool)
+            #x = F.avg_pool2d(x, 8, count_include_pad=False)
             x = x.view(x.size(0), -1)
         return x
 
@@ -306,20 +314,23 @@ class InceptionResnetV2(nn.Module):
         x = self.forward_features(x, pool=True)
         if self.drop_rate > 0:
             x = F.dropout(x, p=self.drop_rate, training=self.training)
-        x = self.classif(x)
+        x = self.last_linear(x)
         return x
 
 
-def inception_resnet_v2(pretrained=False, num_classes=1001, **kwargs):
+def inception_resnet_v2(pretrained=False, num_classes=1000, **kwargs):
     r"""InceptionResnetV2 model architecture from the
     `"InceptionV4, Inception-ResNet..." <https://arxiv.org/abs/1602.07261>`_ paper.
 
     Args:
         pretrained ('string'): If True, returns a model pre-trained on ImageNet
     """
-    model = InceptionResnetV2(num_classes=num_classes, **kwargs)
+    extra_class = 1 if pretrained else 0
+    model = InceptionResnetV2(num_classes=num_classes + extra_class, **kwargs)
     if pretrained:
         print('Loading pretrained from %s' % model_urls['imagenet'])
         model.load_state_dict(model_zoo.load_url(model_urls['imagenet']))
+        model.trim_classifier()
+
     return model
 

models/model_factory.py

@@ -44,7 +44,7 @@ model_config_dict = {
     'dpn68b_extra': {
         'model_name': 'dpn68b', 'num_classes': 1000, 'input_size': 299, 'normalizer': 'dpn'},
     'inception_resnet_v2': {
-        'model_name': 'inception_resnet_v2', 'num_classes': 1001, 'input_size': 299, 'normalizer': 'le'},
+        'model_name': 'inception_resnet_v2', 'num_classes': 1000, 'input_size': 299, 'normalizer': 'le'},
 }
 
 

train.py

@@ -93,34 +93,33 @@ def main():
     batch_size = args.batch_size
     torch.manual_seed(args.seed)
 
-    dataset_train = Dataset(
+    data_mean, data_std = get_model_meanstd(args.model)
-        os.path.join(args.data, 'train'),
-        transform=transforms_imagenet_train())
 
-    loader_train = data.DataLoader(
+    dataset_train = Dataset(os.path.join(args.data, 'train'))
+
+    loader_train = create_loader(
         dataset_train,
+        img_size=args.img_size,
         batch_size=batch_size,
-        shuffle=True,
+        is_training=True,
+        use_prefetcher=True,
+        random_erasing=0.5,
+        mean=data_mean,
+        std=data_std,
         num_workers=args.workers,
-        collate_fn=fast_collate
     )
-    loader_train = PrefetchLoader(
-        loader_train, random_erasing=True,
-        )
 
-    dataset_eval = Dataset(
+    dataset_eval = Dataset(os.path.join(args.data, 'validation'))
-        os.path.join(args.data, 'validation'),
-        transform=transforms_imagenet_eval())
 
-    loader_eval = data.DataLoader(
+    loader_eval = create_loader(
         dataset_eval,
+        img_size=args.img_size,
         batch_size=4 * args.batch_size,
-        shuffle=False,
+        is_training=False,
+        use_prefetcher=True,
+        mean=data_mean,
+        std=data_std,
         num_workers=args.workers,
-        collate_fn=fast_collate,
-    )
-    loader_eval = PrefetchLoader(
-        loader_eval, random_erasing=False,
     )
 
     model = model_factory.create_model(

validate.py

@@ -9,11 +9,9 @@ import torch
 import torch.backends.cudnn as cudnn
 import torch.nn as nn
 import torch.nn.parallel
-import torch.utils.data as data
 
-
+from models import create_model
-from models import create_model, transforms_imagenet_eval
+from data import Dataset, create_loader, get_model_meanstd
-from dataset import Dataset
 
 
 parser = argparse.ArgumentParser(description='PyTorch ImageNet Validation')
@@ -80,14 +78,15 @@ def main():
 
     cudnn.benchmark = True
 
-    dataset = Dataset(
+    data_mean, data_std = get_model_meanstd(args.model)
-        args.data,
+    loader = create_loader(
-        transforms_imagenet_eval(args.model, args.img_size))
+        Dataset(args.data),
-
+        img_size=args.img_size,
-    loader = data.DataLoader(
+        batch_size=args.batch_size,
-        dataset,
+        use_prefetcher=True,
-        batch_size=args.batch_size, shuffle=False,
+        mean=data_mean,
-        num_workers=args.workers, pin_memory=True)
+        std=data_std,
+        num_workers=args.workers)
 
     batch_time = AverageMeter()
     losses = AverageMeter()