Do Better ImageNet Models Transfer Better?

Do Better ImageNet Models Transfer Better?

Simon Kornblith, Jonathon Shlens, and Quoc V. Le
Google Brain
Work done as a member of the Google AI Residency program (

Transfer learning has become a cornerstone of computer vision with the advent of ImageNet features, yet little work has been done to evaluate the performance of ImageNet architectures across different datasets. An implicit hypothesis in modern computer vision research is that models that perform better on ImageNet necessarily perform better on other vision tasks. However, this hypothesis has never been systematically tested. Here, we compare the performance of 13 classification models on 12 image classification tasks in three settings: as fixed feature extractors, fine-tuned, and trained from random initialization. We find that, when networks are used as fixed feature extractors, ImageNet accuracy is only weakly predictive of accuracy on other tasks (). In this setting, ResNets consistently outperform networks that achieve higher accuracy on ImageNet. When networks are fine-tuned, we observe a substantially stronger correlation (). We achieve state-of-the-art performance on eight image classification tasks simply by fine-tuning state-of-the-art ImageNet architectures, outperforming previous results based on specialized methods for transfer learning. Finally, we observe that, on three small fine-grained image classification datasets, networks trained from random initialization perform similarly to ImageNet-pretrained networks. Together, our results show that ImageNet architectures generalize well across datasets, with small improvements in ImageNet accuracy producing improvements across other tasks, but ImageNet features are less general than previously suggested.



1 Introduction

The last decade of computer vision research has pursued academic benchmarks as a measure of progress. No benchmark has been as hotly pursued as ImageNet deng2009imagenet. Network architectures measured against this dataset have fueled much progress in computer vision research across a broad array of problems, including transferring to new datasets donahue2014decaf; razavian2014cnn, object detection huang2016, image segmentation he2017mask; chen2018deeplab and perceptual metrics of images johnson2016perceptual. An implicit assumption behind this progress is that network architectures that perform better on ImageNet necessarily perform better on other vision tasks. Another assumption is that better network architectures learn better features that can be transferred across vision-based tasks. Although previous studies have provided some evidence for these hypotheses (e.g. Chatfield14; simonyan2014very; huang2016; howard2017mobilenets; he2017mask), they have never been systematically explored.

In the present work, we seek to test these hypotheses by investigating the transferability of both ImageNet features and ImageNet classification architectures. Specifically, we conduct a large-scale study of transfer learning across 13 top-performing convolutional neural networks for image classification on 12 image classification datasets in 3 different experimental settings, visualized in Figure 1: as fixed feature extractors donahue2014decaf; razavian2014cnn, fine-tuned from ImageNet initialization agrawal2014; girshick2014rich; Chatfield14, and trained from random initialization. Our main contributions are as follows:

Figure 1: Fine-grained datasets (FGVC Aircraft) benefit significantly from fine-tuning or training from scratch. Datasets similar to ImageNet (CIFAR-10) benefit less. Low dimensional embedding using t-SNE maaten2008visualizing on features from the penultimate layer of Inception v4, for 10 classes from the test set.
  • The best ImageNet models do not provide the best fixed image features. Features from ResNet models he2016deep; he2016identity trained on ImageNet consistently outperform features from networks that achieve higher accuracy on ImageNet.

  • When networks are fine-tuned, ImageNet accuracy is a much stronger indicator of transfer task accuracy (), with a state-of-the-art ImageNet architecture yielding state-of-the-art results across many tasks.

  • Architectures transfer well across tasks even when weights do not. On 3 small fine-grained classification datasets, fine-tuning does not provide a substantial benefit over training from random initialization, but better ImageNet architectures nonetheless obtain higher accuracy.

2 Related work

ImageNet follows in a succession of progressively larger and more realistic benchmark datasets for computer vision. Each successive dataset was designed to address perceived issues with the size and content of previous datasets. \textcitetorralba2011unbiased showed that many early datasets were heavily biased, with classifiers trained to recognize or classify objects on those datasets possessing almost no ability to generalize to images from other datasets.

Early works using convolutional neural networks (CNNs) for transfer learning extracted fixed features from ImageNet-trained networks and used these features to train SVMs and logistic regression classifiers for new tasks donahue2014decaf; razavian2014cnn; Chatfield14. These works found that these features could outperform hand-engineered features even for tasks very distinct from ImageNet classification donahue2014decaf; razavian2014cnn. Several such studies have compared the performance of AlexNet-like CNNs of varying levels of computational complexity in a transfer learning setting with no fine-tuning. \textciteChatfield14 found that, out of three networks, the two more computationally expensive networks performed better on PASCAL VOC. Similar work concluded that deeper networks produce higher accuracy across many transfer tasks, but wider networks produce lower accuracy razavian2016.

A number of studies have compared the accuracy of classifiers trained on fixed image features versus fine-tuning the image representations on a new dataset agrawal2014; Chatfield14; girshick2014rich. Fine-tuning typically achieves higher accuracy, especially for larger datasets or datasets with a larger domain mismatch from the training set yosinski2014transferable; razavian2016; lin2015bilinear; HuhAE16; chu16. In object detection, ImageNet-pretrained networks are used as backbone models for Faster R-CNN and R-FCN detection systems. Classifiers with higher ImageNet accuracy achieve higher overall object detection accuracy huang2016, although variability across network architectures is small compared to variability from other object detection architecture choices. A parallel story likewise appears in image segmentation models chen2018deeplab, although it has not been as systematically explored.

Several authors have investigated how properties of the original training dataset affect transfer accuracy. Examining performance of fixed image features drawn from networks trained on subsets of ImageNet, \textciterazavian2016 reported that the number of classes is more important, while \textciteHuhAE16 reported that the number of images per class is more important, provided that the classes are sampled at random rather than split according to the WordNet hierarchy. \textciteyosinski2014transferable showed that the first layer of AlexNet can be frozen when transferring between natural and manmade subsets of ImageNet without performance impairment, but freezing later layers produces a substantial drop in accuracy.

3 Statistical methods

Much of the analysis in this work requires comparing accuracies across datasets of differing difficulty. When fitting linear models to accuracy values across multiple datasets, we consider effects of model and dataset to be additive. In this context, using untransformed accuracy as a dependent variable is problematic: The meaning of a 1% additive increase in accuracy is different if it is relative to a base accuracy of 50% vs. 99%. Thus, we consider the accuracy after the logit transformation . The logit transformation is the most commonly used transformation for analysis of proportion data, and an additive change in logit-transformed accuracy has a simple interpretation as a multiplicative change in the odds of correct classification:

After the logit transformation, results do not depend upon whether performance is measured in terms of accuracy or error rate, since . We plot results on logit-scaled axes.

We tested for significant differences between pairs of networks on the same dataset using a permutation test or equivalent binomial test of the null hypothesis that the predictions of the two networks are equally likely to be correct; see Appendix A.1.1 for further details. We tested for significant differences between networks in average performance across datasets using a Wilcoxon signed rank test.

When examining the strength of the correlation between ImageNet accuracy and accuracy on transfer datasets, we report for the correlation between the logit-transformed ImageNet accuracy and the logit-transformed transfer accuracy averaged across datasets. We also report the rank correlation (Spearman’s ) in Appendix A.1.2. We computed error bars for model accuracy averaged across datasets using the procedure from \textcitemorey2008 to remove variability due to inherent differences in dataset difficulty. Briefly, given logit-transformed accuracies of model on dataset , we compute adjusted accuracies . For each model, we take the mean and standard error of the adjusted accuracy across datasets, and multiply the latter by a correction factor .

4 Results

We examined 13 networks ranging in ImageNet top-1 accuracy from 69.8% to 82.7%. These networks encompassed the widely used architectures VGG simonyan2014very; Inception v1-v4 and Inception-ResNet v2 szegedy2015going; ioffe2015batch; szegedy2016rethinking; szegedy2017inception; ResNet-50, ResNet-101, and ResNet-152 he2016deep; MobileNet v1 howard2017mobilenets; and the mobile (4 @ 1056) and large (6 @ 4032) variants of NASNet zoph2017learning. Appendix A.3 describes models in further detail and lists the ImageNet top-1 accuracy, parameter count, dimension of the penultimate layer, and input image size for each network. For all experiments, we rescaled images to the same image size as was used for ImageNet training.

Dataset Classes Size (train/test) Accuracy measure Food-101 bossard2014food 101 75,750/25,250 top-1 CIFAR-10 krizhevsky2009learning 10 50,000/10,000 top-1 CIFAR-100 krizhevsky2009learning 10 50,000/10,000 top-1 Birdsnap berg2014birdsnap 500 47,386/2,443 top-1 SUN397 xiao2010sun 397 19,850/19,850 top-1 Stanford Cars krause2013collecting 196 8,144/8,041 top-1 FGVC Aircraft maji13fine-grained 100 6,667/3,333 mean per-class PASCAL VOC 2007 Cls. everingham2010pascal 20 5,011/4,952 11-point mAP Describable Textures (DTD) cimpoi2014describing 47 3,760/1,880 top-1 Oxford-IIIT Pets parkhi2012cats 37 3,680/3,369 mean per-class Caltech-101 fei2004learning 102 3,060/6,084 mean per-class Oxford 102 Flowers nilsback2008automated 102 2,040/6,149 mean per-class
Table 1: Datasets examined in transfer learning

We evaluated models on 12 image classification datasets ranging in training set size from 2,040 to 75,750 images (20 to 5,000 images per class; Table 1). These datasets covered a wide range of image classification tasks, including superordinate-level object classification (CIFAR-10 krizhevsky2009learning, CIFAR-100 krizhevsky2009learning, PASCAL VOC 2007 everingham2010pascal, Caltech-101 fei2004learning); fine-grained object classification (Food-101 bossard2014food, Birdsnap berg2014birdsnap, Stanford Cars krause2013collecting, FGVC Aircraft maji13fine-grained, Oxford-IIIT Pets parkhi2012cats); texture classification (DTD cimpoi2014describing); and scene classification (SUN397 xiao2010sun).

Figure 2: ImageNet accuracy is a strong predictor of transfer accuracy in fine-tuned models. Each set of panels measures correlations between ImageNet accuracy and transfer accuracy across fixed ImageNet features (top), fine-tuned networks (middle) and randomly initialized networks (bottom). Left: Relationship between classification accuracy on transfer datasets (y-axis) and ImageNet top-1 accuracy (x-axis) in different training settings. Axes are logit-scaled (see text). The regression line and a 95% bootstrap confidence interval are plotted in blue. Right: Average log odds of correct classification across datasets, relative to the mean of all classifiers on the dataset. Error bars are standard error. Points corresponding to models not significantly different from the best model () are colored green.

Figure 2 presents correlations between the top-1 accuracy on ImageNet versus the performance of the same model architecture on new image tasks. We measure transfer learning performance in three settings: (1) training a logistic regression classifier on the fixed feature representation from the penultimate layer of the ImageNet-pretrained network, (2) fine-tuning the ImageNet-pretrained network, and (3) training the same CNN architecture from scratch on the new image task.

4.1 ResNets are the best fixed feature extractors

We first examined the performance of different networks when used as fixed feature extractors by training an L2-regularized logistic regression classifier on penultimate layer activations using L-BFGS liu1989limited without data augmentation. In this setting, ImageNet accuracy accounted for only a small fraction of the differences in transfer accuracy among models (). In particular, even though ResNet-50, ResNet-101, and ResNet-152 were the 5th, 6th and 7th best models in terms of ImageNet top-1 accuracy, they were the top three models for transfer learning by logistic regression. ResNets were the best performing architectures on 9 of the 12 datasets, and insignificantly different from the best performing model on another. The surprisingly high performance of ResNets was not directly attributable to the dimension of the penultimate layer, as our experiments included networks with higher ImageNet top-1 performance that had equal (Inception v3), higher (NASNet Large) and lower (Inception-ResNet v2) feature dimension (Appendix A.3).

We performed several control experiments to verify that the superiority of ResNet-152 as a fixed feature extractor was due to the architecture rather than the training procedure or classifier. First, our results remained unchanged when we repeated the logistic regression experiments using ResNet models trained with different image preprocessing szegedy2016rethinking; ioffe2015batch; zoph2017learning. Second, we retrained one network (Inception v3) with less regularization in a setting similar to ResNets (see Appendix A.4). Although the resulting network was more accurate than ResNet-152 on ImageNet, ResNet-152 still performed better on all transfer datasets except for VOC 2007. Third, we trained SVM and k-NN classifiers, and logistic regression classifiers with data augmentation; results were similar (Appendix B).

4.2 ImageNet accuracy is a good predictor of fine-tuning performance

We next examined performance when fine-tuning ImageNet networks. We initialized each network from the ImageNet weights and fine-tuned for 19,531 steps with Nesterov momentum and a cosine decay learning rate schedule at a batch size of 256, searching across learning rate and weight decay on a validation set (for details, see Appendix A.5). Compared to logistic regression on penultimate layer features, fine-tuning produced much stronger correlations between ImageNet accuracy and transfer dataset accuracy (). Correspondingly, the best performing model on ImageNet tested zoph2017learning was the best performing model on 9 transfer datasets and insignificantly different from the best model on the remainder.

When training from random initialization, correlations were more variable, but there was a tendency toward higher performance for models that achieved higher accuracy on ImageNet (). For these experiments, we used a similar training setup as for fine-tuning (see Appendix A.6). The best network in this context was Inception v4, although Inception v3 and NASNet-A Large were insignificantly worse. These are three of the four highest performing networks on ImageNet tested.

4.3 Fine-tuning with better models outperforms specialized methods for transfer learning

Figure 3: Fine-tuning consistently achieves higher accuracy than logistic regression on top of fixed ImageNet features or training from randomly initialized models. The performance of logistic regression on fixed ImageNet features vs. networks trained from random initialization depends heavily on the dataset. Axes are logit scaled.
Figure 4: Fine-tuning achieves state-of-the-art performance. Bars reflect accuracy across models (excluding VGG) for logistic regression, fine-tuning, and training from random initialization. Error bars are standard error. Points represent individual models. Lines represent previous state-of-the-art.

Fine-tuning was substantially more accurate than classifiers trained on fixed features for most datasets. As shown in Figure 3 (left), fine-tuning improved performance over logistic regression in 143 out of 156 dataset and model combinations. For all datasets, the best fine-tuned model was better than the best model fit by logistic regression. When averaged across the tested architectures, fine-tuning yielded significantly better results on all datasets except Caltech-101 (, Wilcoxon signed rank test; Figure 4). The improvement was generally larger for larger datasets. However, fine-tuning produced substantial gains on the smallest dataset, 102 Flowers, with 102 classes and 2,040 training examples.

Previously reported Current work Dataset Acc. Network Acc. Best network Food-101 90.0 Deep layer aggregation deeplayeraggregation 90.1 NASNet-A Large, fine-tuned CIFAR-10 97.9 AmoebaNet amoebanet 98.4a NASNet-A Large, fine-tuned CIFAR-100 87.8 ShakeDrop shakedrop 88.2a NASNet-A Large, fine-tuned Birdsnap 80.2b Mask-CNN WEI2018704 78.5 NASNet-A Large, fine-tuned SUN397 63.2 Places-pretrained VGG zhou2017places 66.5 NASNet-A Large, fine-tuned Stanford Cars 94.1 Deep layer aggregation deeplayeraggregation 93.0 Inception v4, random init FGVC Aircraft 92.9b Deep layer aggregation deeplayeraggregation 89.4 Inception v3, fine-tuned VOC 2007 Cls. 89.7 VGG simonyan2014very 88.4 NASNet-A Large, fine-tuned DTD 75.5 FC+FV-CNN+D-SIFT cimpoi2015deep 76.7 Inception-ResNet v2, fine-tuned Oxford-IIIT Pets 93.8 Object-part attention peng2018object 94.3 NASNet-A Large, fine-tuned Caltech-101 93.4 Spatial pyramid pooling he2014spatial 95.0 NASNet-A Large, fine-tuned Oxford 102 Flowers 97.1 Object-part attention peng2018object 97.7 NASNet-A Large, fine-tuned 11footnotetext: Accuracy excludes images duplicated between the ImageNet training set and CIFAR test sets; see Appendix D.22footnotetext: \textcitekrause2016 achieve 85.4% on Birdsnap and 95.9% on Aircraft using bird and aircraft images collected from Google image search.
Table 2: Performance of best models.

In Table 2, we compare the performance of our models against the best previously reported results on each dataset (see Appendix C for all numerical results). We achieve state-of-the-art performance on 8 of the 12 datasets. For CIFAR-10 and CIFAR-100, the best previously reported results were achieved without auxiliary training data, so the comparison is not strictly fair. All other benchmarks use ImageNet-pretrained networks. Our results suggest that architecture is a critical factor in transfer performance. Several papers have proposed methods to make better use of CNN features and thus improve the efficacy of transfer learning lin2015bilinear; cimpoi2015deep; lin2016visualizing; gao2016compact; yao2016coarse; song2017locally; cui2017kernel; li2017dynamic; peng2018object. On the we datasets examine (Table 2), we outperform all such methods simply by fine-tuning state-of-the-art CNNs. Moreover, in some cases, a better CNN can make up for dataset deficiencies: By fine-tuning ImageNet-pretrained NASNet Large, we obtain state-of-the-art performance on the SUN397 scene dataset, outperforming features extracted from VGG trained on the Places dataset zhou2017places, which more closely matches the domain of SUN397. However, the results on Birdsnap and FGVC Aircraft fall far short of \textcitekrause2016, who augment these datasets with bird and aircraft images collected from the Internet.

Even after excluding methods that perform dataset-specific collection of additional training images, we do not achieve state-of-the-art on the Birdsnap, Cars, Aircraft, or VOC 2007 datasets. The best performing methods on these datasets evaluated at higher image resolutions (i.e., deeplayeraggregation and simonyan2014very) than the fine-tuned networks examined in this work. It is known that networks perform better at larger image sizes at the expense of computational cost, both on ImageNet howard2017mobilenets and in transfer settings cui2018fgvc; lin2016visualizing.

4.4 ImageNet pretraining does not necessarily improve accuracy on fine-grained tasks

Fine-tuning was more accurate than training from random initialization for 129 out of 132 dataset/model combinations, but on Birdsnap, Stanford Cars, and FGVC Aircraft, the improvement was unexpectedly small (Figures 3 and 4). Fine-tuned models were only marginally better than models trained from random initialization (odds ratios of correct classification of 1.11, 1.09, and 1.15 respectively; Figure 4). For Stanford Cars, the best model trained from scratch (Inception v4) outperformed the best fine-tuned model, although the difference was not significant (, binomial test). This model obtained 93.0% accuracy, close to the state-of-the-art result of 94.1% deeplayeraggregation.

ImageNet pretraining has marginal benefits for fine-grained classification tasks where labels are not well-represented in the ILSVRC2012 hierarchy. Stanford Cars and FGVC Aircraft are smaller than most datasets used to train CNNs jeffdean (8,144 training examples/196 classes and 6,667 examples/100 classes, respectively). Three other datasets (Pets, 102 Flowers, and Food-101) require similarly fine-grained classification. However, Pets comprises classes of cats and dogs, which are well-represented within ImageNet, and 102 Flowers is very small, with only 20 training images per class. Performance of training from random initialization on Food-101 approached ImageNet fine-tuning, but there was still a substantial gap (odds ratio 1.31).

4.5 ImageNet pretraining accelerates convergence

Figure 5: Networks pretrained on ImageNet converge faster. Each point represents an independent Inception v4 model trained with optimized hyperparameters. Axes are logit scaled.

Given that fine-tuning and training from random initialization achieved similar performance on Birdsnap, Stanford Cars, and FGVC Aircraft, we next asked whether fine-tuning still posed an advantage in terms of training time. In Figure 5, we examine performance of Inception v4 when fine-tuning or training from random initialization for different numbers of steps. Even when fine-tuning and training from scratch achieved similar final accuracy, we could fine-tune the model to this level of accuracy in an order of magnitude fewer steps. To quantify the acceleration provided by fine-tuning, we computed the number of epochs and steps required to reach 90% of the maximum odds of correct classification achieved at any number of steps, and computed the geometric mean across datasets. Fine-tuning reached this threshold level of accuracy in an average of 26 epochs/1151 steps (inter-quartile ranges 267-4882 steps, 12-58 epochs), whereas training from scratch required 444 epochs/19531 steps (inter-quartile ranges 9765-39062 steps, 208-873 epochs) corresponding to a 17-fold speedup on average.

4.6 Measuring when pretraining improves performance

Figure 6: Pretraining on ImageNet improves performance on fine-grained tasks with small amounts of data. Performance of Inception v4 at different dataset sizes. Error bars reflect standard error over 3 subsets. Note that the maximum dataset size shown is not the full dataset.

We finally examined the behavior of logistic regression, fine-tuning, and training from random initialization in the regime of extremely limited data. In Figure 6, we show the accuracy achieved by these methods for dataset subsets with different numbers of examples per class. When labeled data is sparse (47-800 total examples, or 1-32 per class), logistic regression is a strong baseline, often achieving accuracy comparable to or better than fine-tuning. At larger dataset sizes, fine-tuning achieves higher performance. However, the advantages of ImageNet pretraining can fade surprisingly quickly with more data: On FGVC Aircraft, training from random initialization achieved parity with fine-tuning at 1600 total examples (16 per class).

5 Discussion

The last decade of computer vision research has demonstrated the superiority of image features learned from data over generic, hand-crafted features. Before the rise of convolutional neural networks, most approaches to image understanding relied on manually engineered feature descriptors lowe1999object; dalal2005histograms; bay2008speeded combined with methods to aggregate these descriptors felzenszwalb2010object. \textcitekrizhevsky2012imagenet showed that, given the training data provided by ImageNet deng2009imagenet, features learned by convolutional neural networks could substantially outperform these hand-engineered features. Soon after, it became clear that intermediate representations learned from ImageNet also provided substantial gains over hand-engineered features when transferred to other tasks donahue2014decaf; razavian2014cnn.

Our results reveal clear limits to transferring features, even among natural image datasets. But we also show that, even when features are not transferable, better architectures consistently achieve higher performance. We found that the best fixed image features do not come from the best ImageNet models, as measured by ImageNet accuracies. However, fine-tuning better ImageNet models yields better performance. ImageNet pretraining accelerates convergence and improves performance on many datasets, but its value diminishes with greater training time, more training data, and greater divergence from ImageNet labels. For some fine-grained classification datasets, a few thousand labeled examples, or a few dozen per class, are all that are needed to make training from scratch perform competitively with fine-tuning. Surprisingly, however, the value of architecture persists.

Is the general enterprise of learning widely-useful features doomed to suffer the same fate as feature engineering in computer vision? Given differences among datasets torralba2011unbiased, it is not entirely surprising that features learned on one dataset benefit from some amount of adaptation (i.e., fine-tuning) when applied to another. It is, however, surprising that features learned from a large dataset cannot always be profitably adapted to much smaller datasets. ImageNet weights provide a starting point for features on a new classification task, but perhaps what is needed is a way to learn how to adapt features. This problem is closely related to few-shot learning lake2015human; vinyals2016matching; ravi2016optimization; snell2017prototypical; finn2017model; snell2017prototypical; mishra2017meta, but these methods are typically evaluated with training and test classes from the same distribution. It remains to be seen whether methods can be developed to adapt visual representations learned from ImageNet to provide larger benefits across all natural image tasks.


We thank George Dahl, Sara Hooker, Pieter-jan Kindermans, Jiquan Ngiam, Ruoming Pang, Daiyi Peng, Vishy Tirumalashetty, Vijay Vasudevan, and Emily Xue for comments on the experiments and manuscript, and Aliza Elkin and members of the Google Brain team for support and ideas.


Appendix A Supplementary experimental procedures


a.1 Supplementary statistical methods

a.1.1 Comparison of two models on the same dataset

To test for superiority of one model over another on a given dataset, we constructed permutations where, for each example, we randomly exchanged the predictions of the two networks. For each permutation, we computed the difference in accuracy between the two networks.111For VOC2007, we considered the accuracy of predictions across labels. We computed a p-value as the proportion of permutations where the difference is at least as extreme as the observed difference in accuracy. For top-1 accuracy, this procedure is equivalent to a binomial test sometimes called the "exact McNemar test," and a p-value can be computed exactly. For mean per-class accuracy, we approximated a p-value based on 10,000 permutations. These tests assess whether one trained model performs better than another on data drawn from the test set distribution. However, they are tests between trained models, rather than tests between architectures, since we do not measure variability arising from training networks from different random initializations or from different orderings of the training data.

a.1.2 Measures of correlation

Setting Pearson Pearson Spearman Logistic regression 0.24 0.49 0.59 Fine-tuned 0.86 0.92 0.90 Trained from scratch 0.45 0.67 0.65
Table 3: Correlations between ImageNet accuracy and average transfer accuracy

Table 3 shows the Pearson correlation (as and ) as well as the Spearman rank correlation () in each of the three transfer settings we examine. We believe that Pearson correlation is the more appropriate measure, given that it is less dependent on the specific CNNs chosen for the study and the effects are approximately linear, but our results are similar in either case.

a.2 Datasets

All datasets had a median image size on the shortest side of at least 331 pixels (the highest input image size out of all networks tested), except Caltech-101, for which the median size is 225 on the shortest side and 300 on the longer side, and CIFAR-10 and CIFAR-100, which consist of pixel images.

For datasets with a provided validation set (FGVC Aircraft, VOC2007, DTD, and 102 Flowers), we used this validation set to select hyperparameters. For other datasets, we constructed a validation set by subsetting the original training set. For the DTD and SUN397 datasets, which provide multiple train/test splits, we used only the first provided split. For the Caltech-101 dataset, which specifies no train/test split, we trained on 30 images per class and tested on the remainder, as in previous works donahue2014decaf; simonyan2014very; zeiler2014visualizing; Chatfield14. With the exception of dataset subset results (Figure 6), all results indicate the performance of models retrained on the combined training and validation set.

a.3 Networks

We used publicly available network checkpoints and models from the TF-Slim model repository ( Table 4 lists the ImageNet top-1 accuracy, parameter count, penultimate layer feature dimension, and input image size for each checkpoint. For VGG and ResNets, we used the preprocessing and data augmentation described in \textcitesimonyan2014very. For other networks, we used the preprocessing and data augmentation from \textciteinception_preprocessing.

Model ImageNet top-1222Accuracy of model implementation, which may differ slightly from the original reported result. Parameters333Excludes logits layer. Features Image size VGG-16 (D) simonyan2014very 70.9% 134.3M 4096 224 VGG-19 (E) simonyan2014very 71.0% 139.6M 4096 224 Inception v1444We used Inception models and checkpoints from, which do not match the original papers. szegedy2015going 69.8% 5.6M 1024 224 BN-Inception555This model is called "Inception v2" in TF-Slim model repository, but matches the model described in \textciteioffe2015batch, rather than the model that \textciteszegedy2016rethinking call "Inception v2." ioffe2015batch 74.0% 10.1M 1024 224 Inception v3 szegedy2016rethinking 78.0% 21.8M 2048 299 Inception v4 szegedy2017inception 80.2% 41.1M 1536 299 Inception-ResNet v2 szegedy2017inception 80.4% 54.3M 1536 299 ResNet-50 v1 he2016deep 75.2% 23.5M 2048 224 ResNet-101 v1 he2016deep 76.4% 42.5M 2048 224 ResNet-152 v1 he2016deep 76.8% 58.1M 2048 224 MobileNet v1 howard2017mobilenets 70.7% 3.2M 1024 224 NASNet-A Mobile zoph2017learning 74.0% 4.2M 1056 224 NASNet-A Large zoph2017learning 82.7% 84.7M 4032 331
Table 4: ImageNet classification networks

The majority of our networks were trained with crops after a non-aspect ratio preserving resize, following the data augmentation procedure described in \textciteszegedy2016rethinking. ResNet and VGG models were converted from the original checkpoints of \textcitehe2016deep, and were trained with aspect ratio-preserving crops as in \textcitesimonyan2014very.

a.4 Logistic regression

For each dataset, we extracted features from the penultimate layer of the network and normalized them by their L2 norm as in \textciteChatfield14,simonyan2014very. We trained a multinomial logistic regression classifier using L-BFGS, with an L2 regularization parameter applied to the sum of the per-example losses, selected from a range of 45 logarithmically spaced values from to on the validation set. Since the optimization problem is convex, we used the solution at the previous point along the regularization path can be used as a warm start for the next point, which greatly accelerated the search. For these experiments, we did not perform data augmentation or scale aggregation.

When investigating the effect of preprocessing upon performance, we tested the performance of ResNet-50 v1, ResNet-101 v1, and ResNet-152 v1 models retrained with the preprocessing from \textciteszegedy2015going with label smoothing of 0.1 and an input size of . This model was trained with momentum with an exponentially decaying learning rate schedule, and used an exponential moving average of the weights, as in \textciteszegedy2016rethinking. These models achieved higher performance than non-ResNet models on 9 of 12 datasets.

When investigating the effect of training procedure, we trained an Inception v3 model with regularization and learning rate schedule more comparable to \textcitehe2016deep. Compared to the original Inception v3 model, this model had no auxiliary classifier, no dropout, and no label smoothing, and was trained with momentum and a stepwise learning rate schedule rather than RMSProp with exponential decay. This model achieved an ImageNet top-1 accuracy of 77.2%, 0.8% worse than our original Inception v3 model, but still 0.4% better than ResNet-152, but achieved lower performance than ResNet-152 on all transfer datasets except VOC2007.

a.5 Fine tuning

For all fine-tuning experiments except dataset size experiments, we initialized networks with ImageNet-pretrained weights and trained for 19,531 steps at a batch size of 256 using Nesterov momentum with a momentum parameter of 0.9. We selected the optimal learning rate and weight decay on the validation set by grid search. Our early experiments indicated that the optimal weight decay at a given learning rate varied inversely with the learning rate, as has been recently reported loschilov. Thus, our grid consisted of 7 logarithmically spaced learning rates between 0.0001 and 0.1 and 7 logarithmically spaced weight decay to learning rate ratios between and , as well as zero. We found it useful to decrease the batch normalization momentum parameter from its ImageNet value to where is the number of steps per epoch. We left any other hyperparameters at their ImageNet settings. We found that the maximum performance on the validation set at any step during training was very similar to the maximum performance at the last step (presumably because we searched over both learning rate and weight decay), so we did not perform early stopping.

When examining the effect of dataset size (Section 4.6), we trained for at least 1000 steps or 100 epochs (following guidance from our analysis of training time in Section 4.5) at a batch size of 64, with the learning rate range scaled down by a factor of 4. Because we chose hyperparameters based on a large validation set, the results may not reflect what can be accomplished in practice when training on datasets of this size 46794.

a.6 Training from random initialization

We used a similar training protocol for training from random initialization as for fine tuning, i.e., we trained for 19,531 steps at a batch size of 256 using Nesterov momentum with a momentum parameter of 0.9. Training from random initialization generally achieved optimal performance at higher learning rates and with greater weight decay, so we adjusted the learning rate range to span from 0.001 to 1.0 and the weight decay to learning rate ratio range to span from to .

We found that the best hyperparameters for training VGG tended to be very close to those that caused the network to diverge, so that it was not possible to train the network consistently using the hyperparameters that achieved the best performance on the validation set. We note that implementations of all other networks included batch normalization, which allows the networks to be trained with higher initial learning rates without divergence ioffe2015batch. Rather than report potentially inaccurate results for VGG due to the extent of hyperparameter tuning required, we omitted it from these analyses.

When examining the effect of dataset size (Section 4.6), we trained from random initialization for at least 78,125 steps or 200 epochs at a batch size of 16, with the learning rate range scaled down by a factor of 16. Our investigation of effects of training time (Section 4.5) indicated that training from random initialization always benefited from increased training time, whereas fine tuning did not. Additionally, pilot experiments indicated that training from random initialization, but not fine tuning, benefited from a reduced batch size with very small datasets.

Appendix B Comparison of alternative classifiers

Although a logistic regression classifier trained on the penultimate layer activations has a natural interpretation as retraining the last layer of the neural network, previous studies have typically reported results with SVM donahue2014decaf; razavian2014cnn; simonyan2014very. Thus, we examine performance in this setting as well (Figure 7). SVM and logistic regression results were extremely highly correlated (), and ResNets remained the top performing architectures on 9 datasets, and insignificantly different from the best on one other. SVM achieved higher performance than logistic regression on 102/156 dataset/model pairs, but on average, the improvement was marginal (odds ratio 1.006, , t-test).

We also examined performance of a -nearest neighbor classifier, inspired by olah2014, which reported a reduction in test error for a nearest neighbor classifier trained on the penultimate layer of an MNIST classifier. We selected on the validation set. Performance was universally worse. Nonetheless, ResNets were the top performing architecture on 7 datasets, and insignificantly different from the top performing architecture on one other.

Figure 7: ResNet features are superior for SVM and k-NN classification. Left: Correlation between ImageNet and transfer accuracy, as in Figure 2. Right: Correlation between SVM/k-NN accuracy and logistic regression accuracy.

Finally, we trained a logistic regression classifier with data augmentation, in the same setting we use for fine tuning. We trained for 19,531 steps with Nesterov momentum and a batch size of 256. Because the optimization problem is convex, we did not optimize over learning rate, but instead fixed the learning rate at 0.1. We did optimize over L2 regularization parameters for the final layer, applied to the mean of the per-example losses, selected from a range of 11 logarthmically spaced values between and 1. Results are shown in Figure 8. With data augmentation, ResNets outperformed other architectures on 7 datasets, and were insignificantly different from best performing models on 3 others. Fine-tuning remained clearly superior to logistic regression with data augmentation, achieving better results for 146/156 dataset/model pairs.

Averaged across all model/dataset combinations, logistic regression with data augmentation outperformed logistic regression without data augmentation (odds ratio 1.04), but the superiority was inconsistent. Logistic regression with data augmentation performed better for only 86/156 dataset/model pairs, and the best performing model without data augmentation was better than the best performing model with data augmentation on half of the 12 datasets. One weakness of this comparison is that, since we trained logistic regression without data augmentation in a full-batch setting with a second-order optimizer (albeit at significantly lower computational cost, since we needed to perform only a single forward pass through the network for each image), it is possible that these models reached parameters closer to the minimum of the loss function.

Figure 8: ResNet features are superior for logistic regression trained with data augmentation. Left: Correlation between ImageNet and transfer accuracy, as in Figure 2. Right: Correlation between accuracy of logistic regression without and with data augmentation (top) and between accuracy of logistic regression with data augmentation and fine-tuning (bottom).

Appendix C Numerical performance results

We present the numerical results for logistic regression, fine tuning, and training from random initialization in Table 5. Bold-faced numbers represent best models, or models insignificantly different from the best, in each training setting.

Logistic regression

Network Food CIFAR10 CIFAR100 Birdsnap SUN397 Cars Aircraft VOC2007 DTD Pets Caltech101 Flowers
VGG-16 61.9 86.0 67.1 46.3 54.9 40.2 44.6 82.1 68.4 90.8 89.6 85.1
VGG-19 61.4 87.5 69.2 46.6 54.4 39.7 44.0 81.9 68.8 90.4 89.7 83.5
Inception v1 61.2 87.4 67.5 44.0 54.9 42.2 46.1 80.8 69.0 89.4 90.8 87.5
BN-Inception 62.9 88.4 69.2 43.6 55.9 40.2 42.0 82.7 69.0 91.2 90.1 85.1
Inception v3 68.2 89.1 68.5 49.7 58.7 51.0 49.4 84.1 72.8 92.4 91.4 86.7
Inception v4 68.5 90.7 71.4 50.8 58.8 48.8 46.3 84.4 67.7 92.2 89.8 80.6
Inception-ResNet v2 71.9 92.1 74.3 52.4 60.1 55.5 52.7 84.9 71.8 91.6 91.0 84.8
ResNet-50 v1 71.8 91.6 75.8 51.0 61.1 56.2 51.2 83.0 73.1 91.8 91.8 93.0
ResNet-101 v1 73.7 92.7 76.5 54.9 61.4 56.6 53.5 83.8 73.2 92.0 92.2 92.8
ResNet-152 v1 74.1 93.0 78.0 55.9 61.3 58.4 53.6 84.1 72.2 93.1 92.5 92.8
MobileNet v1 65.4 88.6 69.3 44.2 58.2 57.3 57.8 79.3 68.9 88.4 92.2 91.4
NASNet-A Mobile 58.7 86.6 65.0 38.4 50.9 33.8 36.8 81.1 64.2 88.8 87.0 76.6
NASNet-A Large 70.9 92.8 74.9 52.7 61.3 50.9 49.1 86.1 70.5 92.8 90.8 82.6

Fine tuning

Network Food CIFAR10 CIFAR100 Birdsnap SUN397 Cars Aircraft VOC2007 DTD Pets Caltech101 Flowers
VGG-16 84.9 95.18 80.3 72.6 57.5 88.0 79.1 82.3 70.2 90.5 86.4 95.07
VGG-19 84.6 95.34 81.4 72.6 57.7 88.1 82.1 82.4 71.0 91.2 88.0 95.22
Inception v1 86.0 96.45 82.2 71.9 59.1 91.2 85.6 82.7 71.7 90.7 90.8 96.62
BN-Inception 87.0 97.53 83.8 71.8 58.7 91.9 86.1 85.0 73.0 90.4 92.8 97.03
Inception v3 88.5 97.41 85.5 76.3 62.8 92.2 89.4 86.3 74.9 92.0 93.7 97.19
Inception v4 89.4 97.63 85.2 77.2 62.9 92.7 89.3 84.7 76.0 93.0 93.6 96.44
Inception-ResNet v2 89.1 97.07 86.2 76.0 63.8 90.5 86.7 87.1 76.7 93.2 94.6 96.69
ResNet-50 v1 86.4 96.60 83.1 73.7 61.7 90.0 85.1 84.1 72.2 92.3 90.1 96.94
ResNet-101 v1 86.8 96.87 84.2 73.0 61.0 90.2 85.5 84.6 73.1 93.4 90.9 96.9
ResNet-152 v1 87.0 97.35 84.6 73.4 61.3 89.6 85.6 84.9 73.2 93.5 91.0 97.11
MobileNet v1 85.9 95.78 81.2 70.0 60.2 91.4 84.6 81.7 72.8 88.7 90.0 96.13
NASNet-A Mobile 86.1 96.97 84.2 71.1 60.8 88.6 77.3 84.1 74.3 91.2 91.7 96.34
NASNet-A Large 90.1 98.39 88.4 78.5 66.5 92.7 89.2 88.4 75.7 94.3 95.0 97.74

Random initialization

Network Food CIFAR10 CIFAR100 Birdsnap SUN397 Cars Aircraft VOC2007 DTD Pets Caltech101 Flowers
Inception v1 82.7 94.58 77.5 69.4 51.9 90.3 82.0 66.6 63.0 79.8 74.6 91.9
BN-Inception 84.3 95.32 80.2 70.5 52.6 90.7 83.1 66.4 62.0 79.6 75.0 93.1
Inception v3 86.2 95.92 81.1 76.0 54.9 92.1 86.9 71.3 66.4 81.1 75.8 93.6
Inception v4 86.5 95.47 80.6 76.0 55.0 93.0 88.8 70.8 66.8 83.9 75.8 93.4
Inception-ResNet v2 86.9 95.37 79.5 75.8 53.9 89.9 83.5 65.1 64.4 80.2 71.2 92.6
ResNet-50 v1 83.3 94.27 78.0 69.1 48.9 88.5 81.8 65.0 61.3 79.8 71.2 89.1
ResNet-101 v1 83.7 92.70 78.9 71.2 48.7 88.9 83.5 64.4 61.9 80.6 69.5 87.4
ResNet-152 v1 82.8 94.60 78.8 72.0 48.8 89.7 85.4 63.9 60.8 81.5 70.5 88.7
MobileNet v1 82.2 94.01 77.5 64.7 52.2 89.7 81.9 66.8 63.2 78.4 73.5 91.6
NASNet-A Mobile 81.4 95.11 78.2 68.0 49.7 86.6 79.8 57.3 60.5 74.2 61.0 87.3
NASNet-A Large 86.4 95.79 81.6 76.8 57.6 92.7 88.8 69.4 63.2 82.7 73.7 94.0
Table 5: Model performance

Appendix D Duplicate images

We used a CNN-based duplicate detector trained on synthesized image triplets to detect images that were present in both the ImageNet training set and the datasets we examine. Because the duplicate detector is optimized for speed, it is imperfect. We used a threshold that was conservative based on manual examination, i.e., it resulted in some false positives but very few false negatives. Thus, the results below represent a worst-case scenario for overlap in the datasets examined. Generally, there are relatively few duplicates. For most of these datasets, standard practice is to fine-tune an ImageNet pretrained network without special handling of duplicates, so the presence of duplicates does not affect the comparability of our results to previous work. However, for CIFAR-10 and CIFAR-100, we compare against networks trained from scratch, so we exclude duplicates from the test set.

On CIFAR-10, we achieve an accuracy of 98.39% when fine-tuning NASNet Large (the best model) on the full test set. We also achieve an accuracy of 98.39% on the 9,863 example test set that is disjoint with the ImageNet training set. We achieve an accuracy of 98.54% on the 137 duplicates. On CIFAR-100, we achieve an accuracy of 88.4% on the full test set. We achieve an accuracy of 88.2% on the 9,771 example test set that is disjoint from the ImageNet training set, and an accuracy of 96.94% on the 229 duplicates.

Dataset Train size Test size Train dups Test dups Train dup % Test dup %
Food-101 75,750 25,250 2 1 0.00% 0.00%
CIFAR-10 50,000 10,000 703 137 1.41% 1.37%
CIFAR-100 50,000 10,000 1,134 229 2.27% 2.29%
Birdsnap 47,386 2,443 431 23 0.91% 0.94%
SUN397 19,850 19,850 113 95 0.57% 0.48%
Stanford Cars 8,144 8,041 10 14 0.12% 0.17%
FGVC Aircraft 6,667 3,333 0 1 0.00% 0.03%
VOC2007 5,011 4,952 46 38 0.92% 0.77%
DTD 3,760 1,880 14 9 0.37% 0.48%
Oxford-IIIT Pets 3,680 3,669 227 58 6.17% 1.58%
Caltech-101 3,060 6,084 28 21 0.92% 0.35%
102 Flowers 2,040 6,149 1 0 0.05% 0.00%
Table 6: Prevalence of images duplicated between the ImageNet training set and transfer datasets investigated
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description