EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
2019 · paper · AI and machine learning
Studies how to scale convolutional networks and proposes compound scaling, which uniformly scales depth, width and resolution. The resulting EfficientNet family reaches 84.3% top-1 on ImageNet with a model 8.4× smaller and 6.1× faster than the best prior network.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
11 findings
Substance Table 4 lists "EfficientNet-B1 78.8% @ 0.098s" and "EfficientNet-B7 84.4% @ 3.1s", but Table 2 (and Fig. 1, abstract) give B1 79.1% and B7 84.3%, and Table 8 gives 79.13 and 84.26.
- p. 6 Table 4
78.8% @ 0.098s
- p. 6 Table 4
84.4% @ 3.1s
- p. 6 Table 2
79.1%
- p. 1 Figure 1 caption and inset
Details are in Table 2 and 4.
Only the authors can say what is right here.
Judge's note: Checked on the page 6 image: Table 4 has B1 78.8% and B7 84.4%. Table 2 has B1 79.1% and B7 84.3%. Table 8 on page 9 has val top-1 79.13 and 84.26. The ResNet-152 (77.8%) and GPipe (84.3%) rows in Table 4 match Table 2. The Figure 1 caption sends the reader to Table 4 for the 5.7x/6.1x claims. I found no statement that the latency runs used different checkpoints. Only the authors can say which values are current.
Second check: another model agreed
Report a problem with this finding
Substance Introduction: "our EfficientNet-B4 improves the top-1 accuracy from 76.3% to 83.0% (+6.7%) with similar FLOPS". Table 2 lists ResNet-50 at 76.0% and EfficientNet-B4 at 82.9%, which is a difference of 6.9 points.
- p. 2 Introduction, para 4
from 76.3% to 83.0% (+6.7%) with similar FLOPS
- p. 6 Table 2, ResNet-50 and EfficientNet-B4 rows
76.0%
- p. 6 Table 2, B0 and B4 groups
ResNet-50 (He et al., 2016) 76.0%
- p. 2 Introduction, last paragraph
improves the top-1 accuracy from 76.3% to 83.0% (+6.7%)
Only the authors can say what is right here.
Judge's note: Table 2 (p6) shows ResNet-50 (He et al., 2016) at 76.0% and EfficientNet-B4 at 82.9%. Table 3 (p6) also gives baseline ResNet-50 at 76.0%. Table 8 (p9) gives B4 at 82.89 validation and 82.94 test. Neither 76.3% nor 83.0% appears anywhere else in the paper, and no different evaluation setting is stated.
Second check: another model agreed
Report a problem with this finding
Substance Section 3.3 says "With deeper (d=2.0) and higher resolution (r=2.0), width scaling achieves much better accuracy", but Figure 4 contains only r=1.0 and r=1.3, and its caption gives the last baseline as "(d=2.0, r=1.3)".
- p. 4 Section 3.3, last paragraph
With deeper (d=2.0) and higher resolution (r=2.0)
- p. 4 Figure 4 legend/caption
the last baseline (d=2.0, r=1.3)
- p. 4 Figure 4 legend/caption
d=2.0, r=1.3
- p. 4 Section 3.3
deeper (d=2.0) and higher resolution (r=2.0)
Judge's note: On p4, the Figure 4 legend lists d=1.0/r=1.0, d=1.0/r=1.3, d=2.0/r=1.0 and d=2.0/r=1.3, and the caption states the last baseline is d=2.0, r=1.3 at 299x299. No r=2.0 curve is plotted, so the text should say r=1.3.
Second check: another model agreed
Report a problem with this finding
Substance Table 2 PNASNet row: "23B ... 6.0x". Against EfficientNet-B4's 4.2B this ratio is 5.5x, which is what the AmoebaNet-A row (also 23B) prints.
- p. 6 Table 2, PNASNet row
23B
- p. 6 Table 2, PNASNet row
PNASNet (Liu et al., 2018) 82.9% 96.2% 86M 4.5x 23B 6.0x
Only the authors can say what is right here.
Judge's note: Checked on the p6 image. Either the FLOPS value or the ratio is wrong, and only the authors can say which.
Second check: another model agreed
Report a problem with this finding
Substance The Figure 5 inset lists "NASNet-A (Zoph et al., 2018) 80.7% 24B", but Table 2 and the Figure 1 inset give NASNet-A 82.7%.
- p. 6 Figure 5 inset table
NASNet-A (Zoph et al., 2018) 80.7% 24B
- p. 6 Table 2, NASNet-A row
82.7%
- p. 6 Table 2
NASNet-A (Zoph et al., 2018) 82.7%
- p. 6 Figure 5 legend
NASNet-A (Zoph et al., 2018) 80.7%
Judge's note: Checked in the Figure 5 inset crop and the p6 page image. It is the same model, citation and 24B FLOPS. In Figure 5 itself the NASNet-A point is plotted near 82.7%, so the 80.7% in the inset is inconsistent with the figure as well.
Second check: another model agreed
Report a problem with this finding
Substance Table 7 lists "Baseline model (EfficientNet-B0) 0.4B 77.3%", while Table 2 gives EfficientNet-B0 at 77.1% and Table 8 gives 77.11.
- p. 8 Table 7, baseline row
Baseline model (EfficientNet-B0)
- p. 6 Table 2, EfficientNet-B0 row
77.1%
- p. 8 Table 7, baseline row
Baseline model (EfficientNet-B0) 0.4B 77.3%
- p. 9 Table 8, Val top1 B0
77.11
Only the authors can say what is right here.
Judge's note: All three values checked on the images (pages 6, 8 and 9). The paper does not say that the Section 6 experiments used a different training setup. The Figure 8 baseline mismatch is handled separately in C029.
Second check: another model agreed
Report a problem with this finding
Substance In Table 5, the Oxford-IIIT Pets row gives ResNet-152 at 58M against EfficientNet-B4 at "17M (5.6x)". 58/17 is 3.4x, and the printed Geo-Mean (4.7x) uses 5.6x.
- p. 7 Table 5, Oxford-IIIT Pets row, left block
17M (5.6x)
Only the authors can say what is right here.
Judge's note: Read the Table 5 crop. Every other ratio in the table equals baseline/ours: 85/4=21, 41/28=1.5, 41/10=4.1, 41/17=2.4, 556/64=8.7, 556/41=14. The Pets ratio does not. The geometric mean of the left block is about 4.66 with 5.6 and about 4.38 with 3.4, so the 4.7x was computed from 5.6x. Only the authors can say whether the ratio or a parameter count is wrong.
Second check: another model agreed
Report a problem with this finding
Substance In Figure 8 ("Scaling Up EfficientNet-B0 with Different Methods.") the shared B0 starting point is plotted at about 76.1% top-1 at 0.4B FLOPS, while Table 7 lists "Baseline model (EfficientNet-B0) 0.4B 77.3%".
- p. 8 Figure 8, first point of all four curves at 0.4B FLOPS
Scaling Up EfficientNet-B0 with Different Methods.
- p. 8 Table 7, baseline row
77.3%
Only the authors can say what is right here.
Judge's note: On the page 8 image, the first point of all four curves sits just above the 76 gridline and well below 77. The compound point at 1.8B is about 81.1, matching Table 7. The 1.2-point gap is far larger than plot-reading error. No explanation is given in the paper. This is related to, but distinct from, the Table 7 vs Table 2 mismatch in C025.
Second check: another model agreed
Report a problem with this finding
Reference or label The Figure 5 inset attributes "ResNet-152 (Xie et al., 2017)". Figure 1 and Table 2 cite ResNet-152 as He et al., 2016, and Xie et al., 2017 is the ResNeXt citation in the next row.
- p. 6 Figure 5 inset table
ResNet-152 (Xie et al., 2017)
Judge's note: Checked in the Figure 5 inset crop (p6). Figure 1 (p1) and Table 2 (p6) both give He et al., 2016 for ResNet-152.
Second check: another model agreed
Report a problem with this finding
Reference or label In the Chollet (Xception) reference, the arXiv identifier is printed as a page range: "CVPR, pp. 1610–02357, 2017."
- p. 9 References, Chollet 2017
CVPR, pp. 1610–02357, 2017.
Judge's note: The crop of the p.9 reference list shows "CVPR, pp. 1610–02357, 2017." '1610–02357' has the form of the arXiv identifier 1610.02357 and cannot be a page range, since a page number of 1610 is not plausible for this entry. The citation details are therefore wrong. Fix: give the real CVPR page numbers or the arXiv identifier.
Second check: another model agreed
Report a problem with this finding
Wording The Table 2 caption claims "up to 8.4x parameter reduction and up to 16x FLOPS reduction", but the table lists 18x (ResNeXt-101) and 19x (PolyNet) FLOPS ratios, and the text cites "18x fewer FLOPS".
- p. 6 Table 2 caption
up to 16x FLOPS reduction
- p. 6 Table 2, ResNeXt-101 / PolyNet rows
19x
- p. 6 Table 2 caption
up to 8.4x parameter reduction and up to 16x FLOPS reduction
Judge's note: On p6, the B3 group shows ResNeXt-101 at 32B/18x and PolyNet at 35B/19x, so the largest FLOPS reduction in the table is 19x, not 16x. The 8.4x parameter figure does match the largest parameter ratio printed.
Second check: another model agreed
Report a problem with this finding
18 minor (typos and the like; most checked by the judge only)
Substance GPipe (84.3% top-1, Huang et al., 2018) is listed with "556M" parameters in the Figure 1 inset and Table 5, but "557M" in Related Work ("84.3% using 557M parameters") and Table 2.
- p. 1 Figure 1 inset
556M
- p. 2 Related Work
84.3% using 557M
- p. 6 Table 2 GPipe row
557M
- p. 7 Table 5 GPipe column
556M
Only the authors can say what is right here.
Judge's note: I checked the page images. The Figure 1 inset (p1) shows GPipe 84.3% with 556M. Table 5 (p7) shows 556M in each GPipe row. Table 2 (p6) shows GPipe 84.3% with 557M, and the Related Work paragraph on p2 says 557M. All four places cite the same model, and the paper gives no reason for the 1M difference. The discrepancy is small, but one quantity is printed two ways.
Second check: another model agreed
Report a problem with this finding
Substance SENet (82.7% top-1) has "145M parameters" in Related Work (p2) but "146M" in Table 2 and in the Figure 1 inset.
- p. 2 Related Work, ConvNet Accuracy
82.7% top-1 accuracy with 145M parameters
- p. 6 Table 2 SENet row
146M
- p. 1 Figure 1 inset
146M
- p. 6 Table 2
SENet (Hu et al., 2018) 82.7% 96.2% 146M
Only the authors can say what is right here.
Judge's note: The p2 text says 145M. Table 2 (p6) and the Figure 1 inset (p1) both show 146M for the same SENet model (Hu et al., 2018). The GPipe 556M/557M point some reviewers mention here is handled under C001.
Second check: another model agreed
Report a problem with this finding
Wording The text says "our EfficientNet-B7 surpasses the best existing GPipe accuracy" (p2) and "being more accurate but 8.4x smaller" (p7), but Table 2 and the Figure 1 inset print both models at 84.3% top-1 (and 97.0% top-5).
- p. 2 Introduction, para 4
our EfficientNet-B7 surpasses the best existing GPipe accuracy
- p. 6 Table 2, B7 group
84.3%
- p. 1 Figure 1 inset table
84.3%
- p. 2 Introduction
surpasses the best existing GPipe accuracy
Only the authors can say what is right here.
Judge's note: Table 2 (p6) gives B7 84.3%/97.0% and GPipe 84.3%/97.0%, and the Figure 1 inset also shows 84.3% for both. Table 4 (p6) gives B7 84.4% against GPipe 84.3%, while Table 8 (p9) gives B7 validation top-1 84.26. The paper's own figures therefore do not consistently show B7 above GPipe. The authors should clarify which accuracy supports "surpasses".
Second check: another model agreed
Report a problem with this finding
Wording The symbol w is defined as the width coefficient (Eq. 2-3, footnote 2) and then reused as the trade-off exponent in "ACC(m)×[FLOPS(m)/T]^w ... w=-0.07 is a hyperparameter" (p5).
- p. 5 Section 4, para 2
is a hyperparameter for controlling the trade-off
- p. 3 Section 3.1, after Eq. 2
are coefficients for scaling network width
Judge's note: On p5, Section 4 uses w as the NAS objective exponent. The same page (Eq. 3) and pp. 3-4 use w for network width. The meaning can be recovered from context, but the same symbol carries two meanings.
Second check: another model agreed
Report a problem with this finding
Wording Table 2 gives EfficientNet-B1 top-5 as "94.4%", but Table 8 gives Val top5 B1 as "94.47", which rounds to 94.5.
- p. 6 Table 2, EfficientNet-B1 row
94.4%
- p. 9 Table 8, Val top5, B1
94.47
Only the authors can say what is right here.
Judge's note: Read both tables on the images. All other Table 8 validation values round to the Table 2 entries (for example 95.67 to 95.7, 96.76 to 96.8, 96.97 to 97.0, 83.95 to 84.0), so Table 2 does not use truncation. B1 top-5 is the only exception. The difference is 0.1 point.
Second check: another model agreed
Report a problem with this finding
Wording In Table 5's best-reported block, tied rows bold only the EfficientNet value (FGVC Aircraft DAT 92.9% vs "92.9%" bold; Food-101 GPipe 93.0% vs "93.0%" bold), while elsewhere bold marks the higher entry.
- p. 7 Table 5, right block, FGVC Aircraft and Food-101 rows
93.0%
Judge's note: Image check: GPipe 99.0%, DAT 94.8% and GPipe 95.9% are bold where they are higher. In the two tie rows only the EfficientNet entry is bold, which suggests EfficientNet is ahead when it only ties.
Second check: another model agreed
Report a problem with this finding
Wording Section 5.3 says EfficientNets "still surpass their accuracy in 5 out of 8 datasets", and the Conclusion says "surpassing state-of-the-art accuracy ... on both ImageNet and five commonly used transfer learning datasets". In Table 5, EfficientNet is strictly higher on 3 datasets and ties on 2 (Aircraft 92.9/92.9, Food-101 93.0/93.0).
- p. 8 Section 5.3
still surpass their accuracy in 5 out of 8 datasets
- p. 7 Table 5, right block, FGVC Aircraft and Food-101 rows
92.9%
- p. 9 Conclusion
surpassing state-of-the-art accuracy with an order of magni-tude fewer parameters and FLOPS, on both ImageNet and five commonly used transfer learning datasets
- p. 7 Table 5 caption
new state-of-the-art accuracy for 5 out of 8 datasets
Only the authors can say what is right here.
Judge's note: Read the Table 5 right block on the image. EfficientNet is higher on CIFAR-100 (91.7 vs 91.3), Birdsnap (84.3 vs 83.6) and Flowers (98.8 vs 97.7), ties on Aircraft and Food-101, and is lower on CIFAR-10, Cars and Pets. The caption's 'new state-of-the-art for 5 out of 8' can be read as including ties, but 'surpass' cannot. On ImageNet, B7 at 84.3% also equals GPipe at 84.3%.
Second check: another model agreed
Report a problem with this finding
Wording Section 5.3 (1) says "our EfficientNet models achieve better accuracy with 4.7x average" parameter reduction, but in Table 5's left block FGVC Aircraft B3 is 90.7% vs Inception-v4 at 90.9%, and Flowers ties at 98.5%.
- p. 8 Section 5.3, point (1)
achieve better accuracy with 4.7x average
- p. 7 Table 5 left block, FGVC Aircraft row
90.7%
- p. 8 Sec 5.3 (1)
our EfficientNet models achieve better accuracy with 4.7x average
- p. 7 Table 5, FGVC Aircraft
90.9% 41M EfficientNet-B3 90.7%
Only the authors can say what is right here.
Judge's note: Checked the Table 5 left block on the image. The unqualified 'better accuracy' does not hold for Aircraft (lower) or Flowers (tie).
Second check: another model agreed
Report a problem with this finding
Typo Grammar: "conventional practice that arbitrary scales these factors" (p1) should read "arbitrarily scales", and "we propose a new compound scaling method, which use a compound coefficient" (p5) should read "which uses".
- p. 1 Introduction, last paragraph
Unlike conventional practice that arbitrary scales these fac-
- p. 2 Introduction
our scaling method work well on exist-
- p. 5 Section 3.3
which use a compound coefficient
- p. 1 Introduction
that arbitrary scales these fac-
Judge's note: Both phrases appear as quoted on p1 (Introduction, last paragraph) and p5 (Section 3.3). The "method work well" slip on p2 is handled separately under C007.
Second check: none (judge only)
Report a problem with this finding
Typo "by scaling up a baseline model four time larger" (p1) should read "four times larger".
- p. 1 Introduction, para 1
by scaling up a baseline model four
Judge's note: On the p1 image, the left column ends with "baseline model four" and the right column begins with "time larger."
Second check: none (judge only)
Report a problem with this finding
Typo "We demonstrate that our scaling method work well on existing MobileNets" (p2) should read "works well".
- p. 2 Introduction, para 4
We demonstrate that our scaling method work well on exist-
Judge's note: The p2 image shows the phrase exactly as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 5 misspells AmoebaNet as "AmeobaNet-C (Cubuk et al., 2019)" in the inset table, and as "AmeobaNet-A" in the plot label.
- p. 6 Figure 5 plot label and inset
AmeobaNet-C (Cubuk et al., 2019)
Judge's note: The crop of the Figure 5 inset clearly shows 'AmeobaNet-C'. The text layer also gives 'AmeobaNet-A' for the plot label. Table 2 and Figure 1 spell it 'AmoebaNet'.
Second check: none (judge only)
Report a problem with this finding
Typo Table 2 writes "Inception-resnet-v2 (Szegedy et al., 2017)", while Figures 5 and 6 write 'Inception-ResNet-v2'.
- p. 6 Table 2
Inception-resnet-v2 (Szegedy et al., 2017)
Judge's note: Seen on the page 6 image (Table 2 and the Figure 5 label) and in the Figure 6 legend on page 7. This is an inconsistent capitalisation only.
Second check: none (judge only)
Report a problem with this finding
Typo Section 5.2 has the misspelling "representative CovNets on a real CPU"; it should be 'ConvNets'.
- p. 7 Section 5.2, last paragraph
representative CovNets on a real CPU
Judge's note: Visible on the page 7 image, last paragraph of Section 5.2.
Second check: none (judge only)
Report a problem with this finding
Typo Table 5 writes "Gpipe" in the CIFAR-10 and CIFAR-100 rows but 'GPipe' in the Birdsnap, Pets and Food-101 rows and in the footnote.
- p. 7 Table 5, CIFAR-10/CIFAR-100 rows
Gpipe
Judge's note: Seen in the Table 5 crop.
Second check: none (judge only)
Report a problem with this finding
Typo Grammar error in Section 5.2: "but also computational cheaper" should read 'computationally cheaper'.
- p. 7 Section 5.2
but also computational cheaper
Judge's note: Visible on the page 7 image, right column.
Second check: none (judge only)
Report a problem with this finding
Typo In Section 5.3 the sentence "but using 9.6x fewer parameters" ends the paragraph with no full stop.
- p. 8 Section 5.3, end of second paragraph
but using 9.6x fewer parameters
Judge's note: The crop shows no terminal period before the paragraph that begins 'Figure 6 compares...'.
Second check: none (judge only)
Report a problem with this finding
Typo In the Krause et al. reference, a word is truncated: "Second Workshop on Fine-Grained Visual Categorizatio, 2013." It should read "Categorization".
- p. 10 References, Krause et al. 2013
Fine-Grained Visual Categorizatio
- p. 10 References, Krause et al.
Fine-Grained Visual Categorizatio, 2013
Judge's note: The crop of the p.10 reference list shows 'Categorizatio' followed directly by a comma, with no line break, so this is not an extraction artefact. It is a cosmetic misspelling.
Second check: none (judge only)