Published as a conference paper at ICLR 2019

R ETHINKING THE VALUE OF N ETWORK P RUNING

Zhuang Liu1∗, Mingjie Sun2∗†, Tinghui Zhou1 , Gao Huang2 , Trevor Darrell1
1
University of California, Berkeley 2 Tsinghua University

A BSTRACT
Network pruning is widely used for reducing the heavy inference cost of deep
models in low-resource settings. A typical pruning algorithm is a three-stage
pipeline, i.e., training (a large model), pruning and fine-tuning. During prun-
ing, according to a certain criterion, redundant weights are pruned and important
weights are kept to best preserve the accuracy. In this work, we make several
surprising observations which contradict common beliefs. For all state-of-the-art
structured pruning algorithms we examined, fine-tuning a pruned model only gives
comparable or worse performance than training that model with randomly initial-
ized weights. For pruning algorithms which assume a predefined target network
architecture, one can get rid of the full pipeline and directly train the target net-
work from scratch. Our observations are consistent for multiple network architec-
tures, datasets, and tasks, which imply that: 1) training a large, over-parameterized
model is often not necessary to obtain an efficient final model, 2) learned “impor-
tant” weights of the large model are typically not useful for the small pruned
model, 3) the pruned architecture itself, rather than a set of inherited “important”
weights, is more crucial to the efficiency in the final model, which suggests that in
some cases pruning can be useful as an architecture search paradigm. Our results
suggest the need for more careful baseline evaluations in future research on struc-
tured pruning methods. We also compare with the “Lottery Ticket Hypothesis”
(Frankle & Carbin, 2019), and find that with optimal learning rate, the “winning
ticket” initialization as used in Frankle & Carbin (2019) does not bring improve-
ment over random initialization.

1 I NTRODUCTION
Over-parameterization is a widely-recognized property of deep neural networks (Denton et al., 2014;
Ba & Caruana, 2014), which leads to high computational cost and high memory footprint for infer-
ence. As a remedy, network pruning (LeCun et al., 1990; Hassibi & Stork, 1993; Han et al., 2015;
Molchanov et al., 2016; Li et al., 2017) has been identified as an effective technique to improve the
efficiency of deep networks for applications with limited computational budget. A typical procedure
of network pruning consists of three stages: 1) train a large, over-parameterized model (sometimes
there are pretrained models available), 2) prune the trained large model according to a certain crite-
rion, and 3) fine-tune the pruned model to regain the lost performance.
Generally, there are two common beliefs behind
this pruning procedure. First, it is believed that
starting with training a large, over-parameterized Training Pruning Fine-tuning
network is important (Luo et al., 2017; Carreira-
Perpinán & Idelbayev, 2018), as it provides a high-
performance model (due to stronger representation Figure 1: A typical three-stage network pruning
& optimization power) from which one can safely pipeline.
remove a set of redundant parameters without sig-
nificantly hurting the accuracy. Therefore, this is usually believed, and reported to be superior to
directly training a smaller network from scratch (Li et al., 2017; Luo et al., 2017; He et al., 2017b;
Yu et al., 2018) – a commonly used baseline approach. Second, both the pruned architecture and
its associated weights are believed to be essential for obtaining the final efficient model (Han et al.,
∗
Equal contribution.
†
Work done while visiting UC Berkeley.

1
Published as a conference paper at ICLR 2019

2015). Thus most existing pruning techniques choose to fine-tune a pruned model instead of train-
ing it from scratch. The preserved weights after pruning are usually considered to be critical, as
how to accurately select the set of important weights is a very active research topic in the literature
(Molchanov et al., 2016; Li et al., 2017; Luo et al., 2017; He et al., 2017b; Liu et al., 2017; Suau
et al., 2018).
In this work, we show that both of the beliefs mentioned above are not necessarily true for structured
pruning methods, which prune at the levels of convolution channels or larger. Based on an extensive
empirical evaluation of state-of-the-art pruning algorithms on multiple datasets with multiple net-
work architectures, we make two surprising observations. First, for structured pruning methods with
predefined target network architectures (Figure 2), directly training the small target model from ran-
dom initialization can achieve the same, if not better, performance, as the model obtained from the
three-stage pipeline. In this case, starting with a large model is not necessary and one could instead
directly train the target model from scratch. Second, for structured pruning methods with auto-
discovered target networks, training the pruned model from scratch can also achieve comparable or
even better performance than fine-tuning. This observation shows that for these pruning methods,
what matters more may be the obtained architecture, instead of the preserved weights, despite train-
ing the large model is needed to find that target architecture. Interestingly, for a unstructured pruning
method (Han et al., 2015) that prunes individual parameters, we found that training from scratch can
mostly achieve comparable accuracy with pruning and fine-tuning on smaller-scale datasets, but
fails to do so on the large-scale ImageNet benchmark. Note that in some cases, if a pretrained large
model is already available, pruning and fine-tuning from it can save the training time required to
obtain the efficient model. The contradiction between some of our results and those reported in the
literature might be explained by less carefully chosen hyper-parameters, data augmentation schemes
and unfair computation budget for evaluating baseline approaches.
Our results advocate a rethinking of existing struc- A 4-layer model
tured network pruning algorithms. It seems that
the over-parameterization during the first-stage
training is not as beneficial as previously thought.
Also, inheriting weights from a large model is not
necessarily optimal, and might trap the pruned Predefined: prune Automatic: prune a%,
model into a bad local minimum, even if the x% channels in b%, c%, d% channels
weights are considered “important” by the prun- each layer in each layer
ing criterion. Instead, our results suggest that
the value of automatic structured pruning algo-
rithms sometimes lie in identifying efficient struc-
tures and performing implicit architecture search,
rather than selecting “important” weights. For
most structured pruning methods which prune Figure 2: Difference between predefined and auto-
channels/filters, this corresponds to searching the matically discovered target architectures, in channel
number of channels in each layer. In section 5, pruning as an example. The pruning ratio x is user-
we discuss this viewpoint through carefully de- specified, while a, b, c, d are determined by the prun-
signed experiments, and show the patterns in the ing algorithm. Unstructured sparse pruning can also
be viewed as automatic.
pruned model could provide design guidelines for
efficient architectures.
The rest of the paper is organized as follows: in Section 2, we introduce background and some related
works on network pruning; in Section 3, we describe our methodology for training the pruned model
from scratch; in Section 4 we experiment on various pruning methods and show our main results for
both pruning methods with predefined or automatically discovered target architectures; in Section 5,
we discuss the value of automatic pruning methods in searching efficient network architectures; in
Section 6 we discuss some implications and conclude the paper.

2 BACKGROUND

Recent success of deep convolutional networks (LeCun et al., 1998; Deng et al., 2009; Girshick et al.,
2014; Long et al., 2015; He et al., 2016; 2017a) has been coupled with increased requirement of
computation resources. In particular, the model size, memory footprint, the number of computation
operations (FLOPs) and power usage are major aspects inhibiting the use of deep neural networks

2
Published as a conference paper at ICLR 2019

in some resource-constrained settings. Those large models can be infeasible to store, and run in
real time on embedded systems. To address this issue, many methods have been proposed such as
low-rank approximation of weights (Denton et al., 2014; Lebedev et al., 2014), weight quantization
(Courbariaux et al., 2016; Rastegari et al., 2016), knowledge distillation (Hinton et al., 2014; Romero
et al., 2015) and network pruning (Han et al., 2015; Li et al., 2017), among which network pruning
has gained notable attention due to their competitive performance and compatibility.
One major branch of network pruning methods is individual weight pruning, and it dates back to
Optimal Brain Damage (LeCun et al., 1990) and Optimal Brain Surgeon (Hassibi & Stork, 1993),
which prune weights based on Hessian of the loss function. More recently, Han et al. (2015) pro-
poses to prune network weights with small magnitude, and this technique is further incorporated into
the “Deep Compression” pipeline (Han et al., 2016b) to obtain highly compressed models. Srinivas
& Babu (2015) proposes a data-free algorithm to remove redundant neurons iteratively. Molchanov
et al. (2017) uses Variatonal Dropout (P. Kingma et al., 2015) to prune redundant weights. Louizos
et al. (2018) learns sparse networks through L0 -norm regularization based on stochastic gate. How-
ever, one drawback of these unstructured pruning methods is that the resulting weight matrices are
sparse, which cannot lead to compression and speedup without dedicated hardware/libraries (Han
et al., 2016a).
In contrast, structured pruning methods prune at the level of channels or even layers. Since the
original convolution structure is still preserved, no dedicated hardware/libraries are required to re-
alize the benefits. Among structured pruning methods, channel pruning is the most popular, since it
operates at the most fine-grained level while still fitting in conventional deep learning frameworks.
Some heuristic methods include pruning channels based on their corresponding filter weight norm
(Li et al., 2017) and average percentage of zeros in the output (Hu et al., 2016). Group sparsity is
also widely used to smooth the pruning process after training (Wen et al., 2016; Alvarez & Salz-
mann, 2016; Lebedev & Lempitsky, 2016; Zhou et al., 2016). Liu et al. (2017) and Ye et al. (2018)
impose sparsity constraints on channel-wise scaling factors during training, whose magnitudes are
then used for channel pruning. Huang & Wang (2018) uses a similar technique to prune coarser
structures such as residual blocks. He et al. (2017b) and Luo et al. (2017) minimizes next layer’s
feature reconstruction error to determine which channels to keep. Similarly, Yu et al. (2018) opti-
mizes the reconstruction error of the final response layer and propagates a “importance score” for
each channel. Molchanov et al. (2016) uses Taylor expansion to approximate each channel’s influ-
ence over the final loss and prune accordingly. Suau et al. (2018) analyzes the intrinsic correlation
within each layer and prune redundant channels. Chin et al. (2018) proposes a layer-wise com-
pensate filter pruning algorithm to improve commonly-adopted heuristic pruning metrics. He et al.
(2018a) proposes to allow pruned filters to recover during the training process. Lin et al. (2017);
Wang et al. (2017) prune certain structures in the network based on the current input.
Our work is also related to some recent studies on the characteristics of pruning algorithms. Mittal
et al. (2018) shows that random channel pruning (Anwar & Sung, 2016) can perform on par with
a variety of more sophisticated pruning criteria, demonstrating the plasticity of network models.
In the context of unstructured pruning, The Lottery Ticket Hypothesis (Frankle & Carbin, 2019)
conjectures that certain connections together with their randomly initialized weights, can enable a
comparable accuracy with the original network when trained in isolation. We provide comparisons
between Frankle & Carbin (2019) and this work in Appendix A. Zhu & Gupta (2018) shows that
training a small-dense model cannot achieve the same accuracy as a pruned large-sparse model with
identical memory footprint. In this work, we reveal a different and rather surprising characteristic
of structured network pruning methods: fine-tuning the pruned model with inherited weights is not
better than training it from scratch; the resulting pruned architectures are more likely to be what
brings the benefit.

3 M ETHODOLOGY
In this section, we describe in detail our methodology for training a small target model from scratch.
Target Pruned Architectures. We first divide network pruning methods into two categories. In a
pruning pipeline, the target pruned model’s architecture can be determined by either a human (i.e.,
predefined) or the pruning algorithm (i.e., automatic) (see Figure 2).
When a human predefines the target architecture, a common criterion is the ratio of channels to
prune in each layer. For example, we may want to prune 50% channels in each layer of VGG. In this

3
Published as a conference paper at ICLR 2019

case, no matter which specific channels are pruned, the pruned target architecture remains the same,
because the pruning algorithm only locally prunes the least important 50% channels in each layer. In
practice, the ratio in each layer is usually selected through empirical studies or heuristics. Examples
of predefined structured pruning include Li et al. (2017), Luo et al. (2017), He et al. (2017b) and He
et al. (2018a)
When the target architecture is automatically determined by a pruning algorithm, it is usually based
on a pruning criterion that globally compares the importance of structures (e.g., channels) across
layers. Examples of automatic structured pruning include Liu et al. (2017), Huang & Wang (2018),
Molchanov et al. (2016) and Suau et al. (2018).
Unstructured pruning (Han et al., 2015; Molchanov et al., 2017; Louizos et al., 2018) also falls in
the category of automatic methods, where the positions of pruned weights are determined by the
training process and the pruning algorithm, and it is usually not possible to predefine the positions
of zeros before training starts.
Datasets, Network Architectures and Pruning Methods. In the network pruning literature,
CIFAR-10, CIFAR-100 (Krizhevsky, 2009), and ImageNet (Deng et al., 2009) datasets are the
de-facto benchmarks, while VGG (Simonyan & Zisserman, 2015), ResNet (He et al., 2016) and
DenseNet (Huang et al., 2017) are the common network architectures. We evaluate four predefined
pruning methods, Li et al. (2017), Luo et al. (2017), He et al. (2017b), He et al. (2018a), two auto-
matic structured pruning methods, Liu et al. (2017), Huang & Wang (2018), and one unstructured
pruning method (Han et al., 2015). For the first six methods, we evaluate using the same (target
model, dataset) pairs as presented in the original paper to keep our results comparable. For the last
one (Han et al., 2015), we use the aforementioned architectures instead, since the ones in the origi-
nal paper are no longer state-of-the-art. On CIFAR datasets, we run each experiment with 5 random
seeds, and report the mean and standard deviation of the accuracy.
Training Budget. One crucial question is how long we should train the small pruned model from
scratch. Naively training for the same number of epochs as we train the large model might be
unfair, since the small pruned model requires significantly less computation for one epoch. Alter-
natively, we could compute the floating point operations (FLOPs) for both the pruned and large
models, and choose the number of training epoch for the pruned model that would lead to the same
amount of computation as training the large model. Note that it is not clear how to train the mod-
els to “full convergence” given the stepwise decaying learning rate schedule commonly used in the
CIFAR/ImageNet classification tasks.
In our experiments, we use Scratch-E to denote training the small pruned models for the same
epochs, and Scratch-B to denote training for the same amount of computation budget (on Ima-
geNet, if the pruned model saves more than 2× FLOPs, we just double the number of epochs for
training Scratch-B, which amounts to less computation budget than large model training). When
extending the number of epochs in Scratch-B, we also extend the learning rate decay schedules pro-
portionally. One may argue that we should instead train the small target model for fewer epochs
since it may converge faster. However, in practice we found that increasing the training epochs
within a reasonable range is rarely harmful. In our experiments we found in most times Scratch-E is
enough while in other cases Scratch-B is needed for a comparable accuracy as fine-tuning. Note that
our evaluations use the same computation as large model training without considering the compu-
tation in fine-tuning, since in our evaluated methods fine-tuning does not take too long; if anything
this still favors the pruning and fine-tuning pipeline.
Implementation. In order to keep our setup as close to the original papers as possible, we use the
following protocols: 1) ff a previous pruning method’s training setup is publicly available, e.g. Liu
et al. (2017), Huang & Wang (2018) and He et al. (2018a), we adopt the original implementation; 2)
otherwise, for simpler pruning methods, e.g., Li et al. (2017) and Han et al. (2015), we re-implement
the three-stage pruning procedure and generally achieve similar results as in the original papers; 3)
for the remaining two methods (Luo et al., 2017; He et al., 2017b), the pruned models are publicly
available but without the training setup, thus we choose to re-train both large and small target models
from scratch. Interestingly, the accuracy of our re-trained large model is higher than what is reported
in the original papers. This could be due to the difference in the deep learning frameworks: we used
Pytorch (Paszke et al., 2017) while the original papers used Caffe (Jia et al., 2014). In this case, to

4
Published as a conference paper at ICLR 2019

accommodate the effects of different frameworks and training setups, we report the relative accuracy
drop from the unpruned large model.
We use standard training hyper-parameters and data-augmentation schemes, which are used both
in standard image classification models (He et al., 2016; Huang et al., 2017) and network pruning
methods (Li et al., 2017; Liu et al., 2017; Huang & Wang, 2018; He et al., 2018a). The optimization
method is SGD with Nesterov momentum, using an stepwise decay learning rate schedule. For
random weight initialization, we adopt the scheme proposed in (He et al., 2015). For results of
models fine-tuned from inherited weights, we either use the released models from original papers
(case 3 above) or follow the common practice of fine-tuning the model using the lowest learning rate
when training the large model (Li et al., 2017; He et al., 2017b). For CIFAR, training/fine-tuning
takes 160/40 epochs. For ImageNet, training/fine-tuning takes 90/20 epochs. For reproducing the
results and a more detailed knowledge about the settings, see our code at: https://github.
com/Eric-mingjie/rethinking-network-pruning.

4 E XPERIMENTS
In this section we present our experimental results comparing training pruned models from scratch
and fine-tuning from inherited weights, for both predefined and automatic (Figure 2) structured prun-
ing, as well as a magnitude-based unstructured pruning method (Han et al., 2015). We also include
a comparison with the Lottery Ticket Hypothesis (Frankle & Carbin, 2019), and an experiment on
transfer learning from image classification to object detection in Appendix, due to space limit. We
also put the results and discussions on a pruning method (Soft Filter pruning (He et al., 2018a)) in
Appendix.

4.1 P REDEFINED S TRUCTURED P RUNING

L1 -norm based Filter Pruning (Li et al., 2017) is one of the earliest works on filter/channel pruning
for convolutional networks. In each layer, a certain percentage of filters with smaller L1 -norm will
be pruned. Table 1 shows our results. The Pruned Model column shows the list of predefined target
models (see (Li et al., 2017) for configuration details on each model). We observe that in each
row, scratch-trained models achieve at least the same level of accuracy as fine-tuned models, with
Scratch-B slightly higher than Scratch-E in most cases. On ImageNet, both Scratch-B models are
better than the fine-tuned ones by a noticeable margin.
Dataset Model Unpruned Pruned Model Fine-tuned Scratch-E Scratch-B
VGG-16 93.63 (±0.16) VGG-16-A 93.41 (±0.12) 93.62 (±0.11) 93.78 (±0.15)
ResNet-56-A 92.97 (±0.17) 92.96 (±0.26) 93.09 (±0.14)
ResNet-56 93.14 (±0.12)
CIFAR-10 ResNet-56-B 92.67 (±0.14) 92.54 (±0.19) 93.05 (±0.18)
ResNet-110-A 93.14 (±0.16) 93.25 (±0.29) 93.22 (±0.22)
ResNet-110 93.14 (±0.24)
ResNet-110-B 92.69 (±0.09) 92.89 (±0.43) 93.60 (±0.25)
ResNet-34-A 72.56 72.77 73.03
ImageNet ResNet-34 73.31
ResNet-34-B 72.29 72.55 72.91
Table 1: Results (accuracy) for L1 -norm based filter pruning (Li et al., 2017). “Pruned Model” is the model
pruned from the large model. Configurations of Model and Pruned Model are both from the original paper.

ThiNet (Luo et al., 2017) greedily prunes the channel that has the smallest effect on the next layer’s
activation values. As shown in Table 2, for VGG-16 and ResNet-50, both Scratch-E and Scratch-
B can almost always achieve better performance than the fine-tuned model, often by a significant
margin. The only exception is Scratch-E for VGG-Tiny, where the model is pruned very aggressively
from VGG-16 (FLOPs reduced by 15×), and as a result, drastically reducing the training budget for
Scratch-E. The training budget of Scratch-B for this model is also 7 times smaller than the original
large model, yet it can achieve the same level of accuracy as the fine-tuned model.
Regression based Feature Reconstruction (He et al., 2017b) prunes channels by minimizing the
feature map reconstruction error of the next layer. In contrast to ThiNet (Luo et al., 2017), this
optimization problem is solved by LASSO regression. Results are shown in Table 3. Again, in
terms of relative accuracy drop from the large models, scratch-trained models are better than the
fine-tuned models.

5
Published as a conference paper at ICLR 2019

Dataset Unpruned Strategy Pruned Model

VGG-16 VGG-Conv VGG-GAP VGG-Tiny
71.03 Fine-tuned −1.23 −3.67 −11.61
Scratch-E −2.75 −4.66 −14.36
71.51
Scratch-B +0.21 −2.85 −11.58
ImageNet
ResNet-50 ResNet50-30% ResNet50-50% ResNet50-70%
75.15 Fine-tuned −6.72 −4.13 −3.10
Scratch-E −5.21 −2.82 −1.71
76.13
Scratch-B −4.56 −2.23 −1.01
Table 2: Results (accuracy) for ThiNet (Luo et al., 2017). Names such as “VGG-GAP” and “ResNet50-30%”
are pruned models whose configurations are defined in Luo et al. (2017). To accommodate the effects of
different frameworks between our implementation and the original paper’s, we compare relative accuracy drop
from the unpruned large model. For example, for the pruned model VGG-Conv, −1.23 is relative to 71.03 on
the left, which is the reported accuracy of the unpruned large model VGG-16 in the original paper; −2.75 is
relative to 71.51 on the left, which is VGG-16’s accuracy in our implementation.

Dataset Unpruned Strategy Pruned Model

VGG-16 VGG-16-5x
71.03 Fine-tuned −2.67
Scratch-E −3.46
71.51
Scratch-B −0.51
ImageNet
ResNet-50 ResNet-50-2x
75.51 Fine-tuned −3.25
Scratch-E −1.55
76.13
Scratch-B −1.07
Table 3: Results (accuracy) for Regression based Feature Reconstruction (He et al., 2017b). Pruned models
such as “VGG-16-5x” are defined in He et al. (2017b). Similar to Table 2, we compare relative accuracy drop
from unpruned large models.

4.2 AUTOMATIC S TRUCTURED P RUNING

Network Slimming (Liu et al., 2017) imposes L1 -sparsity on channel-wise scaling factors from
Batch Normalization layers (Ioffe & Szegedy, 2015) during training, and prunes channels with lower
scaling factors afterward. Since the channel scaling factors are compared across layers, this method
produces automatically discovered target architectures. As shown in Table 4, for all networks, the
small models trained from scratch can reach the same accuracy as the fine-tuned models. More
specifically, we found that Scratch-B consistently outperforms (8 out of 10 experiments) the fine-
tuned models, while Scratch-E is slightly worse but still mostly within the standard deviation.

Dataset Model Unpruned Prune Ratio Fine-tuned Scratch-E Scratch-B

VGG-19 93.53 (±0.16) 70% 93.60 (±0.16) 93.30 (±0.11) 93.81 (±0.14)
40% 94.77 (±0.12) 94.70 (±0.11) 94.90 (±0.04)
PreResNet-164 95.04 (±0.16)
CIFAR-10 60% 94.23 (±0.21) 94.58 (±0.18) 94.71 (±0.21)
40% 94.00 (±0.20) 93.68 (±0.18) 94.06 (±0.12)
DenseNet-40 94.10 (±0.12)
60% 93.87 (±0.13) 93.58 (±0.21) 93.85 (±0.25)
VGG-19 72.63 (±0.21) 50% 72.32 (±0.28) 71.94 (±0.17) 73.08 (±0.22)
40% 76.22 (±0.20) 76.36 (±0.32) 76.68 (±0.35)
PreResNet-164 76.80 (±0.19)
CIFAR-100 60% 74.17 (±0.33) 75.05 (± 0.08) 75.73 (±0.29)
40% 73.35 (±0.17) 73.24 (±0.29) 73.19 (±0.26)
DenseNet-40 73.82 (±0.34)
60% 72.46 (±0.22) 72.62 (±0.36) 72.91 (±0.34)
ImageNet VGG-11 70.84 50% 68.62 70.00 71.18

Table 4: Results (accuracy) for Network Slimming (Liu et al., 2017). “Prune ratio” stands for total percentage
of channels that are pruned in the whole network. The same ratios for each model are used as the original paper.

Sparse Structure Selection (Huang & Wang, 2018) also uses sparsified scaling factors to prune
structures, and can be seen as a generalization of Network Slimming. Other than channels, pruning

6
Published as a conference paper at ICLR 2019

can be on residual blocks in ResNet or groups in ResNeXt (Xie et al., 2017). We examine residual
blocks pruning, where ResNet-50 are pruned to be ResNet-41, ResNet-32 and ResNet-26. Table 5
shows our results. On average Scratch-E outperforms pruned models, and for all models Scratch-B
is better than both.
Dataset Model Unpruned Pruned Model Pruned Scratch-E Scratch-B
ResNet-41 75.44 75.61 76.17
ImageNet ResNet-50 76.12 ResNet-32 74.18 73.77 74.67
ResNet-26 71.82 72.55 73.41

Table 5: Results (accuracy) for residual block pruning using Sparse Structure Selection (Huang & Wang, 2018).
In the original paper no fine-tuning is required so there is a “Pruned” column instead of “Fine-tuned” as before.

4.3 U NSTRUCTURED M AGNITUDE - BASED P RUNING (H AN ET AL ., 2015)

Unstructured magnitude-based weight pruning (Han et al., 2015) can also be treated as automatically
discovering architectures, since the positions of exact zeros cannot be determined before training, but
we highlight its differences with structured pruning using another subsection. Because all the net-
work architectures we evaluated are fully-convolutional (except for the last fully-connected layer),
for simplicity, we only prune weights in convolution layers here. Before training the pruned sparse
model from scratch, we re-scale the standard deviation of the Gaussian distribution for weight ini-
tialization, based on how many non-zero weights remain in this layer. This is to keep a constant
scale of backward gradient signal as in (He et al., 2015), which however in our observations does
not bring gains compared with unscaled counterparts.
Dataset Model Unpruned Prune Ratio Fine-tuned Scratch-E Scratch-B
30% 93.51 (±0.05) 93.71 (±0.09) 93.31 (±0.26)
VGG-19 93.50 (±0.11) 80% 93.52 (±0.10) 93.71 (±0.08) 93.64 (±0.09)
95% 93.34 (±0.13) 93.21 (±0.17) 93.63 (±0.18)
30% 95.06 (±0.05) 94.84 (±0.07) 95.11 (±0.09)
CIFAR-10 PreResNet-110 95.04 (±0.15) 80% 94.55 (±0.11) 93.76 (±0.10) 94.52 (±0.13)
95% 92.35 (±0.20) 91.23 (±0.11) 91.55 (±0.34)
30% 95.21 (±0.17) 95.22 (±0.18) 95.23 (±0.14)
DenseNet-BC-100 95.24 (±0.17) 80% 95.04 (±0.15) 94.42 (±0.12) 95.12 (±0.04)
95% 94.19 (±0.15) 92.91 (±0.22) 93.44 (±0.19)
30% 71.96 (±0.36) 72.81 (±0.31) 73.30 (±0.25)
VGG-19 71.70 (±0.31) 50% 71.85 (±0.30) 73.12 (±0.36) 73.77 (±0.23)
95% 70.22 (±0.38) 70.88 (±0.35) 72.08 (±0.15)
30% 76.88 (±0.31) 76.36 (±0.26) 76.96 (±0.31)
CIFAR-100 PreResNet-110 76.96 (±0.34) 50% 76.60 (±0.36) 75.45 (±0.23) 76.42 (±0.39)
95% 68.55 (±0.51) 68.13 (±0.64) 68.99 (±0.32)
30% 77.23 (±0.05) 77.58 (±0.25) 77.97 (±0.31)
DenseNet-BC-100 77.59 (±0.19) 50% 77.41 (±0.14) 77.65 (±0.09) 77.80 (±0.23)
95% 73.67 (±0.03) 71.47 (±0.46) 72.57 (±0.37)
30% 73.68 72.75 74.02
VGG-16 73.37
60% 73.63 71.50 73.42
ImageNet
30% 76.06 74.77 75.70
ResNet-50 76.15
60% 76.09 73.69 74.91

Table 6: Results (accuracy) for unstructured pruning (Han et al., 2015). “Prune Ratio” denotes the percentage
of parameters pruned in the set of all convolutional weights.
As shown in Table 6, on the smaller-scale CIFAR datasets, when the pruned ratio is small (≤ 80%),
Scratch-E sometimes falls short of the fine-tuned results, but Scratch-B is able to perform at least
on par with the latter. However, we observe that in some cases, when the prune ratio is large (95%),
fine-tuning can outperform training from scratch. On the large-scale ImageNet dataset, we note
that the Scratch-B result is mostly worse than fine-tuned result by a noticable margin, despite at
a decent accuracy level. This could be due to the increased difficulty of directly training on the
highly sparse networks (CIFAR), or the scale/complexity of the dataset itself (ImageNet). Another
possible reason is that compared with structured pruning, unstructured pruning significantly changes

7
Published as a conference paper at ICLR 2019

the weight distribution (more details in Appendix G). The difference in scratch-training behaviors
also suggests an important difference between structured and unstructured pruning.

5 N ETWORK P RUNING AS A RCHITECTURE SEARCH

While we have shown that, for structured pruning, the inherited weights in the pruned architecture
are not better than random, the pruned architecture itself turns out to be what brings the efficiency
benefits. In this section, we assess the value of architecture search for automatic network pruning
algorithms (Figure 2) by comparing pruning-obtained models and uniformly pruned models. Note
that the connection between network pruning and architecture learning has also been made in prior
works (Han et al., 2015; Liu et al., 2017; Gordon et al., 2018; Huang et al., 2018), but to our
knowledge we are the first to isolate the effect of inheriting weights and solely compare pruning-
obtained architectures with uniformly pruned ones, by training both of them from scratch.
Parameter Efficiency of Pruned Architectures. In Figure 3(left), we compare the parameter effi-
ciency of architectures obtained by an automatic channel pruning method (Network Slimming (Liu
et al., 2017)), with a naive predefined pruning strategy that uniformly prunes the same percentage of
channels in each layer. All architectures are trained from random initialization for the same number
of epochs. We see that the architectures obtained by Network Slimming are more parameter effi-
cient, as they could achieve the same level of accuracy using 5× fewer parameters than uniformly
pruning architectures. For unstructured magnitude-based pruning (Han et al., 2015), we conducted
a similar experiment shown in Figure 3 (right). Here we uniformly sparsify all individual weights
at a fixed probability, and the architectures obtained this way are much less efficient than the pruned
architectures.

Channel Pruned VGG-16 on CIFAR-10 Weight Sparsified VGG-16 on CIFAR-10

94.0 94.0

93.5
Test Accuracy (%)

Test Accuracy (%)

93.5
93.0
93.0
92.5

92.5 92.0

Network Slimming 91.5 Unstructured Pruning

92.0 Uniform Pruning Uniform Sparsification
91.0
0.2 0.4 0.6 0.8 1.0 1.2 0.2 0.4 0.6 0.8 1.0 1.2
#Parameters ×107 #Parameters ×107

Figure 3: Pruned architectures obtained by different approaches, all trained from scratch, averaged over 5
runs. Architectures obtained by automatic pruning methods (Left: Network Slimming (Liu et al., 2017), Right:
Unstructured pruning (Han et al., 2015)) have better parameter efficiency than uniformly pruning channels or
sparsifying weights in the whole network.

Layer Width Width* Layer Width Width*

1 64 39.0±3.7 8 512 217.3±6.6
2 64 64.0±0.0 9 512 120.0±4.4
3 128 127.8±0.4 10 512 63.0±1.9
4 128 128.0±0.0 11 512 47.8±2.9
Stage 3 Stage 4 Stage 5
5 256 255.0±1.0 12 512 62.0±3.4
6 256 250.5±0.5 13 512 88.8±3.1
7 256 226.0±2.5 Total 4224 1689.2
Figure 4: The average sparsity pattern of all
Table 7: Network architectures obtained by pruning 60% chan- 3×3 convolutional kernels in certain layer
nels on VGG-16 (in total 13 conv-layers) using Network Slim- stages in a unstructured pruned VGG-16.
ming. Width and Width* are number of channels in the original Darker color means higher probability of
and pruned architectures, averaged over 5 runs. weight being kept.

We also found the channel/weight pruned architectures exhibit very consistent patterns (see Table 7
and Figure 4). This suggests the original large models may be redundantly designed for the task and

8
Published as a conference paper at ICLR 2019

the pruning algorithm can help us improve the efficiency. This also confirms the value of automatic
pruning methods for searching efficient models on the architectures evaluated.

Channel Pruned PreResNet-164 on CIFAR-10 Weight Sparsified PreResNet-110 on CIFAR-100 Weight Sparsified DenseNet-40 on CIFAR-100
74
0.950
76 73

Test Accuracy (%)

Test Accuracy (%)
Test Accuracy (%)

0.945
72
0.940 74
71
0.935
72 70
0.930
Network Slimming Unstructured Pruning 69 Unstructured Pruning
0.925 Uniform Pruning 70 Uniform Sparsification Uniform Sparsification
68
0.25 0.50 0.75 1.00 1.25 1.50 0.25 0.50 0.75 1.00 1.25 1.50 0.2 0.4 0.6 0.8 1.0
#Parameters ×106 #Parameters ×106 #Parameters ×106

Figure 5: Pruned architectures obtained by different approaches, all trained from scratch, averaged over 5
runs. Left: Results for PreResNet-164 pruned on CIFAR-10 by Network Slimming (Liu et al., 2017). Middle
and Right: Results for PreResNet-110 and DenseNet-40 pruned on CIFAR-100 by unstructured pruning (Han
et al., 2015).

More Analysis. However, there also exist cases where the architectures obtained by pruning are
not better than uniformly pruned ones. We present such results in Figure 5, where the architectures
obtained by pruning (blue) are not significantly more efficient than uniform pruned architectures
(red). This phenomenon happens more likely on modern architectures like ResNets and DenseNets.
When we investigate the sparsity patterns of those pruned architectures (shown in Table 18, 19 and
20 in Appendix H), we find that they exhibit near-uniform sparsity patterns across stages, which
might be the reason why it can only perform on par with uniform pruning. In contrast, for VGG, the
pruned sparsity patterns can always beat the uniform ones as shown in Figure 3 and Figure 6. We
also show the sparsity patterns of VGG pruned by Network Slimming (Liu et al., 2017) in Table 21
of Appendix H, and they are rather far from uniform. Compared to ResNet and DenseNet, we can
see that VGG’s redundancy is rather imbalanced across layer stages. Network pruning techniques
may help us identify the redundancy better in the such cases.
Generalizable Design Principles from Pruned Architectures. Given that the automatically dis-
covered architectures tend to be parameter efficient on the VGG networks, one may wonder: can
we derive generalizable principles from them on how to design a better architecture? We conduct
several experiments to answer this question.
For Network Slimming, we use the average number of channels in each layer stage (layers with the
same feature map size) from pruned architectures to construct a new set of architectures, and we
call this approach “Guided Pruning”; for magnitude-based pruning, we analyze the sparsity patterns
(Figure 4) in the pruned architectures, and apply them to construct a new set of sparse models,
which we call “Guided Sparsification”. The results are shown in Figure 6. It can be seen that for
both Network Slimming (Figure 6 left) and unstructured pruning (Figure 6 right), guided design of
architectures (green) can perform on par with pruned architectures (blue).
Interestingly, these guided design patterns can sometimes be transferred to a different VGG-variant
and/or dataset. In Figure 6, we distill the patterns of pruned architectures from VGG-16 on CIFAR-
10 and apply them to design efficient VGG-19 on CIFAR-100. These sets of architectures are
denoted as “Transferred Guided Pruning/Sparsification”. We can observe that they (brown) may
sometimes be slightly worse than architectures directly pruned (blue), but are significantly better
than uniform pruning/sparsification (red). In these cases, one does not need to train a large model
to obtain an efficient model as well, as transferred design patterns can help us achieve the efficiency
directly.
Discussions with Conventional Architecture Search Methods. Popular techniques for network
architecture search include reinforcement learning (Zoph & Le, 2017; Baker et al., 2017) and evolu-
tionary algorithms (Xie & Yuille, 2017; Liu et al., 2018a). In each iteration, a randomly initialized
network is trained and evaluated to guide the search, and the search process usually requires thou-
sands of iterations to find the goal architecture. In contrast, using network pruning as architecture
search only requires a one-pass training, however the search space is restricted to the set of all “sub-
networks” inside a large network, whereas traditional methods can search for more variations, e.g.,
activation functions or different layer orders.

9
Published as a conference paper at ICLR 2019

Channel Pruned VGG-19 on CIFAR-100 Weight Sparsified VGG-19 on CIFAR-100

73 74

72
Test Accuracy (%)

Test Accuracy (%)

73
71

70 72

69
Network Slimming 71 Unstructured Pruning
Guided Pruning Guided Sparsification
68
Transferred Guided Pruning Transferred Guided Sparsification
67 Uniform Pruning 70 Uniform Sparsification

0.25 0.50 0.75 1.00 1.25 1.50 0.50 0.75 1.00 1.25 1.50 1.75
#Parameters ×107 #Parameters ×107

Figure 6: Pruned architectures obtained by different approaches, all trained from scratch, averaged over 5 runs.
“Guided Pruning/Sparsification” means using the average sparsity patterns in each layer stage to design the
network; “Transferred Guided Pruning/Sparsification” means using the sparsity patterns obtained by a pruned
VGG-16 on CIFAR-10, to design the network for VGG-19 on CIFAR-100. Following the design guidelines
provided by the pruned architectures, we achieve better parameter efficiency, even when the guidelines are
transferred from another dataset and model.

Recently, Gordon et al. (2018) uses a similar pruning technique to Network Slimming (Liu et al.,
2017) to automate the design of network architectures; He et al. (2018c) prune channels using rein-
forcement learning and automatically compresses the architecture. On the other hand, in the network
architecture search literature, sharing/inheriting trained parameters (Pham et al., 2018; Liu et al.,
2018b) during searching has become a popular approach for reducing the training budgets, but once
the target architecture is found, it is still trained from scratch to maximize the accuracy.

6 D ISCUSSION AND C ONCLUSION

Our results encourage more careful and fair baseline evaluations of structured pruning methods. In
addition to high accuracy, training predefined target models from scratch has the following benefits
over conventional network pruning procedures: a) since the model is smaller, we can train the model
using less GPU memory and possibly faster than training the original large model; b) there is no need
to implement the pruning criterion and procedure, which sometimes requires fine-tuning layer by
layer (Luo et al., 2017) and/or needs to be customized for different network architectures (Li et al.,
2017; Liu et al., 2017); c) we avoid tuning additional hyper-parameters involved in the pruning
procedure.
Our results do support the viewpoint that automatic structured pruning finds efficient architectures in
some cases. However, if the accuracy of pruning and fine-tuning is achievable by training the pruned
model from scratch, it is also important to evaluate the pruned architectures against uniformly pruned
baselines (both training from scratch), to demonstrate the method’s value in identifying efficient
architectures. If the uniformly pruned models are not worse, one could also skip the pipeline and
train them from scratch.
Even if pruning and fine-tuning fails to outperform the mentioned baselines in terms of accuracy,
there are still some cases where using this conventional wisdom can be much faster than training
from scratch: a) when a pre-trained large model is already given and little or no training budget is
available; we also note that pre-trained models can only be used when the method does not require
modifications to the large model training process; b) there is a need to obtain multiple models of
different sizes, or one does not know what the desirable size is, in which situations one can train a
large model and then prune it by different ratios.

R EFERENCES
Jose M Alvarez and Mathieu Salzmann. Learning the number of neurons in deep networks. In NIPS,
2016.

10
Published as a conference paper at ICLR 2019

Sajid Anwar and Wonyong Sung. Compact deep convolutional neural networks with coarse pruning.
arXiv preprint arXiv:1610.09639, 2016.
Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In NIPS, 2014.
Bowen Baker, Otkrist Gupta, Nikhil Naik, and Ramesh Raskar. Designing neural network architec-
tures using reinforcement learning. ICLR, 2017.
Miguel A Carreira-Perpinán and Yerlan Idelbayev. “Learning-compression” algorithms for neural
net pruning. In CVPR, 2018.
Ting-Wu Chin, Cha Zhang, and Diana Marculescu. Layer-compensated pruning for resource-
constrained convolutional neural networks. arXiv preprint arXiv:1810.00518, 2018.
Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized
neural networks: Training deep neural networks with weights and activations constrained to+ 1
or-1. arXiv preprint arXiv:1602.02830, 2016.
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale
hierarchical image database. In CVPR, 2009.
Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear
structure within convolutional networks for efficient evaluation. In NIPS, 2014.
Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural
networks. In International Conference on Learning Representations, 2019. URL https://
openreview.net/forum?id=rJl-b3RcF7.
Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accu-
rate object detection and semantic segmentation. In CVPR, 2014.
Ariel Gordon, Elad Eban, Ofir Nachum, Bo Chen, Hao Wu, Tien-Ju Yang, and Edward Choi. Mor-
phnet: Fast & simple resource-constrained structure learning of deep networks. In CVPR, 2018.
Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for
efficient neural network. In NIPS, 2015.
Song Han, Xingyu Liu, Huizi Mao, Jing Pu, Ardavan Pedram, Mark A Horowitz, and William J
Dally. Eie: efficient inference engine on compressed deep neural network. In Computer Archi-
tecture (ISCA), 2016 ACM/IEEE 43rd Annual International Symposium on, 2016a.
Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks
with pruning, trained quantization and huffman coding. ICLR, 2016b.
Babak Hassibi and David G Stork. Second order derivatives for network pruning: Optimal brain
surgeon. In NIPS, 1993.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing
human-level performance on imagenet classification. In Proceedings of the IEEE international
conference on computer vision, pp. 1026–1034, 2015.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-
nition. In CVPR, 2016.
Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In ICCVs, 2017a.
Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating
deep convolutional neural networks. In IJCAI, 2018a.
Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating
deep convolutional neural networks. https://github.com/he-y/soft-filter-pruning, 2018b.
Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural net-
works. In ICCV, 2017b.

11
Published as a conference paper at ICLR 2019

Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, and Song Han. Amc: Automl for model
compression and acceleration on mobile devices. In ECCV, 2018c.
Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. NIPS
Workshop, 2014.
Hengyuan Hu, Rui Peng, Yu-Wing Tai, and Chi-Keung Tang. Network trimming: A data-driven
neuron pruning approach towards efficient deep architectures. arXiv preprint arXiv:1607.03250,
2016.
Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q Weinberger. Densely connected
convolutional networks. In CVPR, 2017.
Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An
efficient densenet using learned group convolutions. In CVPR, 2018.
Zehao Huang and Naiyan Wang. Data-driven sparse structure selection for deep neural networks.
ECCV, 2018.
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Ser-
gio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embed-
ding. In ACM Multimedia, 2014.
Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009.
Vadim Lebedev and Victor Lempitsky. Fast convnets using group-wise brain damage. In CVPR,
2016.
Vadim Lebedev, Yaroslav Ganin, Maksim Rakhuba, Ivan Oseledets, and Victor Lempitsky.
Speeding-up convolutional neural networks using fine-tuned cp-decomposition. ICLR, 2014.
Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. In NIPS, 1990.
Yann LeCun, Léon Bottou, Yoshua Bengio, Patrick Haffner, et al. Gradient-based learning applied
to document recognition. Proceedings of the IEEE, 1998.
Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for
efficient convnets. In ICLR, 2017.
Ji Lin, Yongming Rao, Jiwen Lu, and Jie Zhou. Runtime neural pruning. In NIPS, 2017.
Hanxiao Liu, Karen Simonyan, Oriol Vinyals, Chrisantha Fernando, and Koray Kavukcuoglu. Hi-
erarchical representations for efficient architecture search. ICLR, 2018a.
Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv
preprint arXiv:1806.09055, 2018b.
Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learn-
ing efficient convolutional networks through network slimming. In ICCV, 2017.
Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic
segmentation. In CVPR, 2015.
Christos Louizos, Max Welling, and Diederik P Kingma. Learning sparse neural networks through
l_0 regularization. ICLR, 2018.
Jian-Hao Luo, Jianxin Wu, and Weiyao Lin. Thinet: A filter level pruning method for deep neural
network compression. In ICCV, 2017.
Deepak Mittal, Shweta Bhardwaj, Mitesh M Khapra, and Balaraman Ravindran. Recovering
from random pruning: On the plasticity of deep convolutional neural networks. arXiv preprint
arXiv:1801.10447, 2018.

12
Published as a conference paper at ICLR 2019

Dmitry Molchanov, Arsenii Ashukha, and Dmitry Vetrov. Variational dropout sparsifies deep neural
networks. In ICML, 2017.
Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional
neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440, 2016.
Diederik P. Kingma, Tim Salimans, and Max Welling. Variational dropout and the local reparame-
terization trick. In NIPS, 2015.
Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito,
Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in
pytorch. In NIPS Workshop, 2017.
Hieu Pham, Melody Y Guan, Barret Zoph, Quoc V Le, and Jeff Dean. Efficient neural architecture
search via parameter sharing. ICML, 2018.
Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet
classification using binary convolutional neural networks. In ECCV, 2016.
Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object
detection with region proposal networks. In NIPS, 2015.
Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and
Yoshua Bengio. Fitnets: Hints for thin deep nets. ICLR, 2015.
Zhiqiang Shen, Zhuang Liu, Jianguo Li, Yu-Gang Jiang, Yurong Chen, and Xiangyang Xue. Dsod:
Learning deeply supervised object detectors from scratch. In ICCV, 2017.
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image
recognition. ICLR, 2015.
Suraj Srinivas and R Venkatesh Babu. Data-free parameter pruning for deep neural networks.
BMVC, 2015.
Xavier Suau, Luca Zappella, Vinay Palakkode, and Nicholas Apostoloff. Principal filter analysis for
guided network compression. arXiv preprint arXiv:1807.10585, 2018.
Xin Wang, Fisher Yu, Zi-Yi Dou, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in
convolutional networks. arXiv preprint arXiv:1711.09485, 2017.
Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in
deep neural networks. In NIPS, 2016.
Lingxi Xie and Alan L Yuille. Genetic cnn. In ICCV, 2017.
Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual trans-
formations for deep neural networks. In CVPR, 2017.
Jianwei Yang, Jiasen Lu, Dhruv Batra, and Devi Parikh. A faster pytorch implementation of faster
r-cnn. https://github.com/jwyang/faster-rcnn.pytorch, 2017.
Jianbo Ye, Xin Lu, Zhe Lin, and James Z Wang. Rethinking the smaller-norm-less-informative
assumption in channel pruning of convolution layers. ICLR, 2018.
Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-
Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation.
In CVPR, 2018.
Hao Zhou, Jose M Alvarez, and Fatih Porikli. Less is more: Towards compact cnns. In ECCV, 2016.
Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for
model compression. ICLR Workshop, 2018.
Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. ICLR, 2017.

13
Published as a conference paper at ICLR 2019

A PPENDIX
A E XPERIMENTS ON THE L OTTERY T ICKET H YPOTHESIS (F RANKLE &
C ARBIN , 2019)
The Lottery Ticket Hypothesis (Frankle & Carbin, 2019) conjectures that inside the large network, a
sub-network together with their initialization makes the training particularly effective, and together
they are termed the “winning ticket”. In this hypothesis, the original initialization of the sub-network
(before large model training) is needed for it to achieve competitive performance when trained in
isolation. Their experiments show that training the sub-network with randomly re-initialized weights
performs worse than training it with the original initialization inside the large network. In contrast,
our work does not require reuse of the original initialization of the pruned model, and shows that
random initialization is enough for the pruned model to achieve competitive performance.
The conclusions seem to be contradictory, but there are several important differences in the evalu-
ation settings: a) Our main conclusion is drawn on structured pruning methods, despite for small-
scale problems (CIFAR) it also holds on unstructured pruning; Frankle & Carbin (2019) only evalu-
ates on unstructured pruning. b) Our evaluated network architectures are all relatively large modern
models used in the original pruning methods, while most of the experiments in Frankle & Carbin
(2019) use small shallow networks (< 6 layers). c) We use momentum SGD with a large initial
learning rate (0.1), which is widely used in prior image classification (He et al., 2016; Huang et al.,
2017) and pruning works (Li et al., 2017; Liu et al., 2017; He et al., 2017b; Luo et al., 2017; He
et al., 2018a; Huang & Wang, 2018) to achieve high accuracy, and is the de facto default optimiza-
tion setting on CIFAR and ImageNet; while Frankle & Carbin (2019) mostly uses Adam with much
lower learning rates. d) Our experiments include the large-scale ImageNet dataset, while Frankle &
Carbin (2019) only considers MNIST and CIFAR.

VGG-16 ResNet-50
94.0 94

93.5 93
Test Accuracy (%)
Test Accuracy (%)

93.0 92

91
92.5
90
92.0 Winning Ticket lr=0.1 Winning Ticket lr=0.1
Random Init lr=0.1 Random Init lr=0.1
89
91.5 Winning Ticket lr=0.01 Winning Ticket lr=0.01
Random Init lr=0.01 88 Random Init lr=0.01
91.0
20 40 60 80 20 40 60 80
%Prune ratio %Prune ratio
(a) Iterative Pruning
VGG-16 ResNet-50
94.0
93
93.5
Test Accuracy (%)
Test Accuracy (%)

92
93.0
91
92.5
90

92.0 Winning Ticket lr=0.1 89 Winning Ticket lr=0.1

Random Init lr=0.1 Random Init lr=0.1
91.5 Winning Ticket lr=0.01 88 Winning Ticket lr=0.01
Random Init lr=0.01 Random Init lr=0.01
91.0 87
20 40 60 80 20 40 60 80
%Prune ratio %Prune ratio
(b) One-shot Pruning
Figure 7: Comparisons with the Lottery Ticket Hypothesis (Frankle & Carbin, 2019) for iterative/one-shot
unstructured pruning (Han et al., 2015) with two initial learning rates 0.1 and 0.01, on CIFAR-10 dataset. Each
point is averaged over 5 runs. Using the winning ticket as initialization only brings improvement when the
learning rate is small (0.01), however such small learning rate leads to a lower accuracy than the widely used
large learning rate (0.1).

14
Published as a conference paper at ICLR 2019

In this section, we show that the difference in learning rate is what causes the seemingly contradict-
ing behaviors between our work and Frankle & Carbin (2019), in the case of unstructured pruning on
CIFAR. For structured pruning, when using both large and small learning rates, the winning ticket
does not outperform random initialization.
We test the Lottery Ticket Hypothesis by comparing the models trained with original initialization
(“winning ticket”) and that trained from randomly re-initialized weights. We experiment with two
choices of initial learning rate (0.1 and 0.01) with a stepwise decay schedule, using momentum SGD.
0.1 is used in our previous experiments and most prior works on CIFAR and ImageNet. Following
Frankle & Carbin (2019), we investigate both iterative pruning (prune 20% in each iteration) and
one-shot pruning for unstructured pruning. We show our results for unstructured pruning (Han et al.,
2015) in Figure 7 and Table 8, and L1 -norm based filter pruning (Li et al., 2017) in Table 9.

Dataset Model Unpruned

Prune Ratio Winning Ticket Random Init
20% 93.66 (±0.20) 93.79 (±0.11)
40% 93.79 (±0.12) 93.77 (±0.10)
VGG-16 93.76 (±0.20) 60% 93.60 (±0.13) 93.72 (±0.11)
80% 93.74 (±0.15) 93.72 (±0.16)
95% 93.18 (±0.12) 93.05 (±0.21)
CIFAR-10
20% 93.38 (±0.18) 93.31 (±0.24)
40% 92.94 (±0.12) 93.06 (±0.22)
ResNet-50 93.48 (±0.20) 60% 92.56 (±0.20) 92.69 (±0.11)
80% 91.83 (±0.20) 91.69 (±0.21)
95% 88.75 (±0.18) 88.59 (±0.09)
(a) One-shot pruning with initial learning rate 0.1

Dataset Model Unpruned

Prune Ratio Winning Ticket Random Init
20% 92.78 (±0.11) 92.52 (±0.15)
40% 92.80 (±0.18) 92.52 (±0.15)
VGG-16 92.69 (±0.12) 60% 92.72 (±0.16) 92.44 (±0.19)
80% 92.75 (±0.07) 92.07 (±0.25)
95% 92.58 (±0.25) 91.83 (±0.11)
CIFAR-10
20% 91.28 (±0.15) 90.93 (±0.34)
40% 91.16 (±0.07) 90.92 (±0.10)
ResNet-50 91.06 (±0.28) 60% 91.00 (±0.15) 90.43 (±0.16)
80% 90.92 (±0.08) 89.71 (±0.18)
95% 87.76 (±0.19) 87.20 (±0.17)
(b) One-shot pruning with initial learning rate 0.01

Table 8: Comparisons with the Lottery Ticket Hypothesis (Frankle & Carbin, 2019) for one-shot unstructured
pruning (Han et al., 2015) with two initial learning rates: 0.1 and 0.01. The same results are visualized in
Figure 7b. Using the winning ticket as initialization only brings improvement when the learning rate is small
(0.01), however such small learning rate leads to a lower accuracy than the widely used large learning rate (0.1).

From Figure 7 and Table 8, we see that for unstructured pruning, using the original initialization as
in (Frankle & Carbin, 2019) only provides advantage over random initialization with small initial
learning rate 0.01. For structured pruning as Li et al. (2017), it can be seen from Table 9 that using
the original initialization is only on par with random initialization for both large and small initial
learning rates. In both cases, we can see that the small learning rate gives lower accuracy than
the widely-used large learning rate. To summarize, in our evaluated settings, the winning ticket
only brings improvement in the case of unstructured pruning, with small initial learning rate, but
this small learning rate yields inferior accuracy compared with the widely-used large learning rate.
Note that Frankle & Carbin (2019) also report in their Section 5, that the winning ticket cannot be
found on ResNet-18/VGG using the large learning rate. The reason why the original initialization is
helpful when the learning rate is small, might be the weights of the final trained model are not too
far from the original initialization due to the small parameter updating stepsize.

15
Published as a conference paper at ICLR 2019

Dataset Model Unpruned Pruned Model Winning Ticket Random Init

VGG-1693.63 (±0.16) VGG-16-A 93.62 (±0.09) 93.60 (±0.15)
ResNet-56-A 92.72 (±0.10) 92.75 (±0.26)
ResNet-56 93.14 (±0.12)
CIFAR-10 ResNet-56-B 92.78 (±0.23) 92.90 (±0.27)
ResNet-110-A 93.21 (±0.09) 93.21 (±0.21)
ResNet-110 93.14 (±0.24)
ResNet-110-B 93.15 (±0.12) 93.37 (±0.29)

(a) Initial learning rate 0.1

Dataset Model Unpruned Pruned Model Winning Ticket Random Init

VGG-1692.64 (±0.05) VGG-16-A 92.65 (±0.18) 92.67 (±0.22)
ResNet-56-A 90.00 (±0.15) 89.87 (±0.25)
ResNet-56 89.81 (±0.27)
CIFAR-10 ResNet-56-B 89.75 (±0.35) 89.81 (±0.24)
ResNet-110-A 89.48 (±0.35) 89.49 (±0.10)
ResNet-110 89.43 (±0.39)
ResNet-110-B 89.36 (±0.30) 89.35 (±0.16)

(b) Initial learning rate 0.01

Table 9: Experiments on the Lottery Ticket Hypothesis (Frankle & Carbin, 2019) on a structured pruning
method (L1 -norm based filter pruning (Li et al., 2017)) with two initial learning rates: 0.1 and 0.01. In both
cases, using winning tickets does not bring improvement on accuracy.

B R ESULTS ON S OFT F ILTER P RUNING (H E ET AL ., 2018 A )

Soft Filter Pruning (SFP) (He et al., 2018a) prunes filters every epoch during training but also
keeps updating the pruned filters, i.e., the pruned weights have the chance to be recovered. In the
original paper, SFP can either run upon a random initialized model or a pretrained model. It falls into
the category of predefined methods (Figure 2). Table 10 shows our results without using pretrained
models and Table 11 shows the results with a pretrained model. We use authors’ code (He et al.,
2018b) for obtaining the results. It can be seen that Scratch-E outperforms pruned models for most
of the time and Scratch-B outperforms pruned models in nearly all cases. Therefore, our conclusion
also holds on this method.

Dataset Model Unpruned Prune Ratio Pruned Scratch-E Scratch-B

10% 92.00 (±0.32) 92.22 (±0.15) 92.13 (±0.10)
ResNet-20 92.41 (±0.12) 20% 91.50 (±0.30) 91.62 (±0.12) 91.67 (±0.15)
30% 90.78 (±0.15) 90.93 (±0.10) 91.07 (±0.23)
10% 93.28 (±0.05) 93.42 (±0.40) 93.08 (±0.13)
ResNet-32 93.22 (±0.16) 20% 92.50 (±0.17) 92.68 (±0.20) 92.96 (±0.11)
30% 92.02 (±0.11) 92.37 (±0.12) 92.56 (±0.06)
CIFAR-10 10% 93.77 (±0.07) 93.42 (±0.40) 93.98 (±0.21)
20% 93.14 (±0.42) 93.44 (±0.05) 93.71 (±0.14)
ResNet-56 93.80 (±0.12)
30% 93.01 (±0.09) 93.19 (±0.20) 93.57 (±0.12)
40% 92.59 (±0.14) 92.80 (±0.25) 93.07 (±0.25)
10% 93.60 (±0.50) 94.21 (±0.39) 94.13 (±0.37)
ResNet-110 93.77 (±0.23) 20% 93.63 (±0.44) 93.52 (±0.18) 94.29 (±0.18)
30% 93.26 (±0.37) 93.70 (±0.16) 93.92 (±0.13)
ResNet-34 73.92 30% 71.83 71.67 72.97
ImageNet
ResNet-50 76.15 30% 74.61 74.98 75.56

Table 10: Results (accuracy) for Soft Filter Pruning (He et al., 2018a) without pretrained models.

16
Published as a conference paper at ICLR 2019

Dataset Model Unpruned

Prune Ratio Pruned Scratch-E Scratch-B
30% 93.51 (±0.26) 94.45 (±0.30) 93.77 (±0.25)
ResNet-56 93.80 (±0.12)
CIFAR-10 40% 93.10 (±0.34) 93.84 (±0.16) 93.41 (±0.08)
ResNet-110 93.77 (±0.23) 30% 93.46 (±0.19) 93.89 (±0.17) 94.37 (±0.24)

Table 11: Results (accuracy) for Soft Filter Pruning (He et al., 2018a) using pretrained models.

C T RANSFER L EARNING TO O BJECT D ETECTION

We have shown that for structured pruning the small pruned model can be trained from scratch to
match the accuracy of the fine-tuned model in classification tasks. To see whether this phenomenon
would also hold for transfer learning to other vision tasks, we evaluate the L1 -norm based filter
pruning method (Li et al., 2017) on the PASCAL VOC object detection task, using Faster-RCNN
(Ren et al., 2015).
Object detection frameworks usually require transferring model weights pre-trained on ImageNet
classification, and one can perform pruning either before or after the weight transfer. More specifi-
cally, the former could be described as “train on classification, prune on classification, fine-tune on
classification, transfer to detection”, while the latter is “train on classification, transfer to detection,
prune on detection, fine-tune on detection”. We call these two approaches Prune-C (classification)
and Prune-D (detection) respectively, and report the results in Table 12. With a slight abuse of nota-
tion, here Scratch-E/B denotes "train the small model on classification, transfer to detection", and is
different from the setup of detection without ImageNet pre-training as in Shen et al. (2017).

Dataset Model
Unpruned Pruned Model Prune-C Prune-D Scratch-E Scratch-B
ResNet34-A 71.47 70.99 71.64 71.90
PASCAL VOC 07 ResNet-34 71.69
ResNet34-B 70.84 69.62 71.68 71.26

Table 12: Results (mAP) for pruning on detection task. The pruned models are chosen from Li et al. (2017).
Prune-C refers to pruning on classifcation pre-trained weights, Prune-D refers to pruning after the weights are
transferred to detection task. Scratch-E/B means pre-training the pruned model from scratch on classification
and transfer to detection.
For this experiment, we adopt the code and default hyper-parameters from Yang et al. (2017), and
use PASCAL VOC 07 trainval/test set as our training/test set. For backbone networks, we evaluate
ResNet-34-A and ResNet-34-B from the L1 -norm based filter pruning (Li et al., 2017), which are
pruned from ResNet-34. Table 12 shows our result, and we can see that the model trained from
scratch can surpass the performance of fine-tuned models under the transfer setting.
Another interesting observation from Table 12 is that Prune-C is able to outperform Prune-D, which
is surprising since if our goal task is detection, directly pruning away weights that are considered
unimportant for detection should presumably be better than pruning on the pre-trained classification
models. We hypothesize that this might be because pruning early in the classification stage makes
the final model less prone to being trapped in a bad local minimum caused by inheriting weights
from the large model. This is in line with our observation that Scratch-E/B, which trains the small
models from scratch starting even earlier at the classification stage, can achieve further performance
improvement.

D AGGRESSIVELY P RUNED M ODELS

It would be interesting to see whether our observation still holds if the model is very aggressively
pruned, since they might not have enough capacity to be trained from scratch and achieve decent
accuracy. Here we provide results using Network Slimming (Liu et al., 2017) and L1 -norm based
filter pruning (Li et al., 2017). From Table 13, Table 14 and Table 15, it can be seen that when the
prune ratio is large, training from scratch is better than fine-tuned models by an even larger margin,
and in this case fine-tuned models are significantly worse than the unpruned models. Note that in
Table 2, the VGG-Tiny model we evaluated for ThiNet (Luo et al., 2017) is also a very aggressively
pruned model (FLOPs reduced by 15× and parameters reduced by 100×).

17
Published as a conference paper at ICLR 2019

Dataset Model Unpruned Prune Ratio Fine-tuned Scratch-E Scratch-B

PreResNet-56
93.69 (±0.07) 80% 74.66 (±0.96) 88.25 (±0.38) 88.65 (±0.32)
80% 91.76 (±0.38) 93.21 (±0.17) 93.49 (±0.20)
CIFAR-10 PreResNet-164 95.04 (±0.16)
90% 82.06 (±0.92) 87.55 (±0.68) 88.44 (±0.19)
DenseNet-40 94.10 (±0.12) 80% 92.64 (±0.12) 93.07 (±0.08) 93.61 (±0.12)
CIFAR-100 DenseNet-40 73.82 (±0.34) 80% 69.60 (±0.22) 71.04 (±0.36) 71.45 (±0.30)

Table 13: Results (accuracy) for Network Slimming (Liu et al., 2017) when the models are aggressively pruned.
“Prune ratio” stands for total percentage of channels that are pruned in the whole network. Larger ratios are
used than the original paper of Liu et al. (2017).

Dataset Model Unpruned Prune Ratio Fine-tuned Scratch-E Scratch-B

CIFAR-10 ResNet-56 93.14 (±0.12) 90% 89.17 (±0.08) 91.02 (±0.12) 91.93 (±0.26)
Table 14: Results (accuracy) for L1 -norm based filter pruning (Li et al., 2017) when the models are aggressively
pruned.

Dataset Model Unpruned Prune Ratio Fine-tuned Scratch-E Scratch-B

CIFAR-10 VGG-19 93.50 (±0.11) 95% 93.34 (±0.13) 93.21 (±0.17) 93.63 (±0.18)
CIFAR-100 VGG-19 71.70 (±0.31) 95% 70.22 (±0.38) 70.88 (±0.35) 72.08 (±0.15)
Table 15: Results (accuracy) for unstructured pruning (Han et al., 2015) when the models are aggressively
pruned.

E E XTENDING F INE - TUNING E POCHS

Generally, pruning algorithms use fewer epochs for fine-tuning than training the large model (Li
et al., 2017; He et al., 2017b; Luo et al., 2017). For example, L1 -norm based filter pruning (Li
et al., 2017) uses 164 epochs for training on CIFAR-10 datasets, and only fine-tunes the pruned
networks for 40 epochs. This is due to that mostly small learning rates are used for fine-tuning to
better preserve the weights from the large model. Here we experiment with fine-tuning for more
epochs (e.g., for the same number of epochs as Scratch-E) and show it does not bring noticeable
performance improvement.

Dataset Model Pruned Model Fine-tune-40 Fine-tune-80 Fine-tune-160 Scratch-E

VGG-16 VGG-16-A 93.40 (±0.12) 93.45 (±0.06) 93.45 (±0.08) 93.62 (±0.11)
ResNet-56-A 92.97 (±0.17) 92.92 (±0.15) 92.94 (±0.16) 92.96 (±0.26)
ResNet-56
CIFAR-10 ResNet-56-B 92.68 (±0.19) 92.67 (±0.14) 92.76 (±0.16) 92.54 (±0.19)
ResNet-110-A 93.14 (±0.16) 93.12 (±0.19) 93.04 (±0.22) 93.25 (±0.29)
ResNet-110
ResNet-110-B 92.69 (±0.09) 92.75 (±0.15) 92.76 (±0.16) 92.89 (±0.43)
Table 16: “Fine-tune-40” stands for fine-tuning 40 epochs and so on. Scratch-E models are trained for 160
epochs. We observe that fine-tuning for more epochs does not help improve the accuracy much, and models
trained from scratch can still perform on par with fine-tuned models.

We use L1 -norm filter pruning (Li et al., 2017) for this experiment. Table 16 shows our results with
different number of epochs for fine-tuning. It can be seen that fine-tuning for more epochs gives
negligible accuracy increase and sometimes small decrease, and Scratch-E models are still on par
with models fine-tuned for enough epochs.

F E XTENDING THE STANDARD TRAINING SCHEDULE

In our experiments, we use the standard training schedule for both CIFAR (160 epochs) and Im-
ageNet (90 epochs). Here we show that our observation still holds after we extend the standard
training schedule. We use L1 -norm based filter pruning (Li et al., 2017) for this experiment. Ta-
ble 17 shows our results when we extend the standard training schedule of CIFAR from 160 to 300
epochs. We observe that scratch trained models still perform at least on par with fine-tuned models.

18
Published as a conference paper at ICLR 2019

Dataset Model Unpruned Pruned Model Fine-tuned Scratch-E Scratch-B

VGG-16 93.79 (±0.05) VGG-16-A 93.67 (±0.11) 93.74 (±0.14) 93.80 (±0.09)
ResNet-56-A 93.44 (±0.15) 93.34 (±0.17) 93.56 (±0.09)
ResNet-56 93.52 (±0.05)
CIFAR-10 ResNet-56-B 93.12 (±0.20) 93.14 (±0.21) 93.30 (±0.17)
ResNet-110-A 93.75 (±0.24) 93.80 (±0.15) 94.10 (±0.12)
ResNet-110 93.82 (±0.32)
ResNet-110-B 93.36 (±0.28) 93.75 (±0.16) 93.90 (±0.17)
Table 17: Results for L1 -norm filter pruning (Li et al., 2017) when the training schedule of the large model is
extended from 160 to 300 epochs.

G W EIGHT D ISTRIBUTIONS

Structured Pruning
Unpruned Fine-tuned Scratch-trained
30000 30000 30000

25000 25000 25000

20000 20000 20000

Count

15000 15000 15000

10000 10000 10000

5000 5000 5000

0 0 0
−0.2 0.0 0.2 −0.2 0.0 0.2 −0.2 0.0 0.2
Weight value

Unstructured Pruning
Unpruned Fine-tuned Scratch-trained
30000 30000 30000

25000 25000 25000

20000 20000 20000

Count

15000 15000 15000

10000 10000 10000

5000 5000 5000

0 0 0
−0.2 0.0 0.2 −0.2 0.0 0.2 −0.2 0.0 0.2
Weight value

Figure 8: Weight distribution of convolutional layers for different pruning methods. We use VGG-16 and
CIFAR-10 for this visualization. We compare the weight distribution of unpruned models, fine-tuned mod-
els and scratch-trained models. Top: Results for Network Slimming (Liu et al., 2017). Bottom: Results for
unstructured pruning (Han et al., 2015).

Accompanying the discussion in subsection 4.3, we show the weight distribution of unpruned large
models, fine-tuned pruned models and scratch-trained pruned models, for two pruning methods:
(structured) Network Slimming (Liu et al., 2017) and unstructured pruning (Han et al., 2015). We
choose VGG-16 and CIFAR-10 for visualization and compare the weight distribution of unpruned
models, fine-tuned models and scratch-trained models. For Network Slimming, the prune ratio is
50%. For unstructured pruning, the prune ratio is 80%. Figure 8 shows our result. We can see that
the weight distribution of fine-tuned models and scratch-trained pruned models are different from
the unpruned large models – the weights that are close to zero are much fewer. This seems to imply
that there are less redundant structures in the found pruned architecture, and support the view of
architecture learning for automatic pruning methods.
For unstructured pruning, the fine-tuned model also has significantly different weight distribution
from the scratch-trained model – it has nearly no close-to-zero values. This might be a potential
reason why sometimes models trained from scratch cannot achieve the accuracy of the fine-tuned
models, as shown in Table 6.

19
Published as a conference paper at ICLR 2019

H M ORE S PARSITY PATTERNS FOR P RUNED A RCHITECTURES

In this section we provide the additional results on sparsity patterns for the pruned models, accom-
panying the discussions of “More Analysis” in Section 5.
10% 20% 30% 40% 50% 60% 70%
Stage 1 0.879 0.729 0.557 0.484 0.421 0.349 0.271
Stage 2 0.959 0.863 0.754 0.651 0.537 0.428 0.320
Stage 3 0.889 0.798 0.716 0.610 0.507 0.403 0.301
Table 18: Sparsity patterns of PreResNet-164 pruned on CIFAR-10 by Network Slimming shown in Figure 5
(left) under different prune ratio. The top row denotes the total prune ratio. The values denote the ratio of
channels to be kept. We can observe that for a certain prune ratio, the sparsity patterns are close to uniform
(across stages).

10% 50% 90%

0.905 0.905 0.909 0.530 0.561 0.538 0.129 0.171 0.133
Stage 1 0.900 0.912 0.899 0.559 0.588 0.551 0.166 0.217 0.176
0.903 0.913 0.902 0.532 0.563 0.547 0.142 0.172 0.163
0.906 0.911 0.906 0.485 0.523 0.503 0.073 0.102 0.085
Stage 2 0.912 0.911 0.915 0.508 0.529 0.525 0.099 0.114 0.111
0.911 0.916 0.912 0.502 0.529 0.519 0.080 0.113 0.096
0.901 0.904 0.900 0.454 0.475 0.454 0.043 0.059 0.048
Stage 3 0.885 0.891 0.889 0.409 0.420 0.415 0.032 0.033 0.035
0.898 0.903 0.902 0.450 0.468 0.458 0.042 0.055 0.046
Table 19: Average sparsity patterns of 3×3 kernels of PreResNet-110 pruned on CIFAR-100 by unstructured
pruning shown in Figure 5 (middle) under different prune ratio. The top row denotes the total prune ratio. The
values denote the ratio of weights to be kept. We can observe that for a certain prune ratio, the sparsity patterns
are close to uniform (across stages).

10% 50% 90%

0.861 0.856 0.858 0.507 0.495 0.510 0.145 0.129 0.142
Stage 1 0.843 0.844 0.851 0.484 0.486 0.479 0.123 0.115 0.126
0.850 0.854 0.857 0.509 0.490 0.511 0.136 0.131 0.147
0.907 0.905 0.906 0.498 0.487 0.499 0.099 0.088 0.100
Stage 2 0.892 0.888 0.892 0.442 0.427 0.444 0.064 0.043 0.065
0.907 0.906 0.905 0.497 0.485 0.493 0.095 0.082 0.098
0.897 0.901 0.899 0.470 0.475 0.472 0.060 0.060 0.064
Stage 3 0.888 0.890 0.889 0.433 0.437 0.435 0.040 0.040 0.042
0.898 0.900 0.899 0.473 0.477 0.473 0.060 0.061 0.063
Table 20: Average sparsity patterns of 3×3 kernels of DenseNet-40 pruned on CIFAR-100 by unstructured
pruning shown in Figure 5 (right) under different prune ratio. The top row denotes the total prune ratio. The
values denote the ratio of weights to be kept. We can observe that for a certain prune ratio, the sparsity patterns
are close to uniform (across stages).

20
Published as a conference paper at ICLR 2019

10% 20% 30% 40% 50% 60%

Stage 1 0.969 0.914 0.883 0.875 0.844 0.836
Stage 2 1.000 1.000 1.000 1.000 1.000 1.000
Stage 3 0.991 0.975 0.966 0.957 0.947 0.947
Stage 4 0.861 0.718 0.575 0.446 0.312 0.258
Stage 5 0.871 0.751 0.626 0.486 0.352 0.132
Table 21: Sparsity patterns of VGG-16 pruned on CIFAR-10 by Network Slimming shown in Figure 3 (left)
under different prune ratio. The top row denotes the total prune ratio. The values denote the ratio of channels
to be kept. For each prune ratio, the latter stages tend to have more redundancy than earlier stages.

21