Skip to content

AdamW for a ResNet56v2 – II – Adam with weight decay vs. AdamW, linear LR-schedule and L2-regularization

This series is about a ResNetv56v2 tested on the CIFAR10 dataset. In the last post

AdamW for a ResNet56v2 – I – a detailed look at results based on the Adam optimizer

we investigated a piecewise constant reduction schedule for the Learning Rate [LR] over 200 epochs. We found that we could reproduce results of R. Atienza, who had claimed validation accuracy values of

  • 0.92 < val_acc < 0.93

with the Adam optimizer. We saw a dependency on the batch size [BS] – and concluded that BS=64 was a good compromise regarding achievable accuracy and runtime.

But we needed typically more than 85 epochs to get safely over an accuracy level of val_acc=0.92. Our objective is to reduce the number of epochs as much as possible to reach this accuracy level. In the present post we try to bring the required number of epochs systematically down to 50 epochs by four different measures:

  1. Target 50 epochs: Reduction of the number of epochs by shorter phases of a piecewise constant LR schedule and using the Adam optimizer
  2. Target 50 epochs: Application of a piecewise linear LR schedule and the Adam optimizer without weight decay
  3. Target 50 epochs: Application of a piecewise linear LR schedule and the Adam optimizer with weight decay
  4. Target 50 epochs: Application of the AdamW optimizer – but keeping L2 regularization active

Three additional approaches

  • Target 50 epochs: AdamW optimizer – without L2-regularization, but with larger values for the weight decay
  • Target 50 epochs: Cosine Annealing schedule and changing the Batch Size [BS] in addition
  • Target 30 epochs: Application of a 1Cycle LR schedule

will be topics of forthcoming posts.

All test runs in this post were done with BS=64 and image augmentation (with shift=0.07), as described in the first post of this series. I did not use the ReduceLROnPlateau() callback which Atienza applied in his experiments. The value for the accuracy reached on the validation set of 10,000 CIFAR10 samples is abbreviated by val_acc.

Layer structure of the ResNet56v2

My ResNet generation program was parameterized such that it created the same layer structure for a ResNet56v2 as the original code of Atienza (see [1]; ). I have checked the layer structure and number of parameters thoroughly against what the code of Atienza produced.


 Total params: 1,673,738 (6.38 MB)
 Trainable params: 1,663,338 (6.35 MB)
 Non-trainable params: 10,400 (40.62 KB)


I have consistently used this type of ResNet56v2 model throughout the test runs described below (with BS=64).

Trial 1: Shortening the first phase of a piecewise constant LR schedule for optimizer Adam

A first simple approach to achieve a shorter training is to reduce the number of epochs during which we keep the first learning rate value LR = 1.e-3 constant. In a first experiment we reduce this phase to 60 epochs. The data of the runs in the previous post let us hope that we get a fast convergence to a good accuracy value between epochs 60 and 70. And indeed:


782/782 ━━━━━━━━━━━━━━━━━━━━ 0s 27ms/step - acc: 0.9634 - loss: 0.3070
Epoch 64: val_acc improved from 0.91970 to 0.92230, saving model to /.../cifar10_ResNet56v2_model.064.keras
782/782 ━━━━━━━━━━━━━━━━━━━━ 24s 30ms/step - acc: 0.9634 - loss: 0.3070 - val_acc: 0.9223 - val_loss: 0.4433 - learning_rate: 1.0000e-04

The solid green line defines the value 0.92, the dotted green line 0.93. The model obviously passed an accuracy value of val_acc=0.922 at epoch 64 and reached a maximum value val_acc=0.9254 at epoch 83. So, we have shortened the number of required epochs to reach the 0.92-level already by around 20 epochs. Without doing much.

However, when we go down to a piecewise constant schedule like this one: {(epoch, LR)} = {(40, 1.e-3), (60, 1.e-4)} we fail and reach

  • epoch 48: val_acc=0.914, epoch 64: val_acc=0.9166,

only. It appears that we have reduced the LR too early.

Adam with weight decay?
The question is whether we can improve the convergence by using Adam with weight decay. And actually, a weight decay factor of

  • wd = 5.e-4 and an ema_momentum = 0.8

helped. Below I specify the used piecewise constant scheme by tuples (epoch, LR). The first element gives a target epoch, the second a learning rate value.

  • LR-schedule: (40, LR=1.e-3), (50, LR=1.e-4), (60, LR=1.e-5)
  • val_acc: epoch 51: val_acc=0.92160 => epoch 54: val_acc=0.92270.

A small adjustment of the LR schedule gave us the following result:

  • LR-schedule: (38, LR=1.e-3), (48, LR=1.e-4), (55, LR=1.e-5), (60, LR=5.e-6)
  • val_acc: epoch 49: val_acc=0.92260 => epoch 53: val_acc=0.92450

For the latter run the following plots show the LR schedule and the evolution of the accuracy, loss, validation accuracy and validation loss:

This success by very simple means proves that we should care about Adam in combination with weight decay. But regarding the piecewise constant scheme we should not reduce the first two phases much more.

By the way: Note the strong variation of val_loss during the first phase in the above experiments. We have discussed this point already in the last post.

Trial 2: Runs with with Adam, L2 regularization, piecewise linear LR schedule, no weight decay

With the next experiments I wanted to reach a level of val_acc=0.92 with a piecewise linear schedule like the following one:

  • LR-schedule: {(20, LR=8.e-4), (40, LR=8.e-5), (50, LR=2.e-5), (60, LR=5.e-6)}

I applied the pure Adam optimizer (as implemented in Keras) with default values and no weight decay.

To make a longer story short: I needed a relatively strong L2 contribution. Otherwise I did not reach my objective within 50 epochs, but only came close to the threshold of 0.92 (val_acc=0.916 to 0.918). But not above it …

#epochacc_valaccepo bestval acc bestshiftl2wdema momreduction model
1510.916200.9956530.911700.071.e-4[8.e-4, 20][4.e-4, 30][1.e-4, 40][1.e-5, 50][4.e-6, 60]
2 ??510.919400.9854570.920900.071.5e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
3500.920400.9842600.920700.071.8e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
4510.92000.9850550.921900.071.85e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
5480.920200.9818600.923000.072.e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
6480.920500.9839570.924200.072.e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
7440.920900.9755590.924200.072.e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
8530.920000.9836580.920500.073.e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
9510.914600.97660.914600.074.e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
10500.917800.9758580.919500.074.5e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
11520.912400.9721570.913800.076.e-4[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]

The evolution of run nr. 5 is shown in the following graphics:

Note: In the runs nr. 3 for l2=1.8 and run nr. 8 for l2=3e-4, respectively, the value of val_acc=0.92 was just touched at epoch 50, but did not stay above 0.92 between epochs [51,59[. During the experiments nr. 9 to nr. 11 we did not even reach the aspired accuracy level. So, it seems that the variation of the L2-parameter “l2” should be kept in the range

  • 1.9e-4 <= l2 < 3.e-4

to reach our goal. This is a rather narrow range.

Interpretation: I. Loshilov and F. Hutter [8] have shown (for another type of network) that convergence to good accuracy values with the Adam optimizer (without weight decay) is possible, only, when the initial learning rate and the L2 parameter are kept in a relatively narrow diagonal band in the plane of (initial LR-values, L2-values). The band indicates a correlation between L2 and LR values (see [8] for details). We observe signs of this correlation here, too.

For further comparisons with [8] please note that we work with a standard (un-branched) ResNet network (of depth 56) in this post series. This model has 1.7 million parameters, only. The authors of [8], instead, employed a branched ResNet (of depth 26) with 11.6 million parameters (according to sect. 6 in [8]). So, we should not be astonished over largely different values of the weight decay parameters in later experiments of this post series.

Trial 3: Adam with weight decay, L2 regularization, piecewise linear LR schedule

It is possible to provide a weight decay factor to the Adam implementation in Keras. However, as the authors of [8] have pointed out – quote: “This is a correct implementation of L2 regularization, but not of weight decay”.

So, it was interesting to see whether weight decay really brought an improvement. Meaning: Would we achieve our goal even with a low value of l2 = 1.e-4?

This was indeed the case:

#epochacc_valaccepo bestval acc bestshiftl2wdema momreduction model
1490.920400.9835540.924600.072.e-44.e-30.99[8.e-4, 20][8.e-5, 40][2.e-5, 50][5.e-6, 60]
2510.920200.9896540.920800.071.e-41.e-50.99[8.e-4, 20][8.e-5, 40][2.e-5, 50][5.e-6, 60]
3420.920000.9801510.924800.071.e-45.e-40.98[8.e-4, 20][8.e-5, 40][2.e-5, 50][5.e-6, 60]
4500.920100.9888530.921300.071.e-41.e-30.98[8.e-4, 20][8.e-5, 40][1.e-5, 50][1.e-5, 60]
5490.921300.9876580.923100.071.e-41.e-40.98[8.e-4, 20][8.e-5, 40][2.e-5, 50][6.e-6, 60]

Run 3 had a remarkable sequence of

  • (0.92, 42), (0.9226, 43), (0.9232, 47), (0.9244, 50), (0.9246, 51), (0.9250, 60).

Such sequences appeared with other settings, too. Below you see a table with data which I got from a sequence of other earlier test runs.

#epochacc_valaccepo bestval acc bestshiftl2wdema momreduction model
1500.923500.9848580.926000.081.e-41.e-50.98[8.e-4, 30][1.5e-5, 40]
2480.920700.9803570.921800.081.e-41.e-50.98[8.e-4, 30][1.e-4, 40][1.e-5, 50][1.e-6, 60]
3500.92100.9870560.921800.081.e-41.e-50.98[8.e-4, 20][4.e-4, 30][4.e-5, 40][4.e-6, 50][1.e-6, 60]
4500.921700.9843520.921900.081.e-41.e-50.98[8.e-4, 20][4.e-4, 30][8.e-5, 40][1.e-5, 50][4.e-6, 60]
5490.924000.9870550.924810.071.e-41.e-50.98[8.e-4, 20][4.e-4, 30][8.e-5, 40][1.e-5, 50][5.e-6, 60]
6500.921800.9878530.923500.071.e-41.e-50.99[8.e-4, 20][4.e-4, 30][8.e-5, 40][1.e-5, 50][5.e-6, 60]
7490.921100.9883560.921800.071.e-41.e-50.95[8.e-4, 20][4.e-4, 30][8.e-5, 40][1.e-5, 50][5.e-6, 60]
8490.922700.9888590.923600.071.e-51.e-50.98[8.e-4, 20][4.e-4, 30][8.e-5, 40][1.e-5, 50][5.e-6, 60]
9490.922300.9870600.922300.071.e-41.e-50.98[8.e-4, 24][4.e-4, 32][8.e-5, 40][1.e-5, 50][5.e-6, 60]
10490.923900.9876560.924100.071.e-47.e-40.98[8.e-4, 20][4.e-4, 30][8.e-5, 40][1.e-5, 50][5.e-6, 60]
11490.923900.9876560.924100.071.e-41.e-40.98[8.e-4, 20][4.e-4, 30][8.e-5, 40][1.e-5, 50][5.e-6, 60]

The plot for run nr. 10 in the latter table looked like:

I forgot to integrate the green border lines at that time. Sorry, …

Trial 4: Runs with AdamW, L2 regularization and piecewise linear LR schedule

Let us switch to the optimizer AdamW.

According to I. Loshilov and F. Hutter [8] and the author of [4] the effects of weight decay are consistently implemented in this optimizer. In addition AdamW provides a regularization equivalent to the effect of the L2-mechanism. This occurs via a similar term in the equation for gradient updates during training and error back propagation. I will have a closer look at the math behind all this in another separate post. For now we just apply AdamW in its Keras implementation.

As long as we keep up the L2 mechanism our network thus deals with two types of regularization – L2 regularization and an equivalent contribution by AdamW’s weight decay. Therefore, I keep the weight decay parameter “wd” relatively small:

  • 1.e-4 <= wd <= 4.e-3.

Note that 4.e-3 is the default value in Keras.

What do we expect? Well, I would say, a slight improvement in comparison to Adam. The following table gives you some results of 6 test runs.

#epochacc_valaccepo bestval acc bestshiftl2wdema momreduction model
1500.919800.9825550.921000.071.e-41.e-40.99[8.e-4, 20][4.e-4, 30][8.e-5, 40][4.e-5, 50][5.e-6, 60]
2500.92050.9838580.923000.072.e-44.e-30.99[8.e-4, 20][8.e-5, 40][4.e-5, 50][4.e-6, 60]
3500.92010.9900590.921000.071.e-44.e-30.99[8.e-4, 20][8.e-5, 40][1.e-5, 50][5.e-6, 60]
4490.922200.9823530.923700.072.e-44.e-30.99[8.e-4, 20][8.e-5, 40][2.e-5, 50][5.e-6, 60]
5440.9221600.9830510.923600.071.e-45.e-40.98[8.e-4, 20][8.e-5, 40][2.e-5, 50][7.e-6, 60]
6460.922700.9846570.925100.071.e-41.e-40.99[8.e-4, 20][8.e-5, 40][2.e-5, 50][5.e-6, 60]

Obviously, it was possible to reproducibly reach acc_val=0.92 latest at epoch 50. But in some cases we even achieved our goal significantly earlier. The plot below depicts the evolution of run nr. 5 for AdamW:

Note that the 5th run for AdamW corresponds to run nr. 5 in the previous table for Adam based runs.

Some experiments showed that it is better to raise the ema_momentum to 0.99 with wd <= 1.e-4. See the plot for run nr. 6 in the table.

So, there is some variation with the the parameters named. However, there are multiple combinations which bring us into our momentary target range of 50 epochs for val_acc=0.92.

The main message is so far is that we do not see any really noteworthy improvement over runs using Adam with a properly set weight decay parameter (in its Keras implementation). This might be due to the still dominant impact of the L2 regularization. We will try a different approach in the next post.

Summary

Observation 1: A reduction of epochs by a faster shrinking learning rate in a piecewise constant schedule is possible
We can reduce the number of epochs for a ResNet56v2 (built according to Atienza [1]) to reach a value of val_acc=0.92 to around epoch 50 just by moving the transition from LR=1.e-3 to LR=1.e-4 in a piecewise schedule to epoch 40. However, we needed Adam with weight decay to achieve our goal.

Observation 2: A piecewise linear LR reduction schedule is working well
A piecewise linear LR reduction after epoch 20 works even better – both with the Adam and the AdamW optimizer. We can get the number of epochs consistently down to 50 for reaching val_acc=0.92.

Observation 3: Adam seems to couple the initial learning rate to L2-values
If you do not use weight decay with the Adam optimizer you may have to raise the level of L2 regularization. However, the range of L2-values is limited with respect to the initial LR.

Observation 4: Weight decay is important – also for the Adam optimizer
We need weight decay to press the number of epochs down below 50 to reach at least a level of val_acc >= 0.92 safely and reproducibly for a variety of hyper-parameter settings.

Observation 5: Adam and AdamW with weight decay and L2-regularization work equally well
As long as L2-regularization is active Adam with weight decay works equally well or better than AdamW with similar wd parameter values. But note: The weight decay parameter was kept small in our experiments (i.e. below 4.e-3) so far.

Observation 6: Do not make the LR too small too fast
All runs which converged to good accuracy values had LR values of LR > 8.e-5 up to epoch 40 and LR > 1.e-5 up to epoch 50.

Conclusion

By some simple tricks as

  • a piecewise linear LR schedule, weight decay, L2 adjustment

we can bring the required number of epochs to reach a validation accuracy level of val_acc >= 0.92 down to or below 50. This is already a substantial improvement in comparison to what R. Atienza achieved in his book.

We should, however, not be too content. Some ideas published in [2] to [8] are worth to check out. A first point is that the authors of [8] have suggested to use AdamW without any L2-regularization. A second point is that they also favored a cosine based LR reduction scheme. In the next post of this series I will explore both options. Afterward we will also study the option of super-convergence.

Links and literature

[1] Rowel Atienza, 2020, “Advanced Deep Learning with Tensorflow 2 and Keras”, 2nd edition, Packt Publishing Ltd., Birmingham, UK

[2] A.Rosebrock, 2019, “Cyclical Learning Rates with Keras and Deep Learning“, pyimagesearch.com

[3] L. Smith, 2018, “A disciplined approach to neural network hyper-parameters: Part 1 — learning rate, batch size, momentum, and weight decay“, arXiv

[4] F.M. Graetz, 2018, “Why AdamW matters“, towards.science.com

[5] L. N.. Smith, N. Topin, 2018, “Super-Convergence: Very Fast Training of Neural Networks Using Large Learning Rates“, arXiv

[6] S. Gugger, J. Howard, 2018, fast.ai – “AdamW and Super-convergence is now the fastest way to train neural nets“, fast.ai

[7] L. Smith, Q.V. Le, 2018, “A Bayesian Perspective on Generalization and Stochastic Gradient Descent

[8] I. Loshchilov, F. Hutter, 2018, “Fixing Weight Decay Regularization in Adam“, arXiv