ImageNet: VGGNet, ResNet, Inception, and Xception with Keras

A few months ago I wrote a tutorial on how to classify images using Convolutional Neural Networks (specifically, VGG16) pre-trained on the ImageNet dataset with Python and the Keras deep learning library.

The pre-trained networks inside of Keras are capable of recognizing 1,000 different object categories, similar to objects we encounter in our day-to-day lives with high accuracy.

Back then, the pre-trained ImageNet models were separate from the core Keras library, requiring us to clone a free-standing GitHub repo and then manually copy the code into our projects.

This solution worked well enough; however, since my original blog post was published, the pre-trained networks (VGG16, VGG19, ResNet50, Inception V3, and Xception) have been fully integrated into the Keras core (no need to clone down a separate repo anymore) — these implementations can be found inside the applications sub-module.

Because of this, I’ve decided to create a new, updated tutorial that demonstrates how to utilize these state-of-the-art networks in your own classification projects.

Specifically, we’ll create a special Python script that can load any of these networks using either a TensorFlow or Theano backend, and then classify your own custom input images.

To learn more about classifying images with VGGNet, ResNet, Inception, and Xception, just keep reading.

Looking for the source code to this post?
Jump right to the downloads section.

VGGNet, ResNet, Inception, and Xception with Keras

In the first half of this blog post I’ll briefly discuss the VGG, ResNet, Inception, and Xception network architectures included in the Keras library.

We’ll then create a custom Python script using Keras that can load these pre-trained network architectures from disk and classify your own input images.

Finally, we’ll review the results of these classifications on a few sample images.

State-of-the-art deep learning image classifiers in Keras

Keras ships out-of-the-box with five Convolutional Neural Networks that have been pre-trained on the ImageNet dataset:

  1. VGG16
  2. VGG19
  3. ResNet50
  4. Inception V3
  5. Xception

Let’s start with a overview of the ImageNet dataset and then move into a brief discussion of each network architecture.

What is ImageNet?

ImageNet is formally a project aimed at (manually) labeling and categorizing images into almost 22,000 separate object categories for the purpose of computer vision research.

However, when we hear the term “ImageNet” in the context of deep learning and Convolutional Neural Networks, we are likely referring to the ImageNet Large Scale Visual Recognition Challenge, or ILSVRC for short.

The goal of this image classification challenge is to train a model that can correctly classify an input image into 1,000 separate object categories.

Models are trained on ~1.2 million training images with another 50,000 images for validation and 100,000 images for testing.

These 1,000 image categories represent object classes that we encounter in our day-to-day lives, such as species of dogs, cats, various household objects, vehicle types, and much more. You can find the full list of object categories in the ILSVRC challenge here.

When it comes to image classification, the ImageNet challenge is the de facto benchmark for computer vision classification algorithms — and the leaderboard for this challenge has been dominated by Convolutional Neural Networks and deep learning techniques since 2012.

The state-of-the-art pre-trained networks included in the Keras core library represent some of the highest performing Convolutional Neural Networks on the ImageNet challenge over the past few years. These networks also demonstrate a strong ability to generalize to images outside the ImageNet dataset via transfer learning, such as feature extraction and fine-tuning.

VGG16 and VGG19

Figure 1: A visualization of the VGG architecture (source).

The VGG network architecture was introduced by Simonyan and Zisserman in their 2014 paper, Very Deep Convolutional Networks for Large Scale Image Recognition.

This network is characterized by its simplicity, using only 3×3 convolutional layers stacked on top of each other in increasing depth. Reducing volume size is handled by max pooling. Two fully-connected layers, each with 4,096 nodes are then followed by a softmax classifier (above).

The “16” and “19” stand for the number of weight layers in the network (columns D and E in Figure 2 below):

Figure 2: Table 1 of Very Deep Convolutional Networks for Large Scale Image Recognition, Simonyan and Zisserman (2014).

In 2014, 16 and 19 layer networks were considered very deep (although we now have the ResNet architecture which can be successfully trained at depths of 50-200 for ImageNet and over 1,000 for CIFAR-10).

Simonyan and Zisserman found training VGG16 and VGG19 challenging (specifically regarding convergence on the deeper networks), so in order to make training easier, they first trained smaller versions of VGG with less weight layers (columns A and C) first.

The smaller networks converged and were then used as initializations for the larger, deeper networks — this process is called pre-training.

While making logical sense, pre-training is a very time consuming, tedious task, requiring an entire network to be trained before it can serve as an initialization for a deeper network.

We no longer use pre-training (in most cases) and instead prefer Xaiver/Glorot initialization or MSRA initialization (sometimes called He et al. initialization from the paper, Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification). You can read more about the importance of weight initialization and the convergence of deep neural networks inside All you need is a good init, Mishkin and Matas (2015).

Unfortunately, there are two major drawbacks with VGGNet:

  1. It is painfully slow to train.
  2. The network architecture weights themselves are quite large (in terms of disk/bandwidth).

Due to its depth and number of fully-connected nodes, VGG is over 533MB for VGG16 and 574MB for VGG19. This makes deploying VGG a tiresome task.

We still use VGG in many deep learning image classification problems; however, smaller network architectures are often more desirable (such as SqueezeNet, GoogLeNet, etc.).


Unlike traditional sequential network architectures such as AlexNet, OverFeat, and VGG, ResNet is instead a form of “exotic architecture” that relies on micro-architecture modules (also called “network-in-network architectures”).

The term micro-architecture refers to the set of “building blocks” used to construct the network. A collection of micro-architecture building blocks (along with your standard CONV, POOL, etc. layers) leads to the macro-architecture (i.e,. the end network itself).

First introduced by He et al. in their 2015 paper, Deep Residual Learning for Image Recognition, the ResNet architecture has become a seminal work, demonstrating that extremely deep networks can be trained using standard SGD (and a reasonable initialization function) through the use of residual modules:

Figure 3: The residual module in ResNet as originally proposed by He et al. in 2015.

Further accuracy can be obtained by updating the residual module to use identify mappings, as demonstrated in their 2016 followup publication, Identity Mappings in Deep Residual Networks:

Figure 4: (Left) The original residual module. (Right) The updated residual module using pre-activation.

That said, keep in mind that the ResNet50 (as in 50 weight layers) implementation in the Keras core is based on the former 2015 paper.

Even though ResNet is much deeper than VGG16 and VGG19, the model size is actually substantially smaller due to the usage of global average pooling rather than fully-connected layers — this reduces the model size down to 102MB for ResNet50.

Inception V3

The “Inception” micro-architecture was first introduced by Szegedy et al. in their 2014 paper, Going Deeper with Convolutions:

Figure 5: The original Inception module used in GoogLeNet.

The goal of the inception module is to act as a “multi-level feature extractor” by computing 1×1, 3×3, and 5×5 convolutions within the same module of the network — the output of these filters are then stacked along the channel dimension and before being fed into the next layer in the network.

The original incarnation of this architecture was called GoogLeNet, but subsequent manifestations have simply been called Inception vN where N refers to the version number put out by Google.

The Inception V3 architecture included in the Keras core comes from the later publication by Szegedy et al., Rethinking the Inception Architecture for Computer Vision (2015) which proposes updates to the inception module to further boost ImageNet classification accuracy.

The weights for Inception V3 are smaller than both VGG and ResNet, coming in at 96MB.


Figure 6: The Xception architecture.

Xception was proposed by none other than François Chollet himself, the creator and chief maintainer of the Keras library.

Xception is an extension of the Inception architecture which replaces the standard Inception modules with depthwise separable convolutions.

The original publication, Xception: Deep Learning with Depthwise Separable Convolutions can be found here.

Xception sports the smallest weight serialization at only 91MB.

What about SqueezeNet?

Figure 7: The “fire” module in SqueezeNet, consisting of a “squeeze” and an “expand”. (Iandola et al., 2016).

For what it’s worth, the SqueezeNet architecture can obtain AlexNet-level accuracy (~57% rank-1 and ~80% rank-5) at only 4.9MB through the usage of “fire” modules that “squeeze” and “expand”.

While leaving a small footprint, SqueezeNet can also be very tricky to train.

That said, I demonstrate how to train SqueezeNet from scratch on the ImageNet dataset inside my upcoming book, Deep Learning for Computer Vision with Python.

Classifying images with VGGNet, ResNet, Inception, and Xception with Python and Keras

Let’s learn how to classify images with pre-trained Convolutional Neural Networks using the Keras library.

Open up a new file, name it , and insert the following code:

Lines 2-13 import our required Python packages. As you can see, most of the packages are part of the Keras library.

Specifically, Lines 2-6 handle importing the Keras implementations of ResNet50, Inception V3, Xception, VGG16, and VGG19, respectively.

Please note that the Xception network is compatible only with the TensorFlow backend (the class will throw an error if you try to instantiate it with a Theano backend).

Line 7 gives us access to the imagenet_utils  sub-module, a handy set of convenience functions that will make pre-processing our input images and decoding output classifications easier.

The remainder of the imports are other helper functions, followed by NumPy for numerical processing and cv2  for our OpenCV bindings.

Next, let’s parse our command line arguments:

We’ll require only a single command line argument, --image , which is the path to our input image that we wish to classify.

We’ll also accept an optional command line argument, --model , a string that specifies which pre-trained Convolutional Neural Network we would like to use — this value defaults to vgg16  for the VGG16 network architecture.

Given that we accept the name of our pre-trained network via a command line argument, we need to define a Python dictionary that maps the model names (strings) to their actual Keras classes:

Lines 25-31 defines our MODELS  dictionary which maps a model name string to the corresponding class.

If the --model  name is not found inside MODELS , we’ll raise an AssertionError  (Lines 34-36).

A Convolutional Neural Network takes an image as an input and then returns a set of probabilities corresponding to the class labels as output.

Typical input image sizes to a Convolutional Neural Network trained on ImageNet are 224×224227×227256×256, and 299×299; however, you may see other dimensions as well.

VGG16, VGG19, and ResNet all accept 224×224 input images while Inception V3 and Xception require 299×299 pixel inputs, as demonstrated by the following code block:

Here we initialize our inputShape  to be 224×224 pixels. We also initialize our preprocess  function to be the standard preprocess_input  from Keras (which performs mean subtraction).

However, if we are using Inception or Xception, we need to set the inputShape  to 299×299 pixels, followed by updating preprocess  to use a separate pre-processing function that performs a different type of scaling.

The next step is to load our pre-trained network architecture weights from disk and instantiate our model:

Line 58 uses the MODELS  dictionary along with the --model  command line argument to grab the correct Network  class.

The Convolutional Neural Network is then instantiated on Line 59 using the pre-trained ImageNet weights;

Note: Weights for VGG16 and VGG19 are > 500MB. ResNet weights are ~100MB, while Inception and Xception weights are between 90-100MB. If this is the first time you are running this script for a given network, these weights will be (automatically) downloaded and cached to your local disk. Depending on your internet speed, this may take awhile. However, once the weights are downloaded, they will not need to be downloaded again, allowing subsequent runs of  to be much faster.

Our network is now loaded and ready to classify an image — we just need to prepare this image for classification:

Line 65 loads our input image from disk using the supplied inputShape  to resize the width and height of the image.

Line 66 converts the image from a PIL/Pillow instance to a NumPy array.

Our input image is now represented as a NumPy array with the shape (inputShape[0], inputShape[1], 3) .

However, we typically train/classify images in batches with Convolutional Neural Networks, so we need to add an extra dimension to the array via np.expand_dims  on Line 72.

After calling np.expand_dims  the image  has the shape (1, inputShape[0], inputShape[1], 3) . Forgetting to add this extra dimension will result in an error when you call .predict  of the model .

Lastly, Line 76 calls the appropriate pre-processing function to perform mean subtraction/scaling.

We are now ready to pass our image through the network and obtain the output classifications:

A call to .predict  on Line 80 returns the predictions from the Convolutional Neural Network.

Given these predictions, we pass them into the ImageNet utility function .decode_predictions  to give us a list of ImageNet class label IDs, “human-readable” labels, and the probability associated with the labels.

The top-5 predictions (i.e., the labels with the largest probabilities) are then printed to our terminal on Lines 85 and 86.

The last thing we’ll do here before we close out our example is load our input image from disk via OpenCV, draw the #1 prediction on the image, and finally display the image to our screen:

To see our pre-trained ImageNet networks in action, take a look at the next section.

VGGNet, ResNet, Inception, and Xception classification results

All examples in this blog post were gathered using Keras >= 2.0 and a TensorFlow backend. If you are using TensorFlow, make sure you are using version >= 1.0, otherwise you will run into errors. I’ve also tested this script with the Theano backend and confirmed that the implementation will work with Theano as well.

Once you have TensorFlow/Theano and Keras installed, make sure you download the source code + example images to this blog post using the “Downloads” section at the bottom of the tutorial.

From there, let’s try classifying an image with VGG16:

Figure 8: Classifying a soccer ball using VGG16 pre-trained on the ImageNet database using Keras (source).

Taking a look at the output, we can see VGG16 correctly classified the image as “soccer ball” with 93.43% accuracy.

To use VGG19, we simply need to change the --network  command line argument:

Figure 9: Classifying a vehicle as “convertible” using VGG19 and Keras (source).

VGG19 is able to correctly classify the the input image as “convertible” with a probability of 91.76%. However, take a look at the other top-5 predictions: sports car with 4.98% probability (which the car is), limousine at 1.06% (incorrect, but still reasonable), and “car wheel” at 0.75% (also technically correct since there are car wheels in the image).

We can see similar levels of top-5 accuracy in the following example where we use the pre-trained ResNet architecture:

Figure 10: Using ResNet pre-trained on ImageNet with Keras + Python (source).

ResNet correctly classifies this image of Clint Eastwood holding a gun as “revolver” with 69.79% accuracy. It’s also interesting to see “rifle” at 7.74% and “assault rifle” at 5.63% included in the top-5 predictions as well. Given the viewing angle of the revolver and the substantial length of the barrel (for a handgun) it’s easy to see how a Convolutional Neural Network would also return higher probabilities for a rifle as well.

This next example attempts to classify the species of dog using ResNet:

Figure 11: Classifying dog species using ResNet, Keras, and Python.

The species of dog is correctly identified as “beagle” with 94.48% confidence.

I then tried classifying the following image of Johnny Depp from the Pirates of the Caribbean franchise:

Figure 12: Classifying a ship wreck with ResNet pre-trained on ImageNet with Keras (source).

While there is indeed a “boat” class in ImageNet, it’s interesting to see that the Inception network was able to correctly identify the scene as a “(ship) wreck” with 96.29% probability. All other predicted labels, including “seashore”, “canoe”, “paddle”, and “breakwater” are all relevant, and in some cases absolutely correct as well.

For another example of the Inception network in action, I took a photo of the couch sitting in my office:

Figure 13: Recognizing various objects in an image with Inception V3, Python, and Keras.

Inception correctly predicts there is a “table lamp” in the image with 69.68% confidence. The other top-5 predictions are also dead-on, including a “studio couch”“window shade” (far right of the image, barely even noticeable), “lampshade”, and “pillow”.

In the context above, Inception wasn’t even used as an object detector, but it was still able to classify all parts of the image within its top-5 predictions. It’s no wonder that Convolutional Neural Networks make for excellent object detectors!

Moving on to Xception:

Figure 14: Using the Xception network architecture to classify an image (source).

Here we have an image of scotch barrels, specifically my favorite scotch, Lagavulin. Xception correctly classifies this image as “barrels”.

This last example was classified using VGG16:

Figure 15: VGG16 pre-trained on ImageNet with Keras.

The image itself was captured a few months ago as I was finishing up The Witcher III: The Wild Hunt (easily in my top-3 favorite games of all time). The first prediction by VGG16 is “home theatre” — a reasonable prediction given that there is a “television/monitor” in the top-5 predictions as well.

As you can see from the examples in this blog post, networks pre-trained on the ImageNet dataset are capable of recognizing a variety of common day-to-day objects. I hope that you can use this code in your own projects!

What now?


You can now recognize 1,000 separate object categories from the ImageNet dataset using pre-trained state-of-the-art Convolutional Neural Networks.

…but what if you wanted to train your own custom deep learning networks from scratch?

How would you go about it?

Do you know where to start?

Let me help:

Whether this is the first time you’ve worked with machine learning and neural networks or you’re already a seasoned deep learning practitioner, my new book is engineered from the ground up to help you reach deep learning expert status.


In today’s blog post we reviewed the five Convolutional Neural Networks pre-trained on the ImageNet dataset inside the Keras library:

  1. VGG16
  2. VGG19
  3. ResNet50
  4. Inception V3
  5. Xception

I then demonstrated how to use each of these architectures to classify your own input images using the Keras library and the Python programming language.

If you are interested in learning more about deep learning and Convolutional Neural Networks (and how to train your own networks from scratch), be sure to take a look at my upcoming book, Deep Learning for Computer Vision with Python, available for pre-order now.


If you would like to download the code and images used in this post, please enter your email address in the form below. Not only will you get a .zip of the code, I’ll also send you a FREE 11-page Resource Guide on Computer Vision and Image Search Engines, including exclusive techniques that I don’t post on this blog! Sound good? If so, enter your email address and I’ll send you the code immediately!

, , , , ,

19 Responses to ImageNet: VGGNet, ResNet, Inception, and Xception with Keras

  1. Aramis March 20, 2017 at 12:24 pm #

    Dear Adrian;

    Thank you for you nice tutorial. I always learn many new points from your tutorials which organized and explained very-well. I have implemented this code and I could figure out how to use these models with keras. I thought now I can use transfer learning with these pre-trained models and train on my own data.

    However, the main problem with my data is that they are medical images and gray-scale. I could follow the tutorial which proposed by FCohelt but I couldn’t figure out how to change the structure of the models to accept 1 channel data.

    I would be glad if you could give some hint for transfer learning with pre-trained models for not RGB but gray-scale images.


    • Adrian Rosebrock March 21, 2017 at 7:20 am #

      These pre-trained networks assume you are using 3 channel images — you won’t be able to modify them to use 1 channel images unless you train them from scratch. Instead, the solution is to turn your 1 channel image into a 3 channel image:

      image = np.dstack([1chan, 1chan, 1chan])

      From there you can pass the image through the network since it’s a 3 channel image (but appears gray).

      I’ll also be discussing transfer learning in great detail in my upcoming book, Deep Learning for Computer Vision with Python.

  2. Parth March 20, 2017 at 12:42 pm #

    Hey Adrian,
    Thanks for the blog.
    I was hoping to do Pedestrian/human detection using Convolutional Neural Networks. I have tried using HoG but it didn’t turn out to be super accurate. The problem I am facing with using CNN with ImageNet trained classifiers is that there is no class/label as ‘person’ or ‘human’ or anything of that sort. What do you suggest I do? Could I try training it with INRIA person dataset or something similar? If yes, how?

    • Adrian Rosebrock March 21, 2017 at 7:18 am #

      I would fine-tune one of the networks on a dataset that is representative of the people you want to detect in images. If that’s INRIA, use it for fine-tuning.

  3. nicho March 20, 2017 at 12:59 pm #


  4. Ashti March 20, 2017 at 5:11 pm #


    Not related to this post.
    But i have a query wrt to keyframe extraction from videos.
    Using python and opencv i have to extract keyframes.
    I tried getting frames for each frame and then subtracting from each other and storing unique one which resulted in huge amoutn of frames.\
    I need to calculate pixel difference of frames and compare it with a threshold value. if PD > threshold store it as keyframe. Can you please give me an example on how can i calculate threshold of images which would be fetch me good amount of keyframes. same would be applied for other videos too…

    • Adrian Rosebrock March 21, 2017 at 7:13 am #

      Hey Ashti — I would kindly ask that comments on a particular blog post be related to the subject matter of the post (otherwise it comes off as a bit rude/presumptive). If you want to learn more about comparing images, try this post. Best of luck with the project.

  5. MiaoDX March 21, 2017 at 3:31 am #

    Aha, not so easy for me to point out a typo since there are so many readers and you’re so careful.

    However, in “VGG16 and VGG19” section, “Due to its depth and number of fully-connected nodes, VGG is over 533MB for VGG16 and 574MB for VGG16. This makes deploying VGG a tiresome task.”

    The latter should be VGG19, I think. ^_^

    • Adrian Rosebrock March 21, 2017 at 7:06 am #

      You are correct, thank you for pointing out the typo! It is fixed now.

  6. Ruben March 22, 2017 at 4:32 am #

    When I import from keras.applications import ResNet50, I have the next error:

    ImportError: cannot import name ‘GlobalAveragePooling2D’

    • Adrian Rosebrock March 22, 2017 at 8:34 am #

      Which version of Keras are you running? And which version of TensorFlow/Theano?

      • Ruben March 22, 2017 at 8:59 am #

        Thanks, all is ok with keras-2.0.2 theano -0.9.0

        • Adrian Rosebrock March 22, 2017 at 9:08 am #

          Congrats on resolving the issue!

  7. Abraham George March 23, 2017 at 11:57 am #

    I need to take live images and label it how do i do it?
    I cannot pre process the obtained frame ,what should i do?

    • Adrian Rosebrock March 25, 2017 at 9:34 am #

      For each frame in your video stream you would pass it through the network and obtain the output class labels.

  8. Abraham George March 23, 2017 at 11:59 am #

    what is the difference between parsing an image and reading it using imread?

    • Adrian Rosebrock March 25, 2017 at 9:34 am #

      I’m not sure what you mean Abraham, can you please elaborate on your comment?

  9. Sunggu kim March 24, 2017 at 11:01 am #

    Thank you for great tutorial.

    i’m always wondering about can i append more class to pre-trained network with my data or should i re-train all things?

    If possible we can save huge time and resources.

    Will it be possible?


    • Adrian Rosebrock March 25, 2017 at 9:22 am #

      The process of changing the output classes of a pre-trained network without having to re-train it from scratch is called fine-tuning. I’ll be covering fine-tuning in detail inside Deep Learning for Computer Vision with Python.

Leave a Reply