Keras: Multiple outputs and multiple losses

A couple weeks ago we discussed how to perform multi-label classification using Keras and deep learning.

Today we are going to discuss a more advanced technique called multi-output classification.

So, what’s the difference between the two? And how are you supposed to keep track of all these terms?

While it can be a bit confusing, especially if you are new to studying deep learning, this is how I keep them straight:

  • In multi-label classification, your network has only one set of fully-connected layers (i.e., “heads”) at the end of the network responsible for classification.
  • But in multi-output classification your network branches at least twice (sometimes more), creating multiple sets of fully-connected heads at the end of the network — your network can then predict a set of class labels for each head, making it possible to learn disjoint label combinations.

You can even combine multi-label classification with multi-output classification so that each fully-connected head can predict multiple outputs!

If this is starting to make your head spin, no worries — I’ve designed today’s tutorial to guide you through multiple output classification with Keras. It’s actually quite easier than it sounds.

That said, this is a more advanced deep learning technique we’re covering today so if you have not already read my first post on Multi-label classification with Keras make sure you do that now.

From there, you’ll be prepared to train your network with multiple loss functions and obtain multiple outputs from the network.

To learn how to use multiple outputs and multiple losses with Keras, just keep reading!

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

Keras: Multiple outputs and multiple losses

Figure 1: Using Keras we can perform multi-output classification where multiple sets of fully-connected heads make it possible to learn disjoint label combinations. This animation demonstrates several multi-output classification results.

In today’s blog post we are going to learn how to utilize:

  • Multiple loss functions
  • Multiple outputs

…using the Keras deep learning library.

As mentioned in the introduction to this tutorial, there is a difference between multi-label and multi-output prediction.

With multi-label classification, we utilize one fully-connected head that can predict multiple class labels.

But with multi-output classification, we have at least two fully-connected heads — each head is responsible for performing a specific classification task.

We can even combine multi-output classification with multi-label classification — in this scenario, each multi-output head would be responsible for computing multiple labels as well!

Your eyes might be starting to gloss over or your might be feeling the first pangs of a headache, so instead of continuing this discussion of multi-output vs. multi-label classification let’s dive into our project. I believe the code presented in this post will help solidify the concept for you.

We’ll start with a review of the dataset we’ll be using to build our multi-output Keras classifier.

From there we’ll implement and train our Keras architecture, FashionNet, which will be used to classify clothing/fashion items using two separate forks in the architecture:

  1. One fork is responsible for classifying the clothing type of a given input image (ex., shirt, dress, jeans, shoes, etc.).
  2. And the second fork is responsible for classifying the color of the clothing (black, red, blue, etc.).

Finally, we’ll use our trained network to classify example images and obtain the multi-output classifications.

Let’s go ahead and get started!

The multi-output deep learning dataset

Figure 2: Our multi-output classification dataset was created using the technique discussed in this post. Notice that our dataset doesn’t contain red/blue shoes or black dresses/shirts. Our multi-output classification with Keras method discussed in this blog post will still be able to make correct predictions for these combinations.

The dataset we’ll be using in today’s Keras multi-output classification tutorial is based on the one from our previous post on multi-label classification with one exception — I’ve added a folder of 358 “black shoes” images.

In total, our dataset consists of 2,525 images across seven color + category combinations, including:

  • Black jeans (344 images)
  • Black shoes (358 images)
  • Blue dress (386 images)
  • Blue jeans (356 images)
  • Blue shirt (369 images)
  • Red dress (380 images)
  • Red shirt (332 images)

I created this dataset using the method described in my previous tutorial on How to (quickly) build a deep learning image dataset.

The entire process of downloading the images and manually removing irrelevant images for each of the seven combinations took approximately 30 minutes. When building your own deep learning image datasets, make sure you follow the tutorial linked above — it will give you a huge jumpstart on building your own datasets.

Our goal today is nearly the same as last time — to predict both the color and clothing type…

…with the added twist of being able to predict the clothing type + color of images our network was not trained on.

For example, given the following image of a “black dress” (again, which our network will not be trained on):

Figure 3: While images of “black dresses” are not included in today’s dataset, we’re still going to attempt to correctly classify them using multi-output classification with Keras and deep learning.

Our goal will be to correctly predict both “black” + “dress” for this image.

Our Keras + deep learning project structure

To work through today’s code walkthrough as well as train + test FashionNet on your own images, scroll to to the “Downloads” section and grab the .zip  associated with this blog post.

From there, unzip  the archive and change directories ( cd ) as shown below. Then, utilizing the tree  command you can view the files and folders in an organized fashion (pun intended):

Above you can find our project structure, but before we move on, let’s first review the contents.

There are 3 notable Python files:

  • pyimagesearch/ : Our multi-output classification network file contains the FashionNet architecture class consisting of three methods: build_category_branch , build_color_branch , and build . We’ll review these methods in detail in the next section.
  • : This script will train the  FashionNet  model and generate all of the files in the output folder in the process.
  • : This script loads our trained network and classifies example images using multi-output classification.

We also have 4 top-level directories:

  • dataset/ : Our fashion dataset which was scraped from Bing Image Search using their API. We introduced the dataset in the previous section. To create your own dataset the same way I did, see How to (quickly) build a deep learning image dataset.
  • examples/ : We have a handful of example images which we’ll use in conjunction with our  script in the last section of this blog post.
  • output/ : Our  script generates a handful of output files:
    • fashion.model : Our serialized Keras model.
    • category_lb.pickle : A serialized LabelBinarizer  object for the clothing categories is generated by scikit-learn. This file can be loaded (and labels recalled) by our  script.
    • color_lb.pickle : A LabelBinarizer  object for colors.
    • output_accs.png : The accuracies training plot image.
    • output_losses.png : The losses training plot image.
  • pyimagesearch/ : This is a Python module containing the FashionNet  class.

A quick review of our multi-output Keras architecture

To perform multi-output prediction with Keras we will be implementing a special network architecture (which I created for the purpose of this blog post) called FashionNet.

The FashionNet architecture contains two special components, including:

  1. A branch early in the network that splits the network into two “sub-networks” — one responsible for clothing type classification and the other for color classification.
  2. Two (disjoint) fully-connected heads at the end of the network, each in charge of its respective classification duty.

Before we start implementing FashionNet, let’s visualize each of these components, the first being the branching:

Figure 4: The top of our multi-output classification network coded in Keras. The clothing category branch can be seen on the left and the color branch on the right. Each branch has a fully-connected head.

In this network architecture diagram, you can see that our network accepts a 96 x 96 x 3  input image.

We then immediately create two branches:

  1. The branch on the left is responsible for classifying the clothing category.
  2. The branch on the right handles classifying the color.

Each branch performs its respective set of convolution, activation, batch normalization, pooling, and dropout operations until we reach the final outputs:

Figure 5: Our deep learning Keras multi-output classification network makes it possible to learn disjoint label combinations.

Notice how these sets of fully-connected (FC) heads look like the FC layers from other architectures we’ve examined on this blog — but now there are two of them, each of them responsible for its given classification task.

The branch on the right-hand side of the network is significantly shallower (not as deep) as the left branch. Predicting color is far easier than predicting clothing category and thus the color branch is shallow in comparison.

To see how we can implement such an architecture, let’s move on to our next section.

Implementing our “FashionNet” architecture

Figure 6: The Keras deep learning library has all of the capability necessary to perform multi-output classification.

Since training a network with multiple outputs using multiple loss functions is more of an advanced technique, I’ll be assuming you understand the fundamentals of CNNs and instead focus on the elements that make multi-output/multi-loss training possible.

If you’re new to the world of deep learning and image classification you should consider working through my book, Deep Learning for Computer Vision with Python, to help you get up to speed.

Ensure you’ve downloaded the files and data from the “Downloads” section before proceeding.

Once you have the downloads in hand, let’s open up  and review:

We begin by importing modules from the Keras library and by importing TensorFlow itself.

Since our network consists of two sub-networks, we’ll define two functions responsible for building each respective branch.

The first, build_category_branch , used to classify clothing type, is defined below:

The build_category_branch  function is defined on Lines 16 and 17 with three notable parameters:

  • inputs : The input volume to our category branch sub-network.
  • numCategories : The number of categories such as “dress”, “shoes”, “jeans”, “shirt”, etc.
  • finalAct : The final activation layer type with the default being a softmax classifier. If you were performing both multi-output and multi-label classification you would want to change this activation to a sigmoid.

Pay close attention to Line 20 where we use a Lambda  layer to convert our image from RGB to grayscale.

Why do this?

Well, a dress is a dress regardless of whether it’s red, blue, green, black, or purple, right?

Thus, we decide to throw away any color information and instead focus on the actual structural components in the image, ensuring our network does not learn to jointly associate a particular color with a clothing type.

Note: Lambdas work differently in Python 3.5 and Python 3.6. I trained this model using Python 3.5 so if you just run the  script to test the model with example images with Python 3.6 you may encounter difficulties. If you run into an error related to the Lambda layer, I suggest you either (a) try Python 3.5 or (b) train and classify on Python 3.6. No changes to the code are necessary.

We then proceed to build our CONV => RELU => POOL  block with dropout on Lines 23-27.

Our first CONV  layer has 32  filters with a 3 x 3  kernel and RELU  activation (Rectified Linear Unit). We apply batch normalization, max pooling, and 25% dropout.

Dropout is the process of randomly disconnecting nodes from the current layer to the next layer. This process of random disconnects naturally helps the network to reduce overfitting as no one single node in the layer will be responsible for predicting a certain class, object, edge, or corner.

This is followed by our two sets of (CONV => RELU) * 2 => POOL  blocks:

The changes in filters, kernels, and pool sizes in this code block work in tandem to progressively reduce the spatial size but increase depth.

Let’s bring it together with a FC => RELU  layer:

The last activation layer is fully connected and has the same number of neurons/outputs as our numCategories .

Take care to notice that we have named our final activation layer "category_output"  on Line 57. This is important as we will reference this layer by name later on in .

Let’s define our second function used to build our multi-output classification network. This one is named build_color_branch , which as the name suggests, is responsible for classifying color in our images:

Our parameters to build_color_branch  are essentially identical to build_category_branch . We distinguish the number of activations in the final layer with numColors  (different from numCategories ).

This time, we won’t apply a Lambda  grayscale conversion layer because we are actually concerned about color in this area of the network. If we converted to grayscale we would lose all of our color information!

This branch of the network is significantly more shallow than the clothing category branch because the task at hand is much simpler. All we’re asking our sub-network to accomplish is to classify color — the sub-network does not have to be as deep.

Just like our category branch, we have a second fully connected head. Let’s build the FC => RELU  block to finish out:

To distinguish the final activation layer for the color branch, I’ve provided the name="color_output"  keyword argument on Line 94. We’ll refer to the name in the training script.

Our final step for building FashionNet  is to put our two branches together and build  the final architecture:

Our build  function is defined on Line 100 and has 5 self-explanatory parameters.

The build  function makes an assumption that we’re using TensorFlow and channels last ordering. This is made clear on Line 105 where our inputShape  tuple is explicitly ordered (height, width, 3) , where the 3 represents the RGB channels.

If you would like to use a backend other than TensorFlow you’ll need to modify the code to: (1) correctly the proper channel ordering for your backend and (2) implement a custom layer to handle the RGB to grayscale conversion.

From there, we define the two branches of the network (Lines 110-113) and then put them together in a Model  (Lines 118-121).

The key takeaway is that our branches have one common input, but two different outputs (the clothing type and color classifications).

Implementing the multi-output and multi-loss training script

Now that we’ve implemented our FashionNet  architecture, let’s train it!

When you’re ready, open up  and let’s dive in:

We begin by importing necessary packages for the script.

From there we parse our command line arguments:

We’ll see how to run the training script soon. For now, just know that --dataset  is the input file path to our dataset and --model , --categorybin , --colorbin  are all three output file paths.

Optionally, you may specify a base filename for the generated accuracy/loss plots using the --plot  argument. I’ll point out these command line arguments again when we encounter them in the script. If Lines 21-32 look greek to you, please see my argparse + command line arguments blog post.

Now, let’s establish four important training variables:

We’re setting the following variables on Lines 36-39:

  • EPOCHS : The number of epochs is set to 50 . Through experimentation I found that 50  epochs yields a model that has low loss and has not overfitted to the training set (or not overfitted as best as we can).
  • INIT_LR : Our initial learning rate is set to 0.001 . The learning rate controls the “step” we make along the gradient. Smaller values indicate smaller steps and larger values indicate bigger steps. We’ll see soon that we’re going to use the Adam optimizer while progressively reducing the learning rate over time.
  • BS : We’ll be training our network in batch sizes of  32  .
  • IMAGE_DIMS : All input images will be resized to 96 x 96  with 3  channels (RGB). We are training with these dimensions and our network architecture input dimensions reflect these as well. When we test our network with example images in a later section, the testing dimensions must match the training dimensions.

Our next step is to grab our image paths and randomly shuffle them. We’ll also initialize lists to hold the images themselves as well as the clothing category and color, respectively:

And subsequently, we’ll loop over the imagePaths , preprocess, and populate the data , categoryLabels , and colorLabels  lists:

We begin looping over our imagePaths  on Line 54.

Inside the loop, we load and resize the image to the IMAGE_DIMS . We also convert our image from BGR ordering to RGB. Why do we do this conversion? Recall back to our FashionNet  class in the build_category_branch  function, where we used TensorFlow’s rgb_to_grayscale  conversion in a Lambda function/layer. Because of this, we first convert to RGB on Line 58, and eventually append the preprocessed image to the data  list.

Next, still inside of the loop, we extract both the color and category labels from the directory name where the current image resides (Line 64).

To see this in action, just fire up Python in your terminal, and provide a sample imagePath  to experiment with like so:

You can of course organize your directory structure any way you wish (but you will have to modify the code). My two favorite methods include (1) using subdirectories for each label or (2) storing all images in a single directory and then creating a CSV or JSON file to map image filenames to their labels.

Let’s convert the three lists to NumPy arrays, binarize the labels, and partition the data into training and testing splits:

Our last preprocessing step — converting to a NumPy array and scaling raw pixel intensities to [0, 1]  — can be performed in one swoop on Line 70.

We also convert the categoryLabels  and colorLabels  to NumPy arrays while we’re at it (Lines 75 and 76). This is necessary as in our next we’re going to binarize the labels using scikit-learn’s LabelBinarizer  which we previously imported (Lines 80-83). Since our network has two separate branches, we can use two independent label binarizers — this is different from multi-label classification where we used the MultiLabelBinarizer  (also from scikit-learn).

Next, we perform a typical 80% training/20% testing split on our dataset (Lines 87-96).

Let’s build the network, define our independent losses, and compile our model:

On Lines 93-96, we instantiate our multi-output FashionNet  model. We dissected the parameters when we created the FashionNet  class and build  function therein, so be sure to take a look at the values we’re actually providing here.

Next, we need to define two losses  for each of the fully-connected heads (Lines 101-104).

Defining multiple losses is accomplished with a dictionary using the names of each of the branch activation layers — this is why we named our output layers in the FashionNet implementation! Each loss will use categorical cross-entropy, the standard loss method used when training networks for classification with > 2 classes.

We also define equal lossWeights  in a separate dictionary (same name keys with equal values) on Line 105. In your particular application, you may wish to weight one loss more heavily than the other.

Now that we’ve instantiated our model and created our losses  + lossWeights  dictionaries, let’s initialize the Adam  optimizer with learning rate decay (Line 109) and compile  our model  (Lines 110 and 111).

Our next block simply kicks off the training process:

Recall back to Lines 87-90 where we split our data into training ( trainX ) and testing ( testX ). On Lines 114-119 we launch the training process while providing the data. Take note on Line 115 where we pass in the labels as a dictionary. The same goes for Lines 116 and 117 where we pass in a 2-tuple for the validation data. Passing in the training and validation labels in this manner is a requirement when performing multi-output classification with Keras. We need to instruct Keras which set of target labels corresponds to which output branch of the network.

Using our command line argument ( args["model"] ), we save the serialized model to disk for future recall.

We’ll also do the same to save our label binarizers as serialized pickle files:

Using the command line argument paths ( args["categorybin"]  and args["colorbin"] ) we write both of our label binarizers ( categoryLB  and colorLB ) to serialized pickle files on disk.

From there it’s all about plotting results in this script:

The above code block is responsible for plotting the loss history for each of the loss functions on separate but stacked plots, including:

  • Total loss
  • Loss for the category output
  • Loss for the color output

Similarly, we’ll plot the accuracies in a separate image file:

Our category accuracy and color accuracy plots are best viewed separately, so they are stacked as separate plots in one image.

Training the multi-output/multi-loss Keras model

Be sure to use the “Downloads” section of this blog post to grab the code and dataset.

Don’t forget: I used Python 3.5 to train the network included in the download for this tutorial. As long as you stay consistent (Python 3.5 or Python 3.6) you shouldn’t have a problem with the Lambda implementation inconsistency. You can probably even run Python 2.7 (haven’t tested this).

Open up terminal. Then paste the following command to kick off the training process (if you don’t have a GPU, you’ll want to grab a beer as well):

For our category output we obtained:

  • 99.31% accuracy on the training set
  • 93.47% accuracy on the testing set

And for the color output we reached:

  • 99.31% accuracy on the training set
  • 97.82% accuracy on the testing set

Below you can find the plots for each of our multiple losses:

Figure 7: Our Keras deep learning multi-output classification training losses are plotted with matplotlib. Our total losses (top), clothing category losses (middle), and color losses (bottom) are plotted independently for analysis.

As well as our multiple accuracies:

Figure 8: FashionNet, a multi-output classification network, is trained with Keras. In order to analyze the training it is best to show the accuracies in separate plots. Clothing category training accuracy plot (top). Color training accuracy plot (bottom).

Further accuracy can likely be obtained by applying data augmentation (covered in my book, Deep Learning for Computer Vision with Python).

Implementing a multi-output classification script

Now that we have trained our network, let’s learn how to apply it to input images not part of our training set.

Open up  and insert the following code:

First, we import our required packages followed by parsing command line arguments:

We have four command line arguments which are required to make this script run in your terminal:

  • --model : The path to the serialized model file we just trained (an output of our previous script).
  • --categorybin : The path to the category label binarizer (an output of our previous script).
  • --colorbin : The path to the color label binarizer (an output of our previous script).
  • --image : Our test image file path — this image will come from our examples/  directory.

From there, we load our image and preprocess it:

Preprocessing our image is required before we run inference. In the above block we load the image, resize it for output purposes, and swap color channels (Lines 24-26) so we can use TensorFlow’s RGB to grayscale function in our Lambda  layer of FashionNet. We then resize the RGB image (recalling IMAGE_DIMS  from our training script), scale it to [0, 1], convert to a NumPy array, and add a dimension (Lines 29-32) for the batch.

It is critical that the preprocessing steps follow the same actions taken in our training script.

Next, let’s load our serialized model and two label binarizers:

Using three of our four command line arguments on Lines 37-39, we load the model , categoryLB , and colorLB .

Now that both the (1) multi-output Keras model and (2) label binarizers are in memory, we can classify an image:

We perform multi-output classification on Line 43 resulting in a probability for both category and color ( categoryProba  and colorProba  respectively).

Note: I didn’t include the include code as it was a bit verbose but you can determine the order in which your TensorFlow + Keras model returns multiple outputs by examining the names of the output tensors. See this thread on StackOverflow for more details.

From there, we’ll extract the indices of the highest probabilities for both category and color (Lines 48 and 49).

Using the high probability indices, we can extract the class names (Lines 50 and 51).

That seems a little too easy, doesn’t it? But that’s really all there is to applying multi-output classification using Keras to new input images!

Let’s display the results to prove it:

We display our results on our output  image (Lines 54-61). It will look a little something like this in green text in the top left corner if we encounter a “red dress”:

  • category: dress (89.04%)
  • color: red (95.07%)

The same information is printed to the terminal on Lines 64 and 65 after which the output  image is shown on the screen (Line 68).

Performing multi-output classification with Keras

Now it’s time for the fun part!

In this section we are going to present our network with five images in the examples  directory which are not part of the training set.

The kicker is that our network has only been specifically trained to recognize two of the example images categories. These first two images (“black jeans” and “red shirt”) should be especially easy for our network to correctly classify both category and color.

The remaining three images are completely foreign to our model — we didn’t train with “red shoes”, “blue shoes”, or “black dresses” but we’re going to attempt multi-output classification and see what happens.

Let’s begin with “black jeans” — this one should be easy as there were plenty of similar images in the training dataset. Be sure to use the four command line arguments like so:

Figure 9: Deep learning multi-label classification allows us to recognize disjoint label combinations such as clothing type and clothing color. Our network has correctly classified this image as both “jeans” and “black”.

As expected, our network correctly classified the image as both “jeans” and “black”.

Let’s try a “red shirt”:

Figure 10: This image of a “red shirt” is a test image which is not in our deep learning image dataset. Our Keras multi-output network has; however, seen other red shirts. It easily classifies this image with both labels at 100% confidence.

With 100% confidence for both class labels, our image definitely contains a “red shirt”. Remember, our network has seen other examples of “red shirts” during the training process.

Now let’s step back.

Take a look at our dataset and recall that it has never seen “red shoes” before, but it has seen “red” in the form of “dresses” and “shirts” as well as “shoes” with “black” color.

Is it possible to make the distinction that this unfamiliar test image contains “shoes” that are “red”?

Let’s find out:

Figure 11: Our deep learning multi-output classification network has never seen the combination of “red” and “shoes” before. During the training process, we did present “shoes” (albeit, “black” ones) and we did present “red” (both “shirts” and “dresses”). Amazingly, our network fires neurons resulting in the correct multi-output labels for this image of “red shoes”. Success!


Looking at the results in the image, we were successful.

We’re off to a good start while presenting unfamiliar multi-output combinations. Our network design + training has paid off and we were able to recognize “red shoes” with high accuracy.

We’re not done yet — let’s present an image containing a “black dress” to our network. Remember, previously this same image did not yield a correct result in our multi-label classification tutorial.

I think we have a great chance at success this time around, so type the following command in your terminal:

Figure 12: While images of “black dresses” are not included in today’s dataset, we’re still able to correctly classify them using multi-output classification with Keras and deep learning.

Check out the class labels on the top-left of the image!

We achieved correct classification in both category and color with both reporting confidence of > 98% accuracy. We’ve accomplished our goal!

For sanity, let’s try one more unfamiliar combination: “blue shoes”. Enter the same command in your terminal, this time changing the --image  argument to examples/blue_shoes.jpg :

Figure 13: While multi-label classification may fail at unfamiliar label combinations, multi-output classification handles the task gracefully.

The same deal is confirmed — our network was not trained on “blue shoes” images but we were able to correctly classify them by using our two sub-networks along with multi-output and multi-loss classification.

Where to now?

If you enjoyed today’s post on multi-output classification and are eager to study more advanced deep learning techniques, including:

  • How to train your own custom object detectors
  • Multi-GPU training
  • Emotion and facial expression recognition
  • Working with large datasets that are too big to fit into memory
  • …and more!

…then you’ll want to be sure to take a look at my new deep learning book. Inside Deep Learning for Computer Vision with Python, I will guide you, step-by-step, on these advanced deep learning techniques, tips, and best practices.

Be sure to take a look — and don’t forget to grab your (free) sample chapters + table of contents PDF while you’re there!


In today’s blog post we learned how to utilize multiple outputs and multiple loss functions in the Keras deep learning library.

To accomplish this task, we defined a Keras architecture that is used for fashion/clothing classification called FashionNet.

The FashionNet architecture contains two forks:

  1. One fork is responsible for classifying the clothing type of a given input image (ex., shirt, dress, jeans, shoes, etc.).
  2. And the second fork is responsible for classifying the color of the clothing (black, red, blue, etc.).

This branch took place early in the network, essentially creating two “sub-networks” that are responsible for each of their respective classification tasks but both contained in the same network.

Most notably, multi-output classification enabled us to solve a nuisance from our previous post on multi-label classification where:

  • We trained our network on six categories, including: black jeans, blue dresses, blue jeans, blue shirts, red dresses, and red shirts…
  • …but we were unable to classify “black dress” as our network had never seen this combination of data before!

By creating two fully-connected heads and associated sub-networks (if necessary) we can train one head to classify clothing type and the other can learn how to recognize color — the end result is a network that can classify “black dress” even though it was never trained on such data!

Keep in mind though, you should always try to provide example training data for each class you want to recognize — deep neural networks, while powerful, are not “magic”!

You need to put in a best effort to train them properly and that includes gathering proper training data first.

I hope enjoyed today’s post on multi-output classification!

To be notified when future posts are published here on PyImageSearch, just enter your email address in the form below!


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 17-page Resource Guide on Computer Vision, OpenCV, and Deep Learning. Inside you'll find my hand-picked tutorials, books, courses, and libraries to help you master CV and DL! Sound good? If so, enter your email address and I’ll send you the code immediately!

, , , , , ,

64 Responses to Keras: Multiple outputs and multiple losses

  1. Alexander June 4, 2018 at 11:24 am #

    Hi Adrian!

    Great tutorial!
    I’m not sure I understand your NN architecture correctly.
    Does the category branch share any parameters with the color branch?
    If it doesn’t, then why call it different branches of the same net, but not 2 different nets?

    Kind regards,
    Alexander Myronov

    • Adrian Rosebrock June 4, 2018 at 12:28 pm #

      Hey Alexander — make sure you refer to my reply to Peng. Don’t get caught up in trying to use two separate networks as that’s specific to this example. Instead, focus on how we are using multiple outputs and multiple loss functions to train the network. There are applications of CNNs that do share parameters with earlier layers in the network but the training procedure will remain the same (that is the point of this tutorial).

  2. Peng June 4, 2018 at 11:25 am #

    Thanks for your blog post! It’s always very exciting to read your articles.

    Just a general question. The network structure in this article seems to be equivalent to 2 separate network (one for clothes and one for color). What is the benefit of designing the system this way? Wouldn’t it be easier to just run two networks in parallel?

    • Adrian Rosebrock June 4, 2018 at 12:27 pm #

      You could certainly use two separate networks if you would like; however, keep in mind that is is just an example.

      To start, depending on the task, it may be significantly less efficient to run two separate networks. Having a single network encapsulating all calculation can add efficiency.

      This efficiency is also compounded if the network shares layers. So while this example immediately branches after the input you could also have a network that branches halfway through the network.

      And a common pattern with LSTMs is to have one input at the top of the network and then another input mid-way through the network.

      Depending on your network architecture and particular task it may be totally impossible to separate the two networks.

      Don’t get too caught up in trying to separate the networks and instead use this as an example and blueprint for your own CNNs where you might have multiple inputs and multiple outputs.

  3. MImranKhan June 4, 2018 at 11:46 am #

    very nice keepit up

    • Adrian Rosebrock June 4, 2018 at 12:23 pm #

      Will do, I have no plans on stopping 😉 I’m glad you liked the post.

  4. Sam June 4, 2018 at 1:34 pm #

    Ran into an error:

    SystemError: unknown opcode

    My environment: Ubuntu 16.04LTS, Tensorflow 1.8.0, keras 2.1.6, cv2 3.4.1

    The whole trace:

    Using TensorFlow backend.
    [INFO] loading network…
    XXX lineno: 20, opcode: 0
    Traceback (most recent call last):
    File “”, line 42, in
    model = load_model(args[“model”], custom_objects={“tf”: tf})

    SystemError: unknown opcode

    • Sam June 4, 2018 at 1:37 pm #

      My error may be as a result of running the code using python 3.6 whereas the model may have been saved with python 3.5

      • Sam June 4, 2018 at 1:56 pm #

        After training the model in my environment…i.e.
        Python 3.6, Ubuntu 16.04LTS, Tensorflow 1.8.0, keras 2.1.6, cv2 3.4.1

        When I classify the example images, most images get classified as shoes with 100%
        Dataset does not seem unbalanced… wonder why the bias towards shoes

        • Sam June 4, 2018 at 2:09 pm #

          Ran training with a different random seed (random.seed(0)) and everything looks good. :o)

          • Adrian Rosebrock June 4, 2018 at 2:57 pm #

            Congrats on resolving the issue, Sam!

  5. Nabila Abraham June 4, 2018 at 3:11 pm #

    Adrian, this is a fantastic post on how to create subnets in a CNN! Thanks so much for putting it out there!
    I have a couple of fundamental questions I hope you don’t mind answering:

    1. What is the point of making a lambda layer? Could we not just convert RBG to grayscale using cv2.cvtColor()? It would still output a 96 x 96 x 1 tensor right? I see lambda layers being used in CNNs but I’m not sure of the point of them.
    2. Can the weights to each loss function be learned somehow ?
    3. How long did it take to train your multiple nets?
    4. Do you see any image segmentation benefits to this sort of structure?

    • Adrian Rosebrock June 5, 2018 at 7:16 am #

      1. The point is that (1) we want the network to be entirely self-contained and (2) more importantly, you may use a Lambda layer in the middle of a network and would still like the gradients backpropagate to the input. If you inserted a cv2.cvtColor conversion in there Keras/TensorFlow would have no idea how to backpropagate that info, hence why we need Lamda layers.

      2. You would typically run cross-validation experiments to tune your hyperparameters.

      3. Check the terminal output from the post. Each epoch was taking ~3 seconds. Over 50 epochs that’s 150 seconds.

      4. Not sure what you mean.

      • Nabila Abraham June 7, 2018 at 11:28 am #

        1. Okay, I think I understand – the intention is to be able to train end-to-end?

        4. I was thinking in terms of semantic segmentation problems using this sort of architecture might be helpful to segment multiple classes of data. Will look into it more but this is a great start!

        • Adrian Rosebrock June 7, 2018 at 2:56 pm #

          Correct. You want to train the network end-to-end. As I mentioned in my first reply imagine a scenario where you needed a grayscale representation in the middle of the network. You would lose the ability to backpropagate the gradients if you did not use the Lamda layer in the network.

  6. anirban ghosh June 4, 2018 at 11:20 pm #


    In the multiclass multi-label classification we had considered each label to be independent of the other and ie P(A intersection B) =0 and we achieved the classification using a binary classifier. Here in this example, we are again doing the same but instead of treating the problem as binary classifier we branched the labels into at the very beginning so that we treat them as independent labels and then used the softmax classifier.

    Can you please provide an example where you would not want to treat the labels as “independent” and what classifier would you use then?

    If I am wrong please feel ok to correct me.

    Anirban Ghosh

    • Adrian Rosebrock June 5, 2018 at 7:14 am #

      Actually, for this example we are using softmax classifiers and categorical cross-entropy which is different than the previous post on multi-label classification which treated the labels independently using a sigmoid activation + binary cross-entropy loss.

  7. Wilf June 6, 2018 at 5:09 am #

    I got an anomalous behavior in my experimentation

    Trained with keras 2.1.6 and tensorflow 1.8.0. Classification of examples all OK
    The classification for black_jeans.jpg was black(77.41%) and jeans(100%)

    VirtEnv2 –
    Trained with keras 2.1.2 and tensorflow 1.4.0-rc0.
    The tensorflow was built to use SSE4.1 and SSE4.2 instructions that accelerates performance.
    The classifications succeed except for black_jeans.jpg which gives blue (94.24%) jeans (100%).

    Do you have any idea about what is happening?

    • Adrian Rosebrock June 6, 2018 at 6:06 am #

      CNNs are stochastic algorithms. You will obtain slightly different results for each training due to the stochastic nature of the algorithms. Secondly, different versions of Keras and TensorFlow may also give slightly different results depending on what exactly was updated. Don’t try to obtain identical probabilities as I did but instead see if your loss and accuracy is close to mine (that is the important part).

  8. Wilf June 6, 2018 at 12:03 pm #

    Point taken regarding the probabilities.
    Still strange that I got blue for color of black_jeans.jpg in my VirtEnv2 above.

    • Wilf June 12, 2018 at 9:05 am #

      I continued my experimentations:

      In a new virtual environment, with an accelerated TensorFlow built from latest code on github. (Keras 2.2.0, TensorFlow 1.9.0-rc0)

      1. for EPOCHS=50: the classifications of the black examples come out as blue.
      2. for EPOCHS=20 and now the classifications of all examples are correct

  9. Anurag June 7, 2018 at 11:58 am #

    Could you write articles about multi modal deep learning.

    • Adrian Rosebrock June 7, 2018 at 2:54 pm #

      Hey Anurag — just to clarify do you mean “multiple models” or “multiple modals” as in “multiple modalities”? There is a difference between the two.

  10. Madhava Jay June 10, 2018 at 10:01 pm #

    Awesome example, as per some other suggestions I would be interested to see how best to determine weight sharing. For example what is the performance when you diverge after several Conv Pool layers in RGB and then applying the grayscaling and classification, while also diverging to determine colour. My guess is that at least the first few conv pool layers should be shared. The performance and memory size should be much lower right?

    • Adrian Rosebrock June 13, 2018 at 5:57 am #

      Whether or not you use weight sharing, and exactly where you share in your architecture, is really dependent on the project and the specific problem you are trying to solve. If you are curious about what happens in performance and memory size I suggest you modify the architecture and run the experiment 🙂

  11. Rizstar June 12, 2018 at 1:21 am #

    Hi Adrian Rosebrock
    I got the following error:
    Please help

    Using TensorFlow backend.
    usage: [-h] -d DATASET -m MODEL -l CATEGORYBIN -c COLORBIN [-p PLOT] error: unrecognized arguments: \

  12. SUMEET SAURAV July 3, 2018 at 5:35 pm #

    Can We have tutorial on single outputs and multiple losses (via Joint Optimization of two or more networks) like the approach use in the paper “Eye state recognition based on deep integrated neural network and transfer learning”.

    • Adrian Rosebrock July 5, 2018 at 6:44 am #

      Thank you for the suggestion, Sumeet. I will consider this topic for the future but I cannot guarantee if and when I will cover it.

  13. Dankam Therese July 8, 2018 at 7:58 pm #

    Thanks very much for the tutorial, can this network structure be used in offline signature recognition?

    • Adrian Rosebrock July 10, 2018 at 8:32 am #

      I’m not sure why you would be using multiple outputs for signature recognition. Perhaps you could elaborate on the project?

  14. Black July 15, 2018 at 11:56 pm #

    Nice post, thank you sir.

    The description in keras functional api “We seek to predict how many retweets and likes a news headline will receive on Twitter. The main input to the model will be the headline itself, as a sequence of words, but to spice things up, our model will also have an auxiliary input, receiving extra data such as the time of day when the headline was posted, etc.” is actually what I am trying to do. And a multi-inputs multi-outputs model do performs better! But I have trouble in interpreting the results.(why such architecture performs better) . Could u recommend some references?

    • Adrian Rosebrock July 17, 2018 at 7:24 am #

      I only cover computer vision on this blog, not text-based analysis. You may want to consider chatting with Jason Brownlee over at Machine Learning Mastery for text-based DL questions.

  15. tam75 August 27, 2018 at 10:52 am #

    Dear Sir, I am trying out if this can handle 2x color and 5x categories (meaning a binary and categorical problem). To do this I removed dataset folder red_dress and red_shirt folders, changed Line107 to binary_crossentropy, and changed Line63 finalAct=”sigmoid. But the error below happened when training.

    I ran out of ideas. Can you please point me to the right direction? Thanks.

    ValueError: Error when checking target: expected color_output to have shape (None, 2) but got array with shape (720, 1)

    • Adrian Rosebrock August 30, 2018 at 9:16 am #

      I think this may be due to a problem parsing your class labels from the file paths. Make sure your labels were parsed correctly. Additionally, the LabelBinarizer class from scikit-learn won’t vectorize a 2-class problem. For that, make sure you use Keras’ .to_categorical function.

  16. Azmi October 29, 2018 at 8:55 pm #

    Hi Adrian,

    This is with regards to class weights.
    Firstly, thank you so much for the code.
    I am implementing the same with my dataset but I see that my dataset is highly imbalanced . so when trying to implement class weights for multiple outputs I would like to know how it is done as I referred this link( for implementing .

    from sklearn.utils import class_weight
    class_weights = class_weight.compute_class_weight(‘balanced’,np.unique(y_train), y_train), y_train, class_weight=class_weights)

    However,the problem is I do not know how to pass two outputs (category and color) for class weights.
    Could you please provide your inputs on how to implement the class weights.

  17. Edward October 30, 2018 at 3:50 am #

    Thanks your a great tutorial. But where I can download the dataset and source code?

    • Adrian Rosebrock November 2, 2018 at 8:31 am #

      You can use the “Downloads” section of the blog post to download the source code.

  18. Edward October 30, 2018 at 4:09 am #

    Hi, one more question, for multiple output CNNs, have you ever compared: separately training vs jointly training

    • Adrian Rosebrock November 2, 2018 at 8:31 am #

      Yes, absolutely. You should always compare training separately vs. jointly training.

  19. Walid December 12, 2018 at 4:02 pm #

    Thanks, it is awesome and open windows to number of applications
    I have a question

    How can I use a lambda layer to resize image inside model to like 32px by 32px?

    Best Regards


  20. Eric January 8, 2019 at 9:01 pm #

    Dear Adrian,

    Thanks for the excellent tutorial.
    If I want to load the pre-trained model in further training, so that I don’t need to restart the training from scratch every time ?


  21. AD January 24, 2019 at 4:35 pm #

    Great article! (I’m new to machine learning, and this was really informative).

    I was curious, how the model would predict and measure multiple outputs within the same branch. i.e. in a photo in which someone was wearing a black dress and red shoes… how would you train the model to predict all four outputs? Would you need multiple models? How would you ensure that the model predicting “black dress” would not be measured as a false negative for red shoes, if black dress was also correct?

  22. Ayush March 20, 2019 at 5:31 am #

    How i can define the value of categorical_crossentropy , which is not define here.

    • Adrian Rosebrock March 22, 2019 at 9:41 am #

      I’m not sure what you mean. Are you getting an error when trying to use categorical cross-entropy?

  23. Emal Khan Ayubi April 18, 2019 at 2:57 am #

    Hello Adrian
    I am getting this error:
    error: (-215:Assertion failed) !ssize.empty() in function ‘resize’

    I compiled opencv from source file but still it is not fixed, I don’t know how to fix it.
    Please help me.

    Thanks in advance.

    • Adrian Rosebrock April 18, 2019 at 6:26 am #

      Double-check your path to the input images. The “cv2.imread” function is returning “None” meaning that the input image path is incorrect.

      • Emal Khan Ayubi April 20, 2019 at 10:07 am #

        I did this before asking here, I tried to load all the images through cv2.imread without resizing it, I did not get any errors, but when I try to resize them I get this error

        • Adrian Rosebrock April 25, 2019 at 9:18 am #

          OpenCV will not return an error if the path to an input image is invalid, it will just return “None”. Check your file paths.

  24. zameer May 7, 2019 at 12:45 pm #

    I am still confused why this network is able to classify the “black dress” correctly.

  25. Rajani K.S May 10, 2019 at 1:39 pm #

    Hi adrian

    I don’t know how to interpret the plots produced by the script.
    would you please tell me how to interpret them, or make a tutorial on it?

    Thanks in advance

  26. Alireza May 30, 2019 at 8:59 am #

    Dea Adrian,
    thanks for your nice blog post.
    Consider I have some colorful object in my training data, and they are not cloth, so it has a color label, but not cloth label.
    I want clothing type fork does not affected loss for this data, and second fork(color) only back-propagate .

    And I also have some clothing with no color label in my train set and do same.

    consider the same test-set with your blog post. and want to use multi output network.

    do you have any idea?

  27. Sheller August 11, 2019 at 10:25 pm #

    Dea Adrian,
    thanks for your nice tutorial.

    I am wondering if you used the batch-normalization layer correctly. The BN layer is designed to avoid internal covariate shift and is usually placed before each nonlinearity. I don’t know it will perform its real effect if you put this layer after activation layer.

    • Adrian Rosebrock August 16, 2019 at 5:50 am #

      That’s a common misconception (and sometimes a hyperparamter choice). The BN layer should typically goes after the nonlinearity. Consider ReLU for example. ReLU squashes any negative values to zero. If you use BN before ReLU then you’ll shift, on average, half your values to the negative range. ReLU then squashes those activations negative to zero. If you place the BN after the ReLU that won’t happen.

      All that said, it’s a hyperparamter and model design choice. Some CNNs that are designed via other DL models may find that a BN before the ReLU works better. Other architectures that have their layer placement studied extensively (such as ResNet) may do even stranger orderings of their layers.

      If you’re interested in learning more about BN layer replacement I would suggest you read Deep Learning for Computer Vision with Python.

  28. Elie Sokhon August 26, 2019 at 10:30 am #

    hi adrian great post it really helped as i have a project that requires this approach but did not know how to do it before reading this post ! also i bought your book online and its great!

    But i have a problem with your code, after training the model i tried to plot the figure using your method, but nothing is displayed. it shows no error though!! plz help me

    • Adrian Rosebrock September 5, 2019 at 10:47 am #

      Hey Elie — can you send me an email with more details, including a screenshot of your terminal running? It’s likely a simple issue but it would be easier to discuss it over email. Thanks!

  29. Sivaramakrishnan Rajaraman October 10, 2019 at 11:05 am #

    Hi, Adrian,
    Nice post as usual. In the script, could you please let know how to modify the script to run the prediction on a folder of images (say, a class of images for those the classifier was not trained), and produce the accuracy and other performance metrics?

    • Adrian Rosebrock October 17, 2019 at 8:09 am #

      1. Use the “paths.list_images” function of imutils to grab all images in an input directory
      2. Loop over those images
      3. Pass them into “model.predict” to obtain your output

  30. Shoaib November 27, 2019 at 12:17 pm #

    Hi Adrian,
    Thanks for making this great tutorial. I have one question, how can we use this with fit generator because when we have a huge data it is not possible to load all in the ram. Kindly help me in this regard.

    • Adrian Rosebrock December 5, 2019 at 10:41 am #

      See this tutorial which will teach you how to create your own custom fit_generator.

Before you leave a comment...

Hey, Adrian here, author of the PyImageSearch blog. I'd love to hear from you, but before you submit a comment, please follow these guidelines:

  1. If you have a question, read the comments first. You should also search this page (i.e., ctrl + f) for keywords related to your question. It's likely that I have already addressed your question in the comments.
  2. If you are copying and pasting code/terminal output, please don't. Reviewing another programmers’ code is a very time consuming and tedious task, and due to the volume of emails and contact requests I receive, I simply cannot do it.
  3. Be respectful of the space. I put a lot of my own personal time into creating these free weekly tutorials. On average, each tutorial takes me 15-20 hours to put together. I love offering these guides to you and I take pride in the content I create. Therefore, I will not approve comments that include large code blocks/terminal output as it destroys the formatting of the page. Kindly be respectful of this space.
  4. Be patient. I receive 200+ comments and emails per day. Due to spam, and my desire to personally answer as many questions as I can, I hand moderate all new comments (typically once per week). I try to answer as many questions as I can, but I'm only one person. Please don't be offended if I cannot get to your question
  5. Do you need priority support? Consider purchasing one of my books and courses. I place customer questions and emails in a separate, special priority queue and answer them first. If you are a customer of mine you will receive a guaranteed response from me. If there's any time left over, I focus on the community at large and attempt to answer as many of those questions as I possibly can.

Thank you for keeping these guidelines in mind before submitting your comment.

Leave a Reply