SALE! Use code blackfriday for 40% off everything!
Hurry, sale ends soon! Click to see the full catalog.

How to Develop a Deep Learning Photo Caption Generator from Scratch

Develop a Deep Learning Model to Automatically
Describe Photographs in Python with Keras, Step-by-Step.

Caption generation is a challenging artificial intelligence problem where a textual description must be generated for a given photograph.

It requires both methods from computer vision to understand the content of the image and a language model from the field of natural language processing to turn the understanding of the image into words in the right order. Recently, deep learning methods have achieved state-of-the-art results on examples of this problem.

Deep learning methods have demonstrated state-of-the-art results on caption generation problems. What is most impressive about these methods is a single end-to-end model can be defined to predict a caption, given a photo, instead of requiring sophisticated data preparation or a pipeline of specifically designed models.

In this tutorial, you will discover how to develop a photo captioning deep learning model from scratch.

After completing this tutorial, you will know:

  • How to prepare photo and text data for training a deep learning model.
  • How to design and train a deep learning caption generation model.
  • How to evaluate a train caption generation model and use it to caption entirely new photographs.

Kick-start your project with my new book Deep Learning for Natural Language Processing, including step-by-step tutorials and the Python source code files for all examples.

Let’s get started.

  • Update Nov/2017: Added note about a bug introduced in Keras 2.1.0 and 2.1.1 that impacts the code in this tutorial.
  • Update Dec/2017: Updated a typo in the function name when explaining how to save descriptions to file, thanks Minel.
  • Update Apr/2018: Added a new section that shows how to train the model using progressive loading for workstations with minimum RAM.
  • Update Feb/2019: Provided direct links for the Flickr8k_Dataset dataset, as the official site was taken down.
  • Update Jun/2019: Fixed typo in dataset name. Fixed minor bug in create_sequences().
  • Update Aug/2020: Update code for API changes in Keras 2.4.3 and TensorFlow 2.3.
  • Update Dec/2020: Added a section for checking library version numbers.
  • Update Dec/2020: Updated progressive loading to fix error “ValueError: No gradients provided for any variable“.
How to Develop a Deep Learning Caption Generation Model in Python from Scratch

How to Develop a Deep Learning Caption Generation Model in Python from Scratch
Photo by Living in Monrovia, some rights reserved.

Tutorial Overview

This tutorial is divided into 6 parts; they are:

  1. Photo and Caption Dataset
  2. Prepare Photo Data
  3. Prepare Text Data
  4. Develop Deep Learning Model
  5. Train With Progressive Loading (NEW)
  6. Evaluate Model
  7. Generate New Captions

Python Environment

This tutorial assumes you have a Python SciPy environment installed, ideally with Python 3.

You must have Keras installed with the TensorFlow backend. The tutorial also assumes you have the libraries NumPy and NLTK installed.

If you need help with your environment, see this tutorial:

I recommend running the code on a system with a GPU. You can access GPUs cheaply on Amazon Web Services. Learn how in this tutorial:

Before we move on, let’s check your deep learning library version.

Run the following script and check your version numbers:

Running the script should show the same library version numbers or higher.

Let’s dive in.

Need help with Deep Learning for Text Data?

Take my free 7-day email crash course now (with code).

Click to sign-up and also get a free PDF Ebook version of the course.

Photo and Caption Dataset

A good dataset to use when getting started with image captioning is the Flickr8K dataset.

The reason is because it is realistic and relatively small so that you can download it and build models on your workstation using a CPU.

The definitive description of the dataset is in the paper “Framing Image Description as a Ranking Task: Data, Models and Evaluation Metrics” from 2013.

The authors describe the dataset as follows:

We introduce a new benchmark collection for sentence-based image description and search, consisting of 8,000 images that are each paired with five different captions which provide clear descriptions of the salient entities and events.

The images were chosen from six different Flickr groups, and tend not to contain any well-known people or locations, but were manually selected to depict a variety of scenes and situations.

Framing Image Description as a Ranking Task: Data, Models and Evaluation Metrics, 2013.

The dataset is available for free. You must complete a request form and the links to the dataset will be emailed to you. I would love to link to them for you, but the email address expressly requests: “Please do not redistribute the dataset“.

You can use the link below to request the dataset (note, this may not work any more, see below):

Within a short time, you will receive an email that contains links to two files:

  • (1 Gigabyte) An archive of all photographs.
  • (2.2 Megabytes) An archive of all text descriptions for photographs.

UPDATE (Feb/2019): The official site seems to have been taken down (although the form still works). Here are some direct download links from my datasets GitHub repository:

Download the datasets and unzip them into your current working directory. You will have two directories:

  • Flickr8k_Dataset: Contains 8092 photographs in JPEG format.
  • Flickr8k_text: Contains a number of files containing different sources of descriptions for the photographs.

The dataset has a pre-defined training dataset (6,000 images), development dataset (1,000 images), and test dataset (1,000 images).

One measure that can be used to evaluate the skill of the model are BLEU scores. For reference, below are some ball-park BLEU scores for skillful models when evaluated on the test dataset (taken from the 2017 paper “Where to put the Image in an Image Caption Generator“):

  • BLEU-1: 0.401 to 0.578.
  • BLEU-2: 0.176 to 0.390.
  • BLEU-3: 0.099 to 0.260.
  • BLEU-4: 0.059 to 0.170.

We describe the BLEU metric more later when we work on evaluating our model.

Next, let’s look at how to load the images.

Prepare Photo Data

We will use a pre-trained model to interpret the content of the photos.

There are many models to choose from. In this case, we will use the Oxford Visual Geometry Group, or VGG, model that won the ImageNet competition in 2014. Learn more about the model here:

Keras provides this pre-trained model directly. Note, the first time you use this model, Keras will download the model weights from the Internet, which are about 500 Megabytes. This may take a few minutes depending on your internet connection.

We could use this model as part of a broader image caption model. The problem is, it is a large model and running each photo through the network every time we want to test a new language model configuration (downstream) is redundant.

Instead, we can pre-compute the “photo features” using the pre-trained model and save them to file. We can then load these features later and feed them into our model as the interpretation of a given photo in the dataset. It is no different to running the photo through the full VGG model; it is just we will have done it once in advance.

This is an optimization that will make training our models faster and consume less memory.

We can load the VGG model in Keras using the VGG class. We will remove the last layer from the loaded model, as this is the model used to predict a classification for a photo. We are not interested in classifying images, but we are interested in the internal representation of the photo right before a classification is made. These are the “features” that the model has extracted from the photo.

Keras also provides tools for reshaping the loaded photo into the preferred size for the model (e.g. 3 channel 224 x 224 pixel image).

Below is a function named extract_features() that, given a directory name, will load each photo, prepare it for VGG, and collect the predicted features from the VGG model. The image features are a 1-dimensional 4,096 element vector.

The function returns a dictionary of image identifier to image features.

We can call this function to prepare the photo data for testing our models, then save the resulting dictionary to a file named ‘features.pkl‘.

The complete example is listed below.

Running this data preparation step may take a while depending on your hardware, perhaps one hour on the CPU with a modern workstation.

At the end of the run, you will have the extracted features stored in ‘features.pkl‘ for later use. This file will be about 127 Megabytes in size.

Prepare Text Data

The dataset contains multiple descriptions for each photograph and the text of the descriptions requires some minimal cleaning.

If you are new to cleaning text data, see this post:

First, we will load the file containing all of the descriptions.

Each photo has a unique identifier. This identifier is used on the photo filename and in the text file of descriptions.

Next, we will step through the list of photo descriptions. Below defines a function load_descriptions() that, given the loaded document text, will return a dictionary of photo identifiers to descriptions. Each photo identifier maps to a list of one or more textual descriptions.

Next, we need to clean the description text. The descriptions are already tokenized and easy to work with.

We will clean the text in the following ways in order to reduce the size of the vocabulary of words we will need to work with:

  • Convert all words to lowercase.
  • Remove all punctuation.
  • Remove all words that are one character or less in length (e.g. ‘a’).
  • Remove all words with numbers in them.

Below defines the clean_descriptions() function that, given the dictionary of image identifiers to descriptions, steps through each description and cleans the text.

Once cleaned, we can summarize the size of the vocabulary.

Ideally, we want a vocabulary that is both expressive and as small as possible. A smaller vocabulary will result in a smaller model that will train faster.

For reference, we can transform the clean descriptions into a set and print its size to get an idea of the size of our dataset vocabulary.

Finally, we can save the dictionary of image identifiers and descriptions to a new file named descriptions.txt, with one image identifier and description per line.

Below defines the save_descriptions() function that, given a dictionary containing the mapping of identifiers to descriptions and a filename, saves the mapping to file.

Putting this all together, the complete listing is provided below.

Running the example first prints the number of loaded photo descriptions (8,092) and the size of the clean vocabulary (8,763 words).

Finally, the clean descriptions are written to ‘descriptions.txt‘.

Taking a look at the file, we can see that the descriptions are ready for modeling. The order of descriptions in your file may vary.

Develop Deep Learning Model

In this section, we will define the deep learning model and fit it on the training dataset.

This section is divided into the following parts:

  1. Loading Data.
  2. Defining the Model.
  3. Fitting the Model.
  4. Complete Example.

Loading Data

First, we must load the prepared photo and text data so that we can use it to fit the model.

We are going to train the data on all of the photos and captions in the training dataset. While training, we are going to monitor the performance of the model on the development dataset and use that performance to decide when to save models to file.

The train and development dataset have been predefined in the Flickr_8k.trainImages.txt and Flickr_8k.devImages.txt files respectively, that both contain lists of photo file names. From these file names, we can extract the photo identifiers and use these identifiers to filter photos and descriptions for each set.

The function load_set() below will load a pre-defined set of identifiers given the train or development sets filename.

Now, we can load the photos and descriptions using the pre-defined set of train or development identifiers.

Below is the function load_clean_descriptions() that loads the cleaned text descriptions from ‘descriptions.txt‘ for a given set of identifiers and returns a dictionary of identifiers to lists of text descriptions.

The model we will develop will generate a caption given a photo, and the caption will be generated one word at a time. The sequence of previously generated words will be provided as input. Therefore, we will need a ‘first word’ to kick-off the generation process and a ‘last word‘ to signal the end of the caption.

We will use the strings ‘startseq‘ and ‘endseq‘ for this purpose. These tokens are added to the loaded descriptions as they are loaded. It is important to do this now before we encode the text so that the tokens are also encoded correctly.

Next, we can load the photo features for a given dataset.

Below defines a function named load_photo_features() that loads the entire set of photo descriptions, then returns the subset of interest for a given set of photo identifiers.

This is not very efficient; nevertheless, this will get us up and running quickly.

We can pause here and test everything developed so far.

The complete code example is listed below.

Running this example first loads the 6,000 photo identifiers in the training dataset. These features are then used to filter and load the cleaned description text and the pre-computed photo features.

We are nearly there.

The description text will need to be encoded to numbers before it can be presented to the model as in input or compared to the model’s predictions.

The first step in encoding the data is to create a consistent mapping from words to unique integer values. Keras provides the Tokenizer class that can learn this mapping from the loaded description data.

Below defines the to_lines() to convert the dictionary of descriptions into a list of strings and the create_tokenizer() function that will fit a Tokenizer given the loaded photo description text.

We can now encode the text.

Each description will be split into words. The model will be provided one word and the photo and generate the next word. Then the first two words of the description will be provided to the model as input with the image to generate the next word. This is how the model will be trained.

For example, the input sequence “little girl running in field” would be split into 6 input-output pairs to train the model:

Later, when the model is used to generate descriptions, the generated words will be concatenated and recursively provided as input to generate a caption for an image.

The function below named create_sequences(), given the tokenizer, a maximum sequence length, and the dictionary of all descriptions and photos, will transform the data into input-output pairs of data for training the model. There are two input arrays to the model: one for photo features and one for the encoded text. There is one output for the model which is the encoded next word in the text sequence.

The input text is encoded as integers, which will be fed to a word embedding layer. The photo features will be fed directly to another part of the model. The model will output a prediction, which will be a probability distribution over all words in the vocabulary.

The output data will therefore be a one-hot encoded version of each word, representing an idealized probability distribution with 0 values at all word positions except the actual word position, which has a value of 1.

We will need to calculate the maximum number of words in the longest description. A short helper function named max_length() is defined below.

We now have enough to load the data for the training and development datasets and transform the loaded data into input-output pairs for fitting a deep learning model.

Defining the Model

We will define a deep learning based on the “merge-model” described by Marc Tanti, et al. in their 2017 papers:

For a gentle introduction to this architecture, see the post:

The authors provide a nice schematic of the model, reproduced below.

Schematic of the Merge Model For Image Captioning

Schematic of the Merge Model For Image Captioning

We will describe the model in three parts:

  • Photo Feature Extractor. This is a 16-layer VGG model pre-trained on the ImageNet dataset. We have pre-processed the photos with the VGG model (without the output layer) and will use the extracted features predicted by this model as input.
  • Sequence Processor. This is a word embedding layer for handling the text input, followed by a Long Short-Term Memory (LSTM) recurrent neural network layer.
  • Decoder (for lack of a better name). Both the feature extractor and sequence processor output a fixed-length vector. These are merged together and processed by a Dense layer to make a final prediction.

The Photo Feature Extractor model expects input photo features to be a vector of 4,096 elements. These are processed by a Dense layer to produce a 256 element representation of the photo.

The Sequence Processor model expects input sequences with a pre-defined length (34 words) which are fed into an Embedding layer that uses a mask to ignore padded values. This is followed by an LSTM layer with 256 memory units.

Both the input models produce a 256 element vector. Further, both input models use regularization in the form of 50% dropout. This is to reduce overfitting the training dataset, as this model configuration learns very fast.

The Decoder model merges the vectors from both input models using an addition operation. This is then fed to a Dense 256 neuron layer and then to a final output Dense layer that makes a softmax prediction over the entire output vocabulary for the next word in the sequence.

The function below named define_model() defines and returns the model ready to be fit.

To get a sense for the structure of the model, specifically the shapes of the layers, see the summary listed below.

We also create a plot to visualize the structure of the network that better helps understand the two streams of input.

Plot of the Caption Generation Deep Learning Model

Plot of the Caption Generation Deep Learning Model

Fitting the Model

Now that we know how to define the model, we can fit it on the training dataset.

The model learns fast and quickly overfits the training dataset. For this reason, we will monitor the skill of the trained model on the holdout development dataset. When the skill of the model on the development dataset improves at the end of an epoch, we will save the whole model to file.

At the end of the run, we can then use the saved model with the best skill on the training dataset as our final model.

We can do this by defining a ModelCheckpoint in Keras and specifying it to monitor the minimum loss on the validation dataset and save the model to a file that has both the training and validation loss in the filename.

We can then specify the checkpoint in the call to fit() via the callbacks argument. We must also specify the development dataset in fit() via the validation_data argument.

We will only fit the model for 20 epochs, but given the amount of training data, each epoch may take 30 minutes on modern hardware.

Complete Example

The complete example for fitting the model on the training data is listed below.

Running the example first prints a summary of the loaded training and development datasets.

After the summary of the model, we can get an idea of the total number of training and validation (development) input-output pairs.

The model then runs, saving the best model to .h5 files along the way.

On my run, the best validation results were saved to the file:

  • model-ep002-loss3.245-val_loss3.612.h5

This model was saved at the end of epoch 2 with a loss of 3.245 on the training dataset and a loss of 3.612 on the development dataset

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Let me know what you get in the comments below.

If you ran the example on AWS, copy the model file back to your current working directory. If you need help with commands on AWS, see the post:

Did you get an error like:

If so, see the next section.

Train With Progressive Loading

Note: If you had no problems in the previous section, please skip this section. This section is for those who do not have enough memory to train the model as described in the previous section (e.g. cannot use AWS EC2 for whatever reason).

The training of the caption model does assume you have a lot of RAM.

The code in the previous section is not memory efficient and assumes you are running on a large EC2 instance with 32GB or 64GB of RAM. If you are running the code on a workstation of 8GB of RAM, you cannot train the model.

A workaround is to use progressive loading. This was discussed in detail in the second-last section titled “Progressive Loading” in the post:

I recommend reading that section before continuing.

If you want to use progressive loading, to train this model, this section will show you how.

The first step is we must define a function that we can use as the data generator.

We will keep things very simple and have the data generator yield one photo’s worth of data per batch. This will be all of the sequences generated for a photo and its set of descriptions.

The function below data_generator() will be the data generator and will take the loaded textual descriptions, photo features, tokenizer and max length. Here, I assume that you can fit this training data in memory, which I believe 8GB of RAM should be more than capable.

How does this work? Read the post I just mentioned above that introduces data generators.

You can see that we are calling the create_sequence() function to create a batch worth of data for a single photo rather than an entire dataset. This means that we must update the create_sequences() function to delete the “iterate over all descriptions” for-loop.

The updated function is as follows:

We now have pretty much everything we need.

Note, this is a very basic data generator. The big memory saving it offers is to not have the unrolled sequences of train and test data in memory prior to fitting the model, that these samples (e.g. results from create_sequences()) are created as needed per photo.

Some off-the-cuff ideas for further improving this data generator include:

  • Randomize the order of photos each epoch.
  • Work with a list of photo ids and load text and photo data as needed to cut even further back on memory.
  • Yield more than one photo’s worth of samples per batch.

I have experienced with these variations myself in the past. Let me know if you do and how you go in the comments.

You can sanity check a data generator by calling it directly, as follows:

Running this sanity check will show what one batch worth of sequences looks like, in this case 47 samples to train on for the first photo.

Finally, we can use the fit_generator() function on the model to train the model with this data generator.

In this simple example we will discard the loading of the development dataset and model checkpointing and simply save the model after each training epoch. You can then go back and load/evaluate each saved model after training to find the one we the lowest loss that you can then use in the next section.

The code to train the model with the data generator is as follows:

That’s it. You can now train the model using progressive loading and save a ton of RAM. This may also be a lot slower.

The complete updated example with progressive loading (use of the data generator) for training the caption generation model is listed below.

Perhaps evaluate each saved model and choose the one final model with the lowest loss on a holdout dataset. The next section may help with this.

Did you use this new addition to the tutorial?
How did you go?

Evaluate Model

Once the model is fit, we can evaluate the skill of its predictions on the holdout test dataset.

We will evaluate a model by generating descriptions for all photos in the test dataset and evaluating those predictions with a standard cost function.

First, we need to be able to generate a description for a photo using a trained model.

This involves passing in the start description token ‘startseq‘, generating one word, then calling the model recursively with generated words as input until the end of sequence token is reached ‘endseq‘ or the maximum description length is reached.

The function below named generate_desc() implements this behavior and generates a textual description given a trained model, and a given prepared photo as input. It calls the function word_for_id() in order to map an integer prediction back to a word.

We will generate predictions for all photos in the test dataset and in the train dataset.

The function below named evaluate_model() will evaluate a trained model against a given dataset of photo descriptions and photo features. The actual and predicted descriptions are collected and evaluated collectively using the corpus BLEU score that summarizes how close the generated text is to the expected text.

BLEU scores are used in text translation for evaluating translated text against one or more reference translations.

Here, we compare each generated description against all of the reference descriptions for the photograph. We then calculate BLEU scores for 1, 2, 3 and 4 cumulative n-grams.

You can learn more about the BLEU score here:

The NLTK Python library implements the BLEU score calculation in the corpus_bleu() function. A higher score close to 1.0 is better, a score closer to zero is worse.

We can put all of this together with the functions from the previous section for loading the data. We first need to load the training dataset in order to prepare a Tokenizer so that we can encode generated words as input sequences for the model. It is critical that we encode the generated words using exactly the same encoding scheme as was used when training the model.

We then use these functions for loading the test dataset.

The complete example is listed below.

Running the example prints the BLEU scores.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

We can see that the scores fit within and close to the top of the expected range of a skillful model on the problem. The chosen model configuration is by no means optimized.

Generate New Captions

Now that we know how to develop and evaluate a caption generation model, how can we use it?

Almost everything we need to generate captions for entirely new photographs is in the model file.

We also need the Tokenizer for encoding generated words for the model while generating a sequence, and the maximum length of input sequences, used when we defined the model (e.g. 34).

We can hard code the maximum sequence length. With the encoding of text, we can create the tokenizer and save it to a file so that we can load it quickly whenever we need it without needing the entire Flickr8K dataset. An alternative would be to use our own vocabulary file and mapping to integers function during training.

We can create the Tokenizer as before and save it as a pickle file tokenizer.pkl. The complete example is listed below.

We can now load the tokenizer whenever we need it without having to load the entire training dataset of annotations.

Now, let’s generate a description for a new photograph.

Below is a new photograph that I chose randomly on Flickr (available under a permissive license).

Photo of a dog at the beach.

Photo of a dog at the beach.
Photo by bambe1964, some rights reserved.

We will generate a description for it using our model.

Download the photograph and save it to your local directory with the filename “example.jpg“.

First, we must load the Tokenizer from tokenizer.pkl and define the maximum length of the sequence to generate, needed for padding inputs.

Then we must load the model, as before.

Next, we must load the photo we which to describe and extract the features.

We could do this by re-defining the model and adding the VGG-16 model to it, or we can use the VGG model to predict the features and use them as inputs to our existing model. We will do the latter and use a modified version of the extract_features() function used during data preparation, but adapted to work on a single photo.

We can then generate a description using the generate_desc() function defined when evaluating the model.

The complete example for generating a description for an entirely new standalone photograph is listed below.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

In this case, the description generated was as follows:

You could remove the start and end tokens and you would have the basis for a nice automatic photo captioning model.

It’s like living in the future guys!

It still completely blows my mind that we can do this. Wow.


This section lists some ideas for extending the tutorial that you may wish to explore.

  • Alternate Pre-Trained Photo Models. A small 16-layer VGG model was used for feature extraction. Consider exploring larger models that offer better performance on the ImageNet dataset, such as Inception.
  • Smaller Vocabulary. A larger vocabulary of nearly eight thousand words was used in the development of the model. Many of the words supported may be misspellings or only used once in the entire dataset. Refine the vocabulary and reduce the size, perhaps by half.
  • Pre-trained Word Vectors. The model learned the word vectors as part of fitting the model. Better performance may be achieved by using word vectors either pre-trained on the training dataset or trained on a much larger corpus of text, such as news articles or Wikipedia.
  • Tune Model. The configuration of the model was not tuned on the problem. Explore alternate configurations and see if you can achieve better performance.

Did you try any of these extensions? Share your results in the comments below.

Further Reading

This section provides more resources on the topic if you are looking go deeper.

Caption Generation Papers

Flickr8K Dataset



In this tutorial, you discovered how to develop a photo captioning deep learning model from scratch.

Specifically, you learned:

  • How to prepare photo and text data ready for training a deep learning model.
  • How to design and train a deep learning caption generation model.
  • How to evaluate a train caption generation model and use it to caption entirely new photographs.

Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.

Develop Deep Learning models for Text Data Today!

Deep Learning for Natural Language Processing

Develop Your Own Text models in Minutes

...with just a few lines of python code

Discover how in my new Ebook:
Deep Learning for Natural Language Processing

It provides self-study tutorials on topics like:
Bag-of-Words, Word Embedding, Language Models, Caption Generation, Text Translation and much more...

Finally Bring Deep Learning to your Natural Language Processing Projects

Skip the Academics. Just Results.

See What's Inside

1,193 Responses to How to Develop a Deep Learning Photo Caption Generator from Scratch

  1. Avatar
    Christian Beckmann November 28, 2017 at 3:21 am #

    Hi Jason,

    thanks for this great article about image caption!

    My results after training were a bit worse (loss 3.566 – val_loss 3.859, then started to overfit) so i decided to try keras.applications.inception_v3.InceptionV3 for the base model. Currently it is still running and i am curious to see if it will do better.

    • Avatar
      Jason Brownlee November 28, 2017 at 8:41 am #

      Let me know how you go Christian.

      • Avatar
        zeeshan August 2, 2019 at 8:44 pm #

        hi jason m recieving this error can u please help me in this

        NameError: name ‘Flickr8k_Dataset’ is not defined

        • Avatar
          Jason Brownlee August 3, 2019 at 8:02 am #

          You may have missed a line of code or the dataset is not in the same directory as the python file.

          • Avatar
            Bhagyashree January 30, 2022 at 7:35 pm #

            Can you provide complete source code link without split code parts?
            please 🙂

          • Avatar
            James Carmichael January 31, 2022 at 10:52 am #

            Hello Bhagyashree…The tutorial contains full code listing that you may utilize.

      • Avatar
        mo December 16, 2020 at 7:54 pm #

        how to solve this , error happen


        6 generator = data_generator(train_descriptions, train_features, tokenizer, max_length, vocab_size)
        7 # fit for one epoch
        —-> 8 model.fit_generator( generator,epochs=1, steps_per_epoch=steps, verbose=1)

        • Avatar
          Jason Brownlee December 17, 2020 at 6:34 am #

          I don’t have enough context to comment, sorry.

          Perhaps these tips will help:

          • Avatar
            sharath May 19, 2021 at 2:23 am #

            Hello Jason
            I,m facing a value error could u help

            ValueError Traceback (most recent call last)
            in ()
            6 image_input=image_input.reshape(2048,)
            7 gen=generate(desc_dict,photo,max_length_of_caption,vocab_size,image_input)
            —-> 8,epochs=1,steps_per_epoch=6000,verbose=1)

            5 frames
            in create_sequence(caption, max_length_of_caption, vocab_size, image_input)
            1 def create_sequence(caption,max_length_of_caption,vocab_size,image_input):
            —-> 2 input_sequence=[],image_sequence=[],output_sequence=[]
            3 for caption in captions:
            4 caption=caption.split(‘ ‘)
            5 caption=[wordtoindex[w] for w in caption if w in vocab]

            ValueError: not enough values to unpack (expected 2, got 0)

        • Avatar
          asd February 2, 2021 at 12:54 am #

          Hey, did a find a solution? I’m facing the same error.

        • Avatar
          Mustafa Dar October 20, 2021 at 12:53 am #

          What accuracy are you getting in your NLP scores?

      • Avatar
        Rajat December 26, 2020 at 4:10 am #

        Hello Jason can you help me with the frontend part I tried using the flask app but failed

    • Avatar
      basil June 21, 2018 at 12:03 am #

      Christian / Jason – instead would Batch normalization help us here. am facing the same issue, over fitting.

      BN should also speed up the training and should also give us more accurate results. any inputs ?

      • Avatar
        Jason Brownlee June 21, 2018 at 6:18 am #

        The model usually does fit in 3-5 epochs.

        You can try batchnorm if you like. Not sure if it will help.

        • Avatar
          basil June 23, 2018 at 4:34 am #

          yep, i agree… not required..thanks..

          am also trying inceptionV3, let you know the results..

          • Avatar
            Jason Brownlee June 23, 2018 at 6:20 am #


          • Avatar
            Ben June 24, 2018 at 8:14 am #

            Hey did anyone try the Inception model? What were the results?

          • Avatar
            abbas November 18, 2018 at 3:37 am #

            hey ben!!!Can you please share the code and results of the inception model?so that we can also try and know more about the inception model.Thanks in advance

    • Avatar
      Shaurya Pratap Singh October 10, 2018 at 7:25 pm #

      can you plz send me the code at

      • Avatar
        Asad March 24, 2019 at 7:36 am #

        did you find code ?

    • Avatar
      Janarddan Sarkar November 24, 2018 at 1:16 am #

      I am getting the same

    • Avatar
      vishal July 6, 2020 at 3:05 am #

      i have tried using the inception v3 but the bleu scores are even than that of vgg16 model.
      BLEU-1: 0.514655
      BLEU-2: 0.266434
      BLEU-3: 0.179374
      BLEU-4: 0.078146

      • Avatar
        Jason Brownlee July 6, 2020 at 6:39 am #

        Nice work!

      • Avatar
        Rohit Kushwaha April 15, 2021 at 1:32 pm #

        i also tried Inception i got BLEU-1 0.571

      • Avatar
        afrid May 17, 2021 at 1:26 am #

        @vishal, can you share the inception v3 code ?

    • Avatar
      Karan Aggarwal June 13, 2021 at 3:55 am #

      Hello Christian Sir,

      To avoid overfit, you used keras.application.inceptionV3, m geeting some error in this line:

      print(‘Extracted Features: %d’ % len(features))

      TypeError Traceback (most recent call last)
      in ()
      —-> 1 print(‘Extracted Features: %d’ % len(features))

      TypeError: object of type ‘NoneType’ has no len()

      Please help in resolving this

    • Avatar
      Nagaraj CL April 12, 2022 at 1:07 pm #

      HI Christian, Please can you share working Inception V3 code, I am not able to make InceptionV3 model working, I am getting following error.

      Incompatible shapes: [47,8,8,256] vs. [47,256]
      [[{{node gradient_tape/model_10/add_7/add/BroadcastGradientArgs}}]] [Op:__inference_train_function_1153371]

  2. Avatar
    Akash November 30, 2017 at 4:56 am #

    Hi Jason,
    Once again great Article.
    I ran into some error while executing the code under “Complete example ” section.
    The error I got was
    ValueError: Error when checking target: expected dense_3 to have shape (None, 7579) but got array with shape (306404, 1)
    Any idea how to fix this?

    • Avatar
      Jason Brownlee November 30, 2017 at 8:26 am #

      Hi Akash, nice catch.

      The fault appears to have been introduced in a recent version of Keras in the to_categorical() function. I can confirm the fault occurs with Keras 2.1.1.

      You can learn more about the fault here:

      There are two options:

      1. Downgrade Keras to 2.0.8


      2. Modify the code, change line 104 in the training code example from:


      I hope that helps.

      • Avatar
        Akash November 30, 2017 at 5:38 pm #

        Thanks Jason. It’s working now.
        Can you suggest the changes to be made to use Inception model and word embedding like word2vec.

    • Avatar
      Gaurav Anand August 3, 2018 at 4:02 pm #

      Hi Akash

      Could you please tell how did you git rid of this problem?

      I am facing

      ValueError: Error when checking input: expected input_1 to have 2 dimensions, but got array with shape (11, 7, 7, 512)

      and after changing input structure to inputs1 = Input(shape=(7, 7, 512,)) I am facing

      ValueError: Error when checking target: expected dense_3 to have 4 dimensions, but got array with shape (11, 3857)

      I have tried with Keras 2.0.8 and latest 2.2.2 versions.
      Any help would be much appreciated.


      • Avatar
        anesh August 7, 2018 at 4:34 pm #

        Did you used different input shape?.If you changed the input shape then you have to flatten it and add fully connected dense layer of 4096 neurons.

        • Avatar
          Gaurav Anand August 14, 2018 at 2:59 pm #

          Should I avoid using “include_top = false” while feature extraction ?
          or keep it as true ?

        • Avatar
          abbas November 18, 2018 at 3:45 am #

          Anesh how to fix this error?

          Error when checking input: expected input_3 to have shape (4096,) but got array with shape (2048,)

          • Avatar
            Jason Brownlee November 18, 2018 at 6:48 am #

            Change the data to meet the model or change the model to meet the data.

  3. Avatar
    Zoltan November 30, 2017 at 11:47 pm #

    Hi Jason,

    Big thumbs up, nicely written, really informative article. I especially like the step by step approach.

    But when I tried to go through it, I got an error in load_poto_features saying that “name ‘load’ not defined”. Which is kinda odd.

    Otherwise everything seems fine.

    • Avatar
      Jason Brownlee December 1, 2017 at 7:35 am #


      Perhaps double check you have the load function imported from pickle?

  4. Avatar
    Bikram Kachari December 1, 2017 at 4:59 pm #

    Hi Jason

    I am a regular follower of your tutorials. They are great. I got to learn a lot. Thank you so much. Please keep up the good work

  5. Avatar
    maibam December 1, 2017 at 7:05 pm #

    Layer (type) Output Shape Param # Connected to
    input_2 (InputLayer) (None, 34) 0
    input_1 (InputLayer) (None, 4096) 0
    embedding_1 (Embedding) (None, 34, 256) 1940224 input_2[0][0]
    dropout_1 (Dropout) (None, 4096) 0 input_1[0][0]
    dropout_2 (Dropout) (None, 34, 256) 0 embedding_1[0][0]
    dense_1 (Dense) (None, 256) 1048832 dropout_1[0][0]
    lstm_1 (LSTM) (None, 256) 525312 dropout_2[0][0]
    add_1 (Add) (None, 256) 0 dense_1[0][0]
    dense_2 (Dense) (None, 256) 65792 add_1[0][0]
    dense_3 (Dense) (None, 7579) 1947803 dense_2[0][0]
    Total params: 5,527,963
    Trainable params: 5,527,963
    Non-trainable params: 0

    ValueError: Error when checking input: expected input_1 to have 2 dimensions, but got array with shape (306404, 7, 7, 512)

    Getting error during[X1train, X2train], ytrain, epochs=20, verbose=2, callbacks=[checkpoint], validation_data=([X1test, X2test], ytest))

    Keras 2.0.8 with tensorflow
    what is wrong ?

    • Avatar
      Jason Brownlee December 2, 2017 at 8:51 am #

      Not sure, did you copy all of the code exactly?

      Is your numpy and tensorflow also up to date?

      • Avatar
        Christian January 16, 2018 at 10:09 pm #

        This looks like he did change the network for feature extraction. When using include_top=False and wheigts=’imagenet” you get this type of data structure.

    • Avatar
      Kingson June 26, 2018 at 9:54 pm #

      @maibam did you find the solution?

      I am getting similar error –
      ValueError: Error when checking input: expected input_1 to have 2 dimensions, but got array with shape (17952, 7, 7, 512)

      Please help me out.

      • Avatar
        Jason Brownlee June 27, 2018 at 8:18 am #

        Ensure your version of Keras is up to date. v2.1.6 or better.

        • Avatar
          Kingson June 27, 2018 at 5:26 pm #

          Layer (type) Output Shape Param # Connected to
          input_2 (InputLayer) (None, 27) 0
          input_1 (InputLayer) (None, 4096) 0
          embedding_1 (Embedding) (None, 27, 256) 1058048 input_2[0][0]
          dropout_1 (Dropout) (None, 4096) 0 input_1[0][0]
          dropout_2 (Dropout) (None, 27, 256) 0 embedding_1[0][0]
          dense_1 (Dense) (None, 256) 1048832 dropout_1[0][0]
          lstm_1 (LSTM) (None, 256) 525312 dropout_2[0][0]
          add_1 (Add) (None, 256) 0 dense_1[0][0]
          dense_2 (Dense) (None, 256) 65792 add_1[0][0]
          dense_3 (Dense) (None, 4133) 1062181 dense_2[0][0]
          Total params: 3,760,165
          Trainable params: 3,760,165
          Non-trainable params: 0
          Traceback (most recent call last):
          File “”, line 179, in
[X1train, X2train], ytrain, epochs=20, verbose=2, callbacks=[checkpoint], validation_data=([X1test, X2test], ytest))

          ValueError: Error when checking input: expected input_1 to have 2 dimensions, but got array with shape (10931, 7, 7, 512)

          keras version is – 2.2.0

          Please help me out.

          • Avatar
            Jason Brownlee June 28, 2018 at 6:13 am #

            Looks like the dimensions of your data do not match the expectations of the model.

            You can change the data or change the model.

          • Avatar
            anesh August 7, 2018 at 4:36 pm #

            If you changed the input shape by include_top=False then you have to flatten it and add two FC dense layer of 4096 neurons.

  6. Avatar
    Vik December 2, 2017 at 7:16 pm #

    Thank you for the article. It is great to see full pipeline.
    Always following your articles with admiration

  7. Avatar
    Gonzalo Gasca Meza December 4, 2017 at 10:42 am #

    In the prepare data section, if using Python 2.7 there is no str.maketrans method.
    To make this work just comment that line and in line 46 do this:
    desc = [w.translate(None, string.punctuation) for w in desc]

    • Avatar
      Jason Brownlee December 4, 2017 at 4:57 pm #

      Thanks Gonzalo!

    • Avatar
      Bani March 8, 2018 at 4:26 am #

      after using the function to_vocabulary()
      I am getting a vocabulary of size 24 which is too less though I have followed the code line by line.
      Can u help?

      • Avatar
        Jason Brownlee March 8, 2018 at 6:36 am #

        Are you able to confirm that your Python is version 3.5+ and that you have the latest version of all libraries installed?

  8. Avatar
    Minel December 11, 2017 at 6:17 pm #

    Hi Jason,
    I am using your code step by step. There is a light mistake :
    you wrote
    # save descriptions
    save_doc(descriptions, ‘descriptions.txt’)

    in fact the right intruction is
    # save descriptions
    save_descriptions(descriptions, ‘descriptions.txt’)

    as you wrote in the final example

  9. Avatar
    Minel December 11, 2017 at 6:34 pm #

    Hi jason
    Another small detail. I had to write
    from pickle import load
    to run the instruction
    all_features = load(open(filename, ‘rb’))


  10. Avatar
    Minel December 11, 2017 at 9:32 pm #

    Hi Jason,
    I met some trouble running your code. I got a MemoryError on the instruction :
    return array(X1), array(X2), array(y)

    I am using a virtual machine with Linux (Debian), Python3, with 32Giga of memory.
    Could you tell me what was the size of the memory on the computer you used to check your program ?


  11. Avatar
    Minel December 12, 2017 at 11:34 pm #

    Thank for the advice.In fact, I upgraded the VM (64Go, 16 cores) and it worked fine (using 45Go of memory)

    • Avatar
      Jason Brownlee December 13, 2017 at 5:35 am #

      Nice! Glad to hear it.

      • Avatar
        Vineeth March 3, 2018 at 12:32 am #

        I get the same error even with 64GB VM :/ What to do

        • Avatar
          Jason Brownlee March 3, 2018 at 8:13 am #

          I’m sorry to hear that, perhaps there is something else going on with your workstation?

          I can confirm the example works on workstations and on EC2 instances with and without GPUs.

          • Avatar
            Vineeth March 3, 2018 at 10:06 pm #

            It’s throwing a Value error for input_1 after sometime. I tried everything i can but i am not able to understand. Can you paste the link of your project so i can compare ?

          • Avatar
            Jason Brownlee March 4, 2018 at 6:03 am #

            Are you able to confirm that your Python environment is up to date?

          • Avatar
            Vineeth March 3, 2018 at 10:26 pm #

            And sir, You said the pickle size must be about 127Mb but mine turns out to be above 700MB what did i do wrong ?

          • Avatar
            Jason Brownlee March 4, 2018 at 6:04 am #

            The size may be different on different platforms (macos/linux/windows).

  12. Avatar
    Josh Ash December 17, 2017 at 9:56 pm #

    Hi Jason – hello from Queensland 🙂
    Your tutorials on applied ML in Python are the best on the net hands down, thanks for putting them together!

  13. Avatar
    Madhivarman December 18, 2017 at 7:12 pm #

    hai Jason.. When i run the script my lap freeze…I don’t know whether its training or not.Did anyone face this issue ?


  14. Avatar
    Muhammad Awais December 20, 2017 at 3:36 pm #

    Thanks for such a great work. I found an error message when running a code
    FileNotFoundError: [Errno 2] No such file or directory: ‘descriptions.txt’
    Please help

    • Avatar
      Jason Brownlee December 20, 2017 at 3:50 pm #

      Ensure you generate the descriptions file before running the prior model – check the tutorial steps again and ensure you execute each in turn.

  15. Avatar
    Daniel F December 21, 2017 at 4:31 am #

    Hi Jason,

    I’m getting a MemoryError when I try to prepare the training sequences:

    Traceback (most recent call last):
    File “C:\Users\Daniel\Desktop\project\”, line 154, in
    X1train, X2train, ytrain = create_sequences(tokenizer, max_length, train_descriptions, train_features)
    File “C:\Users\Daniel\Desktop\project\”, line 104, in create_sequences
    out_seq = to_categorical([out_seq], num_classes=vocab_size)[0]
    File “C:\Program Files\Anaconda3\lib\site-packages\keras\utils\”, line 24, in to_categorical
    categorical = np.zeros((n, num_classes))

    any advice? I have 8GB of RAM.

  16. Avatar
    zonetrooper32 December 28, 2017 at 3:12 am #

    Hi Jason,

    Thank you for this amazing article about image captioning.

    Currently I am trying to re-implement the whole code, except that I am doing it in pure Tensorflow. I’m curious to see if my re-implementation is working as smooth as yours.

    Also a shower thought, it might be better to get a better vector representations for words if using the pretrained word2vec embeddings, for example Glove 6B or GoogleNews. Learning embeddings from scratch with only 8k words might have some performance loss.

    Again thank you for putting everything together, it will take quite some time to implement from scratch without your tutorial.

    • Avatar
      Jason Brownlee December 28, 2017 at 5:26 am #

      Try it and see if it lifts model skill. Let me know how you go.

  17. Avatar
    Sasikanth January 8, 2018 at 5:04 pm #

    Hello Jason,
    Is there a R package to perform modeling of images?


  18. Avatar
    Marco January 16, 2018 at 10:08 pm #

    Hi Jason! Thanks for your amazing tutorial! I have a question. I don’t understand the meaning of the number 1 on this line (extract_features):
    image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))

    Can you explain me what reshape does and the meaning of the arguments?

    Thanks in advance.

  19. Avatar
    junhyung yu January 22, 2018 at 8:54 pm #

    Hi Jason! thank you for your great code.
    but i have one question.

    How long does it take to execute under code?

    # define the model
    model = define_model(vocab_size, max_length)

    This code does not run during the third day.

    I think that “se3 = LSTM(256)(se2)” code in define_model function is causing the problem.

    My computer configuration is like this.

    Intel(R) Core(TM) i7-5820K CPU @ 3.30GHz – 6 core
    Ram 62G
    GeForce GTX TITAN X – 2core

    please help me~~

    • Avatar
      Jason Brownlee January 23, 2018 at 7:55 am #

      Ouch, something is wrong.

      Perhaps try running on AWS?

      Perhaps try other models and test your rig/setup?

      Perhaps try fewer epochs or a smaller model to see if your setup can train the model at all?

      • Avatar
        junhyung yu January 23, 2018 at 3:29 pm #

        1. No. i try running on my indicvdual linux server and using jupyter notebook

        2. No i am using only your code , no other model, no modify

        3.[X1train, X2train], ytrain, epochs=20, verbose=1, callbacks=[checkpoint], validation_data=([X1test, X2test], ytest))

        This code has not yet been executed

        so I do not think epoch is a problem.

        • Avatar
          Jason Brownlee January 24, 2018 at 9:50 am #

          Perhaps run from the command line as a background process without notebook?

          Perhaps check memory usage and cpu/gpu utilization?

  20. Avatar
    krishna January 23, 2018 at 10:41 pm #

    ConnectionResetError: [WinError 10054] An existing connection was forcibly closed by the remote host

    hi sir… I am getting this error above when i run feature extract code.

    • Avatar
      Jason Brownlee January 24, 2018 at 9:55 am #

      Sorry, I have not seen that error.

    • Avatar
      Hiroshi February 26, 2018 at 1:01 pm #

      Hi Krishna,

      I’m also getting this error time to time. Were you able to solve this issue?

    • Avatar
      anesh August 7, 2018 at 4:40 pm #

      You have to connect to the internet to download the vgg network.

  21. Avatar
    Sathiya_Chakra January 28, 2018 at 7:05 am #

    Hi Jason!

    Is it possible to run this neural network on a 8GB RAM laptop with 2GB Graphics card with Intel core i5 processor?

    • Avatar
      Jason Brownlee January 28, 2018 at 8:28 am #


      You might need to adjust it to use progressive loading so that it does not try to hold the entire dataset in RAM.

      • Avatar
        sandhya November 20, 2018 at 4:56 am #

        Hi jason

        Is it possible to run on cpu with progressive loading without any issues??

  22. Avatar
    Ajit Tiwari January 29, 2018 at 10:46 pm #

    Hi Jason,
    Can you provide a link for the tokenizer as well as the model file.
    I Cannot train this model in my system but would like to see if I can use it to create an Android app

  23. Avatar
    Soumya February 1, 2018 at 10:19 pm #

    When I am running

    tokenizer = Tokenizer()

    I am getting error,

    Traceback (most recent call last):
    File “”, line 1, in
    NameError: name ‘Tokenizer’ is not defined

    How to solve this. Any idea please.

  24. Avatar
    Marco February 9, 2018 at 12:41 am #

    Hi Jason, thanks for the tutorial! I want to ask you if you could explain (or send me some links), to better understand, how exactly the fitting works.

    Example description: the girl is …

    The LSTM network during fitting takes the beginning of the sequence of my description (startseq) and it produces a vector with all possible subsequent words. This vector is combined with the vector of the input image features and it is passed within an FF layer where we then take the most probable word (with softmax). it’s right?

    At this point how does the fitting go on? Is the new sequence (e.g startseq – the) passed into the LSTM network, predicts all possible next words, etc.? Continuing this way up to endseq?

    If the network incorrectly generates the next word, what happens? How are the weights arranged? The fitting continues by taking in input “startseq – wrong_word” or continues with the correct one (eg startseq – the)?

    Thanks for your help

  25. Avatar
    Sumit Das February 13, 2018 at 6:10 pm #

    Hi Jason great article on caption generator i think the best till now available online.. i am a newbee in ML(AI). i extracted the features and stored it to features.pkl file but getting an error on create sequence functions memory error and i can see you have suggested progressive loading i do not get that properly could you suggest my how to use the current code modified for progressive loading::

    [‎2/‎13/‎2018 12:34 PM] Sanchawat, Hardik:
    Using TensorFlow backend.
    Dataset: 6000
    Descriptions: train=6000
    Photos: train=6000
    Vocabulary Size: 7579
    Description Length: 34
    Traceback (most recent call last):
    File “C:\Users\hardik.sanchawat\Documents\Scripts\flickr\”, line 154, in
    X1train, X2train, ytrain = create_sequences(tokenizer, max_length, train_descriptions, train_features)
    File “C:\Users\hardik.sanchawat\Documents\Scripts\flickr\”, line 109, in create_sequences
    return array(X1), array(X2), array(y)

    My system configuration is :

    OS: Windows 10
    Processor: AMD A8 PRO-7150B R5, 10 Compute Cores 4C+6G 1.90 GHz
    Memory(RAM): 16 GB (14.9GB Usable)
    System type: 64-bit OS, x64-based processor

  26. Avatar
    Kavya February 14, 2018 at 8:35 am