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
Photo by Living in Monrovia, some rights reserved.
Tutorial Overview
This tutorial is divided into 6 parts; they are:
- Photo and Caption Dataset
- Prepare Photo Data
- Prepare Text Data
- Develop Deep Learning Model
- Train With Progressive Loading (NEW)
- Evaluate Model
- 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:
|
1 2 3 4 5 6 |
# tensorflow version import tensorflow print('tensorflow: %s' % tensorflow.__version__) # keras version import keras print('keras: %s' % keras.__version__) |
Running the script should show the same library version numbers or higher.
|
1 2 |
tensorflow: 2.4.0 keras: 2.4.3 |
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:
- Flickr8k_Dataset.zip (1 Gigabyte) An archive of all photographs.
- Flickr8k_text.zip (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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 |
# extract features from each photo in the directory def extract_features(directory): # load the model model = VGG16() # re-structure the model model = Model(inputs=model.inputs, outputs=model.layers[-2].output) # summarize print(model.summary()) # extract features from each photo features = dict() for name in listdir(directory): # load an image from file filename = directory + '/' + name image = load_img(filename, target_size=(224, 224)) # convert the image pixels to a numpy array image = img_to_array(image) # reshape data for the model image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2])) # prepare the image for the VGG model image = preprocess_input(image) # get features feature = model.predict(image, verbose=0) # get image id image_id = name.split('.')[0] # store feature features[image_id] = feature print('>%s' % name) return 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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 |
from os import listdir from pickle import dump from keras.applications.vgg16 import VGG16 from keras.preprocessing.image import load_img from keras.preprocessing.image import img_to_array from keras.applications.vgg16 import preprocess_input from keras.models import Model # extract features from each photo in the directory def extract_features(directory): # load the model model = VGG16() # re-structure the model model = Model(inputs=model.inputs, outputs=model.layers[-2].output) # summarize print(model.summary()) # extract features from each photo features = dict() for name in listdir(directory): # load an image from file filename = directory + '/' + name image = load_img(filename, target_size=(224, 224)) # convert the image pixels to a numpy array image = img_to_array(image) # reshape data for the model image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2])) # prepare the image for the VGG model image = preprocess_input(image) # get features feature = model.predict(image, verbose=0) # get image id image_id = name.split('.')[0] # store feature features[image_id] = feature print('>%s' % name) return features # extract features from all images directory = 'Flickr8k_Dataset' features = extract_features(directory) print('Extracted Features: %d' % len(features)) # save to file dump(features, open('features.pkl', 'wb')) |
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 |
# load doc into memory def load_doc(filename): # open the file as read only file = open(filename, 'r') # read all text text = file.read() # close the file file.close() return text filename = 'Flickr8k_text/Flickr8k.token.txt' # load descriptions doc = load_doc(filename) |
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 |
# extract descriptions for images def load_descriptions(doc): mapping = dict() # process lines for line in doc.split('\n'): # split line by white space tokens = line.split() if len(line) < 2: continue # take the first token as the image id, the rest as the description image_id, image_desc = tokens[0], tokens[1:] # remove filename from image id image_id = image_id.split('.')[0] # convert description tokens back to string image_desc = ' '.join(image_desc) # create the list if needed if image_id not in mapping: mapping[image_id] = list() # store description mapping[image_id].append(image_desc) return mapping # parse descriptions descriptions = load_descriptions(doc) print('Loaded: %d ' % len(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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 |
import string def clean_descriptions(descriptions): # prepare translation table for removing punctuation table = str.maketrans('', '', string.punctuation) for key, desc_list in descriptions.items(): for i in range(len(desc_list)): desc = desc_list[i] # tokenize desc = desc.split() # convert to lower case desc = [word.lower() for word in desc] # remove punctuation from each token desc = [w.translate(table) for w in desc] # remove hanging 's' and 'a' desc = [word for word in desc if len(word)>1] # remove tokens with numbers in them desc = [word for word in desc if word.isalpha()] # store as string desc_list[i] = ' '.join(desc) # clean descriptions clean_descriptions(descriptions) |
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.
|
1 2 3 4 5 6 7 8 9 10 11 |
# convert the loaded descriptions into a vocabulary of words def to_vocabulary(descriptions): # build a list of all description strings all_desc = set() for key in descriptions.keys(): [all_desc.update(d.split()) for d in descriptions[key]] return all_desc # summarize vocabulary vocabulary = to_vocabulary(descriptions) print('Vocabulary Size: %d' % len(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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 |
# save descriptions to file, one per line def save_descriptions(descriptions, filename): lines = list() for key, desc_list in descriptions.items(): for desc in desc_list: lines.append(key + ' ' + desc) data = '\n'.join(lines) file = open(filename, 'w') file.write(data) file.close() # save descriptions save_descriptions(descriptions, 'descriptions.txt') |
Putting this all together, the complete listing is provided below.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 |
import string # load doc into memory def load_doc(filename): # open the file as read only file = open(filename, 'r') # read all text text = file.read() # close the file file.close() return text # extract descriptions for images def load_descriptions(doc): mapping = dict() # process lines for line in doc.split('\n'): # split line by white space tokens = line.split() if len(line) < 2: continue # take the first token as the image id, the rest as the description image_id, image_desc = tokens[0], tokens[1:] # remove filename from image id image_id = image_id.split('.')[0] # convert description tokens back to string image_desc = ' '.join(image_desc) # create the list if needed if image_id not in mapping: mapping[image_id] = list() # store description mapping[image_id].append(image_desc) return mapping def clean_descriptions(descriptions): # prepare translation table for removing punctuation table = str.maketrans('', '', string.punctuation) for key, desc_list in descriptions.items(): for i in range(len(desc_list)): desc = desc_list[i] # tokenize desc = desc.split() # convert to lower case desc = [word.lower() for word in desc] # remove punctuation from each token desc = [w.translate(table) for w in desc] # remove hanging 's' and 'a' desc = [word for word in desc if len(word)>1] # remove tokens with numbers in them desc = [word for word in desc if word.isalpha()] # store as string desc_list[i] = ' '.join(desc) # convert the loaded descriptions into a vocabulary of words def to_vocabulary(descriptions): # build a list of all description strings all_desc = set() for key in descriptions.keys(): [all_desc.update(d.split()) for d in descriptions[key]] return all_desc # save descriptions to file, one per line def save_descriptions(descriptions, filename): lines = list() for key, desc_list in descriptions.items(): for desc in desc_list: lines.append(key + ' ' + desc) data = '\n'.join(lines) file = open(filename, 'w') file.write(data) file.close() filename = 'Flickr8k_text/Flickr8k.token.txt' # load descriptions doc = load_doc(filename) # parse descriptions descriptions = load_descriptions(doc) print('Loaded: %d ' % len(descriptions)) # clean descriptions clean_descriptions(descriptions) # summarize vocabulary vocabulary = to_vocabulary(descriptions) print('Vocabulary Size: %d' % len(vocabulary)) # save to file save_descriptions(descriptions, 'descriptions.txt') |
Running the example first prints the number of loaded photo descriptions (8,092) and the size of the clean vocabulary (8,763 words).
|
1 2 |
Loaded: 8,092 Vocabulary Size: 8,763 |
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.
|
1 2 3 4 5 6 |
2252123185_487f21e336 bunch on people are seated in stadium 2252123185_487f21e336 crowded stadium is full of people watching an event 2252123185_487f21e336 crowd of people fill up packed stadium 2252123185_487f21e336 crowd sitting in an indoor stadium 2252123185_487f21e336 stadium full of people watch game ... |
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:
- Loading Data.
- Defining the Model.
- Fitting the Model.
- 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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 |
# load doc into memory def load_doc(filename): # open the file as read only file = open(filename, 'r') # read all text text = file.read() # close the file file.close() return text # load a pre-defined list of photo identifiers def load_set(filename): doc = load_doc(filename) dataset = list() # process line by line for line in doc.split('\n'): # skip empty lines if len(line) < 1: continue # get the image identifier identifier = line.split('.')[0] dataset.append(identifier) return set(dataset) |
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 |
# load clean descriptions into memory def load_clean_descriptions(filename, dataset): # load document doc = load_doc(filename) descriptions = dict() for line in doc.split('\n'): # split line by white space tokens = line.split() # split id from description image_id, image_desc = tokens[0], tokens[1:] # skip images not in the set if image_id in dataset: # create list if image_id not in descriptions: descriptions[image_id] = list() # wrap description in tokens desc = 'startseq ' + ' '.join(image_desc) + ' endseq' # store descriptions[image_id].append(desc) return descriptions |
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.
|
1 2 3 4 5 6 7 |
# load photo features def load_photo_features(filename, dataset): # load all features all_features = load(open(filename, 'rb')) # filter features features = {k: all_features[k] for k in dataset} return features |
We can pause here and test everything developed so far.
The complete code example is listed below.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 |
from pickle import load # load doc into memory def load_doc(filename): # open the file as read only file = open(filename, 'r') # read all text text = file.read() # close the file file.close() return text # load a pre-defined list of photo identifiers def load_set(filename): doc = load_doc(filename) dataset = list() # process line by line for line in doc.split('\n'): # skip empty lines if len(line) < 1: continue # get the image identifier identifier = line.split('.')[0] dataset.append(identifier) return set(dataset) # load clean descriptions into memory def load_clean_descriptions(filename, dataset): # load document doc = load_doc(filename) descriptions = dict() for line in doc.split('\n'): # split line by white space tokens = line.split() # split id from description image_id, image_desc = tokens[0], tokens[1:] # skip images not in the set if image_id in dataset: # create list if image_id not in descriptions: descriptions[image_id] = list() # wrap description in tokens desc = 'startseq ' + ' '.join(image_desc) + ' endseq' # store descriptions[image_id].append(desc) return descriptions # load photo features def load_photo_features(filename, dataset): # load all features all_features = load(open(filename, 'rb')) # filter features features = {k: all_features[k] for k in dataset} return features # load training dataset (6K) filename = 'Flickr8k_text/Flickr_8k.trainImages.txt' train = load_set(filename) print('Dataset: %d' % len(train)) # descriptions train_descriptions = load_clean_descriptions('descriptions.txt', train) print('Descriptions: train=%d' % len(train_descriptions)) # photo features train_features = load_photo_features('features.pkl', train) print('Photos: train=%d' % len(train_features)) |
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.
|
1 2 3 |
Dataset: 6,000 Descriptions: train=6,000 Photos: train=6,000 |
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 |
# convert a dictionary of clean descriptions to a list of descriptions def to_lines(descriptions): all_desc = list() for key in descriptions.keys(): [all_desc.append(d) for d in descriptions[key]] return all_desc # fit a tokenizer given caption descriptions def create_tokenizer(descriptions): lines = to_lines(descriptions) tokenizer = Tokenizer() tokenizer.fit_on_texts(lines) return tokenizer # prepare tokenizer tokenizer = create_tokenizer(train_descriptions) vocab_size = len(tokenizer.word_index) + 1 print('Vocabulary Size: %d' % vocab_size) |
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:
|
1 2 3 4 5 6 7 |
X1, X2 (text sequence), y (word) photo startseq, little photo startseq, little, girl photo startseq, little, girl, running photo startseq, little, girl, running, in photo startseq, little, girl, running, in, field photo startseq, little, girl, running, in, field, endseq |
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 |
# create sequences of images, input sequences and output words for an image def create_sequences(tokenizer, max_length, descriptions, photos, vocab_size): X1, X2, y = list(), list(), list() # walk through each image identifier for key, desc_list in descriptions.items(): # walk through each description for the image for desc in desc_list: # encode the sequence seq = tokenizer.texts_to_sequences([desc])[0] # split one sequence into multiple X,y pairs for i in range(1, len(seq)): # split into input and output pair in_seq, out_seq = seq[:i], seq[i] # pad input sequence in_seq = pad_sequences([in_seq], maxlen=max_length)[0] # encode output sequence out_seq = to_categorical([out_seq], num_classes=vocab_size)[0] # store X1.append(photos[key][0]) X2.append(in_seq) y.append(out_seq) return array(X1), array(X2), array(y) |
We will need to calculate the maximum number of words in the longest description. A short helper function named max_length() is defined below.
|
1 2 3 4 |
# calculate the length of the description with the most words def max_length(descriptions): lines = to_lines(descriptions) return max(len(d.split()) for d in lines) |
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:
- Where to put the Image in an Image Caption Generator, 2017.
- What is the Role of Recurrent Neural Networks (RNNs) in an Image Caption Generator?, 2017.
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
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 |
# define the captioning model def define_model(vocab_size, max_length): # feature extractor model inputs1 = Input(shape=(4096,)) fe1 = Dropout(0.5)(inputs1) fe2 = Dense(256, activation='relu')(fe1) # sequence model inputs2 = Input(shape=(max_length,)) se1 = Embedding(vocab_size, 256, mask_zero=True)(inputs2) se2 = Dropout(0.5)(se1) se3 = LSTM(256)(se2) # decoder model decoder1 = add([fe2, se3]) decoder2 = Dense(256, activation='relu')(decoder1) outputs = Dense(vocab_size, activation='softmax')(decoder2) # tie it together [image, seq] [word] model = Model(inputs=[inputs1, inputs2], outputs=outputs) model.compile(loss='categorical_crossentropy', optimizer='adam') # summarize model print(model.summary()) plot_model(model, to_file='model.png', show_shapes=True) return model |
To get a sense for the structure of the model, specifically the shapes of the layers, see the summary listed below.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 |
____________________________________________________________________________________________________ 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] lstm_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 ____________________________________________________________________________________________________ |
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
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.
|
1 2 3 |
# define checkpoint callback filepath = 'model-ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5' checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, save_best_only=True, mode='min') |
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.
|
1 2 |
# fit model model.fit([X1train, X2train], ytrain, epochs=20, verbose=2, callbacks=[checkpoint], validation_data=([X1test, X2test], ytest)) |
Complete Example
The complete example for fitting the model on the training data is listed below.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 |
from numpy import array from pickle import load from keras.preprocessing.text import Tokenizer from keras.preprocessing.sequence import pad_sequences from keras.utils import to_categorical from keras.utils import plot_model from keras.models import Model from keras.layers import Input from keras.layers import Dense from keras.layers import LSTM from keras.layers import Embedding from keras.layers import Dropout from keras.layers.merge import add from keras.callbacks import ModelCheckpoint # load doc into memory def load_doc(filename): # open the file as read only file = open(filename, 'r') # read all text text = file.read() # close the file file.close() return text # load a pre-defined list of photo identifiers def load_set(filename): doc = load_doc(filename) dataset = list() # process line by line for line in doc.split('\n'): # skip empty lines if len(line) < 1: continue # get the image identifier identifier = line.split('.')[0] dataset.append(identifier) return set(dataset) # load clean descriptions into memory def load_clean_descriptions(filename, dataset): # load document doc = load_doc(filename) descriptions = dict() for line in doc.split('\n'): # split line by white space tokens = line.split() # split id from description image_id, image_desc = tokens[0], tokens[1:] # skip images not in the set if image_id in dataset: # create list if image_id not in descriptions: descriptions[image_id] = list() # wrap description in tokens desc = 'startseq ' + ' '.join(image_desc) + ' endseq' # store descriptions[image_id].append(desc) return descriptions # load photo features def load_photo_features(filename, dataset): # load all features all_features = load(open(filename, 'rb')) # filter features features = {k: all_features[k] for k in dataset} return features # covert a dictionary of clean descriptions to a list of descriptions def to_lines(descriptions): all_desc = list() for key in descriptions.keys(): [all_desc.append(d) for d in descriptions[key]] return all_desc # fit a tokenizer given caption descriptions def create_tokenizer(descriptions): lines = to_lines(descriptions) tokenizer = Tokenizer() tokenizer.fit_on_texts(lines) return tokenizer # calculate the length of the description with the most words def max_length(descriptions): lines = to_lines(descriptions) return max(len(d.split()) for d in lines) # create sequences of images, input sequences and output words for an image def create_sequences(tokenizer, max_length, descriptions, photos, vocab_size): X1, X2, y = list(), list(), list() # walk through each image identifier for key, desc_list in descriptions.items(): # walk through each description for the image for desc in desc_list: # encode the sequence seq = tokenizer.texts_to_sequences([desc])[0] # split one sequence into multiple X,y pairs for i in range(1, len(seq)): # split into input and output pair in_seq, out_seq = seq[:i], seq[i] # pad input sequence in_seq = pad_sequences([in_seq], maxlen=max_length)[0] # encode output sequence out_seq = to_categorical([out_seq], num_classes=vocab_size)[0] # store X1.append(photos[key][0]) X2.append(in_seq) y.append(out_seq) return array(X1), array(X2), array(y) # define the captioning model def define_model(vocab_size, max_length): # feature extractor model inputs1 = Input(shape=(4096,)) fe1 = Dropout(0.5)(inputs1) fe2 = Dense(256, activation='relu')(fe1) # sequence model inputs2 = Input(shape=(max_length,)) se1 = Embedding(vocab_size, 256, mask_zero=True)(inputs2) se2 = Dropout(0.5)(se1) se3 = LSTM(256)(se2) # decoder model decoder1 = add([fe2, se3]) decoder2 = Dense(256, activation='relu')(decoder1) outputs = Dense(vocab_size, activation='softmax')(decoder2) # tie it together [image, seq] [word] model = Model(inputs=[inputs1, inputs2], outputs=outputs) model.compile(loss='categorical_crossentropy', optimizer='adam') # summarize model print(model.summary()) plot_model(model, to_file='model.png', show_shapes=True) return model # train dataset # load training dataset (6K) filename = 'Flickr8k_text/Flickr_8k.trainImages.txt' train = load_set(filename) print('Dataset: %d' % len(train)) # descriptions train_descriptions = load_clean_descriptions('descriptions.txt', train) print('Descriptions: train=%d' % len(train_descriptions)) # photo features train_features = load_photo_features('features.pkl', train) print('Photos: train=%d' % len(train_features)) # prepare tokenizer tokenizer = create_tokenizer(train_descriptions) vocab_size = len(tokenizer.word_index) + 1 print('Vocabulary Size: %d' % vocab_size) # determine the maximum sequence length max_length = max_length(train_descriptions) print('Description Length: %d' % max_length) # prepare sequences X1train, X2train, ytrain = create_sequences(tokenizer, max_length, train_descriptions, train_features, vocab_size) # dev dataset # load test set filename = 'Flickr8k_text/Flickr_8k.devImages.txt' test = load_set(filename) print('Dataset: %d' % len(test)) # descriptions test_descriptions = load_clean_descriptions('descriptions.txt', test) print('Descriptions: test=%d' % len(test_descriptions)) # photo features test_features = load_photo_features('features.pkl', test) print('Photos: test=%d' % len(test_features)) # prepare sequences X1test, X2test, ytest = create_sequences(tokenizer, max_length, test_descriptions, test_features, vocab_size) # fit model # define the model model = define_model(vocab_size, max_length) # define checkpoint callback filepath = 'model-ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5' checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, save_best_only=True, mode='min') # fit model model.fit([X1train, X2train], ytrain, epochs=20, verbose=2, callbacks=[checkpoint], validation_data=([X1test, X2test], ytest)) |
Running the example first prints a summary of the loaded training and development datasets.
|
1 2 3 4 5 6 7 8 |
Dataset: 6,000 Descriptions: train=6,000 Photos: train=6,000 Vocabulary Size: 7,579 Description Length: 34 Dataset: 1,000 Descriptions: test=1,000 Photos: test=1,000 |
After the summary of the model, we can get an idea of the total number of training and validation (development) input-output pairs.
|
1 |
Train on 306,404 samples, validate on 50,903 samples |
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:
|
1 |
Memory Error |
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.
|
1 2 3 4 5 6 7 8 9 |
# data generator, intended to be used in a call to model.fit_generator() def data_generator(descriptions, photos, tokenizer, max_length, vocab_size): # loop for ever over images while 1: for key, desc_list in descriptions.items(): # retrieve the photo feature photo = photos[key][0] in_img, in_seq, out_word = create_sequences(tokenizer, max_length, desc_list, photo, vocab_size) yield [in_img, in_seq], out_word |
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:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 |
# create sequences of images, input sequences and output words for an image def create_sequences(tokenizer, max_length, desc_list, photo, vocab_size): X1, X2, y = list(), list(), list() # walk through each description for the image for desc in desc_list: # encode the sequence seq = tokenizer.texts_to_sequences([desc])[0] # split one sequence into multiple X,y pairs for i in range(1, len(seq)): # split into input and output pair in_seq, out_seq = seq[:i], seq[i] # pad input sequence in_seq = pad_sequences([in_seq], maxlen=max_length)[0] # encode output sequence out_seq = to_categorical([out_seq], num_classes=vocab_size)[0] # store X1.append(photo) X2.append(in_seq) y.append(out_seq) return array(X1), array(X2), array(y) |
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:
|
1 2 3 4 5 6 |
# test the data generator generator = data_generator(train_descriptions, train_features, tokenizer, max_length, vocab_size) inputs, outputs = next(generator) print(inputs[0].shape) print(inputs[1].shape) print(outputs.shape) |
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.
|
1 2 3 |
(47, 4096) (47, 34) (47, 7579) |
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:
|
1 2 3 4 5 6 7 8 9 10 |
# train the model, run epochs manually and save after each epoch epochs = 20 steps = len(train_descriptions) for i in range(epochs): # create the data generator generator = data_generator(train_descriptions, train_features, tokenizer, max_length, vocab_size) # fit for one epoch model.fit_generator(generator, epochs=1, steps_per_epoch=steps, verbose=1) # save model model.save('model_' + str(i) + '.h5') |
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 |
from numpy import array from pickle import load from keras.preprocessing.text import Tokenizer from keras.preprocessing.sequence import pad_sequences from keras.utils import to_categorical from keras.utils import plot_model from keras.models import Model from keras.layers import Input from keras.layers import Dense from keras.layers import LSTM from keras.layers import Embedding from keras.layers import Dropout from keras.layers.merge import add from keras.callbacks import ModelCheckpoint # load doc into memory def load_doc(filename): # open the file as read only file = open(filename, 'r') # read all text text = file.read() # close the file file.close() return text # load a pre-defined list of photo identifiers def load_set(filename): doc = load_doc(filename) dataset = list() # process line by line for line in doc.split('\n'): # skip empty lines if len(line) < 1: continue # get the image identifier identifier = line.split('.')[0] dataset.append(identifier) return set(dataset) # load clean descriptions into memory def load_clean_descriptions(filename, dataset): # load document doc = load_doc(filename) descriptions = dict() for line in doc.split('\n'): # split line by white space tokens = line.split() # split id from description image_id, image_desc = tokens[0], tokens[1:] # skip images not in the set if image_id in dataset: # create list if image_id not in descriptions: descriptions[image_id] = list() # wrap description in tokens desc = 'startseq ' + ' '.join(image_desc) + ' endseq' # store descriptions[image_id].append(desc) return descriptions # load photo features def load_photo_features(filename, dataset): # load all features all_features = load(open(filename, 'rb')) # filter features features = {k: all_features[k] for k in dataset} return features # covert a dictionary of clean descriptions to a list of descriptions def to_lines(descriptions): all_desc = list() for key in descriptions.keys(): [all_desc.append(d) for d in descriptions[key]] return all_desc # fit a tokenizer given caption descriptions def create_tokenizer(descriptions): lines = to_lines(descriptions) tokenizer = Tokenizer() tokenizer.fit_on_texts(lines) return tokenizer # calculate the length of the description with the most words def max_length(descriptions): lines = to_lines(descriptions) return max(len(d.split()) for d in lines) # create sequences of images, input sequences and output words for an image def create_sequences(tokenizer, max_length, desc_list, photo, vocab_size): X1, X2, y = list(), list(), list() # walk through each description for the image for desc in desc_list: # encode the sequence seq = tokenizer.texts_to_sequences([desc])[0] # split one sequence into multiple X,y pairs for i in range(1, len(seq)): # split into input and output pair in_seq, out_seq = seq[:i], seq[i] # pad input sequence in_seq = pad_sequences([in_seq], maxlen=max_length)[0] # encode output sequence out_seq = to_categorical([out_seq], num_classes=vocab_size)[0] # store X1.append(photo) X2.append(in_seq) y.append(out_seq) return array(X1), array(X2), array(y) # define the captioning model def define_model(vocab_size, max_length): # feature extractor model inputs1 = Input(shape=(4096,)) fe1 = Dropout(0.5)(inputs1) fe2 = Dense(256, activation='relu')(fe1) # sequence model inputs2 = Input(shape=(max_length,)) se1 = Embedding(vocab_size, 256, mask_zero=True)(inputs2) se2 = Dropout(0.5)(se1) se3 = LSTM(256)(se2) # decoder model decoder1 = add([fe2, se3]) decoder2 = Dense(256, activation='relu')(decoder1) outputs = Dense(vocab_size, activation='softmax')(decoder2) # tie it together [image, seq] [word] model = Model(inputs=[inputs1, inputs2], outputs=outputs) # compile model model.compile(loss='categorical_crossentropy', optimizer='adam') # summarize model model.summary() plot_model(model, to_file='model.png', show_shapes=True) return model # data generator, intended to be used in a call to model.fit_generator() def data_generator(descriptions, photos, tokenizer, max_length, vocab_size): # loop for ever over images while 1: for key, desc_list in descriptions.items(): # retrieve the photo feature photo = photos[key][0] in_img, in_seq, out_word = create_sequences(tokenizer, max_length, desc_list, photo, vocab_size) yield [in_img, in_seq], out_word # load training dataset (6K) filename = 'Flickr8k_text/Flickr_8k.trainImages.txt' train = load_set(filename) print('Dataset: %d' % len(train)) # descriptions train_descriptions = load_clean_descriptions('descriptions.txt', train) print('Descriptions: train=%d' % len(train_descriptions)) # photo features train_features = load_photo_features('features.pkl', train) print('Photos: train=%d' % len(train_features)) # prepare tokenizer tokenizer = create_tokenizer(train_descriptions) vocab_size = len(tokenizer.word_index) + 1 print('Vocabulary Size: %d' % vocab_size) # determine the maximum sequence length max_length = max_length(train_descriptions) print('Description Length: %d' % max_length) # define the model model = define_model(vocab_size, max_length) # train the model, run epochs manually and save after each epoch epochs = 20 steps = len(train_descriptions) for i in range(epochs): # create the data generator generator = data_generator(train_descriptions, train_features, tokenizer, max_length, vocab_size) # fit for one epoch model.fit_generator(generator, epochs=1, steps_per_epoch=steps, verbose=1) # save model model.save('model_' + str(i) + '.h5') |
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.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 |
# map an integer to a word def word_for_id(integer, tokenizer): for word, index in tokenizer.word_index.items(): if index == integer: return word return None |