Consideration-based Picture Captioning with Keras


In picture captioning, an algorithm is given a picture and tasked with producing a smart caption. It’s a difficult process for a number of causes, not the least being that it includes a notion of saliency or relevance. That is why current deep studying approaches largely embody some “consideration” mechanism (typically even multiple) to assist specializing in related picture options.

On this submit, we reveal a formulation of picture captioning as an encoder-decoder downside, enhanced by spatial consideration over picture grid cells. The thought comes from a current paper on Neural Picture Caption Technology with Visible Consideration (Xu et al. 2015), and employs the identical type of consideration algorithm as detailed in our submit on machine translation.

We’re porting Python code from a current Google Colaboratory pocket book, utilizing Keras with TensorFlow keen execution to simplify our lives.

Stipulations

The code proven right here will work with the present CRAN variations of tensorflow, keras, and tfdatasets. Test that you just’re utilizing a minimum of model 1.9 of TensorFlow. If that isn’t the case, as of this writing, this

will get you model 1.10.

When loading libraries, please be sure you’re executing the primary 4 strains on this actual order. We’d like to verify we’re utilizing the TensorFlow implementation of Keras (tf.keras in Python land), and now we have to allow keen execution earlier than utilizing TensorFlow in any manner.

No must copy-paste any code snippets – you’ll discover the whole code (so as essential for execution) right here: eager-image-captioning.R.

The dataset

MS-COCO (“Widespread Objects in Context”) is one in every of, maybe the, reference dataset in picture captioning (object detection and segmentation, too). We’ll be utilizing the coaching pictures and annotations from 2014 – be warned, relying in your location, the obtain can take a lengthy time.

After unpacking, let’s outline the place the photographs and captions are.

annotation_file <- "train2014/annotations/captions_train2014.json"
image_path <- "train2014/train2014"

The annotations are in JSON format, and there are 414113 of them! Fortunately for us we didn’t must obtain that many pictures – each picture comes with 5 completely different captions, for higher generalizability.

annotations <- fromJSON(file = annotation_file)
annot_captions <- annotations[[4]]

num_captions <- size(annot_captions)

We retailer each annotations and picture paths in lists, for later loading.

all_captions <- vector(mode = "listing", size = num_captions)
all_img_names <- vector(mode = "listing", size = num_captions)

for (i in seq_len(num_captions)) {
  caption <- paste0("<begin> ",
                    annot_captions[[i]][["caption"]],
                    " <finish>"
                    )
  image_id <- annot_captions[[i]][["image_id"]]
  full_coco_image_path <- sprintf(
    "%s/COCO_train2014_percent012d.jpg",
    image_path,
    image_id
  )
  all_img_names[[i]] <- full_coco_image_path
  all_captions[[i]] <- caption
}

Relying in your computing atmosphere, you’ll for certain need to prohibit the variety of examples used. This submit will use 30000 captioned pictures, chosen randomly, and put aside 20% for validation.

Beneath, we take random samples, break up into coaching and validation elements. The companion code will even retailer the indices on disk, so you may choose up on verification and evaluation later.

num_examples <- 30000

random_sample <- pattern(1:num_captions, measurement = num_examples)
train_indices <- pattern(random_sample, measurement = size(random_sample) * 0.8)
validation_indices <- setdiff(random_sample, train_indices)

sample_captions <- all_captions[random_sample]
sample_images <- all_img_names[random_sample]
train_captions <- all_captions[train_indices]
train_images <- all_img_names[train_indices]
validation_captions <- all_captions[validation_indices]
validation_images <- all_img_names[validation_indices]

Interlude

Earlier than actually diving into the technical stuff, let’s take a second to mirror on this process. In typical image-related deep studying walk-throughs, we’re used to seeing well-defined issues – even when in some instances, the answer could also be onerous. Take, for instance, the stereotypical canine vs. cat downside. Some canine could seem like cats and a few cats could seem like canine, however that’s about it: All in all, within the ordinary world we stay in, it must be a kind of binary query.

If, however, we ask individuals to explain what they see in a scene, it’s to be anticipated from the outset that we’ll get completely different solutions. Nonetheless, how a lot consensus there may be will very a lot rely on the concrete dataset we’re utilizing.

Let’s check out some picks from the very first 20 coaching objects sampled randomly above.

Figure from MS-COCO 2014

Now this picture doesn’t depart a lot room for determination what to deal with, and obtained a really factual caption certainly: “There’s a plate with one slice of bacon a half of orange and bread.” If the dataset had been all like this, we’d assume a machine studying algorithm ought to do fairly nicely right here.

Selecting one other one from the primary 20:

Figure from MS-COCO 2014

What can be salient data to you right here? The caption supplied goes “A smiling little boy has a checkered shirt.” Is the look of the shirt as essential as that? You may as nicely deal with the surroundings, – and even one thing on a totally completely different degree: The age of the picture, or it being an analog one.

Let’s take a last instance.

From MS-COCO 2014

What would you say about this scene? The official label we sampled right here is “A gaggle of individuals posing in a humorous manner for the digital camera.” Properly …

Please don’t neglect that for every picture, the dataset contains 5 completely different captions (though our n = 30000 samples most likely gained’t). So this isn’t saying the dataset is biased – under no circumstances. As an alternative, we need to level out the ambiguities and difficulties inherent within the process. Really, given these difficulties, it’s all of the extra superb that the duty we’re tackling right here – having a community routinely generate picture captions – must be attainable in any respect!

Now let’s see how we will do that.

For the encoding a part of our encoder-decoder community, we’ll make use of InceptionV3 to extract picture options. In precept, which options to extract is as much as experimentation, – right here we simply use the final layer earlier than the totally linked prime:

image_model <- application_inception_v3(
  include_top = FALSE,
  weights = "imagenet"
)

For a picture measurement of 299×299, the output will likely be of measurement (batch_size, 8, 8, 2048), that’s, we’re making use of 2048 characteristic maps.

See also  Why Information Makes It Totally different – O’Reilly

InceptionV3 being a “huge mannequin,” the place each move by way of the mannequin takes time, we need to precompute options upfront and retailer them on disk. We’ll use tfdatasets to stream pictures to the mannequin. This implies all our preprocessing has to make use of tensorflow capabilities: That’s why we’re not utilizing the extra acquainted image_load from keras under.

Our customized load_image will learn in, resize and preprocess the photographs as required to be used with InceptionV3:

load_image <- perform(image_path) {
  img <-
    tf$read_file(image_path) %>%
    tf$picture$decode_jpeg(channels = 3) %>%
    tf$picture$resize_images(c(299L, 299L)) %>%
    tf$keras$functions$inception_v3$preprocess_input()
  listing(img, image_path)
}

Now we’re prepared to save lots of the extracted options to disk. The (batch_size, 8, 8, 2048)-sized options will likely be flattened to (batch_size, 64, 2048). The latter form is what our encoder, quickly to be mentioned, will obtain as enter.

preencode <- distinctive(sample_images) %>% unlist() %>% kind()
num_unique <- size(preencode)

# adapt this in accordance with your system's capacities  
batch_size_4save <- 1
image_dataset <-
  tensor_slices_dataset(preencode) %>%
  dataset_map(load_image) %>%
  dataset_batch(batch_size_4save)
  
save_iter <- make_iterator_one_shot(image_dataset)
  
until_out_of_range({
  
  save_count <- save_count + batch_size_4save
  batch_4save <- save_iter$get_next()
  img <- batch_4save[[1]]
  path <- batch_4save[[2]]
  batch_features <- image_model(img)
  batch_features <- tf$reshape(
    batch_features,
    listing(dim(batch_features)[1], -1L, dim(batch_features)[4]
  )
                               )
  for (i in 1:dim(batch_features)[1]) {
    np$save(path[i]$numpy()$decode("utf-8"),
            batch_features[i, , ]$numpy())
  }
    
})

Earlier than we get to the encoder and decoder fashions although, we have to maintain the captions.

Processing the captions

We’re utilizing keras text_tokenizer and the textual content processing capabilities texts_to_sequences and pad_sequences to rework ascii textual content right into a matrix.

# we'll use the 5000 most frequent phrases solely
top_k <- 5000
tokenizer <- text_tokenizer(
  num_words = top_k,
  oov_token = "<unk>",
  filters = '!"#$%&()*+.,-/:;[email protected][]^_`~ ')
tokenizer$fit_on_texts(sample_captions)

train_captions_tokenized <-
  tokenizer %>% texts_to_sequences(train_captions)
validation_captions_tokenized <-
  tokenizer %>% texts_to_sequences(validation_captions)

# pad_sequences will use 0 to pad all captions to the identical size
tokenizer$word_index["<pad>"] <- 

# create a lookup dataframe that enables us to go in each instructions
word_index_df <- knowledge.body(
  phrase = tokenizer$word_index %>% names(),
  index = tokenizer$word_index %>% unlist(use.names = FALSE),
  stringsAsFactors = FALSE
)
word_index_df <- word_index_df %>% organize(index)

decode_caption <- perform(textual content) {
  paste(map(textual content, perform(quantity)
    word_index_df %>%
      filter(index == quantity) %>%
      choose(phrase) %>%
      pull()),
    collapse = " ")
}

# pad all sequences to the identical size (the utmost size, in our case)
# may experiment with shorter padding (truncating the very longest captions)
caption_lengths <- map(
  all_captions[1:num_examples],
  perform(c) str_split(c," ")[[1]] %>% size()
  ) %>% unlist()
max_length <- fivenum(caption_lengths)[5]

train_captions_padded <-  pad_sequences(
  train_captions_tokenized,
  maxlen = max_length,
  padding = "submit",
  truncating = "submit"
)

validation_captions_padded <- pad_sequences(
  validation_captions_tokenized,
  maxlen = max_length,
  padding = "submit",
  truncating = "submit"
)

Loading the information for coaching

Now that we’ve taken care of pre-extracting the options and preprocessing the captions, we want a technique to stream them to our captioning mannequin. For that, we’re utilizing tensor_slices_dataset from tfdatasets, passing within the listing of paths to the photographs and the preprocessed captions. Loading the photographs is then carried out as a TensorFlow graph operation (utilizing tf$pyfunc).

The unique Colab code additionally shuffles the information on each iteration. Relying in your {hardware}, this will take a very long time, and given the dimensions of the dataset it isn’t strictly essential to get cheap outcomes. (The outcomes reported under had been obtained with out shuffling.)

batch_size <- 10
buffer_size <- num_examples

map_func <- perform(img_name, cap) {
  p <- paste0(img_name$decode("utf-8"), ".npy")
  img_tensor <- np$load(p)
  img_tensor <- tf$forged(img_tensor, tf$float32)
  listing(img_tensor, cap)
}

train_dataset <-
  tensor_slices_dataset(listing(train_images, train_captions_padded)) %>%
  dataset_map(
    perform(item1, item2) tf$py_func(map_func, listing(item1, item2), listing(tf$float32, tf$int32))
  ) %>%
  # optionally shuffle the dataset
  # dataset_shuffle(buffer_size) %>%
  dataset_batch(batch_size)

Captioning mannequin

The mannequin is mainly the identical as that mentioned within the machine translation submit. Please seek advice from that article for a proof of the ideas, in addition to an in depth walk-through of the tensor shapes concerned at each step. Right here, we offer the tensor shapes as feedback within the code snippets, for fast overview/comparability.

Nevertheless, if you happen to develop your personal fashions, with keen execution you may merely insert debugging/logging statements at arbitrary locations within the code – even in mannequin definitions. So you may have a perform

maybecat <- perform(context, x) {
  if (debugshapes) {
    title <- enexpr(x)
    dims <- paste0(dim(x), collapse = " ")
    cat(context, ": form of ", title, ": ", dims, "n", sep = "")
  }
}

And if you happen to now set

you may hint – not solely tensor shapes, however precise tensor values by way of your fashions, as proven under for the encoder. (We don’t show any debugging statements after that, however the pattern code has many extra.)

Encoder

Now it’s time to outline some some sizing-related hyperparameters and housekeeping variables:

# for encoder output
embedding_dim <- 256
# decoder (LSTM) capability
gru_units <- 512
# for decoder output
vocab_size <- top_k
# variety of characteristic maps gotten from Inception V3
features_shape <- 2048
# form of consideration options (flattened from 8x8)
attention_features_shape <- 64

The encoder on this case is only a totally linked layer, taking within the options extracted from Inception V3 (in flattened type, as they had been written to disk), and embedding them in 256-dimensional house.

cnn_encoder <- perform(embedding_dim, title = NULL) {
    
  keras_model_custom(title = title, perform(self) {
      
    self$fc <- layer_dense(models = embedding_dim, activation = "relu")
      
    perform(x, masks = NULL) {
      # enter form: (batch_size, 64, features_shape)
      maybecat("encoder enter", x)
      # form after fc: (batch_size, 64, embedding_dim)
      x <- self$fc(x)
      maybecat("encoder output", x)
      x
    }
  })
}

Consideration module

Not like within the machine translation submit, right here the eye module is separated out into its personal customized mannequin. The logic is identical although:

attention_module <- perform(gru_units, title = NULL) {
  
  keras_model_custom(title = title, perform(self) {
    
    self$W1 = layer_dense(models = gru_units)
    self$W2 = layer_dense(models = gru_units)
    self$V = layer_dense(models = 1)
      
    perform(inputs, masks = NULL) {
      options <- inputs[[1]]
      hidden <- inputs[[2]]
      # options(CNN_encoder output) form == (batch_size, 64, embedding_dim)
      # hidden form == (batch_size, gru_units)
      # hidden_with_time_axis form == (batch_size, 1, gru_units)
      hidden_with_time_axis <- k_expand_dims(hidden, axis = 2)
        
      # rating form == (batch_size, 64, 1)
      rating <- self$V(k_tanh(self$W1(options) + self$W2(hidden_with_time_axis)))
      # attention_weights form == (batch_size, 64, 1)
      attention_weights <- k_softmax(rating, axis = 2)
      # context_vector form after sum == (batch_size, embedding_dim)
      context_vector <- k_sum(attention_weights * options, axis = 2)
        
      listing(context_vector, attention_weights)
    }
  })
}

Decoder

The decoder at every time step calls the eye module with the options it bought from the encoder and its final hidden state, and receives again an consideration vector. The eye vector will get concatenated with the present enter and additional processed by a GRU and two totally linked layers, the final of which provides us the (unnormalized) chances for the subsequent phrase within the caption.

See also  Does this synthetic intelligence assume like a human? | MIT Information

The present enter at every time step right here is the earlier phrase: the proper one throughout coaching (trainer forcing), the final generated one throughout inference.

rnn_decoder <- perform(embedding_dim, gru_units, vocab_size, title = NULL) {
    
  keras_model_custom(title = title, perform(self) {
      
    self$gru_units <- gru_units
    self$embedding <- layer_embedding(input_dim = vocab_size, 
                                      output_dim = embedding_dim)
    self$gru <- if (tf$check$is_gpu_available()) {
      layer_cudnn_gru(
        models = gru_units,
        return_sequences = TRUE,
        return_state = TRUE,
        recurrent_initializer = 'glorot_uniform'
      )
    } else {
      layer_gru(
        models = gru_units,
        return_sequences = TRUE,
        return_state = TRUE,
        recurrent_initializer = 'glorot_uniform'
      )
    }
      
    self$fc1 <- layer_dense(models = self$gru_units)
    self$fc2 <- layer_dense(models = vocab_size)
      
    self$consideration <- attention_module(self$gru_units)
      
    perform(inputs, masks = NULL) {
      x <- inputs[[1]]
      options <- inputs[[2]]
      hidden <- inputs[[3]]
        
      c(context_vector, attention_weights) %<-% 
        self$consideration(listing(options, hidden))
        
      # x form after passing by way of embedding == (batch_size, 1, embedding_dim)
      x <- self$embedding(x)
        
      # x form after concatenation == (batch_size, 1, 2 * embedding_dim)
      x <- k_concatenate(listing(k_expand_dims(context_vector, 2), x))
        
      # passing the concatenated vector to the GRU
      c(output, state) %<-% self$gru(x)
        
      # form == (batch_size, 1, gru_units)
      x <- self$fc1(output)
        
      # x form == (batch_size, gru_units)
      x <- k_reshape(x, c(-1, dim(x)[[3]]))
        
      # output form == (batch_size, vocab_size)
      x <- self$fc2(x)
        
      listing(x, state, attention_weights)
        
    }
  })
}

Loss perform, and instantiating all of it

Now that we’ve outlined our mannequin (constructed of three customized fashions), we nonetheless want to really instantiate it (being exact: the 2 courses we’ll entry from exterior, that’s, the encoder and the decoder).

We additionally must instantiate an optimizer (Adam will do), and outline our loss perform (categorical crossentropy). Word that tf$nn$sparse_softmax_cross_entropy_with_logits expects uncooked logits as an alternative of softmax activations, and that we’re utilizing the sparse variant as a result of our labels aren’t one-hot-encoded.

encoder <- cnn_encoder(embedding_dim)
decoder <- rnn_decoder(embedding_dim, gru_units, vocab_size)

optimizer = tf$prepare$AdamOptimizer()

cx_loss <- perform(y_true, y_pred) {
  masks <- 1 - k_cast(y_true == 0L, dtype = "float32")
  loss <- tf$nn$sparse_softmax_cross_entropy_with_logits(
    labels = y_true,
    logits = y_pred
  ) * masks
  tf$reduce_mean(loss)
}

Coaching

Coaching the captioning mannequin is a time-consuming course of, and you’ll for certain need to save the mannequin’s weights! How does this work with keen execution?

We create a tf$prepare$Checkpoint object, passing it the objects to be saved: In our case, the encoder, the decoder, and the optimizer. Later, on the finish of every epoch, we’ll ask it to write down the respective weights to disk.

restore_checkpoint <- FALSE

checkpoint_dir <- "./checkpoints_captions"
checkpoint_prefix <- file.path(checkpoint_dir, "ckpt")
checkpoint <- tf$prepare$Checkpoint(
  optimizer = optimizer,
  encoder = encoder,
  decoder = decoder
)

As we’re simply beginning to prepare the mannequin, restore_checkpoint is ready to false. Later, restoring the weights will likely be as simple as

if (restore_checkpoint) {
  checkpoint$restore(tf$prepare$latest_checkpoint(checkpoint_dir))
}

The coaching loop is structured identical to within the machine translation case: We loop over epochs, batches, and the coaching targets, feeding within the right earlier phrase at each timestep. Once more, tf$GradientTape takes care of recording the ahead move and calculating the gradients, and the optimizer applies the gradients to the mannequin’s weights. As every epoch ends, we additionally save the weights.

num_epochs <- 20

if (!restore_checkpoint) {
  for (epoch in seq_len(num_epochs)) {
    
    total_loss <- 
    progress <- 
    train_iter <- make_iterator_one_shot(train_dataset)
    
    until_out_of_range({
      
      batch <- iterator_get_next(train_iter)
      loss <- 
      img_tensor <- batch[[1]]
      target_caption <- batch[[2]]
      
      dec_hidden <- k_zeros(c(batch_size, gru_units))
      
      dec_input <- k_expand_dims(
        rep(listing(word_index_df[word_index_df$word == "<start>", "index"]), 
            batch_size)
      )
      
      with(tf$GradientTape() %as% tape, {
        
        options <- encoder(img_tensor)
        
        for (t in seq_len(dim(target_caption)[2] - 1)) {
          c(preds, dec_hidden, weights) %<-%
            decoder(listing(dec_input, options, dec_hidden))
          loss <- loss + cx_loss(target_caption[, t], preds)
          dec_input <- k_expand_dims(target_caption[, t])
        }
        
      })
      
      total_loss <-
        total_loss + loss / k_cast_to_floatx(dim(target_caption)[2])
      
      variables <- c(encoder$variables, decoder$variables)
      gradients <- tape$gradient(loss, variables)
      
      optimizer$apply_gradients(purrr::transpose(listing(gradients, variables)),
                                global_step = tf$prepare$get_or_create_global_step()
      )
    })
    cat(paste0(
      "nnTotal loss (epoch): ",
      epoch,
      ": ",
      (total_loss / k_cast_to_floatx(buffer_size)) %>% as.double() %>% spherical(4),
      "n"
    ))
    
    checkpoint$save(file_prefix = checkpoint_prefix)
  }
}

Peeking at outcomes

Identical to within the translation case, it’s fascinating to have a look at mannequin efficiency throughout coaching. The companion code has that performance built-in, so you may watch mannequin progress for your self.

The essential perform right here is get_caption: It will get handed the trail to a picture, masses it, obtains its options from Inception V3, after which asks the encoder-decoder mannequin to generate a caption. If at any level the mannequin produces the finish image, we cease early. In any other case, we proceed till we hit the predefined most size.

Anderson et al. 2017) use object detection strategies to bottom-up isolate fascinating objects, and an LSTM stack whereby the primary LSTM computes top-down consideration guided by the output phrase generated by the second.

One other fascinating strategy involving consideration is utilizing a multimodal attentive translator (Liu et al. 2017), the place the picture options are encoded and offered in a sequence, such that we find yourself with sequence fashions each on the encoding and the decoding sides.

One other different is so as to add a discovered matter to the knowledge enter (Zhu, Xue, and Yuan 2018), which once more is a top-down characteristic present in human cognition.

In the event you discover one in every of these, or yet one more, strategy extra convincing, an keen execution implementation, within the fashion of the above, will possible be a sound manner of implementing it.

Anderson, Peter, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. 2017. “Backside-up and High-down Consideration for Picture Captioning and VQA.” CoRR abs/1707.07998. http://arxiv.org/abs/1707.07998.
Liu, Chang, Fuchun Solar, Changhu Wang, Feng Wang, and Alan L. Yuille. 2017. “A Multimodal Attentive Translator for Picture Captioning.” CoRR abs/1702.05658. http://arxiv.org/abs/1702.05658.
Xu, Kelvin, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. 2015. “Present, Attend and Inform: Neural Picture Caption Technology with Visible Consideration.” CoRR abs/1502.03044. http://arxiv.org/abs/1502.03044.
Zhu, Zhihao, Zhan Xue, and Zejian Yuan. 2018. “A Matter-Guided Consideration for Picture Captioning.” CoRR abs/1807.03514v1. https://arxiv.org/abs/1807.03514v1.

Leave a Reply