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 entails 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 concept 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.
Verify that you just’re utilizing not less than 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 to’re executing the primary 4 traces on this precise 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 method.
No have to copy-paste any code snippets – you’ll discover the entire code (so as needed for execution) right here: eager-image-captioning.R.
The dataset
MS-COCO (“Widespread Objects in Context”) is one among, 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 should obtain that many pictures – each picture comes with 5 totally 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 = "record", size = num_captions)
all_img_names <- vector(mode = "record", 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 surroundings, you’ll for certain need to limit the variety of examples used.
This submit will use 30000 captioned pictures, chosen randomly, and put aside 20% for validation.
Under, we take random samples, cut up into coaching and validation elements. The companion code will even retailer the indices on disk, so you possibly can decide up on verification and evaluation later.
num_examples <- 30000
random_sample <- pattern(1:num_captions, dimension = num_examples)
train_indices <- pattern(random_sample, dimension = 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 might seem like cats and a few cats might seem like canine, however that’s about it: All in all, within the traditional world we stay in, it ought to 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 totally different solutions. Nonetheless, how a lot consensus there’s will very a lot rely upon the concrete dataset we’re utilizing.
Let’s check out some picks from the very first 20 coaching objects sampled randomly above.
Now this picture doesn’t go away 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 have been all like this, we’d assume a machine studying algorithm ought to do fairly nicely right here.
Choosing one other one from the primary 20:
What could be salient data to you right here? The caption offered goes “A smiling little boy has a checkered shirt.”
Is the look of the shirt as vital as that? You may as nicely deal with the surroundings, – and even one thing on a totally totally different stage: The age of the picture, or it being an analog one.
Let’s take a ultimate instance.
What would you say about this scene? The official label we sampled right here is “A bunch of individuals posing in a humorous method for the digicam.” Properly …
Please don’t overlook that for every picture, the dataset consists of 5 totally different captions (though our n = 30000 samples most likely received’t).
So this isn’t saying the dataset is biased – by no means. As an alternative, we need to level out the ambiguities and difficulties inherent within the process. Truly, given these difficulties, it’s all of the extra superb that the duty we’re tackling right here – having a community mechanically generate picture captions – ought to be potential 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 dimension of 299×299, the output will likely be of dimension (batch_size, 8, 8, 2048), that’s, we’re making use of 2048 characteristic maps.
InceptionV3 being a “huge mannequin,” the place each go by the mannequin takes time, we need to precompute options prematurely 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 beneath.
Our customized load_image will learn in, resize and preprocess the photographs as required to be used with InceptionV3:
Now we’re prepared to avoid wasting 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 line 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,
record(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 care for the captions.
Processing the captions
We’re utilizing keras text_tokenizer and the textual content processing capabilities texts_to_sequences and pad_sequences to remodel 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 = '!"#$%&()*+.,-/:;=?@[]^_`~ ')
tokenizer$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$word_index["<pad>"] <- 0
# create a lookup dataframe that enables us to go in each instructions
word_index_df <- information.body(
phrase = tokenizer$word_index %>% names(),
index = tokenizer$word_index %>% unlist(use.names = FALSE),
stringsAsFactors = FALSE
)
word_index_df <- word_index_df %>% organize(index)
decode_caption <- perform(textual content) {
paste(map(textual content, perform(quantity)
word_index_df %>%
filter(index == quantity) %>%
choose(phrase) %>%
pull()),
collapse = " ")
}
# pad all sequences to the identical size (the utmost size, in our case)
# may experiment with shorter padding (truncating the very longest captions)
caption_lengths <- map(
all_captions[1:num_examples],
perform(c) str_split(c," ")[[1]] %>% size()
) %>% unlist()
max_length <- fivenum(caption_lengths)[5]
train_captions_padded <- pad_sequences(
train_captions_tokenized,
maxlen = max_length,
padding = "submit",
truncating = "submit"
)
validation_captions_padded <- pad_sequences(
validation_captions_tokenized,
maxlen = max_length,
padding = "submit",
truncating = "submit"
)Loading the info for coaching
Now that we’ve taken care of pre-extracting the options and preprocessing the captions, we want a approach to stream them to our captioning mannequin. For that, we’re utilizing tensor_slices_dataset from tfdatasets, passing within the record of paths to the photographs and the preprocessed captions. Loading the photographs is then carried out as a TensorFlow graph operation (utilizing tf$pyfunc).
The unique Colab code additionally shuffles the info on each iteration. Relying in your {hardware}, this may increasingly take a very long time, and given the dimensions of the dataset it’s not strictly essential to get affordable outcomes. (The outcomes reported beneath have 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$solid(img_tensor, tf$float32)
record(img_tensor, cap)
}
train_dataset <-
tensor_slices_dataset(record(train_images, train_captions_padded)) %>%
dataset_map(
perform(item1, item2) tf$py_func(map_func, record(item1, item2), record(tf$float32, tf$int32))
) %>%
# optionally shuffle the dataset
# dataset_shuffle(buffer_size) %>%
dataset_batch(batch_size)Captioning mannequin
The mannequin is mainly the identical as that mentioned within the machine translation submit. Please consult with that article for an evidence of the ideas, in addition to an in depth walk-through of the tensor shapes concerned at each step. Right here, we offer the tensor shapes as feedback within the code snippets, for fast overview/comparability.
Nevertheless, in the event you develop your personal fashions, with keen execution you possibly can merely insert debugging/logging statements at arbitrary locations within the code – even in mannequin definitions. So you possibly can have a perform
And in the event you now set
you possibly can hint – not solely tensor shapes, however precise tensor values by your fashions, as proven beneath for the encoder. (We don’t show any debugging statements after that, however the pattern code has many extra.)
Encoder
Now it’s time to outline some some sizing-related hyperparameters and housekeeping variables:
# for encoder output
embedding_dim <- 256
# decoder (LSTM) capability
gru_units <- 512
# for decoder output
vocab_size <- top_k
# variety of characteristic maps gotten from Inception V3
features_shape <- 2048
# form of consideration options (flattened from 8x8)
attention_features_shape <- 64The encoder on this case is only a totally linked layer, taking within the options extracted from Inception V3 (in flattened type, as they have been written to disk), and embedding them in 256-dimensional area.
cnn_encoder <- perform(embedding_dim, identify = NULL) {
keras_model_custom(identify = identify, perform(self) {
self$fc <- layer_dense(items = 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$fc(x)
maybecat("encoder output", x)
x
}
})
}Consideration module
Not like within the machine translation submit, right here the eye module is separated out into its personal customized mannequin.
The logic is identical although:
attention_module <- perform(gru_units, identify = NULL) {
keras_model_custom(identify = identify, perform(self) {
self$W1 = layer_dense(items = gru_units)
self$W2 = layer_dense(items = gru_units)
self$V = layer_dense(items = 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$W2(hidden_with_time_axis)))
# attention_weights form == (batch_size, 64, 1)
attention_weights <- k_softmax(rating, axis = 2)
# context_vector form after sum == (batch_size, embedding_dim)
context_vector <- k_sum(attention_weights * options, axis = 2)
record(context_vector, attention_weights)
}
})
}Decoder
The decoder at every time step calls the eye module with the options it bought from the encoder and its final hidden state, and receives again an consideration vector. The eye vector will get concatenated with the present enter and additional processed by a GRU and two totally linked layers, the final of which provides us the (unnormalized) possibilities for the following phrase within the caption.
The present enter at every time step right here is the earlier phrase: the right one throughout coaching (instructor forcing), the final generated one throughout inference.
rnn_decoder <- perform(embedding_dim, gru_units, vocab_size, identify = NULL) {
keras_model_custom(identify = identify, perform(self) {
self$gru_units <- gru_units
self$embedding <- layer_embedding(input_dim = vocab_size,
output_dim = embedding_dim)
self$gru <- if (tf$take a look at$is_gpu_available()) {
layer_cudnn_gru(
items = gru_units,
return_sequences = TRUE,
return_state = TRUE,
recurrent_initializer = 'glorot_uniform'
)
} else {
layer_gru(
items = gru_units,
return_sequences = TRUE,
return_state = TRUE,
recurrent_initializer = 'glorot_uniform'
)
}
self$fc1 <- layer_dense(items = self$gru_units)
self$fc2 <- layer_dense(items = vocab_size)
self$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(record(options, hidden))
# x form after passing by embedding == (batch_size, 1, embedding_dim)
x <- self$embedding(x)
# x form after concatenation == (batch_size, 1, 2 * embedding_dim)
x <- k_concatenate(record(k_expand_dims(context_vector, 2), x))
# passing the concatenated vector to the GRU
c(output, state) %<-% self$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)
record(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 lessons we’ll entry from outdoors, that’s, the encoder and the decoder).
We additionally have to instantiate an optimizer (Adam will do), and outline our loss perform (categorical crossentropy).
Notice that tf$nn$sparse_softmax_cross_entropy_with_logits expects uncooked logits as a substitute of softmax activations, and that we’re utilizing the sparse variant as a result of our labels should not 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$Checkpoint object, passing it the objects to be saved: In our case, the encoder, the decoder, and the optimizer. Later, on the finish of every epoch, we’ll ask it to put in writing the respective weights to disk.
restore_checkpoint <- FALSE
checkpoint_dir <- "./checkpoints_captions"
checkpoint_prefix <- file.path(checkpoint_dir, "ckpt")
checkpoint <- tf$prepare$Checkpoint(
optimizer = optimizer,
encoder = encoder,
decoder = decoder
)As we’re simply beginning to prepare the mannequin, restore_checkpoint is about to false. Later, restoring the weights will likely be as straightforward as
if (restore_checkpoint) {
checkpoint$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$GradientTape takes care of recording the ahead go and calculating the gradients, and the optimizer applies the gradients to the mannequin’s weights.
As every epoch ends, we additionally save the weights.
num_epochs <- 20
if (!restore_checkpoint) {
for (epoch in seq_len(num_epochs)) {
total_loss <- 0
progress <- 0
train_iter <- make_iterator_one_shot(train_dataset)
until_out_of_range({
batch <- iterator_get_next(train_iter)
loss <- 0
img_tensor <- batch[[1]]
target_caption <- batch[[2]]
dec_hidden <- k_zeros(c(batch_size, gru_units))
dec_input <- k_expand_dims(
rep(record(word_index_df[word_index_df$word == "<start>", "index"]),
batch_size)
)
with(tf$GradientTape() %as% tape, {
options <- encoder(img_tensor)
for (t in seq_len(dim(target_caption)[2] - 1)) {
c(preds, dec_hidden, weights) %<-%
decoder(record(dec_input, options, dec_hidden))
loss <- loss + cx_loss(target_caption[, t], preds)
dec_input <- k_expand_dims(target_caption[, t])
}
})
total_loss <-
total_loss + loss / k_cast_to_floatx(dim(target_caption)[2])
variables <- c(encoder$variables, decoder$variables)
gradients <- tape$gradient(loss, variables)
optimizer$apply_gradients(purrr::transpose(record(gradients, variables)),
global_step = tf$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 take a look at mannequin efficiency throughout coaching. The companion code has that performance built-in, so you possibly can watch mannequin progress for your self.
The fundamental perform right here is get_caption: It will get handed the trail to a picture, hundreds 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.
get_caption <-
perform(picture) {
attention_matrix <-
matrix(0, nrow = max_length, ncol = attention_features_shape)
temp_input <- k_expand_dims(load_image(picture)[[1]], 1)
img_tensor_val <- image_model(temp_input)
img_tensor_val <- k_reshape(
img_tensor_val,
record(dim(img_tensor_val)[1], -1, dim(img_tensor_val)[4])
)
options <- encoder(img_tensor_val)
dec_hidden <- k_zeros(c(1, gru_units))
dec_input <-
k_expand_dims(
record(word_index_df[word_index_df$word == "<start>", "index"])
)
outcome <- ""
for (t in seq_len(max_length - 1)) {
c(preds, dec_hidden, attention_weights) %<-%
decoder(record(dec_input, options, dec_hidden))
attention_weights <- k_reshape(attention_weights, c(-1))
attention_matrix[t,] <- attention_weights %>% as.double()
pred_idx <- tf$multinomial(exp(preds), num_samples = 1)[1, 1]
%>% as.double()
pred_word <-
word_index_df[word_index_df$index == pred_idx, "word"]
if (pred_word == "<finish>") {
outcome <-
paste(outcome, pred_word)
attention_matrix <-
attention_matrix[1:length(str_split(result, " ")[[1]]), ,
drop = FALSE]
return (record(outcome, attention_matrix))
} else {
outcome <-
paste(outcome, pred_word)
dec_input <- k_expand_dims(record(pred_idx))
}
}
record(str_trim(outcome), attention_matrix)
}With that performance, now let’s really do this: peek at outcomes whereas the community is studying!
We’ve picked 3 examples every from the coaching and validation units. Right here they’re.
First, our picks from the coaching set:
Let’s see the goal captions:
- a herd of giraffe standing on prime of a grass coated discipline
- a view of playing cards driving down a avenue
- the skateboarding flips his board off of the sidewalk
Curiously, right here we even have an indication of how labeled datasets (like something human) might include errors. (The samples weren’t picked for that; as a substitute, they have been chosen – with out an excessive amount of screening – for being slightly unequivocal of their visible content material.)
Now for the validation candidates.
and their official captions:
- a left handed pitcher throwing the bottom ball
- a girl taking a chunk of a slice of pizza in a restaraunt
- a girl hitting swinging a tennis racket at a tennis ball on a tennis court docket
(Once more, any spelling peculiarities haven’t been launched by us.)
Epoch 1
Now, what does our community produce after the primary epoch? Keep in mind that this implies, having seen every one of many 24000 coaching pictures as soon as.
First then, listed below are the captions for the prepare pictures:
a gaggle of sheep standing within the grass
a gaggle of vehicles driving down a avenue
a person is standing on a avenue
Not solely is the syntax right in each case, the content material isn’t that unhealthy both!
How in regards to the validation set?
a baseball participant is enjoying baseball uniform is holding a baseball bat
a person is holding a desk with a desk with a desk with a desk with a desk with a desk with a desk with a desk with a desk with a desk with a desk with a desk with a desk with a desk
a tennis participant is holding a tennis court docket
This definitely tells that the community has been in a position to generalize over – let’s not name them ideas, however mappings between visible and textual entities, say It’s true that it’s going to have seen a few of these pictures earlier than, as a result of pictures include a number of captions. You could possibly be extra strict organising your coaching and validation units – however right here, we don’t actually care about goal efficiency scores and so, it does probably not matter.
Let’ skip on to epoch 20, our final coaching epoch, and examine for additional enhancements.
Epoch 20
That is what we get for the coaching pictures:
a gaggle of many tall giraffe standing subsequent to a sheep
a view of playing cards and white gloves on a avenue
a skateboarding flips his board
And this, for the validation pictures.
a baseball catcher and umpire hit a baseball sport
a person is consuming a sandwich
a feminine tennis participant is within the court docket
I feel we would agree that this nonetheless leaves room for enchancment – however then, we solely skilled for 20 epochs and on a really small portion of the dataset.
Within the above code snippets, you’ll have seen the decoder returning an attention_matrix – however we weren’t commenting on it.
Now lastly, simply as within the translation instance, take a look what we will make of that.
The place does the community look?
We will visualize the place the community is “wanting” because it generates every phrase by overlaying the unique picture and the eye matrix. This instance is taken from the 4th epoch.
Right here white-ish squares point out areas receiving stronger focus. In comparison with text-to-text translation although, the mapping is inherently much less simple – the place does one “look” when producing phrases like “and,” “the,” or “in?”
Conclusion
It most likely goes with out saying that significantly better outcomes are to be anticipated when coaching on (a lot!) extra information and for far more time.
Aside from that, there are different choices, although. The idea applied right here makes use of spatial consideration over a uniform grid, that’s, the eye mechanism guides the decoder the place on the grid to look subsequent when producing a caption.
Nevertheless, this isn’t the one method, and this isn’t the way it works with people. A way more believable strategy is a mixture of top-down and bottom-up consideration. E.g., (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 data enter (Zhu, Xue, and Yuan 2018), which once more is a top-down characteristic present in human cognition.
If you happen to discover one among these, or yet one more, strategy extra convincing, an keen execution implementation, within the type of the above, will doubtless be a sound method of implementing it.