DeepLarning Couse HSE 2016 fall:
ars.ashuha@gmail.com
,https://vk.com/ars.ashuha
In this seminar you'll be going through the image captioning pipeline.
To begin with, let us download the dataset of image features from a pre-trained GoogleNet.
In [ ]:
###
### Or alternatevely
### !wget https://www.dropbox.com/s/3hj16b0fj6yw7cc/data.tar.gz?dl=1 -O data.tar.gz
### !tar -xvzf data.tar.gz
###
DATA_DIR = '/wrk/trng103/'
In [ ]:
%%time
# Read Dataset
import numpy as np
import os.path as osp
import pickle
img_codes = np.load(osp.join(DATA_DIR, "data/image_codes.npy"))
captions = pickle.load(open(osp.join(DATA_DIR, 'data/caption_tokens.pcl'), 'rb'))
In [ ]:
print ("each image code is a 1000-unit vector:", img_codes.shape)
print (img_codes[0,:10])
print ('\n\n')
print ("for each image there are 5-7 descriptions, e.g.:\n")
print ('\n'.join(captions[0]))
In [ ]:
#split descriptions into tokens
for img_i in range(len(captions)):
for caption_i in range(len(captions[img_i])):
sentence = captions[img_i][caption_i]
captions[img_i][caption_i] = ["#START#"]+sentence.split(' ')+["#END#"]
In [ ]:
# Build a Vocabulary
from collections import Counter
word_counts = Counter()
<Compute word frequencies for each word in captions. See code above for data structure>
In [ ]:
vocab = ['#UNK#', '#START#', '#END#']
vocab += [k for k, v in word_counts.items() if v >= 5]
n_tokens = len(vocab)
assert 10000 <= n_tokens <= 10500
word_to_index = {w: i for i, w in enumerate(vocab)}
In [ ]:
PAD_ix = -1
UNK_ix = vocab.index('#UNK#')
#good old as_matrix for the third time
def as_matrix(sequences,max_len=None):
max_len = max_len or max(map(len,sequences))
matrix = np.zeros((len(sequences),max_len),dtype='int32')+PAD_ix
for i,seq in enumerate(sequences):
row_ix = [word_to_index.get(word,UNK_ix) for word in seq[:max_len]]
matrix[i,:len(row_ix)] = row_ix
return matrix
In [ ]:
#try it out on several descriptions of a random image
as_matrix(captions[1337])
In [ ]:
# network shapes.
CNN_FEATURE_SIZE = img_codes.shape[1]
EMBED_SIZE = 128 #pls change me if u want
LSTM_UNITS = 200 #pls change me if u want
In [ ]:
import theano
import theano.tensor as T
In [ ]:
# Input Variable
sentences = T.imatrix()# [batch_size x time] of word ids
image_vectors = T.matrix() # [batch size x unit] of CNN image features
sentence_mask = T.neq(sentences,PAD_ix)
In [ ]:
import lasagne
from lasagne.layers import *
In [ ]:
#network inputs
l_words = InputLayer((None,None),sentences )
l_mask = InputLayer((None,None),sentence_mask )
#embeddings for words
l_word_embeddings = <apply word embedding. use EMBED_SIZE>
#cudos for using some pre-trained embedding :)
In [ ]:
# input layer for image features
l_image_features = InputLayer((None,CNN_FEATURE_SIZE),image_vectors )
#convert 1000 image features from googlenet to whatever LSTM_UNITS you have set
#it's also a good idea to add some dropout here and there
l_image_features_small = <convert l_image features to a shape equal to rnn hidden state. Also play with dropout/noize>
assert l_image_features_small.output_shape == (None,LSTM_UNITS)
In [ ]:
# Concatinate image features and word embedings in one sequence
decoder = a recurrent layer (gru/lstm) with following checklist:
# * takes word embeddings as an input
# * has LSTM_UNITS units in the final layer
# * has cell_init (or hid init for gru) set to converted image features
# * mask_input = input_mask
# * don't forget the grad clipping (~5-10)
#find out better recurrent architectures for bonus point
In [ ]:
# Decoding of rnn hiden states
from broadcast import BroadcastLayer,UnbroadcastLayer
#apply whatever comes next to each tick of each example in a batch. Equivalent to 2 reshapes
broadcast_decoder_ticks = BroadcastLayer(decoder,(0,1))
print "broadcasted decoder shape = ",broadcast_decoder_ticks.output_shape
In [ ]:
#predict probabilities for next tokens
predicted_probabilities_each_tick = <predict probabilities for each tick, using broadcasted_decoder_shape as an input. No reshaping needed here.>
# maybe a more complicated architecture will work better?
In [ ]:
#un-broadcast back into (batch,tick,probabilities)
predicted_probabilities = UnbroadcastLayer(predicted_probabilities_each_tick,
broadcast_layer=broadcast_decoder_ticks)
print "output shape = ",predicted_probabilities.output_shape
#remove if you know what you're doing (e.g. 1d convolutions or fixed shape)
assert predicted_probabilities.output_shape == (None, None, 10373)
In [ ]:
next_word_probas = <get network output>
predictions_flat = next_word_probas[:,:-1].reshape((-1,n_tokens))
reference_answers = sentences[:,1:].reshape((-1,))
#write symbolic loss function to train NN for
loss = <compute elementwise loss function>
In [ ]:
#trainable NN weights
weights = get_all_params(predicted_probabilities,trainable=True)
updates = <parameter updates using your favorite algoritm>
In [ ]:
#compile a functions for training and evaluation
#please not that your functions must accept image features as FIRST param and sentences as second one
train_step = <function that takes input sentence and image mask, outputs loss and updates weights>
val_step = <function that takes input sentence and image mask and outputs loss>
#for val_step use deterministic=True if you have any dropout/noize
In [ ]:
captions = np.array(captions)
In [ ]:
from random import choice
def generate_batch(images,captions,batch_size,max_caption_len=None):
#sample random numbers for image/caption indicies
random_image_ix = np.random.randint(0,len(images),size=batch_size)
#get images
batch_images = images[random_image_ix]
#5-7 captions for each image
captions_for_batch_images = captions[random_image_ix]
#pick 1 from 5-7 captions for each image
batch_captions = map(choice,captions_for_batch_images)
#convert to matrix
batch_captions_ix = as_matrix(batch_captions,max_len=max_caption_len)
return batch_images, batch_captions_ix
In [ ]:
generate_batch(img_codes,captions,3)
In [ ]:
batch_size=50 #adjust me
n_epochs=100 #adjust me
n_batches_per_epoch = 50 #adjust me
n_validation_batches = 5 #how many batches are used for validation after each epoch
In [ ]:
!pip3 install --user tqdm
In [ ]:
from tqdm import tqdm
for epoch in range(n_epochs):
train_loss=0
for _ in tqdm(range(n_batches_per_epoch)):
train_loss += train_step(*generate_batch(img_codes,captions,batch_size))
train_loss /= n_batches_per_epoch
val_loss=0
for _ in range(n_validation_batches):
val_loss += val_step(*generate_batch(img_codes,captions,batch_size))
val_loss /= n_validation_batches
print('\nEpoch: {}, train loss: {}, val loss: {}'.format(epoch, train_loss, val_loss))
print("Finish :)")
In [ ]:
!pip3 install --user scikit-image
In [ ]:
#the same kind you did last week, but a bit smaller
from pretrained_lenet import build_model,preprocess,MEAN_VALUES
# build googlenet
lenet = build_model()
#load weights
#lenet_weights = pickle.load(open(osp.join(DATA_DIR, 'data/blvc_googlenet.pkl')), encoding='latin1')['param values']
lenet_weights = np.load(osp.join(DATA_DIR, 'data/blvc_googlenet.npz'), encoding='latin1')['param values']
set_all_param_values(lenet["prob"], lenet_weights)
#compile get_features
cnn_input_var = lenet['input'].input_var
cnn_feature_layer = lenet['loss3/classifier']
get_cnn_features = theano.function([cnn_input_var], lasagne.layers.get_output(cnn_feature_layer))
In [ ]:
from matplotlib import pyplot as plt
%matplotlib inline
#sample image
img = plt.imread(osp.join(DATA_DIR, 'data/Dog-and-Cat.jpg'))
img = preprocess(img)
In [ ]:
#deprocess and show, one line :)
from pretrained_lenet import MEAN_VALUES
plt.imshow(np.transpose((img[0] + MEAN_VALUES)[::-1],[1,2,0]).astype('uint8'))
In [ ]:
last_word_probas = <get network-predicted probas at last tick
#TRY OUT deterministic=True if you want more steady results
get_probs = theano.function([image_vectors,sentences], last_word_probas)
#this is exactly the generation function from week5 classwork,
#except now we condition on image features instead of words
def generate_caption(image,caption_prefix = ("START",),t=1,sample=True,max_len=100):
image_features = get_cnn_features(image)
caption = list(caption_prefix)
for _ in range(max_len):
next_word_probs = <obtain probabilities for next words>
assert len(next_word_probs.shape) ==1 #must be one-dimensional
#apply temperature
next_word_probs = next_word_probs**t / np.sum(next_word_probs**t)
if sample:
next_word = np.random.choice(vocab,p=next_word_probs)
else:
next_word = vocab[np.argmax(next_word_probs)]
caption.append(next_word)
if next_word=="#END#":
break
return caption
In [ ]:
for i in range(10):
print ' '.join(generate_caption(img,t=5.)[1:-1])
preprocess
does to your image)
In [ ]:
#apply your network on image sample you found
#
#
In [ ]:
In [ ]: