class FastZeroTagModel(AttalosModel):
"""
Create a tensorflow graph that finds the principal direction of the target word embeddings
(with negative sampling), using the loss function from "Fast Zero-Shot Image Tagging".
"""
def __init__(self, wv_model, datasets, **kwargs):
self.wv_model = wv_model
self.one_hot = OneHot(datasets, valid_vocab=wv_model.vocab)
word_counts = NegativeSampler.get_wordcount_from_datasets(datasets, self.one_hot)
self.negsampler = NegativeSampler(word_counts)
self.w = construct_W(wv_model, self.one_hot.get_key_ordering()).T
self.learning_rate = kwargs.get("learning_rate", 0.0001)
self.optim_words = kwargs.get("optim_words", True)
self.hidden_units = kwargs.get("hidden_units", "200")
self.use_batch_norm = kwargs.get("use_batch_norm",False)
self.opt_type = kwargs.get("opt_type","adam")
if self.hidden_units=='0':
self.hidden_units=[]
else:
self.hidden_units = [int(x) for x in self.hidden_units.split(',')]
self.model_info = dict()
# Placeholders for data
self.model_info['input'] = tf.placeholder(shape=(None, datasets[0].img_feat_size), dtype=tf.float32)
self.model_info['pos_ids'] = tf.placeholder(dtype=tf.int32)
self.model_info['neg_ids'] = tf.placeholder(dtype=tf.int32)
self.model_info['w2v'] = tf.Variable(self.w, dtype=tf.float32)
self.model_info['pos_vecs'] = tf.transpose(tf.nn.embedding_lookup(self.model_info['w2v'],
self.model_info['pos_ids']),
perm=[1,0,2])
self.model_info['neg_vecs'] = tf.transpose(tf.nn.embedding_lookup(self.model_info['w2v'],
self.model_info['neg_ids']),
perm=[1,0,2])
# Construct fully connected layers
layer = self.model_info['input']
layers = []
for i, hidden_size in enumerate(self.hidden_units):
layer = tf.contrib.layers.relu(layer, hidden_size)
layers.append(layer)
if self.use_batch_norm:
layer = tf.contrib.layers.batch_norm(layer)
layers.append(layer)
logger.info("Using batch normalization")
# Output layer should always be linear
layer = tf.contrib.layers.linear(layer, self.w.shape[1])
layers.append(layer)
self.model_info['layers'] = layers
self.model_info['prediction'] = layer
def fztloss( f, pVecs, nVecs ):
"""
Tensorized cost function from Fast Zero-Shot Learning paper
Args:
f: The output from the network, a tensor of shape (# images, word embedding size)
pVecs: The vector embeddings of the ground truth tags, a tensor
of shape (# images, # positive tags, word embedding size)
nVecs: The vector embeddings of negatively sampled tags, a tensor
of shape (# images, # negative samples, word embedding size)
Returns:
Scalar tensor representing the batch cost
"""
posmul = tf.mul(pVecs, f)
negmul = tf.mul(nVecs, f)
tfpos = tf.reduce_sum(posmul, reduction_indices=2)
tfneg = tf.reduce_sum(negmul, reduction_indices=2)
tfpos = tf.transpose(tfpos, [1,0])
tfneg = tf.transpose(tfneg, [1,0])
negexpan = tf.tile( tf.expand_dims(tfneg, -1), [1, 1, tf.shape(tfpos)[1]] )
posexpan = tf.tile( tf.transpose(tf.expand_dims(tfpos, -1), [0,2,1]), [1, tf.shape(tfneg)[1], 1])
differences = tf.sub(negexpan, posexpan)
return tf.reduce_sum(tf.reduce_sum(tf.log(1 + tf.exp(differences)), reduction_indices=[1,2]))
loss = fztloss(self.model_info['prediction'], self.model_info['pos_vecs'], self.model_info['neg_vecs'])
self.model_info['loss'] = loss
if self.opt_type=='sgd':
optimizer=tf.train.GradientDescent
else:
optimizer=tf.train.AdamOptimizer
self.model_info['optimizer'] = optimizer(learning_rate=self.learning_rate).minimize(loss)
def predict_feats(self, sess, x):
return sess.run(self.model_info['prediction'], feed_dict={self.model_info['input']: x})
def _get_ids(self, tag_ids, numSamps=[5, 10], uniform_sampling=False):
"""
Takes a batch worth of text tags and returns positive/negative ids
"""
pos_word_ids = np.ones((len(tag_ids), numSamps[0]), dtype=np.int32)
pos_word_ids.fill(-1)
for ind, tags in enumerate(tag_ids):
if len(tags) > 0:
pos_word_ids[ind] = np.random.choice(tags, size=numSamps[0])
neg_word_ids = None
if uniform_sampling:
neg_word_ids = np.random.randint(0,
self.one_hot.vocab_size,
size=(len(tag_ids), numSamps[1]))
else:
neg_word_ids = np.ones((len(tag_ids), numSamps[1]), dtype=np.int32)
neg_word_ids.fill(-1)
for ind in range(pos_word_ids.shape[0]):
# TODO: Check to see if this benefits from the same bug as negsampling code
neg_word_ids[ind] = self.negsampler.negsamp_ind(pos_word_ids[ind],
numSamps[1])
return pos_word_ids, neg_word_ids
def prep_fit(self, data):
img_feats, text_feats_list = data
text_feat_ids = []
for tags in text_feats_list:
text_feat_ids.append([self.one_hot.get_index(tag) for tag in tags if tag in self.one_hot])
pos_ids, neg_ids = self._get_ids(text_feat_ids)
self.pos_ids = pos_ids
self.neg_ids = neg_ids
fetches = [self.model_info["optimizer"], self.model_info["loss"]]
feed_dict = {
self.model_info["input"]: img_feats,
self.model_info["pos_ids"]: pos_ids,
self.model_info["neg_ids"]: neg_ids
}
return fetches, feed_dict
def prep_predict(self, dataset, cross_eval=False):
if cross_eval:
self.test_one_hot = OneHot([dataset], valid_vocab=self.wv_model.vocab)
self.test_w = construct_W(self.wv_model, self.test_one_hot.get_key_ordering()).T
else:
self.test_one_hot = self.one_hot
self.test_w = self.w
x = []
y = []
for idx in dataset:
image_feats, text_feats = dataset.get_index(idx)
text_feats = self.one_hot.get_multiple(text_feats)
x.append(image_feats)
y.append(text_feats)
x = np.asarray(x)
y = np.asarray(y)
fetches = [self.model_info["prediction"], ]
feed_dict = {
self.model_info["input"]: x
}
truth = y
return fetches, feed_dict, truth
def post_predict(self, predict_fetches, cross_eval=False):
predictions = predict_fetches[0]
if cross_eval and self.test_w is None:
raise Exception("test_w is not set. Did you call prep_predict?")
predictions = np.dot(predictions, self.test_w.T)
return predictions
def get_training_loss(self, fit_fetches):
return fit_fetches[1]
class NegSamplingModel(AttalosModel):
"""
This model performs negative sampling.
"""
def _construct_model_info(self, input_size, output_size, learning_rate, wv_arr,
hidden_units=[200],
optim_words=True,
opt_type='adam',
use_batch_norm=True,
weight_decay=0.0):
model_info = {}
model_info["input"] = tf.placeholder(shape=(None, input_size), dtype=tf.float32)
if optim_words:
model_info["pos_vecs"] = tf.placeholder(dtype=tf.float32)
model_info["neg_vecs"] = tf.placeholder(dtype=tf.float32)
logger.info("Optimization on GPU, word vectors are stored separately.")
else:
model_info["w2v"] = tf.Variable(wv_arr, dtype=tf.float32)
model_info["pos_ids"] = tf.placeholder(dtype=tf.int32)
model_info["neg_ids"] = tf.placeholder(dtype=tf.int32)
model_info["pos_vecs"] = tf.transpose(tf.nn.embedding_lookup(model_info["w2v"],
model_info["pos_ids"]),
perm=[1,0,2])
model_info["neg_vecs"] = tf.transpose(tf.nn.embedding_lookup(model_info["w2v"],
model_info["neg_ids"]),
perm=[1,0,2])
logger.info("Not optimizing word vectors.")
# Construct fully connected layers
layers = []
layer = model_info["input"]
for i, hidden_size in enumerate(hidden_units):
layer = tf.contrib.layers.relu(layer, hidden_size)
layers.append(layer)
if use_batch_norm:
layer = tf.contrib.layers.batch_norm(layer)
layers.append(layer)
# Output layer should always be linear
layer = tf.contrib.layers.linear(layer, wv_arr.shape[1])
layers.append(layer)
model_info["layers"] = layers
model_info["prediction"] = layer
def meanlogsig(predictions, truth):
reduction_indices = 2
return tf.reduce_mean(tf.log(tf.sigmoid(tf.reduce_sum(predictions * truth, reduction_indices=reduction_indices))))
pos_loss = meanlogsig(model_info["prediction"], model_info["pos_vecs"])
neg_loss = meanlogsig(-model_info["prediction"], model_info["neg_vecs"])
model_info["loss"] = -(pos_loss + neg_loss)
# Decide whether or not to use SGD or Adam Optimizers
if self.opt_type == 'sgd':
logger.info("Optimization uses SGD with non-variable rate")
optimizer = tf.train.GradientDescentOptimizer
else:
logger.info("Optimization uses Adam")
optimizer = tf.train.AdamOptimizer
# Are we manually decaying the words? Create a TF variable in that case.
if weight_decay:
logger.info("Learning rate is manually adaptive, dropping every ten (hard coded) epoch")
model_info['learning_rate'] = tf.placeholder(tf.float32, shape=[])
else:
model_info['learning_rate'] = learning_rate
model_info["optimizer"] = optimizer(learning_rate=model_info['learning_rate']).minimize(model_info["loss"])
#model_info["init_op"] = tf.initialize_all_variables()
#model_info["saver"] = tf.train.Saver()
return model_info
def __init__(self, wv_model, datasets, **kwargs):
self.wv_model = wv_model
self.one_hot = OneHot(datasets, valid_vocab=wv_model.vocab)
word_counts = NegativeSampler.get_wordcount_from_datasets(datasets, self.one_hot)
self.negsampler = NegativeSampler(word_counts)
train_dataset = datasets[0] # train_dataset should always be first in datasets
self.w = construct_W(wv_model, self.one_hot.get_key_ordering()).T
scale_words = kwargs.get("scale_words",1.0)
if scale_words == 0.0:
self.w = (self.w.T / np.linalg.norm(self.w,axis=1)).T
else:
self.w *= scale_words
# Optimization parameters
# Starting learning rate, currently default to 0.001. This will change iteratively if decay is on.
self.learning_rate = kwargs.get("learning_rate", 0.0001)
self.weight_decay = kwargs.get("weight_decay", 0.0)
self.optim_words = kwargs.get("optim_words", True)
self.epoch_num = 0
# Sampling methods
self.ignore_posbatch = kwargs.get("ignore_posbatch",False)
self.joint_factor = kwargs.get("joint_factor",1.0)
self.hidden_units = kwargs.get("hidden_units", "200")
if self.hidden_units=='0':
self.hidden_units=[]
else:
self.hidden_units = [int(x) for x in self.hidden_units.split(",")]
self.opt_type = kwargs.get("opt_type", "adam")
self.use_batch_norm = kwargs.get('use_batch_norm',False)
self.model_info = self._construct_model_info(
input_size = train_dataset.img_feat_size,
output_size = self.one_hot.vocab_size,
hidden_units=self.hidden_units,
learning_rate = self.learning_rate,
optim_words = self.optim_words,
use_batch_norm = self.use_batch_norm,
wv_arr = self.w, weight_decay = self.weight_decay
)
self.test_one_hot = None
self.test_w = None
super(NegSamplingModel, self).__init__()
def iter_batches(self, dataset, batch_size):
for x, y in super(NegSamplingModel, self).iter_batches(dataset, batch_size):
yield x, y
# This will decay the learning rate every ten epochs. Hardcoded ten currently...
if self.weight_decay:
if self.epoch_num and self.epoch_num % 15 == 0 and self.learning_rate > 1e-6:
self.learning_rate *= self.weight_decay
logger.info('Learning rate dropped to {}'.format(self.learning_rate))
self.epoch_num+=1
def _get_ids(self, tag_ids, numSamps=[5, 10], uniform_sampling=False):
"""
Takes a batch worth of text tags and returns positive/negative ids
"""
pos_word_ids = np.ones((len(tag_ids), numSamps[0]), dtype=np.int32)
pos_word_ids.fill(-1)
for ind, tags in enumerate(tag_ids):
if len(tags) > 0:
pos_word_ids[ind] = np.random.choice(tags, size=numSamps[0])
neg_word_ids = None
if uniform_sampling:
neg_word_ids = np.random.randint(0,
self.one_hot.vocab_size,
size=(len(tag_ids), numSamps[1]))
else:
neg_word_ids = np.ones((len(tag_ids), numSamps[1]), dtype=np.int32)
neg_word_ids.fill(-1)
for ind in range(pos_word_ids.shape[0]):
if self.ignore_posbatch:
# NOTE: This function call should definitely be pos_word_ids[ind]
# but that results in significantly worse performance
# I wish I understood why.
# I think this means we won't sample any tags that appear in the batch
neg_word_ids[ind] = self.negsampler.negsamp_ind(pos_word_ids, numSamps[1])
else:
neg_word_ids[ind] = self.negsampler.negsamp_ind(pos_word_ids[ind], numSamps[1])
return pos_word_ids, neg_word_ids
def prep_fit(self, data):
img_feats, text_feats_list = data
text_feat_ids = []
for tags in text_feats_list:
text_feat_ids.append([self.one_hot.get_index(tag) for tag in tags if tag in self.one_hot])
pos_ids, neg_ids = self._get_ids(text_feat_ids)
self.pos_ids = pos_ids
self.neg_ids = neg_ids
if not self.optim_words:
fetches = [self.model_info["optimizer"], self.model_info["loss"]]
feed_dict = {
self.model_info["input"]: img_feats,
self.model_info["pos_ids"]: pos_ids,
self.model_info["neg_ids"]: neg_ids
}
else:
pvecs = np.zeros((pos_ids.shape[0], pos_ids.shape[1], self.w.shape[1]))
nvecs = np.zeros((neg_ids.shape[0], neg_ids.shape[1], self.w.shape[1]))
for i, ids in enumerate(pos_ids):
pvecs[i] = self.w[ids]
for i, ids in enumerate(neg_ids):
nvecs[i] = self.w[ids]
pvecs = pvecs.transpose((1, 0, 2))
nvecs = nvecs.transpose((1, 0, 2))
fetches = [self.model_info["optimizer"], self.model_info["loss"], self.model_info["prediction"]]
feed_dict = {
self.model_info["input"]: img_feats,
self.model_info["pos_vecs"]: pvecs,
self.model_info["neg_vecs"]: nvecs
}
if self.weight_decay:
feed_dict[self.model_info['learning_rate']] = self.learning_rate
return fetches, feed_dict
def _updatewords(self, vpindex, vnindex, vin):
for i, (vpi, vni) in enumerate(zip(vpindex, vnindex)):
self.w[vpi] += self.joint_factor*self.learning_rate * np.outer(1 - sigmoid(self.w[vpi].dot(vin[i])), vin[i])
self.w[vni] -= self.joint_factor*self.learning_rate * np.outer(sigmoid(self.w[vni].dot(vin[i])), vin[i])
def fit(self, sess, fetches, feed_dict):
fit_fetches = super(NegSamplingModel, self).fit(sess, fetches, feed_dict)
if self.optim_words:
if self.pos_ids is None or self.neg_ids is None:
raise Exception("pos_ids or neg_ids is not set; cannot update word vectors. Did you run prep_fit()?")
_, _, prediction = fit_fetches
self._updatewords(self.pos_ids, self.neg_ids, prediction)
return fit_fetches
def prep_predict(self, dataset, cross_eval=False):
if cross_eval:
self.test_one_hot = OneHot([dataset], valid_vocab=self.wv_model.vocab)
self.test_w = construct_W(self.wv_model, self.test_one_hot.get_key_ordering()).T
else:
self.test_one_hot = self.one_hot
self.test_w = self.w
x = []
y = []
for idx in dataset:
image_feats, text_feats = dataset.get_index(idx)
text_feats = self.one_hot.get_multiple(text_feats)
x.append(image_feats)
y.append(text_feats)
x = np.asarray(x)
y = np.asarray(y)
fetches = [self.model_info["prediction"], ]
feed_dict = {
self.model_info["input"]: x
}
truth = y
return fetches, feed_dict, truth
def post_predict(self, predict_fetches, cross_eval=False):
predictions = predict_fetches[0]
if cross_eval and self.test_w is None:
raise Exception("test_w is not set. Did you call prep_predict?")
predictions = np.dot(predictions, self.test_w.T)
return predictions
def get_training_loss(self, fit_fetches):
return fit_fetches[1]
class NaiveSumModel(AttalosModel):
"""
This model performs linear regression via NN using the naive sum of word vectors as targets.
"""
def _construct_model_info(self, input_size, output_size, learning_rate,
hidden_units=[200]):
logger.info("Input size: %s" % input_size)
logger.info("Output size: %s" % output_size)
model_info = {}
model_info["input"] = tf.placeholder(shape=(None, input_size), dtype=tf.float32)
model_info["y_truth"] = tf.placeholder(shape=(None, output_size), dtype=tf.float32)
layers = []
layer = model_info["input"]
for i, hidden_size in enumerate(hidden_units):
layer = tf.contrib.layers.relu(layer, hidden_size)
layers.append(layer)
model_info["predictions"] = tf.contrib.layers.fully_connected(layer,
output_size,
activation_fn=None)
model_info["loss"] = tf.reduce_sum(tf.square(model_info["predictions"] - model_info["y_truth"]))
model_info["optimizer"] = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(model_info["loss"])
return model_info
def __init__(self, wv_model, datasets, **kwargs):
self.wv_model = wv_model
#self.cross_eval = kwargs.get("cross_eval", False)
#self.one_hot = OneHot([train_dataset] if self.cross_eval else [train_dataset, test_dataset],
# valid_vocab=wv_model.vocab)
self.one_hot = OneHot(datasets, valid_vocab=wv_model.vocab)
self.wv_transformer = NaiveW2V.create_from_vocab(wv_model, self.one_hot, vocab=self.one_hot.get_key_ordering())
train_dataset = datasets[0] # train_dataset should always be first in datasets iterable
self.learning_rate = kwargs.get("learning_rate", 0.0001)
self.hidden_units = kwargs.get("hidden_units", "200")
if self.hidden_units == '0':
self.hidden_units = []
else:
self.hidden_units = [int(x) for x in self.hidden_units.split(",")]
self.model_info = self._construct_model_info(
input_size = train_dataset.img_feat_size,
output_size = self.wv_model.get_word_vector_shape()[0],
learning_rate = self.learning_rate,
hidden_units=self.hidden_units,
)
self.test_one_hot = None
super(NaiveSumModel, self).__init__()
def prep_fit(self, data):
img_feats_list, text_feats_list = data
img_feats = np.array(img_feats_list)
text_feats = [self.one_hot.get_multiple(text_feats) for text_feats in text_feats_list]
text_feats = np.array(text_feats)
text_feats = self.wv_transformer.transform(text_feats)
fetches = [self.model_info["optimizer"], self.model_info["loss"]]
feed_dict = {
self.model_info["input"]: img_feats,
self.model_info["y_truth"]: text_feats
}
return fetches, feed_dict
def prep_predict(self, dataset, cross_eval=False):
if self.test_one_hot is None or self.test_dataset is dataset:
self.test_dataset = dataset
self.test_one_hot = OneHot([dataset], valid_vocab=self.wv_model.vocab)
self.test_w = construct_W(self.wv_model, self.test_one_hot.get_key_ordering()).T
x = []
y = []
for idx in dataset:
image_feats, text_feats = dataset.get_index(idx)
text_feats = self.test_one_hot.get_multiple(text_feats)
x.append(image_feats)
y.append(text_feats)
x = np.asarray(x)
truth = np.asarray(y)
fetches = [self.model_info["predictions"], ]
feed_dict = {
self.model_info["input"]: x
}
return fetches, feed_dict, truth
def post_predict(self, predict_fetches, cross_eval=False):
if self.test_one_hot is None:
raise Exception("test_one_hot is not set. Did you call prep_predict to initialize it?")
predictions = predict_fetches[0]
predictions = np.dot(predictions, construct_W(self.wv_model, self.test_one_hot.get_key_ordering()))
return predictions
def get_training_loss(self, fit_fetches):
_, loss = fit_fetches
return loss
class WDVModel(AttalosModel):
"""
This model performs logistic regression via NN using word distribution vectors (correlation vectors) as targets.
"""
def _construct_model_info(self, input_size, output_size, learning_rate,
hidden_units=[200]):
logger.info("Input size: %s" % input_size)
logger.info("Output size: %s" % output_size)
model_info = {}
model_info["input"] = tf.placeholder(shape=(None, input_size), dtype=tf.float32)
model_info["y_truth"] = tf.placeholder(shape=(None, output_size), dtype=tf.float32)
layers = []
layer = model_info["input"]
for i, hidden_size in enumerate(hidden_units):
layer = tf.contrib.layers.relu(layer, hidden_size)
layers.append(layer)
model_info["predictions"] = tf.contrib.layers.fully_connected(layer,
output_size,
activation_fn=tf.sigmoid)
model_info["loss"] = tf.reduce_sum(tf.square(model_info["predictions"] - model_info["y_truth"]))
model_info["optimizer"] = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(model_info["loss"])
return model_info
def __init__(self, wv_model, datasets, **kwargs): #train_dataset, test_dataset, **kwargs):
self.wv_model = wv_model
#self.cross_eval = kwargs.get("cross_eval", False)
#self.one_hot = OneHot([train_dataset] if self.cross_eval else [train_dataset, test_dataset],
# valid_vocab=wv_model.vocab)
self.one_hot = OneHot(datasets, valid_vocab=self.wv_model.vocab)
train_dataset = datasets[0] # train_dataset should always be first in datasets iterable
self.wv_transformer = WDV.create_from_vocab(wv_model, vocab1=self.one_hot.get_key_ordering(), preprocess_fn=WDV.preprocess)
self.learning_rate = kwargs.get("learning_rate", 0.0001)
self.hidden_units = kwargs.get("hidden_units", "200")
if self.hidden_units == '0':
self.hidden_units = []
else:
self.hidden_units = [int(x) for x in self.hidden_units.split(",")]
self.model_info = self._construct_model_info(
input_size = train_dataset.img_feat_size,
output_size = self.one_hot.vocab_size, #self.wv_model.get_word_vector_shape()[0],
learning_rate = self.learning_rate,
hidden_units=self.hidden_units,
)
self.transform_cache = {}
self.test_one_hot = None
self.test_wv_transformer = None
super(WDVModel, self).__init__()
def generate_key(self, word_list):
return " ".join(sorted(word_list))
def add_cache(self, word_list, val):
key = self.generate_key(word_list)
self.transform_cache[key] = val
def get_cache(self, word_list):
key = self.generate_key(word_list)
if key in self.transform_cache:
return self.transform_cache[key]
else:
return None
def prep_fit(self, data):
img_feats_list, text_feats_list = data
img_feats = np.array(img_feats_list)
# normalize img_feats
# new_img_feats = (new_img_feats.T / np.linalg.norm(new_img_feats, axis=1)).T
new_text_feats_list = []
for text_feats in text_feats_list:
new_text_feats = self.get_cache(text_feats)
if new_text_feats is None:
new_text_feats = [self.one_hot.get_multiple(text_feats)]
new_text_feats = np.array(new_text_feats)
new_text_feats = self.wv_transformer.transform(new_text_feats, postprocess_fn=WDV.postprocess)
new_text_feats = new_text_feats[0] # new_text_feats is a list; get first element
self.add_cache(text_feats, new_text_feats)
new_text_feats_list.append(new_text_feats)
text_feats = np.array(new_text_feats_list)
fetches = [self.model_info["optimizer"], self.model_info["loss"]]
feed_dict = {
self.model_info["input"]: img_feats,
self.model_info["y_truth"]: text_feats
}
return fetches, feed_dict
def prep_predict(self, dataset, cross_eval=False):
x = []
y = []
if cross_eval:
self.test_one_hot = OneHot([dataset], valid_vocab=self.wv_model.vocab)
self.test_wv_transformer = WDV.create_from_vocab(self.wv_model,
vocab1=self.one_hot.get_key_ordering(),
vocab2=self.test_one_hot.get_key_ordering(),
preprocess_fn=scale3)
else:
self.test_one_hot = self.one_hot
self.test_wv_transformer = None
for idx in dataset:
image_feats, text_feats = dataset.get_index(idx)
text_feats = self.test_one_hot.get_multiple(text_feats)
x.append(image_feats)
y.append(text_feats)
x = np.asarray(x)
y = np.asarray(y)
fetches = [self.model_info["predictions"], ]
feed_dict = {
self.model_info["input"]: x
}
truth = y
return fetches, feed_dict, truth
def post_predict(self, predict_fetches, cross_eval=False):
predictions = predict_fetches[0]
if cross_eval:
if self.test_wv_transformer is None:
raise Exception("test_wv_transformers is not set. Did you call prep_predict?")
predictions = np.dot(predictions, self.test_wv_transformer.wdv_arr)
return predictions
def get_training_loss(self, fit_fetches):
_, loss = fit_fetches
return loss
"""
class NegSamplingModel(AttalosModel):
"""
This model performs negative sampling.
"""
def _construct_model_info(self,
input_size,
output_size,
learning_rate,
wv_arr,
hidden_units=[200, 200],
optim_words=True,
use_batch_norm=True):
model_info = {}
model_info["input"] = tf.placeholder(shape=(None, input_size),
dtype=tf.float32)
if optim_words:
model_info["pos_vecs"] = tf.placeholder(dtype=tf.float32)
model_info["neg_vecs"] = tf.placeholder(dtype=tf.float32)
logger.info(
"Optimization on GPU, word vectors are stored separately.")
else:
model_info["w2v"] = tf.Variable(wv_arr, dtype=tf.float32)
model_info["pos_ids"] = tf.placeholder(dtype=tf.int32)
model_info["neg_ids"] = tf.placeholder(dtype=tf.int32)
model_info["pos_vecs"] = tf.transpose(tf.nn.embedding_lookup(
model_info["w2v"], model_info["pos_ids"]),
perm=[1, 0, 2])
model_info["neg_vecs"] = tf.transpose(tf.nn.embedding_lookup(
model_info["w2v"], model_info["neg_ids"]),
perm=[1, 0, 2])
logger.info("Not optimizing word vectors.")
# Construct fully connected layers
layers = []
layer = model_info["input"]
for i, hidden_size in enumerate(hidden_units[:-1]):
layer = tf.contrib.layers.relu(layer, hidden_size)
layers.append(layer)
if use_batch_norm:
layer = tf.contrib.layers.batch_norm(layer)
layers.append(layer)
# Output layer should always be linear
layer = tf.contrib.layers.linear(layer, wv_arr.shape[1])
layers.append(layer)
model_info["layers"] = layers
model_info["prediction"] = layer
def meanlogsig(predictions, truth):
reduction_indices = 2
return tf.reduce_mean(
tf.log(
tf.sigmoid(
tf.reduce_sum(predictions * truth,
reduction_indices=reduction_indices))))
pos_loss = meanlogsig(model_info["prediction"], model_info["pos_vecs"])
neg_loss = meanlogsig(-model_info["prediction"],
model_info["neg_vecs"])
model_info["loss"] = -(pos_loss + neg_loss)
model_info["optimizer"] = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(model_info["loss"])
#model_info["init_op"] = tf.initialize_all_variables()
#model_info["saver"] = tf.train.Saver()
return model_info
def __init__(self, wv_model, datasets, **kwargs):
self.wv_model = wv_model
self.one_hot = OneHot(datasets, valid_vocab=wv_model.vocab)
word_counts = NegativeSampler.get_wordcount_from_datasets(
datasets, self.one_hot)
self.negsampler = NegativeSampler(word_counts)
train_dataset = datasets[
0] # train_dataset should always be first in datasets
self.w = construct_W(wv_model, self.one_hot.get_key_ordering()).T
self.learning_rate = kwargs.get("learning_rate", 0.0001)
self.optim_words = kwargs.get("optim_words", True)
self.hidden_units = kwargs.get("hidden_units", "200,200")
self.hidden_units = [int(x) for x in self.hidden_units.split(",")]
self.model_info = self._construct_model_info(
input_size=train_dataset.img_feat_size,
output_size=self.one_hot.vocab_size,
hidden_units=self.hidden_units,
learning_rate=self.learning_rate,
optim_words=self.optim_words,
wv_arr=self.w)
self.test_one_hot = None
self.test_w = None
super(NegSamplingModel, self).__init__()
def _get_ids(self, tag_ids, numSamps=[5, 10], uniform_sampling=False):
"""
Takes a batch worth of text tags and returns positive/negative ids
"""
pos_word_ids = np.ones((len(tag_ids), numSamps[0]), dtype=np.int32)
pos_word_ids.fill(-1)
for ind, tags in enumerate(tag_ids):
if len(tags) > 0:
pos_word_ids[ind] = np.random.choice(tags, size=numSamps[0])
neg_word_ids = None
if uniform_sampling:
neg_word_ids = np.random.randint(0,
self.one_hot.vocab_size,
size=(len(tag_ids), numSamps[1]))
else:
neg_word_ids = np.ones((len(tag_ids), numSamps[1]), dtype=np.int32)
neg_word_ids.fill(-1)
for ind in range(pos_word_ids.shape[0]):
if self.optim_words:
neg_word_ids[ind] = self.negsampler.negsamp_ind(
pos_word_ids[ind], numSamps[1])
else:
# NOTE: This function call should definitely be pos_word_ids[ind]
# but that results in significantly worse performance
# I wish I understood why.
# I think this means we won't sample any tags that appear in the batch
neg_word_ids[ind] = self.negsampler.negsamp_ind(
pos_word_ids, numSamps[1])
return pos_word_ids, neg_word_ids
def prep_fit(self, data):
img_feats, text_feats_list = data
text_feat_ids = []
for tags in text_feats_list:
text_feat_ids.append([
self.one_hot.get_index(tag) for tag in tags
if tag in self.one_hot
])
pos_ids, neg_ids = self._get_ids(text_feat_ids)
self.pos_ids = pos_ids
self.neg_ids = neg_ids
if not self.optim_words:
fetches = [self.model_info["optimizer"], self.model_info["loss"]]
feed_dict = {
self.model_info["input"]: img_feats,
self.model_info["pos_ids"]: pos_ids,
self.model_info["neg_ids"]: neg_ids
}
else:
pvecs = np.zeros(
(pos_ids.shape[0], pos_ids.shape[1], self.w.shape[1]))
nvecs = np.zeros(
(neg_ids.shape[0], neg_ids.shape[1], self.w.shape[1]))
for i, ids in enumerate(pos_ids):
pvecs[i] = self.w[ids]
for i, ids in enumerate(neg_ids):
nvecs[i] = self.w[ids]
pvecs = pvecs.transpose((1, 0, 2))
nvecs = nvecs.transpose((1, 0, 2))
fetches = [
self.model_info["optimizer"], self.model_info["loss"],
self.model_info["prediction"]
]
feed_dict = {
self.model_info["input"]: img_feats,
self.model_info["pos_vecs"]: pvecs,
self.model_info["neg_vecs"]: nvecs
}
return fetches, feed_dict
def _updatewords(self, vpindex, vnindex, vin):
for i, (vpi, vni) in enumerate(zip(vpindex, vnindex)):
self.w[vpi] += self.learning_rate * np.outer(
1 - sigmoid(self.w[vpi].dot(vin[i])), vin[i])
self.w[vni] -= self.learning_rate * np.outer(
sigmoid(self.w[vni].dot(vin[i])), vin[i])
def fit(self, sess, fetches, feed_dict):
fit_fetches = super(NegSamplingModel,
self).fit(sess, fetches, feed_dict)
if self.optim_words:
if self.pos_ids is None or self.neg_ids is None:
raise Exception(
"pos_ids or neg_ids is not set; cannot update word vectors. Did you run prep_fit()?"
)
_, _, prediction = fit_fetches
self._updatewords(self.pos_ids, self.neg_ids, prediction)
return fit_fetches
def prep_predict(self, dataset, cross_eval=False):
if cross_eval:
self.test_one_hot = OneHot([dataset],
valid_vocab=self.wv_model.vocab)
self.test_w = construct_W(self.wv_model,
self.test_one_hot.get_key_ordering()).T
else:
self.test_one_hot = self.one_hot
self.test_w = self.w
x = []
y = []
for idx in dataset:
image_feats, text_feats = dataset.get_index(idx)
text_feats = self.one_hot.get_multiple(text_feats)
x.append(image_feats)
y.append(text_feats)
x = np.asarray(x)
y = np.asarray(y)
fetches = [
self.model_info["prediction"],
]
feed_dict = {self.model_info["input"]: x}
truth = y
return fetches, feed_dict, truth
def post_predict(self, predict_fetches, cross_eval=False):
predictions = predict_fetches[0]
if cross_eval and self.test_w is None:
raise Exception("test_w is not set. Did you call prep_predict?")
predictions = np.dot(predictions, self.test_w.T)
return predictions
def get_training_loss(self, fit_fetches):
return fit_fetches[1]
class FastZeroTagModel(AttalosModel):
"""
Create a tensorflow graph that finds the principal direction of the target word embeddings
(with negative sampling), using the loss function from "Fast Zero-Shot Image Tagging".
"""
def __init__(self, wv_model, datasets, **kwargs):
self.wv_model = wv_model
self.one_hot = OneHot(datasets, valid_vocab=wv_model.vocab)
word_counts = NegativeSampler.get_wordcount_from_datasets(datasets, self.one_hot)
self.negsampler = NegativeSampler(word_counts)
self.w = construct_W(wv_model, self.one_hot.get_key_ordering()).T
self.learning_rate = kwargs.get("learning_rate", 0.0001)
self.optim_words = kwargs.get("optim_words", True)
self.hidden_units = kwargs.get("hidden_units", "200,200")
self.hidden_units = [int(x) for x in self.hidden_units.split(",")]
self.model_info = dict()
# Placeholders for data
self.model_info['input'] = tf.placeholder(shape=(None, datasets[0].img_feat_size), dtype=tf.float32)
self.model_info['pos_ids'] = tf.placeholder(dtype=tf.int32)
self.model_info['neg_ids'] = tf.placeholder(dtype=tf.int32)
self.model_info['w2v'] = tf.Variable(self.w, dtype=tf.float32)
self.model_info['pos_vecs'] = tf.transpose(tf.nn.embedding_lookup(self.model_info['w2v'],
self.model_info['pos_ids']),
perm=[1,0,2])
self.model_info['neg_vecs'] = tf.transpose(tf.nn.embedding_lookup(self.model_info['w2v'],
self.model_info['neg_ids']),
perm=[1,0,2])
# Construct fully connected layers
layers = []
for i, hidden_size in enumerate(self.hidden_units[:-1]):
if i == 0:
layer = tf.contrib.layers.relu(self.model_info['input'], hidden_size)
else:
layer = tf.contrib.layers.relu(layer, hidden_size)
layers.append(layer)
layer = tf.contrib.layers.batch_norm(layer)
layers.append(layer)
# Output layer should always be linear
layer = tf.contrib.layers.linear(layer, self.hidden_units[-1])
layers.append(layer)
self.model_info['layers'] = layers
self.model_info['prediction'] = layer
def fztloss( f, pVecs, nVecs ):
"""
Tensorized cost function from Fast Zero-Shot Learning paper
Args:
f: The output from the network, a tensor of shape (# images, word embedding size)
pVecs: The vector embeddings of the ground truth tags, a tensor
of shape (# images, # positive tags, word embedding size)
nVecs: The vector embeddings of negatively sampled tags, a tensor
of shape (# images, # negative samples, word embedding size)
Returns:
Scalar tensor representing the batch cost
"""
posmul = tf.mul(pVecs, f)
negmul = tf.mul(nVecs, f)
tfpos = tf.reduce_sum(posmul, reduction_indices=2)
tfneg = tf.reduce_sum(negmul, reduction_indices=2)
tfpos = tf.transpose(tfpos, [1,0])
tfneg = tf.transpose(tfneg, [1,0])
negexpan = tf.tile( tf.expand_dims(tfneg, -1), [1, 1, tf.shape(tfpos)[1]] )
posexpan = tf.tile( tf.transpose(tf.expand_dims(tfpos, -1), [0,2,1]), [1, tf.shape(tfneg)[1], 1])
differences = tf.sub(negexpan, posexpan)
return tf.reduce_sum(tf.reduce_sum(tf.log(1 + tf.exp(differences)), reduction_indices=[1,2]))
loss = fztloss(self.model_info['prediction'], self.model_info['pos_vecs'], self.model_info['neg_vecs'])
self.model_info['loss'] = loss
self.model_info['optimizer'] = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(loss)
def predict_feats(self, sess, x):
return sess.run(self.model_info['prediction'], feed_dict={self.model_info['input']: x})
def _get_ids(self, tag_ids, numSamps=[5, 10], uniform_sampling=False):
"""
Takes a batch worth of text tags and returns positive/negative ids
"""
pos_word_ids = np.ones((len(tag_ids), numSamps[0]), dtype=np.int32)
pos_word_ids.fill(-1)
for ind, tags in enumerate(tag_ids):
if len(tags) > 0:
pos_word_ids[ind] = np.random.choice(tags, size=numSamps[0])
neg_word_ids = None
if uniform_sampling:
neg_word_ids = np.random.randint(0,
self.one_hot.vocab_size,
size=(len(tag_ids), numSamps[1]))
else:
neg_word_ids = np.ones((len(tag_ids), numSamps[1]), dtype=np.int32)
neg_word_ids.fill(-1)
for ind in range(pos_word_ids.shape[0]):
# TODO: Check to see if this benefits from the same bug as negsampling code
neg_word_ids[ind] = self.negsampler.negsamp_ind(pos_word_ids[ind],
numSamps[1])
return pos_word_ids, neg_word_ids
def prep_fit(self, data):
img_feats, text_feats_list = data
text_feat_ids = []
for tags in text_feats_list:
text_feat_ids.append([self.one_hot.get_index(tag) for tag in tags if tag in self.one_hot])
pos_ids, neg_ids = self._get_ids(text_feat_ids)
self.pos_ids = pos_ids
self.neg_ids = neg_ids
fetches = [self.model_info["optimizer"], self.model_info["loss"]]
feed_dict = {
self.model_info["input"]: img_feats,
self.model_info["pos_ids"]: pos_ids,
self.model_info["neg_ids"]: neg_ids
}
return fetches, feed_dict
def prep_predict(self, dataset, cross_eval=False):
if cross_eval:
self.test_one_hot = OneHot([dataset], valid_vocab=self.wv_model.vocab)
self.test_w = construct_W(self.wv_model, self.test_one_hot.get_key_ordering()).T
else:
self.test_one_hot = self.one_hot
self.test_w = self.w
x = []
y = []
for idx in dataset:
image_feats, text_feats = dataset.get_index(idx)
text_feats = self.one_hot.get_multiple(text_feats)
x.append(image_feats)
y.append(text_feats)
x = np.asarray(x)
y = np.asarray(y)
fetches = [self.model_info["prediction"], ]
feed_dict = {
self.model_info["input"]: x
}
truth = y
return fetches, feed_dict, truth
def post_predict(self, predict_fetches, cross_eval=False):
predictions = predict_fetches[0]
if cross_eval and self.test_w is None:
raise Exception("test_w is not set. Did you call prep_predict?")
predictions = np.dot(predictions, self.test_w.T)
return predictions
def get_training_loss(self, fit_fetches):
return fit_fetches[1]