def initialize(self, hidden_size, projection_size, in_vocabulary, out_vocabulary, batch_size, hidden_activation="Tanh", bptt_steps=5, use_pauses=False):
self.hidden_size = hidden_size
self.projection_size = projection_size
self.bptt_steps = bptt_steps
self.batch_size = batch_size
self.use_pauses = use_pauses
self.in_vocabulary = in_vocabulary
self.out_vocabulary = out_vocabulary
self.hidden_activation_name = hidden_activation
self.hidden_activation = getattr(activation_functions, hidden_activation)
self.We = self.weights(get_vocabulary_size(self.in_vocabulary), self.projection_size)
self.Wp = self.weights(1, self.projection_size)
self.W = self.weights(self.projection_size, self.hidden_size*4)
self.Wip = self.weights(1, self.hidden_size)
self.Wfp = self.weights(1, self.hidden_size)
self.Wop = self.weights(1, self.hidden_size)
self.Wr = self.weights(self.hidden_size, self.hidden_size*4)
self.Wy = self.weights(self.hidden_size, get_vocabulary_size(self.out_vocabulary))
# AdaGread sum of squares of per feature historical gradients
for p in self.params:
setattr(self, p+"_hg", np.zeros_like(getattr(self, p)))
self.reset_state()
self.initialized = True
def train(model_name, p1_train_data, p1_dev_data, p2_train_data, p2_dev_data):
### PHASE 1 ###
training_data = np.load(p1_train_data)
validation_data = np.load(p1_dev_data)
assert training_data["batch_size"] == validation_data["batch_size"]
assert training_data["vocabulary"] == validation_data["vocabulary"]
assert training_data["punctuations"] == validation_data["punctuations"]
print "1st phase data loaded..."
print "Vocabulary size is %d" % utils.get_vocabulary_size(validation_data["vocabulary"])
print "Training set size is %d" % training_data["total_size"]
print "Validation set size is %d" % validation_data["total_size"]
net = models.T_LSTM()
net.initialize(hidden_size=conf.PHASE1["HIDDEN_SIZE"],
projection_size=conf.PHASE1["PROJECTION_SIZE"],
in_vocabulary=training_data["vocabulary"],
out_vocabulary=training_data["punctuations"],
batch_size=training_data["batch_size"],
hidden_activation=conf.PHASE1["HIDDEN_ACTIVATION"],
bptt_steps=conf.PHASE1["BPTT_STEPS"],
use_pauses=False)
_train(net, training_data, validation_data, model_name, conf.PHASE1["LEARNING_RATE"], conf.PHASE1["MAX_EPOCHS"], conf.PHASE1["MIN_IMPROVEMENT"])
### PHASE 2 ###
if not os.path.isfile(p2_train_data) or not os.path.isfile(p2_train_data):
print "No second phase data."
return
training_data = np.load(p2_train_data)
validation_data = np.load(p2_dev_data)
assert training_data["batch_size"] == validation_data["batch_size"] == net.batch_size
assert training_data["vocabulary"] == validation_data["vocabulary"] == net.in_vocabulary
assert training_data["punctuations"] == validation_data["punctuations"] == net.out_vocabulary
print "2nd phase data loaded..."
print "Training set size is %d" % training_data["total_size"]
print "Validation set size is %d" % validation_data["total_size"]
print "Trainging %s pause durations." % ("with" if conf.PHASE2["USE_PAUSES"] else "without")
t_lstm = net
net = models.TA_LSTM()
net.initialize(hidden_size=conf.PHASE2["HIDDEN_SIZE"],
t_lstm=t_lstm,
out_vocabulary=training_data["punctuations"],
batch_size=training_data["batch_size"],
hidden_activation=conf.PHASE2["HIDDEN_ACTIVATION"],
bptt_steps=conf.PHASE2["BPTT_STEPS"],
use_pauses=conf.PHASE2["USE_PAUSES"])
_train(net, training_data, validation_data, model_name, conf.PHASE2["LEARNING_RATE"], conf.PHASE2["MAX_EPOCHS"], conf.PHASE2["MIN_IMPROVEMENT"])
def train(model_name, p1_train_data, p1_dev_data, p2_train_data, p2_dev_data):
### PHASE 1 ###
training_data = np.load(p1_train_data)
validation_data = np.load(p1_dev_data)
assert training_data["batch_size"] == validation_data["batch_size"]
assert training_data["vocabulary"] == validation_data["vocabulary"]
assert training_data["punctuations"] == validation_data["punctuations"]
print("1st phase data loaded...")
print("Vocabulary size is %d" %
utils.get_vocabulary_size(validation_data["vocabulary"]))
print("Training set size is %d" % training_data["total_size"])
print("Validation set size is %d" % validation_data["total_size"])
net = models.T_LSTM()
net.initialize(hidden_size=conf.PHASE1["HIDDEN_SIZE"],
projection_size=conf.PHASE1["PROJECTION_SIZE"],
in_vocabulary=training_data["vocabulary"],
out_vocabulary=training_data["punctuations"],
batch_size=training_data["batch_size"],
hidden_activation=conf.PHASE1["HIDDEN_ACTIVATION"],
gate_activation=conf.GATE_ACTIVATION,
bptt_steps=conf.PHASE1["BPTT_STEPS"],
use_pauses=False)
_train(net, training_data, validation_data, model_name,
conf.PHASE1["LEARNING_RATE"], conf.PHASE1["MAX_EPOCHS"],
conf.PHASE1["MIN_IMPROVEMENT"])
def initialize(self,
hidden_size,
projection_size,
in_vocabulary,
out_vocabulary,
batch_size,
hidden_activation="Tanh",
gate_activation="Sigmoid",
bptt_steps=5,
use_pauses=False):
self.hidden_size = hidden_size
self.projection_size = projection_size
self.bptt_steps = bptt_steps
self.batch_size = batch_size
self.use_pauses = use_pauses
self.in_vocabulary = in_vocabulary
self.out_vocabulary = out_vocabulary
self.hidden_activation_name = hidden_activation
self.hidden_activation = getattr(activation_functions,
hidden_activation)
self.gate_activation_name = gate_activation
self.gate_activation = getattr(activation_functions, gate_activation)
self.We = self.weights(get_vocabulary_size(self.in_vocabulary),
self.projection_size)
self.Wp = self.weights(1, self.projection_size)
self.W = self.weights(self.projection_size, self.hidden_size * 4)
self.Wip = self.weights(1, self.hidden_size)
self.Wfp = self.weights(1, self.hidden_size)
self.Wop = self.weights(1, self.hidden_size)
self.Wr = self.weights(self.hidden_size, self.hidden_size * 4)
self.Wy = self.weights(self.hidden_size,
get_vocabulary_size(self.out_vocabulary))
# AdaGread sum of squares of per feature historical gradients
for p in self.params:
setattr(self, p + "_hg", np.zeros_like(getattr(self, p)))
self.reset_state()
self.initialized = True
def initialize(self,
hidden_size,
t_lstm,
out_vocabulary,
batch_size,
hidden_activation="Tanh",
bptt_steps=5,
use_pauses=False):
assert isinstance(t_lstm, T_LSTM)
self.hidden_size = hidden_size
self.t_lstm = t_lstm
self.bptt_steps = bptt_steps
self.batch_size = batch_size
self.use_pauses = use_pauses
self.in_vocabulary = self.t_lstm.in_vocabulary
self.out_vocabulary = out_vocabulary
self.hidden_activation_name = hidden_activation
self.hidden_activation = getattr(activation_functions,
hidden_activation)
self.W = self.weights(self.t_lstm.hidden_size, self.hidden_size * 4)
self.Wp = self.weights(1, self.hidden_size * 4)
self.Wy = self.weights(self.hidden_size,
get_vocabulary_size(self.out_vocabulary))
self.Wip = self.weights(1, self.hidden_size)
self.Wfp = self.weights(1, self.hidden_size)
self.Wop = self.weights(1, self.hidden_size)
self.Wr = self.weights(self.hidden_size, self.hidden_size * 4)
# AdaGread sum of squares of per feature historical gradients
for p in self.params:
setattr(self, p + "_hg", np.zeros_like(getattr(self, p)))
self.reset_state()
self.initialized = True
SEQUENCE_LENGTH = 250
EMBEDDING_DIM = 100
HIDDEN_SIZE = 150
ATTENTION_SIZE = 50
KEEP_PROB = 0.8
BATCH_SIZE = 256
NUM_EPOCHS = 3 # Model easily overfits without pre-trained words embeddings, that's why train for a few epochs
DELTA = 0.5
MODEL_PATH = './model'
# Load the data set
(X_train, y_train), (X_test, y_test) = imdb.load_data(num_words=NUM_WORDS,
index_from=INDEX_FROM)
# Sequences pre-processing
vocabulary_size = get_vocabulary_size(X_train)
X_test = fit_in_vocabulary(X_test, vocabulary_size)
X_train = zero_pad(X_train, SEQUENCE_LENGTH)
X_test = zero_pad(X_test, SEQUENCE_LENGTH)
# Different placeholders
with tf.name_scope('Inputs'):
batch_ph = tf.placeholder(tf.int32, [None, SEQUENCE_LENGTH],
name='batch_ph')
target_ph = tf.placeholder(tf.float32, [None], name='target_ph')
seq_len_ph = tf.placeholder(tf.int32, [None], name='seq_len_ph')
keep_prob_ph = tf.placeholder(tf.float32, name='keep_prob_ph')
# Embedding layer
with tf.name_scope('Embedding_layer'):
embeddings_var = tf.Variable(tf.random_uniform(
def read_sparse_input(lang_code, concepts_to_include, validation_set):
'''Reads data for langauge lang_code from a file in the sparse matrix folder.
All of the concepts not found in the set concepts_to_include are excluded.
Aditionaly the concepts found in the validation_set are also excluded from the read data (even if they are found in concepts_to_include).'''
rows = []
columns = []
data = []
concept_id_2_index = {}
intersection_with_concept_validation_set = set() # Sanity check
with gzip.open(utils.get_sparse_matrix_file_path(lang_code), 'r') as f:
# Used to check for repeated assignment of same cell (which should not happen)
non_zero = set()
# Used to check for word ID conformation
vocabulary_size = utils.get_vocabulary_size(lang_code)
documents_counter = 0
for line in codecs.getreader("utf-8")(f):
parts = line.split()
if len(parts) != 3:
raise Exception(
"Line formatting not as expected! Line: %s" % line)
concept = parts[0]
word_id = int(parts[1])
if concept not in concepts_to_include:
continue
if concept in validation_set:
intersection_with_concept_validation_set.add(concept)
continue
# Keeps track of which concept is described in the given document
if concept not in concept_id_2_index:
if documents_counter % 100000 == 0:
print("%d documents processed" % documents_counter)
sys.stdout.flush()
concept_id_2_index[concept] = documents_counter
documents_counter = documents_counter + 1
r = concept_id_2_index[concept]
if word_id > vocabulary_size:
raise Exception('Word_ID higher then vocabulary size!')
c = word_id
d = int(parts[2])
field = (r, c)
if field in non_zero:
# raise Exception('Repeated cell assignment!')
print('Repeated cell assignment!')
print(field)
else:
non_zero.add(field)
rows.append(r)
columns.append(c)
data.append(d)
index_2_concept_id = {
index: concept_id
for concept_id, index in concept_id_2_index.items()
}
# Generating the word index
word_id_2_index = {}
# Removes zero columns, by first generating a new index for the nonzero columns, and then updating the document representation to comply
# with the new index which contain only nonzero_columns
c = 0
non_zero_word_ids = sorted(set(columns))
for word_id in non_zero_word_ids:
word_id_2_index[word_id] = c
c += 1
for i in range(len(columns)):
word_id = columns[i]
columns[i] = word_id_2_index[word_id]
index_2_word_id = {
index: word_id
for word_id, index in word_id_2_index.items()
}
print("Intersection with validation set in %s counts %d elements" %
(lang_code,
len(intersection_with_concept_validation_set))) # Sanity check
sys.stdout.flush()
temp_processed_data = {'rows' : np.array(rows), 'columns' : np.array(columns), 'data' : np.array(data), 'concept_id_2_index' : concept_id_2_index, \
'index_2_concept_id' : index_2_concept_id, 'word_id_2_index' : word_id_2_index, 'index_2_word_id' : index_2_word_id }
return temp_processed_data
def train_RNN(SEQUENCE_LENGTH, EMBEDDING_DIM, HIDDEN_SIZE, ATTENTION_SIZE,
KEEP_PROB, BATCH_SIZE, NUM_EPOCHS, DELTA, LEARNING_RATE,
ALPHA_DIVIDER, MAXIMUM_DATA_NUM):
t1 = time.time()
# Load preprocessed data
with open('raw_sums', 'rb') as fp:
raw_sums = pickle.load(fp)
with open('raw_texts', 'rb') as fp:
raw_texts = pickle.load(fp)
with open('vocab_limit', 'rb') as fp:
vocab_limit = pickle.load(fp)
#removing stopwords for the y_in of the arnn
y_trains = []
for line in raw_sums:
line = [
word for word in line if word not in open('english.txt').read()
]
y_trains.append(line)
print(y_trains[0])
# Embedd vocabulary(important for our int2word later to have only one embedding)
embd_vocab = []
max_words = len(vocab_limit)
tokenizer = Tokenizer(num_words=max_words)
# This builds the word index
tokenizer.fit_on_texts(vocab_limit)
# This turns strings into lists of integer indices.
embd_vocab = tokenizer.texts_to_sequences(vocab_limit)
embd_vocab = list(itertools.chain(*embd_vocab))
#embd_vocab.append(encoded_docs)
print('TRAIN_ARNN: vocab embedded.')
print('Länge EMBD_VOCAB: ' + str(len(embd_vocab)))
# Saving
with open('tokenizer.pickle', 'wb') as handle:
pickle.dump(tokenizer, handle, protocol=pickle.HIGHEST_PROTOCOL)
keys = vocab_limit[:]
values = embd_vocab[:]
dictionary = dict(zip(keys, values))
def word2int(texts):
embd_texts = []
for each_text in texts:
embd_text = []
for word in each_text:
for key in dictionary.keys():
if key == word:
embd_text.append(dictionary.get(key))
embd_texts.append(embd_text)
return embd_texts
print('TRAIN_ARNN: ' + str(texts) + ' embedded.')
embd_texts = word2int(raw_texts)
embd_sums = word2int(raw_sums)
embd_summaries = word2int(y_trains)
#%% ADDED BY MKLOENNE from summarization model -- Creating train and test sets
train_len = int((.7) * len(embd_sums))
train_texts = embd_texts[0:train_len]
train_summaries = embd_sums[0:train_len]
train_sums = embd_summaries[0:train_len]
val_len = int((.15) * len(embd_sums))
'''
val_texts = embd_texts[train_len:train_len+val_len]
val_summaries = embd_sums[train_len:train_len+val_len]
'''
test_texts = embd_texts[train_len + val_len:len(embd_sums)]
test_summaries = embd_sums[train_len + val_len:len(embd_sums)]
test_sums = embd_summaries[train_len + val_len:len(embd_summaries)]
# Load the dataset
(X_train, y_train), (X_test, y_tests) = (train_texts,
train_summaries), (test_texts,
test_sums)
y_trains = train_sums
# Convert text lists into numpy arrays to get shape like in the imdb dataset for train and test data
def convert2array(listtoconvert, comparing):
converted_arr = []
maxList = max(max(len(x) for x in listtoconvert),
max(len(x) for x in comparing))
pre_array = np.asarray([np.asarray(x) for x in listtoconvert])
for arr in pre_array:
arr = np.lib.pad(arr, (0, maxList - len(arr)),
'constant',
constant_values=0)
text_with_zero = np.concatenate([[1], arr])
text_with_zero = np.concatenate([text_with_zero, [0]])
converted_arr.append(text_with_zero)
converted_arr = np.asarray(converted_arr)
return converted_arr
max_y = max(max(len(x)
for x in train_sums), max(len(x) for x in test_sums)) + 2
y_trains = convert2array(train_sums, test_sums)
y_tests = convert2array(test_sums, train_sums)
X_test = convert2array(test_texts, train_texts)
X_train = convert2array(train_texts, test_texts)
y_test = convert2array(test_summaries, train_summaries)
y_train = convert2array(train_summaries, test_summaries)
print(str(X_train[0]))
#%%
def normalize(i, mini, maxi):
norm = (np.float64(i) - mini) / (maxi - mini)
return norm
vocabulary_size = int(get_vocabulary_size(X_train))
print('SIZE X_TRAIN:' + str(vocabulary_size))
# Different placeholders
batch_ph = tf.placeholder(tf.int32, [None, None], name="X")
target_ph = tf.placeholder(tf.float32, [None, None],
name="Y") #adapted for y
seq_len_ph = tf.placeholder(tf.int32, [None], name="SEQ_LEN")
keep_prob_ph = tf.placeholder(tf.float32, name="KEEP_PROB")
with tf.name_scope("Embedding-Layer"):
# Embedding layer
embeddings_var = tf.Variable(tf.random_uniform(
[vocabulary_size, EMBEDDING_DIM], 0.0, 1.0),
trainable=True,
name="embedding")
batch_embedded = tf.nn.embedding_lookup(embeddings_var, batch_ph)
with tf.name_scope("RNN"):
# (Bi-)RNN layer(-s)
rnn_outputs, _ = bi_rnn(GRUCell(HIDDEN_SIZE),
GRUCell(HIDDEN_SIZE),
inputs=batch_embedded,
sequence_length=seq_len_ph,
dtype=tf.float32)
#, _ = rnn(GRUCell(HIDDEN_SIZE), inputs=batch_embedded, sequence_length=seq_len_ph, dtype=tf.float32)
with tf.name_scope("Attention-Layer"):
# Attention layer
attention_output, alphas = attention(rnn_outputs,
ATTENTION_SIZE,
return_alphas=True)
with tf.name_scope("Dropout"):
# Dropout
drop = tf.nn.dropout(attention_output, keep_prob_ph)
with tf.name_scope("Dense"):
# Fully connected layer
W = tf.Variable(
tf.truncated_normal(shape=[HIDDEN_SIZE * 2, max_y], stddev=0.1),
name="weights") # Hidden size is multiplied by 2 for Bi-RNN
b = tf.Variable(tf.constant(0., shape=[max_y]), name="bias")
y_hat = tf.nn.xw_plus_b(drop, W, b)
y_hat = tf.squeeze(y_hat)
with tf.name_scope("loss"):
# Cross-entropy loss and optimizer initialization
loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=y_hat,
labels=target_ph))
# = tf.reduce_mean(tf.nn.nce_loss(weights=W, biases=b, labels=target_ph, inputs=y_hat))
tf.summary.scalar('loss', loss)
optimizer = tf.train.AdamOptimizer(
learning_rate=LEARNING_RATE).minimize(loss)
with tf.name_scope("Accuracy"):
# Accuracy metric
accuracy = tf.reduce_mean(
tf.cast(tf.equal(tf.round(tf.sigmoid(y_hat)), target_ph),
tf.float32))
tf.summary.scalar('Accuracy', accuracy)
# rec, rec_op = tf.metrics.recall(labels=target_ph, predictions=y_hat)
# pre, pre_op = tf.metrics.precision(labels=target_ph, predictions=y_hat)
# Batch generators
train_batch_generator = batch_generator(X_train, y_trains, BATCH_SIZE)
test_batch_generator = batch_generator(X_test, y_tests, BATCH_SIZE)
#%%
print(type([x for x in X_train[0]]))
saver = tf.train.Saver()
# Train session
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train_writer = tf.summary.FileWriter('./logs/rnn_logs' + '/train')
train_writer.add_graph(sess.graph)
test_writer = tf.summary.FileWriter('./logs/rnn_logs' + '/test')
test_writer.add_graph(sess.graph)
merged = tf.summary.merge_all()
losses = []
print("TRAIN_ARNN: Start learning...")
for epoch in range(NUM_EPOCHS):
loss_train = 0
loss_test = 0
accuracy_train = 0
accuracy_test = 0
print("epoch: {}\t".format(epoch), end="")
# Training
num_batches = X_train.shape[0] // BATCH_SIZE
for b in range(num_batches):
x_batch, y_batch = next(train_batch_generator)
maxValue_y = max([max(sublist) for sublist in y_batch])
minValue_y = min([min(sublist) for sublist in y_batch])
j = 0
y_batch_normed = []
for each_y_batch in y_batch:
norm_y = [
normalize(i, minValue_y, maxValue_y)
for i in each_y_batch
]
y_batch_normed.append(norm_y)
j += 1
seq_len = np.array([list(x).index(0) + 1 for x in x_batch
]) # actual lengths of sequences
summary, loss_tr, acc, _ = sess.run(
[merged, loss, accuracy, optimizer],
feed_dict={
batch_ph: x_batch,
target_ph: y_batch_normed,
seq_len_ph: seq_len,
keep_prob_ph: KEEP_PROB
})
accuracy_train += acc
loss_train = loss_tr * DELTA + loss_train * (1 - DELTA)
train_writer.add_summary(summary, epoch * num_batches + b)
if accuracy_train != 0:
accuracy_train /= num_batches
saver.save(
sess,
"/home/kloe_mr/AnacondaProjects/tf-rnn-attention/Bachelorarbeit/BA/buffering-model-rnn.ckpt"
)
# Testing
print(X_test.shape)
num_batches = X_test.shape[0] // BATCH_SIZE
for b in range(num_batches):
x_batch, y_batch = next(test_batch_generator)
maxValue_y = max([max(sublist) for sublist in y_batch])
minValue_y = min([min(sublist) for sublist in y_batch])
j = 0
y_batch_normed = []
for each_y_batch in y_batch:
norm_y = [
normalize(i, minValue_y, maxValue_y)
for i in each_y_batch
]
y_batch_normed.append(norm_y)
j += 1
seq_len = np.array([list(x).index(0) + 1 for x in x_batch
]) # actual lengths of sequences
summary, loss_test_batch, acc = sess.run(
[merged, loss, accuracy],
feed_dict={
batch_ph: x_batch,
target_ph: y_batch_normed,
seq_len_ph: seq_len,
keep_prob_ph: 1.0
})
accuracy_test += acc
loss_test += loss_test_batch
test_writer.add_summary(summary, epoch * num_batches + b)
if loss_test == 0:
print('loss_test is zero!')
break
if accuracy_train != 0:
accuracy_test /= num_batches
loss_test /= num_batches
losses.append(loss_test)
print(
"loss: {:.3f}, val_loss: {:.3f}, acc: {:.3f}, val_acc: {:.3f}".
format(loss_train, loss_test, accuracy_train, accuracy_test))
#saver.save(sess, "H:/AnacondaProjects/tf-rnn-attention/tf-rnn-attention-master/model.ckpt") # "model")
saver.save(
sess,
"/home/kloe_mr/AnacondaProjects/tf-rnn-attention/tf-rnn-attention-master/model.ckpt"
) # "model")
#%% Restore session
with tf.Session() as sess:
saver.restore(
sess,
"/home/kloe_mr/AnacondaProjects/tf-rnn-attention/tf-rnn-attention-master/model.ckpt"
) #"model")
#saver.restore(sess, "H:/AnacondaProjects/tf-rnn-attention/tf-rnn-attention-master/model.ckpt")#"model")
max_y = max(max(len(x) for x in train_summaries),
max(len(x) for x in test_summaries)) + 2
x_batch_train, y_batch_train = X_train[:len(
X_train)], y_trains[:len(y_trains)]
seq_len_test = np.array(
[list(x).index(0) + 1 for x in x_batch_train])
maxValue_y = max([max(sublist) for sublist in y_batch])
minValue_y = min([min(sublist) for sublist in y_batch])
j = 0
y_batch_normed = []
for each_y_batch in y_batch_train:
norm_y = [
normalize(i, minValue_y, maxValue_y) for i in each_y_batch
]
y_batch_normed.append(norm_y)
j += 1
alphas_train = sess.run(
[alphas],
feed_dict={
batch_ph: x_batch_train,
target_ph: y_batch_normed,
seq_len_ph: seq_len_test,
keep_prob_ph: 1.0
})
#%% Restore session
with tf.Session() as sess:
saver.restore(
sess,
"/home/kloe_mr/AnacondaProjects/tf-rnn-attention/tf-rnn-attention-master/model.ckpt"
) #"model")
#saver.restore(sess, "H:/AnacondaProjects/tf-rnn-attention/tf-rnn-attention-master/model.ckpt")#"model")
max_y = max(max(len(x) for x in train_summaries),
max(len(x) for x in test_summaries)) + 2
x_batch_test, y_batch_test = X_test[:len(X_test
)], y_tests[:len(y_tests)]
seq_len_test = np.array(
[list(x).index(0) + 1 for x in x_batch_test])
maxValue_y = max([max(sublist) for sublist in y_batch])
minValue_y = min([min(sublist) for sublist in y_batch])
j = 0
y_batch_normed = []
for each_y_batch in y_batch_test:
norm_y = [
normalize(i, minValue_y, maxValue_y) for i in each_y_batch
]
y_batch_normed.append(norm_y)
j += 1
alphas_test = sess.run(
[alphas],
feed_dict={
batch_ph: x_batch_test,
target_ph: y_batch_normed,
seq_len_ph: seq_len_test,
keep_prob_ph: 1.0
})
#%% setting up the generated outputs by choosing the most weighted higlights out of the text for the LSTM input
alphas_values = alphas_train[:][0]
y_batch = y_train[:len(y_train)]
# Save visualization as HTML
rnn_outs = []
#with open("visualization.html", "w") as html_file:
for words, alphas in zip(x_batch_train, alphas_values):
rnn_out = []
Largest_alphas = len(words) // ALPHA_DIVIDER
if Largest_alphas == 0:
continue
min_value = min(heapq.nlargest(Largest_alphas, alphas, key=None))
for word, alpha in zip(
words, alphas): #_values):# / alphas_values.max()):
if alpha > min_value:
if word not in rnn_out: #think about wether it makes sense to remove 2nd appearance, maybe check for frequency in listheap
rnn_out.append(word)
if word == 0:
break
#html_file.write('<font style="background: rgba(255, 255, 0, %f)">%s</font>\n' % (alpha, word))
rnn_outs.append(rnn_out)
print('RNN_OUTS: ' + str(len(rnn_outs)))
alphas_values = alphas_test[:][0]
#y_batch = y_test[:len(y_test)]
#print(alphas_values[0])
# Save visualization as HTML
rnntest_outs = []
#with open("visualization.html", "w") as html_file:
for words, alphas in zip(x_batch_test, alphas_values):
rnn_out = []
Largest_alphas = len(words) // ALPHA_DIVIDER
if Largest_alphas == 0:
continue
min_value = min(heapq.nlargest(Largest_alphas, alphas, key=None))
for word, alpha in zip(
words, alphas): #_values):# / alphas_values.max()):
if alpha > min_value:
if word not in rnn_out: #think about wether it makes sense to remove 2nd appearance, maybe check for frequency in listheap
rnn_out.append(word)
if word == 0:
break
#html_file.write('<font style="background: rgba(255, 255, 0, %f)">%s</font>\n' % (alpha, word))
rnntest_outs.append(rnn_out)
# Creates list of lists with reembedded words
def int2word(data):
lists = []
for paragraph in data:
words = []
for word in paragraph:
if word == 1:
continue
if word == 0:
break
words.append(
list(dictionary.keys())[list(
dictionary.values()).index(word)])
lists.append(words)
return lists
rnn_out = int2word(rnn_outs)
batch_words = int2word(y_batch_train)
y_batch = int2word(y_batch)
print(rnn_out[0])
print(batch_words[0])
mean_testloss = np.mean(losses)
avg_testloss = np.mean(losses)
std_testloss = np.std(losses)
##EVENTUELL AUCH FÜR TEST DATA(????)
#in csv datei packen
t2 = time.time()
t = t2 - t1
r = Rouge()
def rouge_score(rnn_outs, rnn_out, batch_words):
d = []
for i in range(len(rnn_outs)):
system_generated_summary = rnn_out[i]
manual_summmary = batch_words[i]
try:
[precision, recall,
f_score] = r.rouge_l([system_generated_summary],
[manual_summmary])
except ZeroDivisionError:
continue
d.append(
(batch_words[i], rnn_out[i], precision, recall, f_score, t,
EMBEDDING_DIM, HIDDEN_SIZE, ATTENTION_SIZE, KEEP_PROB,
BATCH_SIZE, NUM_EPOCHS, DELTA, LEARNING_RATE, ALPHA_DIVIDER,
MAXIMUM_DATA_NUM, mean_testloss, avg_testloss, std_testloss))
print("Summary" + str(i) + "\nPrecision is :" + str(precision) +
"\nRecall is :" + str(recall) + "\nF Score is :" +
str(f_score))
return d
d = rouge_score(rnn_outs, rnn_out, batch_words)
df = pd.DataFrame(
d,
columns=('Original Sum', 'Predicted Sum', 'Precision', 'Recall',
'F_Score', 'RUNTIME/sec', 'EMBEDDING_DIM', 'HIDDEN_SIZE',
'ATTENTION_SIZE', 'KEEP_PROB', 'BATCH_SIZE', 'EPOCHS',
'DELTA', 'LEARNING_RATE', 'ALPHA_DIVIDER', 'MAXIMUM_DATA_NUM',
'MEAN_TESTLOSS', 'AVG_TESTLOSS', 'STD_TESTLOSS'))
print(df.head(2))
df.to_csv("arnn_train_1" + str(BATCH_SIZE) + "_batches" + str(NUM_EPOCHS) +
"_epochs.csv",
sep=',')
#%%
# saving
with open('rnn_out.pickle', 'wb') as handle:
pickle.dump(rnn_out, handle, protocol=pickle.HIGHEST_PROTOCOL)
# saving
with open('batch_words.pickle', 'wb') as handle:
pickle.dump(y_batch, handle, protocol=pickle.HIGHEST_PROTOCOL)
rnntest_out = int2word(rnntest_outs)
batch_words = int2word(y_batch_test)
d = rouge_score(rnntest_outs, rnntest_out, batch_words)
df = pd.DataFrame(
d,
columns=('Original Sum', 'Predicted Sum', 'Precision', 'Recall',
'F_Score', 'RUNTIME/sec', 'EMBEDDING_DIM', 'HIDDEN_SIZE',
'ATTENTION_SIZE', 'KEEP_PROB', 'BATCH_SIZE', 'EPOCHS',
'DELTA', 'LEARNING_RATE', 'ALPHA_DIVIDER', 'MAXIMUM_DATA_NUM',
'MEAN_TESTLOSS', 'AVG_TESTLOSS', 'STD_TESTLOSS'))
print(df.head(2))
df.to_csv("arnn_test_1" + str(BATCH_SIZE) + "_batches" + str(NUM_EPOCHS) +
"_epochs.csv",
sep=',')
def train(model_name, p1_train_data, p1_dev_data, p2_train_data, p2_dev_data):
### PHASE 1 ###
training_data = np.load(p1_train_data)
validation_data = np.load(p1_dev_data)
assert training_data["batch_size"] == validation_data["batch_size"]
assert training_data["vocabulary"] == validation_data["vocabulary"]
assert training_data["punctuations"] == validation_data["punctuations"]
print "1st phase data loaded..."
print "Vocabulary size is %d" % utils.get_vocabulary_size(
validation_data["vocabulary"])
print "Training set size is %d" % training_data["total_size"]
print "Validation set size is %d" % validation_data["total_size"]
net = models.T_LSTM()
net.initialize(hidden_size=conf.PHASE1["HIDDEN_SIZE"],
projection_size=conf.PHASE1["PROJECTION_SIZE"],
in_vocabulary=training_data["vocabulary"],
out_vocabulary=training_data["punctuations"],
batch_size=training_data["batch_size"],
hidden_activation=conf.PHASE1["HIDDEN_ACTIVATION"],
bptt_steps=conf.PHASE1["BPTT_STEPS"],
use_pauses=False)
_train(net, training_data, validation_data, model_name,
conf.PHASE1["LEARNING_RATE"], conf.PHASE1["MAX_EPOCHS"],
conf.PHASE1["MIN_IMPROVEMENT"])
### PHASE 2 ###
if not os.path.isfile(p2_train_data) or not os.path.isfile(p2_train_data):
print "No second phase data."
return
training_data = np.load(p2_train_data)
validation_data = np.load(p2_dev_data)
assert training_data["batch_size"] == validation_data[
"batch_size"] == net.batch_size
assert training_data["vocabulary"] == validation_data[
"vocabulary"] == net.in_vocabulary
assert training_data["punctuations"] == validation_data[
"punctuations"] == net.out_vocabulary
print "2nd phase data loaded..."
print "Training set size is %d" % training_data["total_size"]
print "Validation set size is %d" % validation_data["total_size"]
print "Trainging %s pause durations." % (
"with" if conf.PHASE2["USE_PAUSES"] else "without")
t_lstm = net
net = models.TA_LSTM()
net.initialize(hidden_size=conf.PHASE2["HIDDEN_SIZE"],
t_lstm=t_lstm,
out_vocabulary=training_data["punctuations"],
batch_size=training_data["batch_size"],
hidden_activation=conf.PHASE2["HIDDEN_ACTIVATION"],
bptt_steps=conf.PHASE2["BPTT_STEPS"],
use_pauses=conf.PHASE2["USE_PAUSES"])
_train(net, training_data, validation_data, model_name,
conf.PHASE2["LEARNING_RATE"], conf.PHASE2["MAX_EPOCHS"],
conf.PHASE2["MIN_IMPROVEMENT"])