diff --git a/.idea/deployment.xml b/.idea/deployment.xml
index a0101359350c91a85fc7a058e2211bbc993ccb65..242dd10ff439bfb0dc8cfd560ff76be8c1b1def9 100644
--- a/.idea/deployment.xml
+++ b/.idea/deployment.xml
@@ -1,6 +1,6 @@
<?xml version="1.0" encoding="UTF-8"?>
<project version="4">
- <component name="PublishConfigData" autoUpload="Always" serverName="guppi" createEmptyFolders="true" autoUploadExternalChanges="true">
+ <component name="PublishConfigData" autoUpload="Always" serverName="guppi" createEmptyFolders="true" persistUploadOnCheckin="false" autoUploadExternalChanges="true">
<serverData>
<paths name="guppi">
<serverdata>
diff --git a/code_jurica/classificationICD10_attention.py b/code_jurica/classificationICD10_attention.py
new file mode 100644
index 0000000000000000000000000000000000000000..c8d794fe0e31333ce9bd7bdc6ee34979b45efe01
--- /dev/null
+++ b/code_jurica/classificationICD10_attention.py
@@ -0,0 +1,127 @@
+# from comet_ml import Experiment
+# experiment=Experiment(api_key="hSd9vTj0EfMu72569YnVEvtvj")
+
+# from loader import *
+from util import *
+import numpy as np
+import random
+import traceback
+from sklearn.preprocessing import LabelEncoder
+from sklearn.model_selection import train_test_split
+
+from keras.preprocessing.sequence import pad_sequences
+from keras.preprocessing.text import Tokenizer
+from keras.optimizers import Adam
+from keras.callbacks import EarlyStopping, ModelCheckpoint, CSVLogger
+from keras.layers import Embedding, Input, LSTM, Dense, Bidirectional
+from keras.models import Model
+from keras.utils import multi_gpu_model, np_utils
+
+import tensorflow as tf
+
+from _layers import AttentionWithContext, Attention
+
+###################################
+# TensorFlow wizardry
+config=tf.ConfigProto()
+
+# Don't pre-allocate memory; allocate as-needed
+config.gpu_options.allow_growth=True
+config.gpu_options.allocator_type='BFC'
+
+callbacks_list=[
+ EarlyStopping(
+ monitor='val_loss',
+ patience=2,
+ ),
+ ModelCheckpoint(
+ filepath='models/icd10Classification_attention.h5',
+ monitor='val_loss',
+ save_best_only=True,
+ ),
+ CSVLogger(
+ append=True,
+ filename='logs/icd10Classification_attention_{}.csv'.format(date_label),
+ )
+]
+
+latent_dim = 512
+epochs = 100
+batch_size = 1000
+
+tokenizer=TokenizePreprocessor()
+kerasTokenizer = Tokenizer()
+dataLoader=prepareData()
+corpora=dataLoader.prepareDictionaries()
+print("Extracted {} data points".format(len(corpora)))
+# input('size of corpora')
+
+#prepareing the texts for input in RNN
+corpus=[x[0] for x in corpora]
+tokens=tokenizer.transform([x for x in corpus])
+tmp=[item for item in list(set(flatten(tokens))) if item.strip()]
+vocabulary={item.strip():i+1 for i,item in enumerate(tmp)}
+index_to_word_dict={i+1:item.strip() for i,item in enumerate(tmp)}
+kerasTokenizer.word_index=vocabulary
+# saving
+with open('models/icd10_tokenizer.p', 'wb') as handle:
+ pickle.dump(kerasTokenizer, handle)
+
+source_word_sequence=kerasTokenizer.texts_to_sequences(corpus)
+max_sequence = max([len(x) for x in source_word_sequence])
+word_sequence = pad_sequences(source_word_sequence, maxlen=max_sequence, padding='post')
+
+embedding_matrix=embedding_matrix(vocabulary)
+embedding_layer = Embedding(
+ embedding_matrix.shape[0],
+ embedding_matrix.shape[1],
+ weights=[embedding_matrix],
+ input_length=max_sequence,
+ trainable=True,
+ mask_zero=True)
+
+#preparing the labels as one hot encoding vector
+labels=[x[1] for x in corpora]
+num_labels=len(list(set(labels)))
+print("Labels: {}\tUnique labels:{}".format(len(labels), num_labels))
+encoder = LabelEncoder()
+encoder.fit(labels)
+with open('models/icd10_mappings.p', 'wb') as handle:
+ pickle.dump(encoder, handle)
+
+encoded_Y = encoder.transform(labels)
+
+# convert integers to dummy variables (i.e. one hot encoded)
+labels_one_hot = np_utils.to_categorical(encoded_Y)
+
+X_train, X_test, Y_train, Y_test = train_test_split(word_sequence, labels_one_hot, test_size=0.05, random_state=777)
+print("Prepared data: ", len(X_train), len(Y_train), len(X_test), len(Y_test))
+
+try:
+ # LAYERS
+ print("Creating Model...")
+ inputs = Input(shape=(max_sequence,))
+ embedding = embedding_layer(inputs)
+ decoder_LSTM = Bidirectional(LSTM(latent_dim, return_sequences=True))
+ decoder_out = decoder_LSTM(embedding) #, initial_state=encoder_states)
+ attention = Attention()(decoder_out)
+ decoder_dense = Dense(num_labels, activation='softmax')
+ decoder_out = decoder_dense(attention)
+
+ #MODEL
+ model = Model(inputs=inputs, outputs=decoder_out)
+ adam = Adam(lr=1e-4, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
+ model.compile(optimizer=adam, loss='categorical_crossentropy', metrics=['accuracy'])
+ model.summary()
+ print("Traning Model...")
+ model.fit(X_train, Y_train,
+ batch_size=batch_size,
+ epochs=epochs,
+ callbacks=callbacks_list,
+ validation_split=0.25
+ )
+
+except Exception as e:
+ print(e)
+ traceback.print_exc()
+
diff --git a/code_jurica/classificationICD10_attention_char.py b/code_jurica/classificationICD10_attention_char.py
new file mode 100644
index 0000000000000000000000000000000000000000..9bf696abfc64b0ef4b6a095aee7a608fc3783d73
--- /dev/null
+++ b/code_jurica/classificationICD10_attention_char.py
@@ -0,0 +1,141 @@
+# from comet_ml import Experiment
+# experiment=Experiment(api_key="hSd9vTj0EfMu72569YnVEvtvj")
+
+# from loader import *
+from util import *
+import numpy as np
+import random
+import traceback
+from sklearn.preprocessing import LabelEncoder
+from sklearn.model_selection import train_test_split
+
+from keras.preprocessing.sequence import pad_sequences
+from keras.preprocessing.text import Tokenizer
+from keras.optimizers import Adam
+from keras.callbacks import EarlyStopping, ModelCheckpoint, CSVLogger
+from keras.layers import Embedding, Input, LSTM, Dense, Bidirectional
+from keras.models import Model
+from keras.utils import multi_gpu_model, np_utils
+
+import tensorflow as tf
+
+from _layers import AttentionWithContext, Attention
+
+###################################
+# TensorFlow wizardry
+config=tf.ConfigProto()
+
+# Don't pre-allocate memory; allocate as-needed
+config.gpu_options.allow_growth=True
+config.gpu_options.allocator_type='BFC'
+
+callbacks_list=[
+ EarlyStopping(
+ monitor='val_loss',
+ patience=2,
+ ),
+ ModelCheckpoint(
+ filepath='models/icd10Classification_attention_char.h5',
+ monitor='val_loss',
+ save_best_only=True,
+ ),
+ CSVLogger(
+ append=True,
+ filename='logs/icd10Classification_attention_char_{}.csv'.format(date_label),
+ )
+]
+
+latent_dim = 256
+epochs = 100
+batch_size = 1000
+
+sentences = []
+labels=[]
+chars = set()
+
+# tokenizer=TokenizePreprocessor()
+kerasTokenizer = Tokenizer(char_level=True, filters=None)
+dataLoader=prepareData()
+corpora=dataLoader.prepareDictionaries()
+print("Extracted {} data points".format(len(corpora)))
+# input('size of corpora')
+
+for line in corpora:
+
+ labels.append(line[1])
+ sentences.append(line[0])
+
+ for ch in line[0]:
+ if (ch not in chars):
+ chars.add(ch)
+
+chars = sorted(list(chars))
+
+char_to_index_dict={item.strip():i+1 for i,item in enumerate(chars)}
+index_to_char_dict={i+1:item.strip() for i,item in enumerate(chars)}
+
+kerasTokenizer.word_index=char_to_index_dict
+# saving
+with open('models/icd10_char_tokenizer.p', 'wb') as handle:
+ pickle.dump(kerasTokenizer, handle)
+
+char_sequence=kerasTokenizer.texts_to_sequences(sentences)
+# print(char_sequence)
+max_sequence = max([len(x) for x in char_sequence])
+word_sequence = pad_sequences(char_sequence, maxlen=max_sequence, padding='post')
+
+embedding_matrix=embedding_matrix(char_to_index_dict)
+embedding_layer = Embedding(
+ embedding_matrix.shape[0],
+ embedding_matrix.shape[1],
+ weights=[embedding_matrix],
+ input_length=max_sequence,
+ trainable=True,
+ mask_zero=True)
+
+#preparing the labels as one hot encoding vector
+
+num_labels=len(list(set(labels)))
+print("Labels: {}\tUnique labels:{}".format(len(labels), num_labels))
+encoder = LabelEncoder()
+encoder.fit(labels)
+with open('models/icd10_char_mappings.p', 'wb') as handle:
+ pickle.dump(encoder, handle)
+
+encoded_Y = encoder.transform(labels)
+
+# convert integers to dummy variables (i.e. one hot encoded)
+labels_one_hot = np_utils.to_categorical(encoded_Y)
+
+X_train, X_test, Y_train, Y_test = train_test_split(word_sequence, labels_one_hot, test_size=0.05, random_state=777)
+print("Prepared data: ", len(X_train), len(Y_train), len(X_test), len(Y_test))
+
+try:
+ # LAYERS
+ print("Creating Model...")
+ inputs = Input(shape=(max_sequence,))
+ embedding = embedding_layer(inputs)
+ decoder_LSTM = Bidirectional(LSTM(latent_dim, return_sequences=True))
+ decoder_out = decoder_LSTM(embedding) #, initial_state=encoder_states)
+ attention = Attention()(decoder_out)
+ decoder_dense = Dense(num_labels, activation='softmax')
+ decoder_out = decoder_dense(attention)
+
+ #MODEL
+ model = Model(inputs=inputs, outputs=decoder_out)
+ # model = multi_gpu_model(tmp_model, gpus=2)
+ adam = Adam(lr=1e-4, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
+ model.compile(optimizer=adam, loss='categorical_crossentropy', metrics=['accuracy'])
+ model.summary()
+ print("Traning Model...")
+ model.fit(X_train, Y_train,
+ batch_size=batch_size,
+ epochs=epochs,
+ callbacks=callbacks_list,
+ validation_split=0.25
+ )
+
+except Exception as e:
+ print(e)
+ traceback.print_exc()
+
diff --git a/code_jurica/seq2seq.py b/code_jurica/seq2seq.py
index 41f589bba289c58c20afdb3b8c6e4de7239b4b51..1b4087c37a5be3c91e0214c19bb8f7148af5c62e 100644
--- a/code_jurica/seq2seq.py
+++ b/code_jurica/seq2seq.py
@@ -19,14 +19,12 @@ config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.allocator_type = 'BFC'
-
# LOAD ICD 10 CLASSIFICATION MODEL
-
try:
- icd10_model = keras_load_model('models/classificationICD10.h5')
+ icd10_model = keras_load_model('models/icd10Classification.h5')
except OSError:
from classificationICD10 import *
- icd10_model = keras_load_model('models/classificationICD10.h5')
+ icd10_model = keras_load_model('models/icd10Classification.h5')
with open('models/icd10_tokenizer.p', 'rb') as handle:
kerasTokenizer = pickle.load(handle)