def test_testing_convolution(text_vectorizer, class_vectorizer, classif_level, classif_type, dataset_location):
load_standard_sets = True
root_location = fh.get_root_location('data/convolutional_outcome/')
sets_location = fh.join_paths(root_location, "model_sets")
checkpoint_path = fh.join_paths(root_location, "model_checkpoints")
model_path = fh.link_paths(checkpoint_path, 'convolution_model')
weights_path = fh.link_paths(checkpoint_path, 'convolution_weights')
# load sets and settings
model_name = text_vectorizer+'/'+class_vectorizer+'/NN'
# train_data, test_data, val_data, train_labels, test_labels, val_labels, settings = ch.load_sets(sets_location, '2020-01-08T22:28:44.757410')
# classes, n_classes, vocab_processor, len_vocabulary = settings
# only for test
train_data, test_data, val_data, train_labels, test_labels, val_labels, _ = ch.load_sets(sets_location)
# it could be that a label is only in the test/data data, might be a problem
sequence_length = train_data.shape[1]
# calculates metrics with testing data
model, predictions = pmh.run_cnn_test(_,
train_data, train_labels, test_data, test_labels, val_data, val_labels,
model_path, weights_path, False)
binary_predictions = mh.get_binary_0_5(predictions)
# display testing metrics
metrics = mh.get_sequential_metrics(test_labels, predictions, binary_predictions)
mh.display_sequential_metrics('testing convolution sequence', metrics)
classifier_name, layers = ch.get_sequential_classifier_information(model)
ch.save_results(classifier_name+' Test', metrics, layers, model_name, classif_level, classif_type, dataset_location)
def save_data_frame(script_key, data_frame, csvfile):
index = settings[script_key]["data_frame_file_name"].rfind('/')
path_to_csv = fh.get_root_location(settings[script_key]["data_frame_file_name"][:index])
if csvfile:
output_path = fh.link_paths(path_to_csv, csvfile)
else:
output_path = fh.link_paths(path_to_csv, settings[script_key]["data_frame_file_name"][index+1:])
data_frame.to_csv(output_path, index=False, sep=',', header=False, quoting=csv.QUOTE_NONE, quotechar="", escapechar=" ")
def apply_df_vectorizer(data_frame, type_, classification_type, model_name):
model_path = fh.get_root_location('data/vectorizer_models/')
# type_ = "doc2vec"
# classification_type = "multi_label"
index = type_.rfind('/')
if index != -1:
train, test = data_frame
data_frame = pd.concat([train, test])
type_ = type_[index+1:]
if type_ == 'doc2vec':
X_train, X_test, y_train, y_test, classes, n_classes, patent_ids = apply_doc2vec(data_frame, classification_type, model_path)
# array - not numpy
print('finisched apply doc2vec')
vocab_processor, len_vocabulary = None, 0
elif type_ == 'standard':
standard_results = apply_standard_vectorizer(data_frame, classification_type)
X_train, X_test, y_train, y_test, classes, n_classes, vocab_processor, len_vocabulary = standard_results
patent_ids = data_frame['patent_id']
else:
Y_vect, classes, n_classes = apply_classification_vectorizer(classification_type, data_frame)
vocab_processor, len_vocabulary = None, 0
if type_ == 'tfidf':
X_vect = pd.DataFrame({'text' : data_frame['text'], 'patent_id' : data_frame['patent_id']})
X_train, X_test, y_train, y_test = get_train_test_from_data(X_vect, Y_vect)
patent_ids = X_train['patent_id']
vectorizer = apply_tfidf_vectorizer_fit(X_vect, model_path)
# print('prima: ', X_test.iloc[0])
X_train = apply_tfidf_vectorizer_transform(vectorizer, X_train)
X_test = apply_tfidf_vectorizer_transform(vectorizer, X_test)
# print('dopo: ', X_test[0])
# X_train = apply_tfidf_vectorizer_fit_transform(vectorizer, X_train)
# X_test = apply_tfidf_vectorizer_fit_transform(vectorizer, X_test)
# csr matrix
elif type_ == 'word2vec': # required google pre-trained vectors, it starts downloading if you don't have
X_vect = apply_word2vec_word_averaging(data_frame)
# ndarray
print('finished word2vec word averaging')
text_indexes = data_frame.index.values
X_vect_temp = pd.DataFrame({'text' : text_indexes, 'patent_id' : data_frame['patent_id']})
X_train_temp, X_test_temp, y_train, y_test = get_train_test_from_data(X_vect_temp, Y_vect)
X_train, patent_ids = th.get_set_from_index(X_train_temp, X_vect)
X_test, _ = th.get_set_from_index(X_test_temp, X_vect)
print('finished word2vec vectorizer')
else:
data_frame.drop(columns=['classification'], inplace=True)
X_train, X_test, y_train, y_test = get_train_test_from_data(data_frame, Y_vect)
return X_train, X_test, y_train, y_test, classes, n_classes, vocab_processor, len_vocabulary
save_training_set(patent_ids, model_name)
return X_train, X_test, y_train, y_test, classes, n_classes, vocab_processor, len_vocabulary
def second_attempt_from_web(data_frame, text_vectorizer, class_vectorizer, classif_level, classif_type, dataset_location):
root_location = fh.get_root_location('data/convolutional_outcome/')
# imdb = keras.datasets.imdb
# (train_data, train_labels), (test_data, test_labels) = imdb.load_data(num_words=10000)
# explore data
# print("Training entries: {}, labels: {}".format(len(train_data), len(train_labels)))
# print("how data looks like: ", train_data[0]) # [1, 14, 34...]
# print("how labels looks like: ", train_labels[0]) # 0
# preprocess data
# A dictionary mapping words to an integer index
# word_index = imdb.get_word_index()
# The first indices are reserved, thus it increases all the indices by 3
# word_index = {k : (v + 3) for k, v in word_index.items()}
# word_index["<PAD>"] = 0
# word_index["<START>"] = 1
# word_index["<UNK>"] = 2 # unknown
# word_index["<UNUSED>"] = 3
# make both the train and the test dataset the same length
# train_data = keras.preprocessing.sequence.pad_sequences(train_data,
# value=word_index["<PAD>"],
# padding='post',
# maxlen=256)
# test_data = keras.preprocessing.sequence.pad_sequences(test_data,
# value=word_index["<PAD>"],
# padding='post',
# maxlen=256)
# print(train_data[0]) # [1 14 34 0 0 0] - with little difference: PAD,START,UNK,UNUSED
model_name = text_vectorizer+'/'+class_vectorizer+'/NN'
standard_results = ch.apply_df_vectorizer(data_frame, text_vectorizer, class_vectorizer, model_name)
train_data, test_data, train_labels, test_labels, classes, n_classes, vocab_processor, len_vocabulary = standard_results
train_data, val_data, train_labels, val_labels = ch.get_train_test_from_data(train_data, train_labels)
# print("Training entries: {}, labels: {}".format(len(train_data), len(train_labels)))
# print("how data looks like: ", train_data[0]) # [1 14 34 0 0 0]
# print("how labels looks like: ", train_labels[0:5]) # list of lists [[0 1 0 0 0 1 0 1], ...]
model = pmh.get_text_convolutional_from_web(len_vocabulary, n_classes)
metrics, predictions = pmh.run_text_cnn_model(model, train_data, train_labels, test_data, test_labels)
classifier_name, layers = ch.get_sequential_classifier_information(model)
mh.display_convolutional_metrics(classifier_name, metrics[0], metrics[1], metrics[2], test_labels, predictions)
ch.save_results(classifier_name, metrics, layers, model_name, classif_level, classif_type, dataset_location)
def test_basic_LSTM(data_frame, text_vectorizer, classif_level, classif_type, dataset_location):
print('### LSTM Doing Testing ###')
root_location = fh.get_root_location('data/lstm_outcome/')
doc2vec_model_location = fh.link_paths(fh.join_paths(root_location, "doc2vec_model/vocab_model/"), "doc2vec_model")
model_location = link_paths(join_paths(root_location, "lstm_model"), "lstm_model")
sets_location = join_paths(root_location, "model_sets")
save_results = True
load_model = True
# date problem
X_train, X_test, X_val, y_train, y_test, y_val, settings = load_sets(sets_location)
classes, n_classes, vocab_processor, len_vocabulary = settings
training_docs_list = X_train['patent_id']
test_docs_list = X_test['patent_id']
val_docs_list = X_val['patent_id']
sequence_size, embedding_size = 1, 150
X_data, Xv_data, Xt_data = ch.get_df_data(3, training_docs_list, val_docs_list, test_docs_list, sequence_size, embedding_size, doc2vec_model_location)
#####
print(X_data.shape)
input_size, output_size = X_data.shape[2], n_classes
if load_model:
model = tf.keras.models.load_model(model_location)
min_delta, patience = 0.00001, 15
early_stopper = keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=min_delta, patience=patience, verbose=1, mode='auto')
metrics_callback = MetricsCNNCallback(Xv_data, y_val, patience)
metrics, predictions, val_metrics = run_lstm(model,
X_data, y_train, Xt_data, y_test, Xv_data, y_val,
batch_size, [early_stopper, metrics_callback], queue_size,
training_docs_list, val_docs_list)
binary_predictions = mh.get_binary_0_5(predictions)
print('\nGenerating Testing Metrics')
metrics = mh.get_sequential_metrics(y_val, predictions, binary_predictions)
mh.display_sequential_metrics(metrics)
if save_results:
classifier_name, parameters = ch.get_sequential_classifier_information(model)
model_name = text_vectorizer+'/'+class_vectorizer+'/'+classifier_name
ch.save_results(classifier_name+'_LSTM', metrics, parameters, model_name, classif_level, classif_type, dataset_location)
print('end testing step')
def load_data_frame(script_key, csvfile):
index = settings[script_key]["data_frame_file_name"].rfind('/')
path_to_csv = fh.get_root_location(settings[script_key]["data_frame_file_name"][:index])
if csvfile:
input_path = fh.link_paths(path_to_csv, csvfile)
else:
input_path = fh.link_paths(path_to_csv, settings[script_key]["data_frame_file_name"][index+1:])
print('input_path: ', input_path)
data_frame = pd.read_csv(input_path, sep=',', quoting=csv.QUOTE_NONE, quotechar="", escapechar=" ", header=None, engine='python')
data_frame.columns = ['patent_id', 'text', 'classification']
# data_frame['text'] = data_frame['text'].apply(lambda text : clean_text(text))
data_frame = check_out_empty_texts_and_wrong_classcodes(data_frame)
classification_df = pd.DataFrame(columns=['class', 'count'])
# data_frame['classification'].apply(lambda classcode : th.calculate_class_distribution(classcode, classification_df))
return data_frame, classification_df
def train_testing_convolution(data_frame, text_vectorizer, class_vectorizer):
save_standard_sets = True
root_location = fh.get_root_location('data/convolutional_outcome/')
sets_location = fh.join_paths(root_location, "model_sets")
checkpoint_path = fh.join_paths(root_location, "model_checkpoints")
model_path = fh.link_paths(checkpoint_path, 'convolution_model')
weights_path = fh.link_paths(checkpoint_path, 'convolution_weights')
# get sets
model_name = text_vectorizer+'/'+class_vectorizer+'/NN'
standard_results = ch.apply_df_vectorizer(data_frame, text_vectorizer, class_vectorizer, model_name)
train_data, test_data, train_labels, test_labels, classes, n_classes, vocab_processor, len_vocabulary = standard_results
train_data, val_data, train_labels, val_labels = ch.get_train_test_from_data(train_data, train_labels)
# save sets
# ch.save_sets(sets_location, train_data, test_data, val_data, train_labels, test_labels, val_labels,
# [classes, n_classes, vocab_processor, len_vocabulary])
# this is for test
train_data, test_data, val_data, train_labels, test_labels, val_labels, _ = ch.load_sets(sets_location)
# it could be that a label is only in the test/data data, might be a problem
sequence_length = train_data.shape[1]
# define the model
model = pmh.get_cnn_test(len_vocabulary, n_classes, sequence_length)
# calculates metrics with validating data
model, val_predictions = pmh.run_cnn_test(model,
train_data, train_labels, val_data, val_labels, val_data, val_labels,
model_path, weights_path, True)
binary_val_predictions = mh.get_binary_0_5(val_predictions)
print(val_labels.shape)
print(val_predictions.shape)
# display validation metrics
metrics = mh.get_sequential_metrics(val_labels, val_predictions, binary_predictions)
mh.display_sequential_metrics('validation convolution sequence', metrics)
def test_LSTM_from_web(data_frame, text_vectorizer, class_vectorizer, classif_level, classif_type, dataset_location):
print('### LSTM Doing Testing ###')
root_location = fh.get_root_location('data/lstm_outcome/')
nn_parameter_search_location = fh.join_paths(root_location, "nn_fhv_parameter_search")
doc2vec_model_location = fh.link_paths(fh.join_paths(root_location, "doc2vec_model/vocab_model/"), "doc2vec_model")
save_results = True
sequence_size = 1
EMBEDDING_SIZE = 150
model_name = text_vectorizer+'/'+class_vectorizer+'/LSTM'
results = ch.apply_df_vectorizer(data_frame, text_vectorizer, class_vectorizer, model_name)
X_train, X_test, y_train, y_test, classes, n_classes, vocab_processor, len_vocabulary = results
X_train, X_val, y_train, y_val = ch.get_train_test_from_data(X_train, y_train)
training_docs_list = X_train['patent_id']
test_docs_list = X_test['patent_id']
val_docs_list = X_val['patent_id']
X_data, Xv_data, Xt_data = ch.get_df_data(3, training_docs_list, val_docs_list, test_docs_list, sequence_size, EMBEDDING_SIZE, doc2vec_model_location)
GLOBAL_VARS.DOC2VEC_MODEL_NAME, GLOBAL_VARS.MODEL_NAME = wmh.set_parameters_lstm_doc2vec(nn_parameter_search_location, classif_level, classif_type)
NN_INPUT_NEURONS, NN_SEQUENCE_SIZE, NN_OUTPUT_NEURONS = pmh.get_lstm_shapes(X_data, n_classes)
NN_BATCH_SIZE, PARTS_LEVEL, NN_MAX_EPOCHS, QUEUE_SIZE = pmh.get_lstm_basic_parameters()
params = pmh.get_lstm_testing_parameters()
lstm_output_size,w_dropout_do,u_dropout_do, stack_layers, conv_size, conv_filter_length, conv_max_pooling_length = params
EARLY_STOPPER_MIN_DELTA, EARLY_STOPPER_PATIENCE = pmh.get_early_stopping_parameters()
TEST_METRICS_FILENAME = '{}_level_{}_standard_nn_test_metrics_dict.pkl'
test_metrics_dict = dict()
test_metrics_path = fh.link_paths(fh.link_paths(nn_parameter_search_location, GLOBAL_VARS.MODEL_NAME), TEST_METRICS_FILENAME.format(classif_type, PARTS_LEVEL))
param_results_path = fh.link_paths(fh.link_paths(nn_parameter_search_location, GLOBAL_VARS.MODEL_NAME), NN_PARAMETER_SEARCH_PREFIX.format(classif_type, classif_level, NN_BATCH_SIZE))
param_results_dict = pickle.load(open(param_results_path, 'rb'))
GLOBAL_VARS.NN_MODEL_NAME = 'lstm_size_{}_w-drop_{}_u-drop_{}_stack_{}_conv_{}'.format(lstm_output_size,
w_dropout_do,
u_dropout_do,
stack_layers,
str(conv_size)
)
if conv_size:
GLOBAL_VARS.NN_MODEL_NAME += '_conv-filter-length_{}_max-pooling-size_{}'.format(conv_filter_length,
conv_max_pooling_length)
if GLOBAL_VARS.NN_MODEL_NAME not in param_results_dict.keys():
print("Can't find model: {}".format(GLOBAL_VARS.NN_MODEL_NAME))
raise Exception()
if fh.ensure_exists_path_location(test_metrics_path):
test_metrics_dict = pickle.load(open(fh.link_paths(fh.link_paths(nn_parameter_search_location, GLOBAL_VARS.MODEL_NAME), TEST_METRICS_FILENAME.format(classif_type,PARTS_LEVEL)), 'rb'))
if GLOBAL_VARS.NN_MODEL_NAME in test_metrics_dict.keys():
print("Test metrics already exist for: {}".format(GLOBAL_VARS.NN_MODEL_NAME))
test_metrics = test_metrics_dict[GLOBAL_VARS.NN_MODEL_NAME]
print("** Test Metrics: Cov Err: {:.3f}, Avg Labels: {:.3f}, \n\t\t Top 1: {:.3f}, Top 3: {:.3f}, Top 5: {:.3f}, \n\t\t F1 Micro: {:.3f}, F1 Macro: {:.3f}".format(
test_metrics['coverage_error'], test_metrics['average_num_of_labels'],
test_metrics['top_1'], test_metrics['top_3'], test_metrics['top_5'],
test_metrics['f1_micro'], test_metrics['f1_macro']))
raise Exception()
print('***************************************************************************************')
print(GLOBAL_VARS.NN_MODEL_NAME)
model = pmh.get_keras_rnn_model(NN_INPUT_NEURONS, NN_SEQUENCE_SIZE, NN_OUTPUT_NEURONS,
lstm_output_size, w_dropout_do, u_dropout_do, stack_layers, conv_size,
conv_filter_length, conv_max_pooling_length)
# get model best weights
weights = param_results_dict[GLOBAL_VARS.NN_MODEL_NAME]['best_weights']
model.set_weights(weights)
print('Evaluating on Test Data using best weights')
_, ytp, ytp_binary = pmh.predict_generator(None, model, Xt_data, y_test, NN_BATCH_SIZE, QUEUE_SIZE, test_docs_list)
print('Generating Test Metrics')
test_metrics = mh.get_sequential_metrics(y_test, ytp, ytp_binary)
mh.display_sequential_metrics(test_metrics)
if save_results:
classifier_name, parameters = ch.get_sequential_classifier_information(model)
ch.save_results(classifier_name+'_LSTM', test_metrics, parameters, model_name, classif_level, classif_type, dataset_location)
test_metrics_dict[GLOBAL_VARS.NN_MODEL_NAME] = test_metrics
pickle.dump(test_metrics_dict, open(test_metrics_path, 'wb'))
def train_LSTM_from_web(data_frame, text_vectorizer, class_vectorizer, classif_level, classif_type, dataset_location):
print('### LSTM Doing Training ###')
root_location = fh.get_root_location('data/lstm_outcome/')
exports_location = fh.join_paths(root_location, "exported_data/")
matrices_save_location = fh.join_paths(root_location, "fhv_matrices/")
nn_parameter_search_location = fh.join_paths(root_location, "nn_fhv_parameter_search")
doc2vec_model_location = fh.link_paths(fh.join_paths(root_location, "doc2vec_model/vocab_model/"), "doc2vec_model")
load_existing_results = True # it was True
save_results = True
sequence_size = 1
EMBEDDING_SIZE = 150
model_name = text_vectorizer+'/'+class_vectorizer+'/LSTM'
results = ch.apply_df_vectorizer(data_frame, text_vectorizer, class_vectorizer, model_name)
X_train, X_val, y_train, y_val, classes, n_classes, vocab_processor, len_vocabulary = results
training_docs_list = X_train['patent_id']
val_docs_list = X_val['patent_id']
X_data, Xv_data, _ = ch.get_df_data(2, training_docs_list, val_docs_list, None, sequence_size, EMBEDDING_SIZE, doc2vec_model_location)
GLOBAL_VARS.DOC2VEC_MODEL_NAME, GLOBAL_VARS.MODEL_NAME = wmh.set_parameters_lstm_doc2vec(nn_parameter_search_location, classif_level, classif_type)
# print(X_data.shape) # 64, 1, 200
# print(Xt_data.shape) # 20, 1, 200
# print(Xv_data.shape) # 16, 1, 200
NN_INPUT_NEURONS, NN_SEQUENCE_SIZE, NN_OUTPUT_NEURONS = pmh.get_lstm_shapes(X_data, n_classes)
NN_BATCH_SIZE, PARTS_LEVEL, NN_MAX_EPOCHS, QUEUE_SIZE = pmh.get_lstm_basic_parameters()
param_sampler = pmh.get_lstm_training_parameters()
EARLY_STOPPER_MIN_DELTA, EARLY_STOPPER_PATIENCE = pmh.get_early_stopping_parameters()
param_results_dict = dict()
param_results_path = fh.link_paths(fh.join_paths(nn_parameter_search_location, GLOBAL_VARS.MODEL_NAME), NN_PARAMETER_SEARCH_PREFIX.format(classif_type, classif_level, NN_BATCH_SIZE))
index = param_results_path.rfind('/')
fh.create_folder(fh.link_paths(fh.join_paths(nn_parameter_search_location, GLOBAL_VARS.MODEL_NAME), NN_PARAMETER_SEARCH_PREFIX.format(classif_type, classif_level, NN_BATCH_SIZE))[:index])
###########
# print(X_data.shape)
# input_size, sequence_size, output_size = X_data.shape[2], X_data.shape[0], n_classes
# NN_BATCH_SIZE, PARTS_LEVEL, NN_MAX_EPOCHS, QUEUE_SIZE = 1, 2, 200, 100
# lstm_output_sizes = [500, 1000]
# w_dropout_options = [0., 0.5]
# u_dropout_options = [0., 0.5]
# stack_layers_options = [1, 2, 3]
# lstm_output_size, w_dropout_do, u_dropout_do, stack_layers = 500, 0.0, 0.0, 1
# conv_size, conv_filter_length, max_pooling_length = 128, 2, 2
# EARLY_STOPPER_MIN_DELTA, EARLY_STOPPER_PATIENCE = 0.00001, 15
# import tensorflow as tf
# from tensorflow import keras
# from keras.layers import Input, Dense, Dropout, Activation
# from keras.models import Model, Sequential
# from keras.layers.convolutional import MaxPooling1D, Convolution1D
# from keras.layers.recurrent import LSTM
# model = Sequential()
# # model.add(Convolution1D(nb_filter=conv_size, input_shape=(sequence_size, input_size),
# # filter_length=conv_filter_length, border_mode='same', activation='relu'))
# # model.add(MaxPooling1D(pool_length=max_pooling_length))
# model.add(LSTM(lstm_output_size, activation='tanh', recurrent_activation='hard_sigmoid', use_bias=False,
# input_dim=input_size, dropout_W=w_dropout_do, dropout_U=u_dropout_do,
# implementation=1,
# return_sequences=False if 1 == stack_layers else True,
# go_backwards=False, stateful=False, unroll=False,
# name='lstm_{}_w-drop_{}_u-drop_{}_layer_{}'.format(lstm_output_size, str(u_dropout_do),
# str(w_dropout_do), str(1))))
# # model.add(LSTM(lstm_output_size, activation='tanh', recurrent_activation='hard_sigmoid', use_bias=False,
# # input_dim=input_size, dropout_W=w_dropout_do, dropout_U=u_dropout_do,
# # implementation=1,
# # return_sequences=False if 2 == stack_layers else True,
# # go_backwards=False, stateful=False, unroll=False,
# # name='lstm_{}_w-drop_{}_u-drop_{}_layer_{}'.format(lstm_output_size, str(u_dropout_do),
# # str(w_dropout_do), str(2))))
# # model.add(LSTM(lstm_output_size, activation='tanh', recurrent_activation='hard_sigmoid', use_bias=False,
# # input_dim=input_size, dropout_W=w_dropout_do, dropout_U=u_dropout_do,
# # implementation=1,
# # return_sequences=False if 3 == stack_layers else True,
# # go_backwards=False, stateful=False, unroll=False,
# # name='lstm_{}_w-drop_{}_u-drop_{}_layer_{}'.format(lstm_output_size, str(u_dropout_do),
# # str(w_dropout_do), str(3))))
# model.add(Dense(output_size, activation='sigmoid', name='sigmoid_output'))
# model.compile(optimizer='rmsprop', loss='binary_crossentropy')
# input_matrix = fh.join_paths(matrices_save_location, GLOBAL_VARS.MODEL_NAME)
# early_stopper = keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=EARLY_STOPPER_MIN_DELTA,
# patience=EARLY_STOPPER_PATIENCE, verbose=1, mode='auto')
# metrics_callback = mh.MetricsCallback(input_matrix, classif_type, PARTS_LEVEL, NN_BATCH_SIZE, is_mlp=False)
# history, yvp, yvp_binary = pmh.fit_predict_generator(model, X_data, y_train, Xv_data, y_val, training_docs_list, val_docs_list, early_stopper, metrics_callback, NN_BATCH_SIZE, NN_MAX_EPOCHS, QUEUE_SIZE)
# validation_metrics = mh.get_sequential_metrics(y_val, yvp, yvp_binary)
# mh.display_sequential_metrics(validation_metrics)
# ###########
# useful to skip all the already tested models
if load_existing_results:
param_results_path = fh.link_paths(fh.join_paths(nn_parameter_search_location, GLOBAL_VARS.MODEL_NAME),
NN_PARAMETER_SEARCH_PREFIX.format(classif_type, classif_level, NN_BATCH_SIZE))
if fh.ensure_exists_path_location(param_results_path):
print('Loading Previous results in {}'.format(param_results_path))
param_results_dict = pickle.load(open(param_results_path, 'rb'))
else:
print('No Previous results exist in {}'.format(param_results_path))
for params in param_sampler:
start_time = time.time()
lstm_output_size = params['lstm_output_size']
w_dropout_do = params['w_dropout']
u_dropout_do = params['u_dropout']
stack_layers = params['stack_layers']
conv_size = params['conv_size']
conv_filter_length = params['conv_filter_length']
conv_max_pooling_length = params['conv_max_pooling_length']
GLOBAL_VARS.NN_MODEL_NAME = 'lstm_size_{}_w-drop_{}_u-drop_{}_stack_{}_conv_{}'.format(
lstm_output_size, w_dropout_do, u_dropout_do, stack_layers, str(conv_size))
if conv_size:
GLOBAL_VARS.NN_MODEL_NAME += '_conv-filter-length_{}_max-pooling-size_{}'.format(conv_filter_length,
conv_max_pooling_length)
if GLOBAL_VARS.NN_MODEL_NAME in param_results_dict.keys():
print("skipping: {}".format(GLOBAL_VARS.NN_MODEL_NAME))
continue
# creating the actual keras model
model = pmh.get_keras_rnn_model(NN_INPUT_NEURONS, NN_SEQUENCE_SIZE, NN_OUTPUT_NEURONS,
lstm_output_size, w_dropout_do, u_dropout_do, stack_layers, conv_size,
conv_filter_length, conv_max_pooling_length)
classifier_name, parameters = ch.get_sequential_classifier_information(model)
model_name = text_vectorizer+'/'+class_vectorizer+'/'+classifier_name
input_matrix = fh.join_paths(matrices_save_location, GLOBAL_VARS.MODEL_NAME)
early_stopper = keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=EARLY_STOPPER_MIN_DELTA,
patience=EARLY_STOPPER_PATIENCE, verbose=1, mode='auto')
metrics_callback = mh.MetricsCallback(input_matrix, classif_type, PARTS_LEVEL, NN_BATCH_SIZE, is_mlp=False)
history, yvp, yvp_binary = pmh.fit_predict_generator(model, X_data, y_train, Xv_data, y_val, training_docs_list, val_docs_list, early_stopper, metrics_callback, NN_BATCH_SIZE, NN_MAX_EPOCHS, QUEUE_SIZE)
print('\nGenerating Validation Metrics')
validation_metrics = mh.get_sequential_metrics(y_val, yvp, yvp_binary)
mh.display_sequential_metrics(classifier_name, validation_metrics)
param_results_dict[GLOBAL_VARS.NN_MODEL_NAME] = dict()
# param_results_dict[GLOBAL_VARS.NN_MODEL_NAME]['best_validation_metrics'] = best_validation_metrics
param_results_dict[GLOBAL_VARS.NN_MODEL_NAME]['epochs'] = len(history.history['val_loss'])
param_results_dict[GLOBAL_VARS.NN_MODEL_NAME]['best_weights'] = metrics_callback.best_weights
param_results_dict[GLOBAL_VARS.NN_MODEL_NAME]['best_val_loss'] = metrics_callback.best_val_loss
param_results_dict[GLOBAL_VARS.NN_MODEL_NAME]['training_loss'] = metrics_callback.losses
param_results_dict[GLOBAL_VARS.NN_MODEL_NAME]['validation_loss'] = metrics_callback.val_losses
duration = time.time() - start_time
param_results_dict[GLOBAL_VARS.NN_MODEL_NAME]['duration'] = duration
ch.delete_variables(history, metrics_callback)
ch.save_results(classifier_name+'_LSTM', validation_metrics, parameters, model_name, classif_level, classif_type, dataset_location)
if save_results:
file = open(param_results_path, 'wb')
pickle.dump(param_results_dict, file)
def preprocessing_data_for_fasttext(data_frame, text_vectorizer,
class_vectorizer):
root_location = fh.get_root_location('data/fasttext_outcome/')
data_frame['text'] = data_frame['text'].replace(
'\n', ' ', regex=True).replace('\t', ' ', regex=True)
model_name = text_vectorizer + '/' + class_vectorizer + '/FastText'
try:
X_train, X_test, Y_train, Y_test, _, _, _, _ = ch.apply_df_vectorizer(
data_frame, text_vectorizer, class_vectorizer, model_name)
X_train, X_val, Y_train, Y_val = ch.get_train_test_from_data(
X_train, Y_train)
# self.data_vectors = pd.DataFrame(columns=range(vectors_size), index=range(corpus_size))
if not isinstance(X_train, pd.DataFrame):
train = pd.DataFrame(data=X_train)
test = pd.DataFrame(data=X_test)
val = pd.DataFrame(data=X_val)
# test_labels = pd.DataFrame(columns=[''])
else:
train = X_train
test = X_test
val = X_val
train.loc[:, 1] = Y_train
test.loc[:, 1] = Y_test
val.loc[:, 1] = Y_val
train.drop(columns=['patent_id'], inplace=True)
test.drop(columns=['patent_id'], inplace=True)
val.drop(columns=['patent_id'], inplace=True)
data_frame.to_csv(fh.link_paths(root_location, 'dataframe.csv'),
index=False,
sep=' ',
header=False,
quoting=csv.QUOTE_NONE,
quotechar="",
escapechar=" ")
train.to_csv(fh.link_paths(root_location, 'training set.csv'),
index=False,
sep=' ',
header=False,
quoting=csv.QUOTE_NONE,
quotechar="",
escapechar=" ")
test.to_csv(fh.link_paths(root_location, 'testing set.csv'),
index=False,
sep=',',
header=False,
quoting=csv.QUOTE_NONE,
quotechar="",
escapechar=" ")
except:
print('a problem occurred while trying to store the dataframes')
X_train, X_test, Y_train, Y_test, _, _, _, _ = ch.apply_df_vectorizer(
data_frame, text_vectorizer, class_vectorizer, model_name)
X_train, X_val, Y_train, Y_val = ch.get_train_test_from_data(
X_train, Y_train)
val = pd.DataFrame({'text': X_val, 'classification': Y_val})
train = pd.DataFrame({'text': X_train, 'classification': Y_train})
test = pd.DataFrame({'text': X_test, 'classification': Y_test})
data_frame.to_csv(fh.link_paths(root_location, 'dataframe.csv'),
index=False,
sep=' ',
header=False,
quoting=csv.QUOTE_NONE,
quotechar="",
escapechar=" ")
val.to_csv(fh.link_paths(root_location, 'validating set.csv'),
index=False,
sep=' ',
header=False,
quoting=csv.QUOTE_NONE,
quotechar="",
escapechar=" ")
train.to_csv(fh.link_paths(root_location, 'training set.csv'),
index=False,
sep=' ',
header=False,
quoting=csv.QUOTE_NONE,
quotechar="",
escapechar=" ")
test.to_csv(fh.link_paths(root_location, 'testing set.csv'),
index=False,
sep=',',
header=False,
quoting=csv.QUOTE_NONE,
quotechar="",
escapechar=" ")
def apply_fasttext(data_frame, text_vectorizer, class_vectorizer,
classif_level, classif_type, dataset_location):
root_location = fh.get_root_location('data/fasttext_outcome/')
fasttext_location = fh.join_paths(root_location, 'fasttext_models/')
word2vec_location = fh.join_paths(root_location, 'word2vec_models/')
timestamp = datetime.datetime.now().isoformat()
#############################
# unsupervised model
# model = fasttext.train_unsupervised(input=fh.link_paths(root_location, 'training set.csv'),
# autotuneValidationFile=fh.link_paths(root_location, 'training set.csv'))
#############################
# supervised model
# model = fasttext.train_supervised(input=fh.link_paths(root_location, 'training set.csv'),
# autotuneValidationFile=fh.link_paths(root_location, 'validating set.csv'), verbose=3, autotuneDuration=5000)
#############################
model, parameters = pmh.get_fasttext(
fh.link_paths(root_location, 'training set.csv'))
pmh.save_fasttext_model(
model, fh.link_paths(fasttext_location, 'model ' + timestamp + '.bin'))
# pmh.load_fasttext_model(fh.link_paths(fasttext_location, 'model '+timestamp+'.bin'))
test_labels, predictions, results = pmh.predict_test_fasttext(
model, fh.link_paths(root_location, 'testing set.csv'))
# print('predicted labels {}, probabilities of the labels {}'.format(predictions[0], predictions[1]))
result_top_15 = model.test(fh.link_paths(root_location, 'testing set.csv'),
k=15)
result_top_8 = model.test(fh.link_paths(root_location, 'testing set.csv'),
k=8)
result_top_5 = model.test(fh.link_paths(root_location, 'testing set.csv'),
k=5)
result_top_3 = model.test(fh.link_paths(root_location, 'testing set.csv'),
k=3)
result_top_1 = model.test(fh.link_paths(root_location, 'testing set.csv'),
k=1)
classifier_name = ch.get_fasttext_classifier_information(str(model))
model_name = text_vectorizer + '/' + class_vectorizer + '/' + classifier_name
mh.display_directly_metrics(
'k=-1 ' + classifier_name, -1, results[1], results[2],
2 * (results[1] * results[2]) / (results[1] + results[2]))
mh.display_directly_metrics(
'k= 15 ' + classifier_name, -1, result_top_15[1], result_top_15[2],
2 * (result_top_15[1] * result_top_15[2]) /
(result_top_15[1] + result_top_15[2]))
mh.display_directly_metrics(
'k= 8 ' + classifier_name, -1, result_top_8[1], result_top_8[2],
2 * (result_top_8[1] * result_top_8[2]) /
(result_top_8[1] + result_top_8[2]))
mh.display_directly_metrics(
'k= 5 ' + classifier_name, -1, result_top_5[1], result_top_5[2],
2 * (result_top_5[1] * result_top_5[2]) /
(result_top_5[1] + result_top_5[2]))
mh.display_directly_metrics(
'k= 3 ' + classifier_name, -1, result_top_3[1], result_top_3[2],
2 * (result_top_3[1] * result_top_3[2]) /
(result_top_3[1] + result_top_3[2]))
mh.display_directly_metrics(
'k= 1 ' + classifier_name, -1, result_top_1[1], result_top_1[2],
2 * (result_top_1[1] * result_top_1[2]) /
(result_top_1[1] + result_top_1[2]))
ch.save_results(classifier_name, results, parameters, model_name,
classif_level, classif_type, dataset_location)
manual_metrics = mh.calculate_manual_metrics(model_name, test_labels,
predictions)
none_average, binary_average, micro_average, macro_average = manual_metrics
# print(model.test_label()) # path is missing
# list_ = model.words
# new_list = []
# for token in list_:
# if len(token) in [0,1,2]:
# new_list.append(token)
# elif len(token) > 29:
# new_list.append(token)
# # print(new_list)
# print(len(new_list))
print(model.labels)
def first_attempt_based_on_text_classification_paper(data_frame, text_vectorizer, class_vectorizer):
# Parameters - can be placed at the beginning of the script
# ==================================================
# Data loading params
tf.flags.DEFINE_float("dev_sample_percentage", .2, "Percentage of the training data to use for validation")
tf.flags.DEFINE_string("positive_data_file", "./data/rt-polaritydata/rt-polarity.pos", "Data source for the positive data.")
tf.flags.DEFINE_string("negative_data_file", "./data/rt-polaritydata/rt-polarity.neg", "Data source for the negative data.")
# Model Hyperparameters
tf.flags.DEFINE_integer("embedding_dim", 128, "Dimensionality of character embedding (default: 128)")
tf.flags.DEFINE_string("filter_sizes", "3,4,5", "Comma-separated filter sizes (default: '3,4,5')")
tf.flags.DEFINE_integer("num_filters", 128, "Number of filters per filter size (default: 128)")
tf.flags.DEFINE_float("dropout_keep_prob", 0.5, "Dropout keep probability (default: 0.5)")
tf.flags.DEFINE_float("l2_reg_lambda", 0.0, "L2 regularization lambda (default: 0.0)")
# Training parameters
tf.flags.DEFINE_integer("batch_size", 64, "Batch Size (default: 64)")
tf.flags.DEFINE_integer("num_epochs", 200, "Number of training epochs (default: 200)")
tf.flags.DEFINE_integer("evaluate_every", 100, "Evaluate model on dev set after this many steps (default: 100)")
tf.flags.DEFINE_integer("checkpoint_every", 100, "Save model after this many steps (default: 100)")
tf.flags.DEFINE_integer("num_checkpoints", 5, "Number of checkpoints to store (default: 5)")
# Misc Parameters
tf.flags.DEFINE_boolean("allow_soft_placement", True, "Allow device soft device placement")
tf.flags.DEFINE_boolean("log_device_placement", False, "Log placement of ops on devices")
FLAGS = tf.flags.FLAGS
# FLAGS._parse_flags()
# print("\nParameters:")
# for attr, value in sorted(FLAGS.__flags.items()):
# print("{}={}".format(attr.upper(), value))
# print("")
dev_sample_percentage = FLAGS.dev_sample_percentage
# x_train, y_train, vocab_processor, len_vocabulary, x_dev, y_dev, classe, n_classes = preprocess(data_frame, dev_sample_percentage)
model_name = text_vectorizer+'/'+class_vectorizer+'/TextCNN'
standard_results = ch.apply_df_vectorizer(data_frame, 'standard', 'multi_label', model_name)
x_train, x_dev, y_train, y_dev, classes, n_classes, vocab_processor, len_vocabulary = standard_results
with tf.Graph().as_default():
session_conf = tf.ConfigProto(
allow_soft_placement = FLAGS.allow_soft_placement,
log_device_placement = FLAGS.log_device_placement)
sess = tf.Session(config = session_conf)
with sess.as_default():
# Code that operates on the default graph and session comes here…
cnn = pmh.TextCNN(
sequence_length = x_train.shape[1],
num_classes = n_classes,
vocab_size = (len_vocabulary),
embedding_size = FLAGS.embedding_dim,
filter_sizes = list(map(int, FLAGS.filter_sizes.split(","))),
num_filters = FLAGS.num_filters)
# define training procedure
global_step = tf.Variable(0, name="global_step", trainable=False)
optimizer = tf.train.AdamOptimizer(1e-4)
grads_and_vars = optimizer.compute_gradients(cnn.loss)
train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step)
# output directory for models and summaries
root_location = fh.get_root_location('data/convolutional_runs/')
timestamp = datetime.datetime.now().isoformat()
out_dir = fh.link_paths(root_location, timestamp)
print("Writing to {}\n".format(out_dir))
# summaries for loss and accuracy
loss_summary = tf.summary.scalar("loss", cnn.loss)
acc_summary = tf.summary.scalar("accuracy", cnn.accuracy)
# train summaries
train_summary_op = tf.summary.merge([loss_summary, acc_summary])
train_summary_dir = fh.join_paths(fh.link_paths(out_dir, 'summaries'), 'train')
train_summary_writer = tf.summary.FileWriter(train_summary_dir, sess.graph_def)
# dev summaries
dev_summary_op = tf.summary.merge([loss_summary, acc_summary])
dev_summary_dir = fh.join_paths(fh.link_paths(out_dir, 'summaries'), 'dev')
dev_summary_writer = tf.summary.FileWriter(dev_summary_dir, sess.graph_def)
# checkpointing
checkpoint_dir = fh.link_paths(out_dir, 'checkpoints')
checkpoint_prefix = fh.join_paths(checkpoint_dir, 'model')
saver = tf.train.Saver(tf.all_variables())
# write vocabulary
try:
vocab_processor.save(os.path.join(out_dir, "vocab"))
except:
pass
# initialize all variables
sess.run(tf.global_variables_initializer())
def train_step(x_batch, y_batch):
"""
A single training step
"""
feed_dict = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: FLAGS.dropout_keep_prob
}
_, step, summaries, loss, accuracy = sess.run(
[train_op, global_step, train_summary_op, cnn.loss, cnn.accuracy],
feed_dict)
time_str = datetime.datetime.now().isoformat()
print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))
train_summary_writer.add_summary(summaries, step)
def dev_step(x_batch, y_batch, writer=None):
"""
Evaluates model on a dev set
"""
feed_dict = {
cnn.input_x: x_batch,
cnn.input_y: y_batch,
cnn.dropout_keep_prob: 1.0
}
step, summaries, loss, accuracy = sess.run(
[global_step, dev_summary_op, cnn.loss, cnn.accuracy],
feed_dict)
time_str = datetime.datetime.now().isoformat()
print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))
if writer:
writer.add_summary(summaries, step)
# generate batches
batches = batch_iter(
list(zip(x_train, y_train)), FLAGS.batch_size, FLAGS.num_epochs)
# training loop. For each batch…
for batch in batches:
x_batch, y_batch = zip(*batch)
train_step(x_batch, y_batch)
current_step = tf.train.global_step(sess, global_step)
if current_step % FLAGS.evaluate_every == 0:
print("\nEvaluation:")
dev_step(x_dev, y_dev, writer=dev_summary_writer)
if current_step % FLAGS.checkpoint_every == 0:
path = saver.save(sess, checkpoint_prefix, global_step=current_step)
print("Saved model checkpoint to {}\n".format(path))
def get_csv_path(model_key):
index = settings[model_key]["results_file_name"].rfind('/')
path_to_csv = fh.get_root_location(settings[model_key]["results_file_name"][:index])
return fh.link_paths(path_to_csv, settings[model_key]["results_file_name"][index+1:])
def save_training_set(training_set, model_name):
print('### saving_training_set ###')
root_location = fh.get_root_location('data/saved_training_set/')
model_name = model_name.replace("/", "-")
path = fh.link_paths(root_location, 'training '+model_name+' '+str(datetime.now())[:-10]+'.npy')
np.save(path, training_set)
def train_doc2vec(data_frame, patent_ids, classif_level, classif_type):
root_location = fh.get_root_location("data/lstm_outcome/")
doc2vec_model_save_location = fh.join_paths(root_location, "doc2vec_model/")
preprocessed_location = fh.join_paths(root_location, "preprocessed_data/separated_datasets/")
training_preprocessed_files_prefix = fh.join_paths(preprocessed_location, "training_docs_data_preprocessed/")
validation_preprocessed_files_prefix = fh.join_paths(preprocessed_location, "validation_docs_data_preprocessed/")
test_preprocessed_files_prefix = fh.join_paths(preprocessed_location, "test_docs_data_preprocessed/")
vocab_path = fh.join_paths(doc2vec_model_save_location, "vocab_model")
training_docs_iterator = create_tuple_array(data_frame, patent_ids, text_batch_size=10000)
#####
tagged_data = training_docs_iterator
cores = multiprocessing.cpu_count()
model_dbow = Doc2Vec(dm=1, vector_size=200, window=2, negative=10, sample=1e-8, hs=0, min_count=50,
alpha=0.25, min_alpha=0.05, dbow_words=0, seed=1234, concat=0, workers=cores)
model_dbow.build_vocab([x for x in tqdm(tagged_data)])
for epoch in range(30):
# model_dbow.train(utils_shuffle_rows([x for x in tqdm(tagged_data)]), total_examples=len(tagged_data), epochs=1)
model_dbow.train(utils_shuffle_rows([x for x in tqdm(tagged_data)]), total_examples=len(tagged_data), epochs=1)
model_dbow.alpha -= 0.002
model_dbow.min_alpha = model_dbow.alpha
date = datetime.datetime.now().isoformat()
model_dbow.save(fh.link_paths(vocab_path, 'doc2vec_vocab_30_epochs'))
#####
params = wmh.get_parameters_lstm_doc2vec()
GLOBAL_VARS.DOC2VEC_MODEL_NAME, placeholder_model_name, doc2vec_model = wmh.get_lstm_doc2vec(params, classif_level, classif_type)
# yields a list of sentences id, text as a tuple or (id, tuple)
# training_docs_iterator = lrh.BatchWrapper(training_preprocessed_files_prefix, text_batch_size=10000, level=classif_level,
# level_type=classif_type)
doc2vec_model.build_vocab(documents=training_docs_iterator, progress_per=params[13])
doc2vec_model.save(fh.link_paths(vocab_path, "doc2vec_vocab"))
DOC2VEC_ALPHA_DECREASE = wmh.set_alpha_parameters_lstm_doc2vec(doc2vec_model)
start_epoch = 1
# for epoch in range(1, params[11] + 1):
# GLOBAL_VARS.MODEL_NAME = placeholder_model_name.format(epoch)
# doc2vec_folder_path = fh.join_paths(doc2vec_model_save_location, GLOBAL_VARS.MODEL_NAME)
# if fh.ensure_exists_path_location(fh.link_paths(doc2vec_folder_path, "doc2vec_model")):
# start_epoch = epoch
# if start_epoch > 1:
# GLOBAL_VARS.MODEL_NAME = placeholder_model_name.format(start_epoch)
# doc2vec_folder_path = fh.join_paths(doc2vec_model_save_location, GLOBAL_VARS.MODEL_NAME)
# # if a model of that epoch already exists, we load it and proceed to the next epoch
# doc2vec_model = Doc2Vec.load(fh.link_paths(doc2vec_folder_path, "doc2vec_model"))
# start_epoch += 1
## The Actual Training
for epoch in range(start_epoch, params[11] + 1):
print("### epoch "+str(epoch)+" ###")
# set new filename/path to include the epoch
GLOBAL_VARS.MODEL_NAME = placeholder_model_name.format(epoch)
doc2vec_folder_path = fh.join_paths(doc2vec_model_save_location, GLOBAL_VARS.MODEL_NAME)
# train the doc2vec model
# training_docs_iterator = lrh.BatchWrapper(training_preprocessed_files_prefix, text_batch_size=10000, level=classif_level,
# level_type=classif_type) # yields a list of sentences id, text as a tuple or (id, tuple)
doc2vec_model.train(documents=training_docs_iterator, total_examples=len(training_docs_iterator),
report_delay=params[12], epochs=params[10])
doc2vec_model.alpha -= DOC2VEC_ALPHA_DECREASE # decrease the learning rate
doc2vec_model.min_alpha = doc2vec_model.alpha # fix the learning rate, no decay
doc2vec_model.save(fh.link_paths(doc2vec_folder_path, "doc2vec_model"))
if epoch != params[11]:
print("still training epochs missing: " + str(epoch))
sys.exit(1)