# Random Forest
rf.fit(train_features_df.values.tolist(), train_labels_df['cat'].tolist())
predicted_labels = []
for index, vector in enumerate(test_features_df.values):
predicted_labels.append(str(rf.predict(vector.reshape(1, -1))[0]))
tree_confusion_matrices["Random Forest"].append(tree.plot_confusion_matrix(test_labels_df['cat'].values.astype(str), predicted_labels)) # Bit hacky to use the tree method
train_features_df = (train_features_df - train_features_df.mean()) / (train_features_df.max() - train_features_df.min())
train_features_df = train_features_df.reset_index(drop=True)
test_features_df = (test_features_df - test_features_df.mean()) / (test_features_df.max() - test_features_df.min())
test_features_df = test_features_df.reset_index(drop=True)
# Neural Network
model = build_nn(nr_features=len(train_features_df.columns))
model.initialize()
layer_info = PrintLayerInfo()
layer_info(model)
y_train = np.reshape(np.asarray(train_labels_df, dtype='int32'), (-1, 1)).ravel()
model.fit(train_features_df.values, np.add(y_train, -1))
predicted_labels = []
for index, vector in enumerate(test_features_df.values):
predicted_labels.append(str(model.predict(vector.reshape(1, -1))[0]+1))
tree_confusion_matrices["Neural Network"].append(tree.plot_confusion_matrix(test_labels_df['cat'].values.astype(str), predicted_labels)) # Bit hacky to use the tree method
#Bayesian Network
train_features_df, test_features_df = features_df.iloc[train_index,:].copy(), features_df.iloc[test_index,:].copy()
train_labels_df, test_labels_df = labels_df.iloc[train_index,:].copy(), labels_df.iloc[test_index,:].copy()
train_features_df = train_features_df.reset_index(drop=True)
test_features_df = test_features_df.reset_index(drop=True)
train_labels_df = train_labels_df.reset_index(drop=True)
def get_layer_info():
"""Created 04/11/2016"""
from nolearn.lasagne import PrintLayerInfo
layer_info = PrintLayerInfo()
return layer_info
def local_test(self, images, results, feature_extractors, model, k=2, size=64):
kf = KFold(len(images), n_folds=k, shuffle=True, random_state=1337)
# kf = KFold(500, n_folds=k, shuffle=True, random_state=1337)
train_errors = []
test_errors = []
for train, validation in kf:
# Divide the train_images in a training and validation set (using KFold)
train_set = [images[i % len(images)] for i in train]
validation_set = [images[i % len(images)] for i in validation]
train_set_results = [results[i % len(images)] for i in train]
validation_set_results = [results[i % len(images)] for i in validation]
# Create an empty feature_vectors array and set the codebook of the sift extractor if there is any
feature_vectors = []
sift_extractor = temp_extractor = next(
(extractor for extractor in feature_extractors if type(extractor) == SiftFeatureExtractor), None)
if (sift_extractor != None):
sift_extractor.set_codebook(train_set)
feature_extractors[feature_extractors.index(temp_extractor)] = sift_extractor
# Iterate over the train_set, extract the features from each image and append them to feature_vectors
for image in train_set:
print("Training ", image, "...")
preprocessed_color_image = self.preprocess_image(image, size)
feature_vector = []
if feature_extractors != []:
for feature_extractor in feature_extractors:
if type(feature_extractor) != SiftFeatureExtractor:
feature_vector = append(feature_vector,
feature_extractor.extract_feature_vector(preprocessed_color_image))
else:
feature_vector = append(feature_vector, feature_extractor.extract_feature_vector(image))
else:
feature_vector = np.asarray(resize(cv2.imread(image), (48, 48, 3)).transpose(2,0,1).reshape(3, 48, 48))
feature_vectors.append(feature_vector)
# Logistic Regression for feature selection, higher C = more features will be deleted
clf2 = LogisticRegression(penalty='l1', dual=False, tol=0.0001, C=4)
# Feature selection/reduction
if(model != "conv"):
new_feature_vectors = clf2.fit_transform(feature_vectors, train_set_results)
if(model == "neural"):
model = self.build_nn(nr_features=len(new_feature_vectors[0]))
new_feature_vectors = np.asarray(new_feature_vectors)
train_set_results = np.asarray(train_set_results)
model.initialize()
layer_info = PrintLayerInfo()
layer_info(model)
# Fit our model
model.fit(new_feature_vectors, train_set_results)
else:
model = self.build_conv()
# Fit our model
model.fit(np.asarray(feature_vectors), np.asarray(train_set_results))
train_prediction_object = Prediction()
counter=0
for im in train_set:
print("predicting train image ", counter)
counter+=1
preprocessed_color_image = self.preprocess_image(im, size)
validation_feature_vector = []
if feature_extractors != []:
for feature_extractor in feature_extractors:
if type(feature_extractor) != SiftFeatureExtractor:
validation_feature_vector = append(validation_feature_vector, feature_extractor.extract_feature_vector(preprocessed_color_image))
else:
validation_feature_vector = append(validation_feature_vector, feature_extractor.extract_feature_vector(im))
validation_feature_vector = clf2.transform(validation_feature_vector)
else:
validation_feature_vector = np.asarray(resize(cv2.imread(image), (48, 48, 3)).transpose(2,0,1).reshape(3, 48, 48))
train_prediction_object.addPrediction(model.predict_proba(validation_feature_vector)[0])
print("predicting test images")
test_prediction_object = Prediction()
counter=0
for im in validation_set:
print("predicting test image ", counter)
counter+=1
preprocessed_color_image = self.preprocess_image(im, size)
validation_feature_vector = []
if feature_extractors != []:
for feature_extractor in feature_extractors:
if type(feature_extractor) != SiftFeatureExtractor:
validation_feature_vector = append(validation_feature_vector, feature_extractor.extract_feature_vector(preprocessed_color_image))
else:
validation_feature_vector = append(validation_feature_vector, feature_extractor.extract_feature_vector(im))
validation_feature_vector = clf2.transform(validation_feature_vector)
else:
validation_feature_vector = np.asarray(resize(cv2.imread(image), (48, 48, 3)).transpose(2,0,1).reshape(3, 48, 48))
test_prediction_object.addPrediction(model.predict_proba(validation_feature_vector)[0])
train_errors.append(train_prediction_object.evaluate(train_set_results))
test_errors.append(test_prediction_object.evaluate(validation_set_results))
return [train_errors, test_errors]
def make_submission(self, train_images, train_results, test_images, output_file_path, feature_extractors, model, size=64):
# Create a vector of feature vectors and initialize the codebook of sift extractor
feature_vectors = []
sift_extractor = temp_extractor = next((extractor for extractor in feature_extractors if type(extractor) == SiftFeatureExtractor), None)
if(sift_extractor != None):
sift_extractor.set_codebook(train_images)
feature_extractors[feature_extractors.index(temp_extractor)] = sift_extractor
# Extract features from every image
for image in train_images:
print("Training ", image, "...")
preprocessed_color_image = self.preprocess_image(image, size)
feature_vector = []
if feature_extractors != []:
for feature_extractor in feature_extractors:
if type(feature_extractor) != SiftFeatureExtractor:
feature_vector = append(feature_vector,
feature_extractor.extract_feature_vector(preprocessed_color_image))
else:
feature_vector = append(feature_vector, feature_extractor.extract_feature_vector(image))
else:
feature_vector = np.asarray(resize(cv2.imread(image), (48, 48, 3)).transpose(2,0,1).reshape(3, 48, 48))
feature_vectors.append(feature_vector)
# Logistic Regression for feature selection, higher C = more features will be deleted
clf2 = LogisticRegression(penalty='l1', dual=False, tol=0.0001, C=4)
# Feature selection/reduction
if(model != "conv"):
print("Old feature vector shape = ", len(feature_vectors), len(feature_vectors[0]))
new_feature_vectors = clf2.fit_transform(feature_vectors, train_results)
print("New feature vector shape = ", len(new_feature_vectors), len(new_feature_vectors[0]))
if(model == "neural"):
model = self.build_nn(nr_features=len(new_feature_vectors[0]))
new_feature_vectors = np.asarray(new_feature_vectors)
train_results = np.asarray(train_results)
model.initialize()
layer_info = PrintLayerInfo()
layer_info(model)
# Fit our model
model.fit(new_feature_vectors, train_results)
else:
model = self.build_conv()
# Fit our model
model.fit(np.asarray(feature_vectors), np.asarray(train_results))
# Iterate over the test images and add their prediction to a prediction object
prediction_object = Prediction()
for im in test_images:
print("Predicting ", im)
preprocessed_color_image = self.preprocess_image(im, size)
validation_feature_vector = []
if feature_extractors != []:
for feature_extractor in feature_extractors:
if type(feature_extractor) != SiftFeatureExtractor:
validation_feature_vector = append(validation_feature_vector, feature_extractor.extract_feature_vector(preprocessed_color_image))
else:
validation_feature_vector = append(validation_feature_vector, feature_extractor.extract_feature_vector(im))
validation_feature_vector = clf2.transform(validation_feature_vector)
else:
validation_feature_vector = np.asarray(resize(cv2.imread(image), (48, 48, 3)).transpose(2,0,1).reshape(3, 48, 48))
prediction_object.addPrediction(model.predict_proba(validation_feature_vector)[0])
# Write out the prediction object
FileParser.write_CSV(output_file_path, prediction_object)
def make_memnn(vocab_size,
cont_sl,
cont_wl,
quest_wl,
answ_wl,
rnn_size,
rnn_type='LSTM',
pool_size=4,
answ_n=4,
dence_l=[100],
dropout=0.5,
batch_size=16,
emb_size=50,
grad_clip=40,
init_std=0.1,
num_hops=3,
rnn_style=False,
nonlin=LN.softmax,
init_W=None,
rng=None,
art_pool=4,
lr=0.01,
mom=0,
updates=LU.adagrad,
valid_indices=0.2,
permute_answ=False,
permute_cont=False):
def select_rnn(x):
return {
'RNN': LL.RecurrentLayer,
'LSTM': LL.LSTMLayer,
'GRU': LL.GRULayer,
}.get(x, LL.LSTMLayer)
# dence = dence + [1]
RNN = select_rnn(rnn_type)
#-----------------------------------------------------------------------weights
tr_variables = {}
tr_variables['WQ'] = theano.shared(
init_std * np.random.randn(vocab_size, emb_size).astype('float32'))
tr_variables['WA'] = theano.shared(
init_std * np.random.randn(vocab_size, emb_size).astype('float32'))
tr_variables['WC'] = theano.shared(
init_std * np.random.randn(vocab_size, emb_size).astype('float32'))
tr_variables['WTA'] = theano.shared(
init_std * np.random.randn(cont_sl, emb_size).astype('float32'))
tr_variables['WTC'] = theano.shared(
init_std * np.random.randn(cont_sl, emb_size).astype('float32'))
tr_variables['WAnsw'] = theano.shared(
init_std * np.random.randn(vocab_size, emb_size).astype('float32'))
#------------------------------------------------------------------input layers
layers = [(LL.InputLayer, {
'name': 'l_in_q',
'shape': (batch_size, 1, quest_wl),
'input_var': T.itensor3('l_in_q_')
}),
(LL.InputLayer, {
'name': 'l_in_a',
'shape': (batch_size, answ_n, answ_wl),
'input_var': T.itensor3('l_in_a_')
}),
(LL.InputLayer, {
'name': 'l_in_q_pe',
'shape': (batch_size, 1, quest_wl, emb_size)
}),
(LL.InputLayer, {
'name': 'l_in_a_pe',
'shape': (batch_size, answ_n, answ_wl, emb_size)
}),
(LL.InputLayer, {
'name': 'l_in_cont',
'shape': (batch_size, cont_sl, cont_wl),
'input_var': T.itensor3('l_in_cont_')
}),
(LL.InputLayer, {
'name': 'l_in_cont_pe',
'shape': (batch_size, cont_sl, cont_wl, emb_size)
})]
#------------------------------------------------------------------slice layers
# l_qs = []
# l_cas = []
l_a_names = ['l_a_{}'.format(i) for i in range(answ_n)]
l_a_pe_names = ['l_a_pe{}'.format(i) for i in range(answ_n)]
for i in range(answ_n):
layers.extend([(LL.SliceLayer, {
'name': l_a_names[i],
'incoming': 'l_in_a',
'indices': slice(i, i + 1),
'axis': 1
})])
for i in range(answ_n):
layers.extend([(LL.SliceLayer, {
'name': l_a_pe_names[i],
'incoming': 'l_in_a_pe',
'indices': slice(i, i + 1),
'axis': 1
})])
#------------------------------------------------------------------MEMNN layers
#question----------------------------------------------------------------------
layers.extend([(EncodingFullLayer, {
'name': 'l_emb_f_q',
'incomings': ('l_in_q', 'l_in_q_pe'),
'vocab_size': vocab_size,
'emb_size': emb_size,
'W': tr_variables['WQ'],
'WT': None
})])
l_mem_names = ['ls_mem_n2n_{}'.format(i) for i in range(num_hops)]
layers.extend([(MemoryLayer, {
'name': l_mem_names[0],
'incomings': ('l_in_cont', 'l_in_cont_pe', 'l_emb_f_q'),
'vocab_size': vocab_size,
'emb_size': emb_size,
'A': tr_variables['WA'],
'C': tr_variables['WC'],
'AT': tr_variables['WTA'],
'CT': tr_variables['WTC'],
'nonlin': nonlin
})])
for i in range(1, num_hops):
if i % 2:
WC, WA = tr_variables['WA'], tr_variables['WC']
WTC, WTA = tr_variables['WTA'], tr_variables['WTC']
else:
WA, WC = tr_variables['WA'], tr_variables['WC']
WTA, WTC = tr_variables['WTA'], tr_variables['WTC']
layers.extend([(MemoryLayer, {
'name':
l_mem_names[i],
'incomings': ('l_in_cont', 'l_in_cont_pe', l_mem_names[i - 1]),
'vocab_size':
vocab_size,
'emb_size':
emb_size,
'A':
WA,
'C':
WC,
'AT':
WTA,
'CT':
WTC,
'nonlin':
nonlin
})])
#answers-----------------------------------------------------------------------
l_emb_f_a_names = ['l_emb_f_a{}'.format(i) for i in range(answ_n)]
for i in range(answ_n):
layers.extend([(EncodingFullLayer, {
'name': l_emb_f_a_names[i],
'incomings': (l_a_names[i], l_a_pe_names[i]),
'vocab_size': vocab_size,
'emb_size': emb_size,
'W': tr_variables['WAnsw'],
'WT': None
})])
#------------------------------------------------------------concatenate layers
layers.extend([(LL.ConcatLayer, {
'name': 'l_qma_concat',
'incomings': l_mem_names + l_emb_f_a_names
})])
#--------------------------------------------------------------------RNN layers
layers.extend([(
RNN,
{
'name': 'l_qa_rnn_f',
'incoming': 'l_qma_concat',
# 'mask_input': 'l_qamask_concat',
'num_units': rnn_size,
'backwards': False,
'only_return_final': False,
'grad_clipping': grad_clip
})])
layers.extend([(
RNN,
{
'name': 'l_qa_rnn_b',
'incoming': 'l_qma_concat',
# 'mask_input': 'l_qamask_concat',
'num_units': rnn_size,
'backwards': True,
'only_return_final': False,
'grad_clipping': grad_clip
})])
layers.extend([(LL.SliceLayer, {
'name': 'l_qa_rnn_f_sl',
'incoming': 'l_qa_rnn_f',
'indices': slice(-answ_n, None),
'axis': 1
})])
layers.extend([(LL.SliceLayer, {
'name': 'l_qa_rnn_b_sl',
'incoming': 'l_qa_rnn_b',
'indices': slice(-answ_n, None),
'axis': 1
})])
layers.extend([(LL.ElemwiseMergeLayer, {
'name': 'l_qa_rnn_conc',
'incomings': ('l_qa_rnn_f_sl', 'l_qa_rnn_b_sl'),
'merge_function': T.add
})])
#-----------------------------------------------------------------pooling layer
# layers.extend([(LL.DimshuffleLayer, {'name': 'l_qa_rnn_conc_',
# 'incoming': 'l_qa_rnn_conc', 'pattern': (0, 'x', 1)})])
layers.extend([(LL.Pool1DLayer, {
'name': 'l_qa_pool',
'incoming': 'l_qa_rnn_conc',
'pool_size': pool_size,
'mode': 'max'
})])
#------------------------------------------------------------------dence layers
l_dence_names = ['l_dence_{}'.format(i) for i, _ in enumerate(dence_l)]
if dropout:
layers.extend([(LL.DropoutLayer, {
'name': 'l_dence_do',
'p': dropout
})])
for i, d in enumerate(dence_l):
if i < len(dence_l) - 1:
nonlin = LN.tanh
else:
nonlin = LN.softmax
layers.extend([(LL.DenseLayer, {
'name': l_dence_names[i],
'num_units': d,
'nonlinearity': nonlin
})])
if i < len(dence_l) - 1 and dropout:
layers.extend([(LL.DropoutLayer, {
'name': l_dence_names[i] + 'do',
'p': dropout
})])
if isinstance(valid_indices, np.ndarray) or isinstance(
valid_indices, list):
train_split = TrainSplit_indices(valid_indices=valid_indices)
else:
train_split = TrainSplit(eval_size=valid_indices, stratify=False)
if permute_answ or permute_cont:
batch_iterator_train = PermIterator(batch_size, permute_answ,
permute_cont)
else:
batch_iterator_train = BatchIterator(batch_size=batch_size)
def loss(x, t):
return LO.aggregate(
LO.categorical_crossentropy(T.clip(x, 1e-6, 1. - 1e-6), t))
# return LO.aggregate(LO.squared_error(T.clip(x, 1e-6, 1. - 1e-6), t))
nnet = NeuralNet(
y_tensor_type=T.ivector,
layers=layers,
update=updates,
update_learning_rate=lr,
# update_epsilon=1e-7,
objective_loss_function=loss,
regression=False,
verbose=2,
batch_iterator_train=batch_iterator_train,
batch_iterator_test=BatchIterator(batch_size=batch_size / 2),
# batch_iterator_train=BatchIterator(batch_size=batch_size),
# batch_iterator_test=BatchIterator(batch_size=batch_size),
#train_split=TrainSplit(eval_size=eval_size)
train_split=train_split,
on_batch_finished=[zero_memnn])
nnet.initialize()
PrintLayerInfo()(nnet)
return nnet
input_shape=(None, num_features),
dense_num_units=64,
narrow_num_units=48,
denseReverse1_num_units=64,
denseReverse2_num_units=128,
output_num_units=128,
#input_nonlinearity = None, #nonlinearities.sigmoid,
#dense_nonlinearity = nonlinearities.tanh,
narrow_nonlinearity=nonlinearities.softplus,
#denseReverse1_nonlinearity = nonlinearities.tanh,
denseReverse2_nonlinearity=nonlinearities.softplus,
output_nonlinearity=nonlinearities.linear, #nonlinearities.softmax,
#dropout0_p=0.1,
dropout1_p=0.01,
dropout2_p=0.001,
regression=True,
verbose=1)
ae.initialize()
PrintLayerInfo()(ae)
maybe_this_is_a_history = ae.fit(Z, Z)
#learned_parameters = ae.get_all_params_values()
#np.save("task4/learned_parameter.npy", learned_parameters)
#SaveWeights(path='task4/koebi_train_history_AE')(ae, maybe_this_is_a_history)
ae.save_params_to('task4/koebi_train_history_AE2')
def local_test(feature_vectors_df, labels_df, k=2):
# labeltjes = [None] * labels_df.values.shape[0]
labeltjes = labels_df.values
print labeltjes.shape
labeltjes = labeltjes
labeltjes -= 1
labeltjes = labeltjes.ravel().tolist()
kf = StratifiedKFold(labeltjes, n_folds=k, shuffle=True)
# kf = StratifiedKFold(len(feature_vectors_df.index), n_folds=k, shuffle=True)
# kf = KFold(500, n_folds=k, shuffle=True, random_state=1337)
confusion_matrices_folds = []
for train, test in kf:
# Divide the train_images in a training and validation set (using KFold)
X_train = feature_vectors_df.values[train, :]
X_test = feature_vectors_df.values[test, :]
y_train = [labeltjes[i] for i in train]
y_test = [labeltjes[i] for i in test]
# Logistic Regression for feature selection, higher C = more features will be deleted
# Feature selection/reduction
model = build_nn(nr_features=X_train.shape[1])
model.initialize()
layer_info = PrintLayerInfo()
layer_info(model)
# Fit our model
y_train = np.reshape(np.asarray(y_train, dtype='int32'),
(-1, 1)).ravel()
# print y_train
# print "Train feature vectors shape: " + X_train.shape.__str__()
# print "Train labels shape:" + len(y_train).__str__()
#
# print "X_train as array shape: " + str(X_train.shape)
# print "y_train as array shape: " + str(np.reshape(np.asarray(y_train), (-1, 1)).shape)
model.fit(X_train, np.reshape(np.asarray(y_train), (-1, 1)).ravel())
preds = model.predict(X_test)
c = []
[
c.append(preds[i]) if preds[i] == y_test[i] else None
for i in range(min(len(y_test), len(preds)))
]
# checks = len([i for i, j in zip(preds, np.reshape(np.asarray(y_train), (-1, 1))) if i == j])
model = None
del model
# Save the confusion matrix for this fold and plot it
confusion_matrix = sklearn.metrics.confusion_matrix(y_test, preds)
confusion_matrices_folds.append(confusion_matrix)
# print preds.tolist()
# print "number of ones: " + str(sum(preds))
# print y_test
# print c
print "Accuracy for fold: " + str(
((len(c) * 1.0) / (len(y_test) * 1.0)
)) + "\n\n\n\n\n-----------------------------\n\n\n"
# Let's plot the confusion matrix of the avarage confusion matrix
sum = confusion_matrices_folds[0] * 1.0
for i in range(1, len(confusion_matrices_folds)):
sum += confusion_matrices_folds[i]
sum /= len(confusion_matrices_folds)
metrics.plot_confusion_matrix(sum)
def make_grnn(
batch_size,
emb_size,
g_hidden_size,
word_n,
wc_num,
dence,
wsm_num=1,
rnn_type='LSTM',
rnn_size=12,
dropout_d=0.5, # pooling='mean',
quest_na=4,
gradient_steps=-1,
valid_indices=None,
lr=0.05,
grad_clip=10):
def select_rnn(x):
return {
'RNN': LL.RecurrentLayer,
'LSTM': LL.LSTMLayer,
'GRU': LL.GRULayer,
}.get(x, LL.LSTMLayer)
# dence = dence + [1]
RNN = select_rnn(rnn_type)
#------------------------------------------------------------------input layers
layers = [
(LL.InputLayer, {
'name': 'l_in_se_q',
'shape': (None, word_n, emb_size)
}),
(LL.InputLayer, {
'name': 'l_in_se_a',
'shape': (None, quest_na, word_n, emb_size)
}),
(LL.InputLayer, {
'name': 'l_in_mask_q',
'shape': (None, word_n)
}),
(LL.InputLayer, {
'name': 'l_in_mask_a',
'shape': (None, quest_na, word_n)
}),
(LL.InputLayer, {
'name': 'l_in_mask_ri_q',
'shape': (None, word_n)
}),
(LL.InputLayer, {
'name': 'l_in_mask_ri_a',
'shape': (None, quest_na, word_n)
}),
(LL.InputLayer, {
'name': 'l_in_wt_q',
'shape': (None, word_n, word_n)
}),
(LL.InputLayer, {
'name': 'l_in_wt_a',
'shape': (None, word_n, quest_na, word_n)
}),
(LL.InputLayer, {
'name': 'l_in_act_',
'shape': (None, word_n, g_hidden_size)
}),
(LL.InputLayer, {
'name': 'l_in_act__',
'shape': (None, word_n, word_n, g_hidden_size)
}),
]
#------------------------------------------------------------------slice layers
# l_qs = []
# l_cas = []
l_ase_names = ['l_ase_{}'.format(i) for i in range(quest_na)]
l_amask_names = ['l_amask_{}'.format(i) for i in range(quest_na)]
l_amask_ri_names = ['l_amask_ri_{}'.format(i) for i in range(quest_na)]
l_awt_names = ['l_awt_{}'.format(i) for i in range(quest_na)]
for i in range(quest_na):
layers.extend([(LL.SliceLayer, {
'name': l_ase_names[i],
'incoming': 'l_in_se_a',
'indices': i,
'axis': 1
})])
for i in range(quest_na):
layers.extend([(LL.SliceLayer, {
'name': l_amask_names[i],
'incoming': 'l_in_mask_a',
'indices': i,
'axis': 1
})])
for i in range(quest_na):
layers.extend([(LL.SliceLayer, {
'name': l_amask_ri_names[i],
'incoming': 'l_in_mask_ri_a',
'indices': i,
'axis': 1
})])
for i in range(quest_na):
layers.extend([(LL.SliceLayer, {
'name': l_awt_names[i],
'incoming': 'l_in_wt_a',
'indices': i,
'axis': 1
})])
#-------------------------------------------------------------------GRNN layers
WC = theano.shared(
np.random.randn(wc_num, g_hidden_size,
g_hidden_size).astype('float32'))
# WC = LI.Normal(0.1)
WSM = theano.shared(
np.random.randn(emb_size, g_hidden_size).astype('float32'))
b = theano.shared(np.ones(g_hidden_size).astype('float32'))
# b = lasagne.init.Constant(1.0)
layers.extend([(GRNNLayer, {
'name':
'l_q_grnn',
'incomings':
['l_in_se_q', 'l_in_mask_q', 'l_in_wt_q', 'l_in_act_', 'l_in_act__'],
'emb_size':
emb_size,
'hidden_size':
g_hidden_size,
'word_n':
word_n,
'wc_num':
wc_num,
'wsm_num':
wsm_num,
'only_return_final':
False,
'WC':
WC,
'WSM':
WSM,
'b':
b
})])
l_a_grnns_names = ['l_a_grnn_{}'.format(i) for i in range(quest_na)]
for i, l_a_grnns_name in enumerate(l_a_grnns_names):
layers.extend([(GRNNLayer, {
'name':
l_a_grnns_name,
'incomings': [
l_ase_names[i], l_amask_names[i], l_awt_names[i], 'l_in_act_',
'l_in_act__'
],
'emb_size':
emb_size,
'hidden_size':
g_hidden_size,
'word_n':
word_n,
'wc_num':
wc_num,
'wsm_num':
wsm_num,
'only_return_final':
False,
'WC':
WC,
'WSM':
WSM,
'b':
b
})])
#------------------------------------------------------------concatenate layers
layers.extend([(LL.ConcatLayer, {
'name': 'l_qa_concat',
'incomings': ['l_q_grnn'] + l_a_grnns_names
})])
layers.extend([(LL.ConcatLayer, {
'name': 'l_qamask_concat',
'incomings': ['l_in_mask_ri_q'] + l_amask_ri_names
})])
#--------------------------------------------------------------------RNN layers
layers.extend([(RNN, {
'name': 'l_qa_rnn_f',
'incoming': 'l_qa_concat',
'mask_input': 'l_qamask_concat',
'num_units': rnn_size,
'backwards': False,
'only_return_final': True,
'grad_clipping': grad_clip
})])
layers.extend([(RNN, {
'name': 'l_qa_rnn_b',
'incoming': 'l_qa_concat',
'mask_input': 'l_qamask_concat',
'num_units': rnn_size,
'backwards': True,
'only_return_final': True,
'grad_clipping': grad_clip
})])
layers.extend([(LL.ElemwiseSumLayer, {
'name': 'l_qa_rnn_conc',
'incomings': ['l_qa_rnn_f', 'l_qa_rnn_b']
})])
##-----------------------------------------------------------------pooling layer
## l_qa_pool = layers.extend([(LL.ExpressionLayer, {'name': 'l_qa_pool',
## 'incoming': l_qa_rnn_conc,
## 'function': lambda X: X.mean(-1),
## 'output_shape'='auto'})])
#------------------------------------------------------------------dence layers
l_dence_names = ['l_dence_{}'.format(i) for i, _ in enumerate(dence)]
if dropout_d:
layers.extend([(LL.DropoutLayer, {
'name': 'l_dence_do' + 'do',
'p': dropout_d
})])
for i, d in enumerate(dence):
if i < len(dence) - 1:
nonlin = LN.tanh
else:
nonlin = LN.softmax
layers.extend([(LL.DenseLayer, {
'name': l_dence_names[i],
'num_units': d,
'nonlinearity': nonlin
})])
if i < len(dence) - 1 and dropout_d:
layers.extend([(LL.DropoutLayer, {
'name': l_dence_names[i] + 'do',
'p': dropout_d
})])
def loss(x, t):
return LO.aggregate(
LO.categorical_crossentropy(T.clip(x, 1e-6, 1. - 1e-6), t))
# return LO.aggregate(LO.squared_error(T.clip(x, 1e-6, 1. - 1e-6), t))
if isinstance(valid_indices, np.ndarray) or isinstance(
valid_indices, list):
train_split = TrainSplit_indices(valid_indices=valid_indices)
else:
train_split = TrainSplit(eval_size=valid_indices, stratify=False)
nnet = NeuralNet(
y_tensor_type=T.ivector,
layers=layers,
update=LU.adagrad,
update_learning_rate=lr,
# update_epsilon=1e-7,
objective_loss_function=loss,
regression=False,
verbose=2,
batch_iterator_train=PermIterator(batch_size=batch_size),
batch_iterator_test=BatchIterator(batch_size=batch_size / 2),
# batch_iterator_train=BatchIterator(batch_size=batch_size),
# batch_iterator_test=BatchIterator(batch_size=batch_size),
#train_split=TrainSplit(eval_size=eval_size)
train_split=train_split)
nnet.initialize()
PrintLayerInfo()(nnet)
return nnet
narrow_nonlinearity=nonlinearities.softplus,
reverse_nonlinearity=nonlinearities.sigmoid,
coutput_nonlinearity=nonlinearities.softmax,
#dropout0_p=0.1,
dropout1_p=0.01,
#regression=True,
regression=False,
verbose=1)
nn.initialize()
nn.load_params_from('task4/koebi_train_history_AE')
PrintLayerInfo()(nn)
nn.fit(X, Y)
test = pd.read_hdf("task4/test.h5", "test")
id_col = test.index
test_data = np.array(test)
test_data = skpre.StandardScaler().fit_transform(test_data)
test_prediction = nn.predict(test_data)
# Write prediction and it's linenumber into a csv file
with open('task4/' + result_file_name + '.csv', 'wb') as csvfile:
fieldnames = ['Id', 'y']
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
# print test_prediction
writer.writeheader()
