def fit(self, X_train, y_train, X_test, y_test):
self.input_dim = X_train.shape[1]
X_train = self.convertToArray(X_train)
X_test = self.convertToArray(X_test)
y_train = self.convertToArray(y_train)
y_test = self.convertToArray(y_test)
print("Start training. Configurations count: %i" % (len(self.param_grid)))
i = 1
best_score = 0
self.check_keys(['hLayers','activation','optimizer'])
for curParams in self.param_grid:
if ('hLayers' not in curParams.keys()):
raise Exception("'hLayers' must be one of the grid parameters")
else:
hLayers = curParams.get('hLayers')
if ('activation' not in curParams.keys()):
activation = self.default_activation
else:
activation = curParams.get('activation')
if ('optimizer' not in curParams.keys()):
optimizer = self.default_optimizer
else:
optimizer = curParams.get('optimizer')
filePath = self.log_files_path + self.getFileName(hLayers, activation, optimizer)
dnn = self.createDNN(hLayers, activation, optimizer)
lhdnn, chkp, es = self.createCallBacks(X_train, y_train, X_test, y_test, filePath)
dnn.fit(X_train, y_train, nb_epoch=self.epoch_count, batch_size=self.batch_size,
validation_data = [X_test, y_test], verbose=2, callbacks=[lhdnn, chkp, es])
if self.score_func == 'accuracy':
true_test = np_utils.probas_to_classes(y_test)
pred_test = np_utils.probas_to_classes(dnn.predict(X_test))
cur_score = accuracy_score(true_test, pred_test)
if (cur_score > best_score):
best_score = cur_score
best_dnn = dnn
self.writeInfoToFile(lhdnn, filePath)
elif self.score_func == 'r2_score':
cur_score = r2_score(y_test, dnn.predict(X_test))
if (cur_score > best_score):
best_score = cur_score
best_dnn = dnn
self.writeInfoToFile(lhdnn, filePath)
print("Configuration #%i. Completed." % (i))
i += 1
return dnn
def test(self, data):
y_predicted = np_utils.probas_to_classes(self.model.predict(data.x))
y_actual = np_utils.probas_to_classes(data.y)
error = (np.ravel(y_predicted) != np.ravel(y_actual)).sum().astype(float)/y_actual.shape[0]
print("PREDICTED: class 0: {0}, class 1: {1}, class 2: {2}".format(
np.sum(np.ravel(y_predicted) == 0),
np.sum(np.ravel(y_predicted) == 1),
np.sum(np.ravel(y_predicted) == 2)))
print("ACTUAL: class 0: {0}, class 1: {1}, class 2: {2}".format(
np.sum(np.ravel(y_actual) == 0),
np.sum(np.ravel(y_actual) == 1),
np.sum(np.ravel(y_actual) == 2)))
print("ERROR RATE: ", error)
return error
def train(label_set='full', drop_unk=False,
word_vecs=None, setup_only=False, layer_sizes=[512,256],
pool_mode='sum'):
print "Loading data..."
df = sentences_df(SENTENCES_CSV, labels=label_set, drop_unk=drop_unk)
X, y, word2idx, l_enc = load_dataset(df, pad=True)
print "X shape:", X.shape
y_orig = y
y_binary = to_categorical(y)
labels = np.unique(y_orig)
nb_labels = labels.shape[0]
if drop_unk:
label_set_str = label_set + ' (-unk)'
else:
label_set_str = label_set
print "Number of labels: %i [%s]" % (nb_labels, label_set_str)
if nb_labels > 2:
y = y_binary
maxlen = X.shape[1]
vocab_size = len(word2idx) + 1 # 0 masking
if pretrained_embeddings is True:
word_vectors = load_bin_vec(word_vecs, word2idx)
add_unknown_words(word_vectors, word2idx)
embedding_weights = np.zeros((vocab_size+1, emb_dim))
for word, index in word2idx.items():
embedding_weights[index,:] = word_vectors[word]
else:
embedding_weights = None
print "Data loaded."
skf = StratifiedKFold(y_orig, n_folds=10, shuffle=True, random_state=0)
cv_scores = []
for i, (train, test) in enumerate(skf):
start_time = time.time()
nn = None
nn = EnsembleNN(vocab_size, nb_labels, emb_dim, maxlen,
embedding_weights, filter_hs, nb_filters,
dropout_p, trainable_embeddings, pretrained_embeddings,
layer_sizes, pool_mode)
if i == 0:
print_summary(nn.model.layers)
acc = train_and_test_model(nn, X[train], y[train], X[test], y[test],
batch_size, nb_epoch,
lr, beta_1, beta_2, epsilon)
cv_scores.append(acc)
train_time = time.time() - start_time
print('\nLabel frequencies in y[test]')
print_label_frequencies((y_orig[test], l_enc))
y_pred = nn.model.predict(X[test])
y_pred = probas_to_classes(y_pred)
c = Counter(y_pred)
total = float(len(y_pred))
print('\nLabel frequencies in predict(y[test])')
for label, count in c.most_common():
print l_enc.inverse_transform(label), count, count / total
print "fold %i/10 - time: %.2f s - acc: %.4f on %i samples" % \
(i+1, train_time, acc, len(test))
print "Avg cv accuracy: %.4f" % np.mean(cv_scores)
def predict(self, X):
if not isinstance(X,np.ndarray):
X = X.values
ypred = self.model.predict(X, batch_size=self.batch_size, verbose=self.verbose)
if self.nb_classes>1:
ypred = np_utils.probas_to_classes(ypred)
else:
ypred = ypred.flatten()
return ypred
def on_epoch_end(self, epoch, logs={}):
self.train_losses.append(mse(self.y_train, self.model.predict(self.X_train)))
self.val_losses.append(logs.get('val_loss'))
if self.score_func == 'accuracy':
true_train = np_utils.probas_to_classes(self.y_train)
pred_train = np_utils.probas_to_classes(self.model.predict(self.X_train))
self.add_train_scores.append(accuracy_score(true_train, pred_train))
true_test = np_utils.probas_to_classes(self.y_test)
pred_test = np_utils.probas_to_classes(self.model.predict(self.X_test))
val_score = accuracy_score(true_test, pred_test)
self.add_val_scores.append(val_score)
elif self.score_func == 'r2_score':
val_score = r2_score(self.y_test, self.model.predict(self.X_test))
self.add_val_scores.append(val_score)
self.add_train_scores.append(r2_score(self.y_train, self.model.predict(self.X_train)))
self.best_score = max(self.best_score, val_score)
self.printCurrentStage(epoch)
def recognize(chars):
N = len(chars)
width, height = chars[0].shape
chars = numpy.asarray(chars, dtype=numpy.float32).reshape((N, 1, width, height)) / 255.0
# chinese_classes = np_utils.probas_to_classes(chinese_model.predict(chars[:1], batch_size=1))
alnum_classes = np_utils.probas_to_classes(alnum_model.predict(chars[1:], batch_size=6))
return ['浙'] + \
[alnum_labels[cls] for cls in alnum_classes]
def predict(model, x, y, ix, output_dir):
"""
Store predictions in a CSV file and predicted probabilities in an NPZ file.
"""
y_proba_pred = model.predict(x)
np.savez_compressed(output_dir + '/predictions_proba.npz',
y_proba_pred=y_proba_pred)
df = pd.DataFrame({
'y_pred': np_utils.probas_to_classes(y_proba_pred),
'y_true': np_utils.categorical_probas_to_classes(y)})
df['accurate'] = df['y_true'] == df['y_pred']
df['split'] = ''
for key, indexes in ix.items():
df.ix[indexes, 'split'] = key
df = df[['split', 'y_true', 'y_pred', 'accurate']]
df.to_csv(output_dir + '/predictions.csv', index=None)
return y_proba_pred
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--csv_in', required=True, help='CSV input filename')
parser.add_argument('--csv_out', required=True, help='CSV output filename--do NOT confuse with csv_in')
parser.add_argument('--video_in', required=True, help='Video input filename')
parser.add_argument('--num_frames', type=int, default=1000, help='Number of frames to use to form training and test set. \
Default: 1000')
parser.add_argument('--frame_interval', type=int, default=30, help='Number of frames to skip when creating training and \
test sets. Must be greater than 0. Default: 30')
parser.add_argument('--object', required=True, help='Object(s) to classify. Multiple values should be separated by a comma \
(e.g., person,car)')
parser.add_argument('--scale', type=float, default=0.1, help='Scale factor applied to each frame. Default: 0.1')
parser.add_argument('--sample', dest='sample_data', action='store_true')
parser.add_argument('--no_sample', dest='sample_data', action='store_false')
parser.set_defaults(sample_data=True)
parser.add_argument('--test_ratio', type=float, default=0.5, help='Ratio between test and training set. Default: 0.5')
parser.add_argument('--models', default='lsvm',
help='Type of model: Logistic Regression (lr), Linear SVM (lsvm), rbf SVM (rsvm), or polynomial SVM (psvm). \
Multiple values must be separated by comma (e.g., lr,lsvm). Default: lsvm')
parser.add_argument('--class_weight_factor', type=float, default=1.0,
help='How much to increase weight of positive training examples, and decrease weight of negative examples. \
Higher -> fewer false negatives, more false positives. Default: 1.0')
args = parser.parse_args()
csv_out = args.csv_out
if args.frame_interval <= 0:
import sys
print '--frame_interval must be greater than 0'
sys.exit(1)
print args
models_to_try = args.models.strip().split(',')
args_dict = args.__dict__
del(args_dict['models'])
del(args_dict['csv_out'])
init_header, init_row = zip(*sorted(list(args_dict.iteritems())))
init_header, init_row = list(init_header), list(init_row)
print 'Retrieving %d frames from %s' % (args.num_frames, args.video_in)
video_frames = VideoUtils.get_all_frames(args.num_frames, args.video_in, scale=args.scale,
interval=args.frame_interval)
print 'Retrieving %d labels from %s' % (args.num_frames, args.csv_in)
Y = DataUtils.get_binary(args.csv_in, [args.object], limit=args.num_frames, interval=args.frame_interval)
Y = Y.flatten()
if args.sample_data:
print 'Partitioning training and test sets using random sampling'
train_ind, test_ind, Y_train, Y_test = train_test_split(np.arange(len(Y)), Y, test_size=args.test_ratio)
else:
print 'Partitioning training and test sets using simple strategy: first \
%0.2f are training, remaining %0.2f are test' % (1-args.test_ratio, args.test_ratio)
split_index = int(len(Y)*(1 - args.test_ratio))
inds = np.arange(len(Y))
train_ind, test_ind, Y_train, Y_test = inds[:split_index], inds[split_index:], Y[:split_index], Y[split_index:]
print '(train) positive examples: %d, total examples: %d' % \
(np.count_nonzero(np_utils.probas_to_classes(Y_train)),
len(Y_train))
print '(test) positive examples: %d, total examples: %d' % \
(np.count_nonzero(np_utils.probas_to_classes(Y_test)),
len(Y_test))
class_weights = DataUtils.get_class_weights(Y_train, args.class_weight_factor)
print 'Class weights:', class_weights
rows = []
print 'Getting features....'
X = get_features(video_frames)
X_train, X_test = X[train_ind], X[test_ind]
for model_type in models_to_try:
headers = init_header[:]
row = init_row[:]
headers.append('model')
row.append(model_type)
print model_type
model = get_model(model_type, X_train, Y_train, class_weights)
print 'evaluating on training set'
train_metrics = evaluate_model(model, X_train, Y_train)
for key, val in train_metrics:
headers.append('train ' + key)
row.append(val)
print train_metrics
print 'evaluating on test set'
test_metrics = evaluate_model(model, X_test, Y_test)
for key, val in train_metrics:
headers.append('test ' + key)
row.append(val)
print test_metrics
rows.append(row)
output_csv(csv_out, np.array(rows), np.array(headers))
def create_model(word_to_id,
train_data,
dev_data,
test_data,
embedding_layer=None):
id_to_word = {jj: ii for ii, jj in word_to_id.items()}
word_num = len(id_to_word)
X = []
Y = []
for sentence in train_data:
# app = False
# for wordtag in sentence['tags']:
# if wordtag!=1:
# app = True
# break
# if app:
X.append(sentence['words'])
Y.append(sentence['tags'])
newY = []
for sent in Y:
sent2 = []
for wtag in sent:
xx = np.zeros(4)
xx[wtag - 1] = 1
sent2.append(xx)
newY.append(sent2)
X, Y = np.array(X, dtype='int32'), np.array(newY, dtype='int32')
# X, X_test, Y, Y_test = cross_validation.train_test_split(X, Y, test_size=0.1, random_state=123)
print(X.shape[0])
print(Y.shape[0])
seqlen = 50
# Y = Y.reshape((-1, seqlen, 1))
seq_input = Input(shape=(seqlen, ), dtype='int32')
if not embedding_layer:
embedded = embedding_layer(seq_input)
print('Loading pretrained embedding')
else:
print('Default embedding')
embedded = Embedding(len(id_to_word) + 1,
50,
input_length=seqlen,
dropout=0.2)(seq_input)
forwards = LSTM(128, return_sequences=True)(embedded)
backwards = LSTM(128, return_sequences=True, go_backwards=True)(embedded)
merged = merge([forwards, backwards], mode='concat', concat_axis=-1)
after_dp = Dropout(0.2)(merged)
## Convoluntion1D Layer
half_window_size = 2
paddinglayer = ZeroPadding1D(padding=half_window_size)(embedded)
conv = Conv1D(nb_filter=50,
filter_length=(2 * half_window_size + 1),
border_mode='valid')(paddinglayer)
conv_d = Dropout(0.1)(conv)
dense_conv = TimeDistributed(Dense(50))(conv_d)
## Concat Bi-LSTM Layer and Convolution Layer
rnn_cnn_merge = merge([after_dp, dense_conv], mode='concat', concat_axis=2)
outputs = TimeDistributed(Dense(4, activation='softmax'))(rnn_cnn_merge)
#output = TimeDistributed(Dense(1,activation='sigmod'))
model = Md(input=seq_input, output=outputs)
with open('model/model_cnn.json', 'w') as fout:
fout.write(model.to_json())
modelfile = './model/model_cnn.h5'
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
with open('model/model_cnn.yaml', 'w') as fout:
fout.write(model.to_yaml())
if X.shape[0] > 5000: nb_epoch = 200
if X.shape[0] > 10000: nb_epoch = 150
if X.shape[0] > 40000: nb_epoch = 10
if X.shape[0] > 70000: nb_epoch = 7
if X.shape[0] > 100000: nb_epoch = 4
model.fit(X,
Y,
batch_size=256,
callbacks=[ModelCheckpoint(modelfile, save_best_only=True)],
validation_split=0.1,
nb_epoch=nb_epoch)
print("########### dev ##############")
X_dev = []
Y_dev = []
for sentence in dev_data:
X_dev.append(sentence['words'])
Y_dev.append(sentence['tags'])
tempY = []
for sent in Y_dev:
sent2 = []
for wtag in sent:
xx = np.zeros(4)
xx[wtag - 1] = 1
sent2.append(xx)
tempY.append(sent2)
X_dev, Y_dev = np.array(X_dev, dtype='int32'), np.array(tempY,
dtype='int32')
print(X_dev.shape[0])
print(Y_dev.shape[0])
loss_and_metrics = model.evaluate(X_dev, Y_dev, batch_size=128)
print(loss_and_metrics)
X_test = []
Y_test = []
for sentence in test_data:
X_test.append(sentence['words'])
Y_test.append(sentence['tags'])
teY = []
for sent in Y_test:
sent2 = []
for wtag in sent:
xx = np.zeros(4)
xx[wtag - 1] = 1
sent2.append(xx)
teY.append(sent2)
X_test, Y_test = np.array(X_test, dtype='int32'), np.array(teY,
dtype='int32')
print("########### test ##############")
print(X_test.shape[0])
print(Y_test.shape[0])
Z = model.predict(X_test, batch_size=256, verbose=1)
print(Z.shape)
tZ = np_utils.probas_to_classes(Z.reshape((-1, 1)))
tY = np_utils.probas_to_classes(Y_test.reshape((-1, 1)))
TP = (tZ + tY == 2).sum()
PP = (tZ == 1).sum()
RR = (tY == 1).sum()
prec, reca = TP / PP, TP / RR
F1 = 2 * prec * reca / (prec + reca)
ret = 'P=%d/%d %.5f\tR=%d/%d %.5f\tF1=%.5f' % (TP, PP, prec, TP, RR, reca,
F1)
print(ret)
show = True
if show:
amount = 0
entityNum = 0
findentityNum = 0
score = 0
with open('./results/ner_cnn.txt', 'w', encoding='utf-8') as fout:
for x, y, z in zip(X_test, Y_test, Z):
for xx, yy, zz in zip(x, y, z):
tagN = 0
for i in range(4):
if yy[i] == 1:
tagN = i + 1
break
tagY = id_to_tag[tagN]
maxpro = 0
otagN = 0
for j in range(4):
if zz[j] > maxpro:
maxpro = zz[j]
otagN = j + 1
tagZ = id_to_tag[otagN]
word = id_to_word[xx]
if word != '<PAD>':
amount = amount + 1
if ((tagY == 'ORG') or (tagY == 'LOC')
or (tagY == 'PER')):
entityNum = entityNum + 1
if tagY == tagZ:
score = score + 1
if ((tagY == 'ORG') or (tagY == 'LOC')
or (tagY == 'PER')):
findentityNum = findentityNum + 1
# #wd = ee.id2wd[wid]
# wd = tt
# if zz > 0.5: wd = wd + '@%.3f' % zz
# es.append(wd)
# if wid == 0: break
fout.writelines([
str(word) + ' ' + str(tagY) + ' ' + str(tagZ) +
'\n'
])
print('Num of Character : ' + str(amount))
print('Num of Character NER : ' + str(score))
print('Precison ' + str(score / amount))
print('Num of entity : ' + str(entityNum))
print('Num of entity NER : ' + str(findentityNum))
print('NER accuracy : ' + str(findentityNum / entityNum))
model.add(Dropout(0.5))
## Second hidden layer.
model.add(Dense(second_layer_width))
model.add(Activation("relu"))
if dropout_rate > 0:
model.add(Dropout(0.5))
model.add(Dense(second_layer_width))
model.add(Activation("relu"))
if dropout_rate > 0:
model.add(Dropout(0.5))
model.add(Dense(y_train.shape[1]))
## For classification, the activation is softmax
model.add(Activation('softmax'))
## Define optimizer. In this tutorial/codelab, we select SGD.
## You can also use other methods, e.g., opt = RMSprop()
opt = SGD(lr=learning_rate, clipnorm=5.)
## Define loss function = 'categorical_crossentropy' or 'mean_squared_error'
model.compile(loss="categorical_crossentropy", optimizer=opt, metrics=["accuracy"])
model.fit(scaled_x_train, y_train, nb_epoch=20, batch_size=32)
# Final test predict
test_proba = model.predict(scaled_test)
print(test_proba.shape)
print(test_proba[:5])
test_classes = np_utils.probas_to_classes(test_proba)
print(test_classes.shape)
print(test_classes[:5])
def _predict(self, X_test): proba = self.predict_proba(X_test) return np_utils.probas_to_classes(proba)
rms_optimizer = RMSprop()
model.compile(loss='categorical_crossentropy', optimizer=rms_optimizer, class_mode="categorical", theano_mode='FAST_COMPILE')
print("Train...")
model.fit(X_train, y_train_cat, batch_size=batch_size, nb_epoch=nb_epoch, validation_split=0.1, show_accuracy=True)
score = model.evaluate(X_test, y_test_cat, batch_size=batch_size)
print('Test score:', score)
classes_proba = model.predict_proba(X_test, batch_size=batch_size)
for i in range(5):
probs = sorted(zip(range(len(classes_proba)), classes_proba[i].tolist()), key=lambda x: x[1], reverse=True)
print('Test sample %d (Correct label: %s)' % (i, id_to_word[y_test[i]]))
for j, p in probs[:5]:
print(id_to_word[j].ljust(20) + ': ' + str(p))
classes = np_utils.probas_to_classes(classes_proba)
correct, wrong = 0, 0
for (i,q) in enumerate(test_questions):
options = q[5]
options_probs = classes_proba[i][options]
best_idx = np.argmax(options_probs)
predicted = options[best_idx]
print('Test sample %d (Correct label: %s)' % (i, id_to_word[y_test[i]]))
for k in range(len(options)):
print(id_to_word[options[k]].ljust(20) + ': ' + str(options_probs[k]))
if predicted == y_test[i]:
correct += 1
else:
wrong += 1
h = F.relu(self.l1(x))
return self.l2(h)
class Classifier(Chain):
def __init__(self, predictor):
super(Classifier, self).__init__(predictor=predictor)
def __call__(self, x, t):
y = self.predictor(x)
self.loss = F.softmax_cross_entropy(y, t)
self.accuracy = F.accuracy(y, t)
return self.loss
trY, teY = np_utils.probas_to_classes(trY).astype(np.int32), np_utils.probas_to_classes(teY).astype(np.int32)
# net = MNISTNet()
# x = chainer.Variable(trX)
# print net(x)
model = L.Classifier(MNISTNet())
optimizer = optimizers.SGD()
optimizer.setup(model)
# x = chainer.Variable(trX)
# t = chainer.Variable(trY)
# print model(x, t)
teX = chainer.Variable(teX)
teY = chainer.Variable(teY)
# use decoder to decode latent space information into images
z_in = layers.Input(shape=(L_SIZE, ))
y_in = layers.Input(shape=(OUT_SIZE, ))
decoder_in = layers.merge([z_in, y_in], mode='concat', concat_axis=1)
decoder_hidden = P_(decoder_in)
decoder_out = f_(decoder_hidden)
DECODER = Model([z_in, y_in], decoder_out)
"""
PLOTTING
"""
# ensure that encoder and decoder have a .predict() function on each.
x_encoded = ENCODER.predict([X_train, Y_train], batch_size=BATCH_SIZE)
f, ax = utils.data_on_latent_space(x_encoded,
np_utils.probas_to_classes(Y_train))
# f.colorbar()
f.savefig('latent_viz.cvae.pdf')
n_digits = 10
n = 5
latent_range = (0.05, 0.95)
digit_shape = (28, 28)
figure = np.zeros((digit_shape[0] * n_digits, digit_shape[1] * n))
for i in range(1, n_digits + 1):
digit_code = np.zeros(
n_digits) # encode the digit we are trying to generate
digit_code[i - 1] = 1
classifier.summary()
classifier.fit(X_train,
to_categorical(y_train,
len(y_values) + 1),
batch_size=BATCH_SIZE,
class_weight=class_weight,
verbose=1,
nb_epoch=EPOCHS)
logger.debug(classifier)
logger.info("Reading test data")
X_dev, y_dev_gold, _, entities = read_and_map(dev_src, mapper, y_values)
logger.info("Testing")
y_dev_pred = probas_to_classes(classifier.predict(X_dev))
# output entities
with open(os.path.join(output_dir, "entities.txt"), "w") as f:
for entity in entities:
f.write(str(entity) + '\n')
# output gold labels
with open(os.path.join(output_dir, "gold.txt"), "w") as f:
for i in xrange(len(y_dev_gold)):
f.write(y_values[y_dev_gold[i]] + '\n')
# output predicted labels
with open(os.path.join(output_dir, "pred.txt"), "w") as f:
for i in xrange(len(y_dev_pred)):
f.write(y_values[y_dev_pred[i]] + '\n')
def train(model_type='parallel', label_set='full', drop_unk=False,
word_vecs=None, setup_only=False):
print "Loading data..."
df = sentences_df(SENTENCES_CSV, labels=label_set, drop_unk=drop_unk)
X, y, word2idx, l_enc = load_dataset(df, pad=True)
print "X shape:", X.shape
y_orig = y
y_binary = to_categorical(y)
labels = np.unique(y_orig)
nb_labels = labels.shape[0]
if drop_unk:
label_set_str = label_set + ' (-unk)'
else:
label_set_str = label_set
print "Number of labels: %i [%s]" % (nb_labels, label_set_str)
if nb_labels > 2:
y = y_binary
maxlen = X.shape[1]
vocab_size = len(word2idx) + 1 # 0 masking
if pretrained_embeddings is True:
word_vectors = load_bin_vec(word_vecs, word2idx)
add_unknown_words(word_vectors, word2idx)
embedding_weights = np.zeros((vocab_size+1, emb_dim))
for word, index in word2idx.items():
embedding_weights[index,:] = word_vectors[word]
else:
embedding_weights = None
print "Data loaded."
if setup_only:
cnn = create_model(vocab_size, nb_labels, emb_dim, maxlen,
embedding_weights, filter_hs, nb_filters,
dropout_p, trainable_embeddings,
pretrained_embeddings, model_type=model_type)
return {'X': X,
'y': y,
'word2idx': word2idx,
'l_enc': l_enc,
'y_binary': y_binary,
'labels': labels,
'nb_labels': nb_labels,
'maxlen': maxlen,
'emb_dim': emb_dim,
'vocab_size': vocab_size,
'embedding_weights': embedding_weights,
'cnn': cnn}
params = [('filter_hs',filter_hs), ('nb_filters',nb_filters),
('dropout_p',dropout_p),
('trainable_embeddings',trainable_embeddings),
('pretrained_embeddings',pretrained_embeddings),
('batch_size',batch_size), ('nb_epoch',nb_epoch),
('lr',lr), ('beta_1',beta_1), ('beta_2',beta_2),
('epsilon',epsilon)]
print "\nModel type: %s" % model_type
for (name, value) in params:
print name + ':', value
skf = StratifiedKFold(y_orig, n_folds=10, shuffle=True, random_state=0)
cv_scores = []
for i, (train, test) in enumerate(skf):
start_time = time.time()
cnn = None
cnn = create_model(vocab_size,
nb_labels,
emb_dim,
maxlen,
embedding_weights,
filter_hs,
nb_filters,
dropout_p,
trainable_embeddings,
pretrained_embeddings,
model_type=model_type)
if i == 0:
print_summary(cnn.model.layers)
acc = train_and_test_model(cnn, X[train], y[train], X[test], y[test],
batch_size, nb_epoch,
lr, beta_1, beta_2, epsilon)
cv_scores.append(acc)
train_time = time.time() - start_time
print('\nLabel frequencies in y[test]')
print_label_frequencies((y_orig[test], l_enc))
y_pred = cnn.model.predict(X[test])
y_pred = probas_to_classes(y_pred)
c = Counter(y_pred)
total = float(len(y_pred))
print('\nLabel frequencies in predict(y[test])')
for label, count in c.most_common():
print l_enc.inverse_transform(label), count, count / total
print "fold %i/10 - time: %.2f s - acc: %.4f on %i samples" % \
(i+1, train_time, acc, len(test))
print "Avg cv accuracy: %.4f" % np.mean(cv_scores)
model.add(Dense(out_embedding_size, num_words))
model.add(Activation('softmax'))
sgd_optimizer = SGD(lr=0.006, momentum=0.9, decay=0.99, nesterov=True)
adg_optimizer = Adagrad()
rms_optimizer = RMSprop()
model.compile(loss='categorical_crossentropy', optimizer=rms_optimizer, class_mode="categorical", theano_mode='FAST_COMPILE')
print("Train...")
model.fit(X_train, y_train_cat, batch_size=batch_size, nb_epoch=nb_epoch, validation_split=0.1, show_accuracy=True)
score = model.evaluate(X_test, y_test_cat, batch_size=batch_size)
print('Test score:', score)
classes_proba = model.predict_proba(X_test, batch_size=batch_size)
for i in range(5):
probs = sorted(zip(range(len(classes_proba)), classes_proba[i].tolist()), key=lambda x: x[1], reverse=True)
print('Test sample %d (Correct label: %s)' % (i, id_to_word[y_test[i]]))
for j, p in probs[:5]:
print(id_to_word[j].ljust(20) + ': ' + str(p))
classes = np_utils.probas_to_classes(classes_proba)
acc = np_utils.accuracy(classes, y_test)
print('Test accuracy:', acc)
# print(classes.shape)
# print(classes[0])
# print(y_test[0])
# classes_list = classes.tolist()
# print(map(lambda x: id_to_word[x], classes_list[:25]))
# finally, an output layer with one node per class
model.add(Dense(num_labels))
model.add(Activation('softmax'))
# use the Adam optimiser
adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
# now compile the model, Keras will take care of the Tensorflow boilerplate
model.compile(loss='categorical_crossentropy',
metrics=['accuracy'],
optimizer=adam)
# for quicker training, just using one epoch, you can experiment with more
model.fit(train_x,
train_y,
validation_data=(valid_x, valid_y),
batch_size=32,
nb_epoch=1)
# finally, evaluate the model using the withheld test dataset
# determine the ROC AUC score
y_prob = model.predict_proba(test_x, verbose=0)
y_pred = np_utils.probas_to_classes(y_prob)
y_true = np.argmax(test_y, 1)
roc = metrics.roc_auc_score(test_y, y_prob)
print "ROC:", round(roc, 3)
# determine the classification accuracy
score, accuracy = model.evaluate(test_x, test_y, batch_size=32)
print("\nAccuracy = {:.2f}".format(accuracy))
h = F.relu(self.l1(x))
return self.l2(h)
class Classifier(Chain):
def __init__(self, predictor):
super(Classifier, self).__init__(predictor=predictor)
def __call__(self, x, t):
y = self.predictor(x)
self.loss = F.softmax_cross_entropy(y, t)
self.accuracy = F.accuracy(y, t)
return self.loss
trY, teY = np_utils.probas_to_classes(trY).astype(
np.int32), np_utils.probas_to_classes(teY).astype(np.int32)
# net = MNISTNet()
# x = chainer.Variable(trX)
# print net(x)
model = L.Classifier(MNISTNet())
optimizer = optimizers.SGD()
optimizer.setup(model)
# x = chainer.Variable(trX)
# t = chainer.Variable(trY)
# print model(x, t)
teX = chainer.Variable(teX)
teY = chainer.Variable(teY)
def train_model(model_generator, feature_set):
# TODO get data function that returns dict (options for one hot or not, val or not)
# load dataset
paths = u.load_paths('PATHS.yaml') # get paths from file
train_x, val_x, test_x = u.load_data(paths['extracted_data'] +
'features_%s.p' % feature_set)
train_y, val_y, test_y = u.load_data(paths['extracted_data'] +
'labels_%s.p' % feature_set)
model_name, model, training = model_generator(n_dim=train_x.shape[1:],
n_labels=test_y.shape[1])
run_id = '%s_%s' % (model_name, datetime.datetime.now().isoformat())
print('\nTrain and Evaluate: %s' % model_name)
# callbacks
earlystop = EarlyStopping(monitor='val_loss',
patience=training.early_stop_patience,
verbose=1,
mode='auto')
log_dir = os.path.join(paths['tensorboard_logs'], run_id)
tensorboard = TensorBoard(log_dir=log_dir,
histogram_freq=3,
write_graph=True)
t0 = time.time()
history = model.fit(train_x,
train_y,
validation_data=(val_x, val_y),
callbacks=[earlystop, tensorboard],
nb_epoch=training.n_epoch,
batch_size=training.batch_size)
training_time = time.time() - t0
# test
y_prob = model.predict_proba(test_x, verbose=0)
y_pred = np_utils.probas_to_classes(y_prob)
y_true = np.argmax(test_y, 1)
# evaluate the model's accuracy
t0 = time.time()
score, accuracy = model.evaluate(test_x,
test_y,
batch_size=training.batch_size)
testing_time = time.time() - t0
cm = confusion_matrix(y_true, y_pred, labels=None)
# p, r, f, s = precision_recall_fscore_support(y_true, y_pred, average='micro')
# roc = roc_auc_score(test_y, y_prob)
# print("F-Score:", round(f, 2)) # similar value to the accuracy score, but useful for cross-checking
# print("ROC:", round(roc, 3))
# print results
print("\nclassifier: %s" % model_name)
print("training time: %0.4fs" % training_time)
print("testing time: %0.4fs" % testing_time)
print("accuracy: %0.4f" % accuracy)
print("confusion matrix:\n%s" % cm)
# print model.summary()
# plot and save results
fname = paths['model_save'] + model_name + '_accuracy_%0.2f' % accuracy
u.plot_keras_loss(fname, history) # saves plot png
model.save(fname + '.h5')
cm_path = './confusion_plots/%s' % model_name
cm_title = '%s (Accuracy: %0.2f)' % (model_name, accuracy)
u.plot_confusion_matrix(cm, cm_path, title=cm_title)
def predict(self, Xtest):
ypred = self.model.predict(Xtest, batch_size=self.batch_size, verbose=self.verbose)
return np_utils.probas_to_classes(ypred)
def print_class_numbers(Y, nb_classes):
classes = np_utils.probas_to_classes(Y)
for i in xrange(nb_classes):
print 'class %d: %d' % (i, np.sum(classes == i))
def predict_class_label(self, audio_file):
x_features = self.load_features(audio_file)
instrument_class = np_utils.probas_to_classes(self.model.predict(x_features, verbose=0))[0]
label = self.instr_family_le.inverse_transform(instrument_class)
return label
model.save('digitRecognitionModel.h5') # creates a HDF5 file 'my_model.h5'
#pre=x_test[0]
#==============================================================================
# pre=np.array([pre])
# #print pre.shape[0]
# predictedValue =model.predict(pre)
# print predictedValue
# rounded = [round(x) for x in predictedValue[0]]
# print rounded
# predictedClass=np_utils.probas_to_classes(predictedValue)
# print predictedClass
#==============================================================================
#######################prediction of image file#########################
originalImage = Image.open('digitsImages/5.bmp')
reducedImage = originalImage.resize((28,28))
#plt.imshow(reducedImage ,cmap=plt.get_cmap('gray'))
imageArray= np.array(reducedImage)
numPixels=imageArray.shape[0]*imageArray.shape[1]
imageArray=imageArray.reshape(1,numPixels).astype('float32')
predictedValue =model.predict(imageArray)
predictedClass=np_utils.probas_to_classes(predictedValue)
print "Digit is ",predictedClass
]
input_test = [
subject_idx_test, question_idx_test, comment_idx_test,
subj_overlap_idx_test, ques_overlap_idx_test, comm_overlap_idx_test,
overlap_feat_test
]
print('Load Trained Model')
model = model_from_json(open(base_model_path).read())
model.load_weights(current_trained_model_path)
y_pred_train = model.predict(input_train)
y_pred_test = model.predict(input_test)
y_pred_test_cls = probas_to_classes(y_pred_test)
ofile = open('data_folder/semeval2016-task3-taskA-pred.tsv', 'wb')
for i, (qid, cid) in enumerate(zip(relq_id_test, relc_id_test)):
p = y_pred_test_cls[i]
if p == 0:
label = 'false'
elif p == 1:
label = 'true'
proba = y_pred_test[i][1]
outline = '{}\t{}\t{}\t{}\t{}\n'.format(qid, cid, 0, proba, label)
ofile.write(outline)
ofile.close()
train_results = defaultdict(lambda: [])
def predict_class_label(self, audio_file):
x_features = self.load_features(audio_file)
instrument_class = np_utils.probas_to_classes(
self.model.predict(x_features, verbose=0))[0]
label = self.instr_family_le.inverse_transform(instrument_class)
return label
print('Load Supervised Model')
model.load_weights(super_weight_path)
for tids, X_test, y_test, name in test_sets:
raw_data = open(os.path.join('semeval', '{}.tsv'.format(name)),
'r').readlines()
raw_data = map(lambda x: x.replace('\n', '').split('\t'), raw_data)
raw_tweets = map(lambda x: (x[0], x[-1]), raw_data)
raw_lables = map(lambda x: (x[0], x[-2]), raw_data)
raw_data_dict = dict(raw_tweets)
raw_lables_dict = dict(raw_lables)
ofile = open(os.path.join(res_dir, name), 'w')
y_pred = model.predict(X_test)
y_pred = probas_to_classes(y_pred)
score = semeval_f1_taskA(y_test, y_pred)
scores = f1_score(y_test, y_pred, average=None)
output += '{}: {}\t'.format(name, score)
output += 'neg_f1: {}\t neut_f1: {}\t pos_f1: {}'.format(
scores[0], scores[1], scores[2])
for tid, label in zip(tids, y_pred):
tweet = raw_data_dict[tid].replace('\n', '')
truth = raw_lables_dict[tid]
l = {0: 'negative', 1: 'neutral', 2: 'positive'}.get(label)
outline = '{}\t{}\t{}\n'.format(tweet, truth, l)
ofile.write(outline)
open(os.path.join('results', 'results_log.tsv'), 'a').write(output + '\n')
optimizer='sgd',
metrics=['accuracy'])
model.fit(trainData,
y_train,
batch_size=batch_size,
nb_epoch=nb_epoch,
verbose=1,
validation_data=(testData, y_test))
score = model.evaluate(testData, y_test, verbose=0)
print('Test score:', score[0])
print('Test accuracy:', score[1])
W, b = model.layers[0].get_weights()
# print('Weights=', W, '\n biases=', b)
y_test_pred = model.predict(testData)
print(
np_utils.probas_to_classes(y_test)[:10],
np_utils.probas_to_classes(y_test_pred)[:10])
'''test '''
predictFileName = "./data/test.csv"
x_pre = np.array(pd.read_csv(predictFileName))
print("x_pre shape", x_pre.shape)
y_pred = model.predict(x_pre)
pre_class = np_utils.probas_to_classes(y_pred)
print("predict class :", pre_class)
index = np.linspace(1, len(pre_class), len(pre_class))
print("index", index)
np.savetxt("./data/pre.csv",
list(zip(index, pre_class)),
delimiter=',',
fmt='%10.5f')
for train_index,valid_index in kfold.split(np.array([0]*3000),np.array([0]*3000)):
print('<<<<<COUNT>>>>> '+str(count))
model = Model(input = [input1], output=[output])
# load model weight
CNN_model = model.load_weights('model(0.9375).h5')
model.compile(loss='binary_crossentropy', optimizer = "adam", metrics=['accuracy'])
loss, accuracy = model.evaluate(train[valid_index], train_label[valid_index], verbose=0)
valid_score.extend([accuracy])
print('loss: '+str(loss)+', acc: '+str(accuracy))
count = count+1
print("valid: %.2f%% (+/- %.2f%%)" % (np.mean(valid_score), np.std(valid_score)))
"""
"""predict"""
predictions = model.predict({'input1':test})
POS_predict_keras = np_utils.probas_to_classes(predictions)
pre = np.sum(test_label[:,1]==POS_predict_keras)/len(POS_predict_keras)
print(pre)
"""save result"""
model.save('CNN_model.h5')
import csv
f = open("POS_predict_keras.csv", "w")
writer = csv.writer(f, lineterminator="\n")
writer.writerow(POS_predict_keras)
f.close
#######
'''