def create_GRU_stack_model(images_shape, dict_size, sentence_len, settings, pretrained_emb):
# input (None, 224, 224, 3), outputs (None, sentence_len, 512)
image_model = create_image_model_resnet50(images_shape, sentence_len)
# outputs (None, sentence_len, 128)
sentence_model = create_sentence_model(dict_size, sentence_len, pretrained_emb, 160)
combined_model = Sequential()
combined_model.add(Merge([image_model, sentence_model], mode='concat', concat_axis=-1))
combined_model.add(GRU(160, return_sequences=True, dropout_U=0.25, dropout_W=0.25))
combined_model2 = Sequential()
combined_model2.add(Merge([image_model, combined_model], mode='concat', concat_axis=-1))
combined_model2.add(GRU(256, return_sequences=False, dropout_U=0.25, dropout_W=0.25))
addPredictionLayers(combined_model2, dict_size, settings)
# input words are 1-indexed and 0 index is used for masking!
# but result words are 0-indexed and will go into [0, ..., dict_size-1] !!!
combined_model2.compile(loss='sparse_categorical_crossentropy', optimizer=create_optimizer(settings))
return combined_model2
def create_GRU_xception_model(images_shape, dict_size, sentence_len, settings, pretrained_emb):
image_model = create_image_model_xception(images_shape, sentence_len)
sentence_model = create_sentence_model(dict_size, sentence_len, pretrained_emb)
combined_model = Sequential()
combined_model.add(Merge([image_model, sentence_model], mode='concat', concat_axis=-1))
combined_model.add(GRU(256, return_sequences=False, dropout_U=0.2, dropout_W=0.2))
addPredictionLayers(combined_model, dict_size, settings)
# input words are 1-indexed and 0 index is used for masking!
# but result words are 0-indexed and will go into [0, ..., dict_size-1] !!!
combined_model.compile(loss='sparse_categorical_crossentropy', optimizer=create_optimizer(settings))
return combined_model
#Left Image
model_left = Sequential()
model_left.add(Cropping2D(cropping=((70,25),(0,0)), input_shape=(160,320,3)))
model_left.add(Lambda(lambda x: x / 255.0 - 0.5))
model_left.add(Convolution2D(24,5,5, subsample=(1,1), activation="relu"))
#Right Image
model_right = Sequential()
model_right.add(Cropping2D(cropping=((70,25),(0,0)), input_shape=(160,320,3)))
model_right.add(Lambda(lambda x: x / 255.0 - 0.5))
model_right.add(Convolution2D(24,5,5, subsample=(1,1), activation="relu"))
#Nvidia
model = Sequential()
model.add(Merge([model_center, model_left, model_right], mode='concat'))
#model.add(Convolution2D(24,5,5, subsample=(2,2), activation="relu"))
model.add(Convolution2D(36,5,5, subsample=(2,2), activation="relu"))
model.add(Convolution2D(48,5,5, subsample=(2,2), activation="relu"))
model.add(Convolution2D(64,5,5, subsample=(2,2), activation="relu"))
#model.add(Convolution2D(64,5,5, subsample=(2,2), activation="relu"))
model.add(Flatten())
model.add(Dense(100))
model.add(Dense(50))
model.add(Dense(1))
model.compile(loss='mse', optimizer='adam')
print("Fitting")
model.fit([X_train, X_train, X_train], y_train, validation_split=0.2, shuffle=True, nb_epoch=5, verbose=1)
model.save('model.h5')
def main():
index_name = [
'end', 'ignore', 'approach', 'move', 'grasp_left', 'grasp_right',
'ungrasp_left', 'ungrasp_right', 'twist', 'push', 'neutral', 'pull',
'pinch', 'unpinch'
]
training_data, training_current_action, training_next_action, testing_data, testing_current_action, testing_next_action = load_data(
index_name)
# model = Sequential()
# model.add(Dense(64, input_dim=training_data.shape[1]+len(index_name), activation='relu'))
# model.add(Dropout(0.3))
# model.add(Dense(64, activation='relu'))
# model.add(Dropout(0.3))
# model.add(Dense(len(index_name), activation='softmax'))
#
# model.compile(loss='categorical_crossentropy',
# optimizer='adadelta',
# metrics=['accuracy'])
#
# model.fit(np.hstack((training_data, training_current_action)),
# training_next_action,
# nb_epoch=500,
# validation_split=0.2,
# batch_size=16) # starts training
#
# model.save('model1.h5')
# score = model.fit(np.hstack((testing_data, testing_current_action)), testing_next_action, batch_size=16)
# print score
left_branch = Sequential()
left_branch.add(
Dense(64, input_dim=training_data.shape[1], activation='relu'))
left_branch.add(Dropout(0.3))
left_branch.add(Dense(64, activation='relu'))
left_branch.add(Dropout(0.3))
left_branch.add(Dense(64, activation='relu'))
left_branch.add(Dropout(0.3))
left_branch.add(Dense(64, activation='relu'))
right_branch = Sequential()
right_branch.add(Dense(8, input_dim=len(index_name), activation='relu'))
merged = Merge([left_branch, right_branch], mode='concat')
model = Sequential()
model.add(merged)
model.add(Dense(32, activation='relu'))
# model.add(Dense(32, activation='relu'))
model.add(Dense(14, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='adadelta',
metrics=['accuracy'])
model.fit([training_data, training_current_action],
training_next_action,
nb_epoch=20000,
validation_split=0.2,
batch_size=16) # starts training
model.save('model2.h5')
score = model.evaluate([testing_data, testing_current_action],
testing_next_action,
batch_size=16)
print score
def predictOneShop_ANN_LSTM(shopid, all_data, trainAsTest=False):
"""
用ANN预测某一个商店,2个网络分别模拟近期趋势和中期趋势,隐藏层合并,1层隐藏层,近期趋势用LSTM,速度慢,而且效果有时候也不好
:param shopid: 预测商店id
:param trainAsTest: 是否使用训练集后14天作为测试集
:return:
"""
part_data = all_data[all_data.shopid == shopid]
last_14_real_y = None
# 取出一部分做训练集
if trainAsTest: #使用训练集后14天作为测试集的话，训练集为前面部分
last_14_real_y = part_data[len(part_data) - 14:]["count"].values
part_data = part_data[0:len(part_data) - 14]
# print last_14_real_y
verbose = 2
rnn_nb_epoch = 10
skipNum = 28
day_backNum = 7
sameday_backNum = 3
week_backnum = 3
learnrate = 0.01
sameday = extractBackSameday(part_data, sameday_backNum, skipNum, nan_method_sameday_mean)
day = extractBackDay(part_data,day_backNum,skipNum,nan_method_sameday_mean)
count = extractCount(part_data, skipNum)
train_x = getOneWeekdayFomExtractedData(sameday)
train_x2 = getOneWeekdayFomExtractedData(day)
train_y = getOneWeekdayFomExtractedData(count)
other_features = [statistic_functon_mean,statistic_functon_median]
# other_features = []
for feature in other_features:
value = getOneWeekdayFomExtractedData(extractBackWeekValue(part_data, week_backnum, skipNum, nan_method_sameday_mean, feature))
train_x = np.append(train_x, value, axis=1)
# '''添加周几'''
# extract_weekday = getOneWeekdayFomExtractedData(extractWeekday(part_data, skipNum))
# train_x = np.append(train_x, extract_weekday, axis=1)
# ''''''
'''将t标准化'''
x_scaler = MinMaxScaler().fit(train_x)
x2_scaler = MinMaxScaler().fit(train_x2)
y_scaler = MinMaxScaler().fit(train_y)
train_x = x_scaler.transform(train_x)
train_x2 = x2_scaler.transform(train_x2)
train_x2 = train_x2.reshape((train_x2.shape[0],
train_x2.shape[1], 1))
train_y = y_scaler.transform(train_y)
'''标准化结束'''
# train_x = train_x.reshape((train_x.shape[0],
# train_x.shape[1], 1))
model1 = Sequential()
model2 = Sequential()
final_model = Sequential()
# print getrefcount(model1)
model1.add(Dense(32, input_dim=train_x.shape[1], activation="sigmoid")) #sigmoid
# model1.add(Dense(1, activation='linear'))
'''近期趋势'''
model2.add(LSTM(32, input_shape=(train_x2.shape[1],train_x2.shape[2]), activation="sigmoid"))
final_model.add(Merge([model1, model2],mode="concat",concat_axis=1))
final_model.add(Dense(1, activation='linear'))
#, W_regularizer=l2(0.01), activity_regularizer=activity_l2(0.01)
# print getrefcount(model1)
# 设置优化器（除了学习率外建议保持其他参数不变）
rms=RMSprop(lr=0.05)
# sgd=SGD(lr=0.1, momentum=0.9, nesterov=True)
final_model.compile(loss="mse", optimizer=rms)
print final_model.summary()
# print model1.summary()
# print getrefcount(model1)
# print model1.summary()
final_model.fit([train_x, train_x2], train_y, nb_epoch=rnn_nb_epoch, batch_size=1, verbose=verbose)
# print model1.get_weights()
# part_counts = []
# for i in range(7):
# weekday = i + 1
# part_count = getOneWeekdayFomExtractedData(count, weekday)
# part_counts.append(part_count)
# print getrefcount(model1)
format = "%Y-%m-%d"
if trainAsTest:
startTime = datetime.datetime.strptime("2016-10-18", format)
else:
startTime = datetime.datetime.strptime("2016-11-1", format)
timedelta = datetime.timedelta(1)
preficts = []
for i in range(14):
currentTime = startTime + timedelta * i
strftime = currentTime.strftime(format)
# index = getWeekday(strftime) - 1
# part_count = part_counts[index]
#取前{sameday_backNum}周同一天的值为特征进行预测
part_data = part_data.append({"count":0, "shopid":shopid, "time":strftime, "weekday":getWeekday(strftime)}, ignore_index=True)
x = getOneWeekdayFomExtractedData(extractBackSameday(part_data,sameday_backNum,part_data.shape[0] - 1, nan_method_sameday_mean))
x2 = getOneWeekdayFomExtractedData(extractBackDay(part_data,day_backNum,part_data.shape[0]-1,nan_method_sameday_mean))
for feature in other_features:
x_value = getOneWeekdayFomExtractedData(extractBackWeekValue(part_data, week_backnum, part_data.shape[0]-1, nan_method_sameday_mean, feature))
x = np.append(x, x_value, axis=1)
# '''添加周几'''
# x = np.append(x, getOneWeekdayFomExtractedData(extractWeekday(part_data, part_data.shape[0]-1)), axis=1)
# ''''''
x = x_scaler.transform(x)
x2 = x2_scaler.transform(x2)
x2 = x2.reshape((x2.shape[0],x2.shape[1],1))
# for j in range(sameday_backNum):
# x.append(train_y[len(train_y) - (j+1)*7][0])
# x = np.array(x).reshape((1, sameday_backNum))
# print x
# x = x.reshape(1, sameday_backNum, 1)
predict = final_model.predict([x,x2])
predict = y_scaler.inverse_transform(predict)[0][0]
if(predict <= 0):
predict == 1
preficts.append(predict)
part_data.set_value(part_data.shape[0]-1, "count", predict)
# preficts.append(predict)
# part_counts[index] = np.append(part_count, predict).reshape((part_count.shape[0] + 1, 1))
preficts = (removeNegetive(toInt(np.array(preficts)))).astype(int)
# preficts = np.array(preficts)
if trainAsTest:
last_14_real_y = (removeNegetive(toInt(np.array(last_14_real_y)))).astype(int)
# print preficts,last_14_real_y
print str(shopid)+',score:', scoreoneshop(preficts, last_14_real_y)
return [preficts, last_14_real_y]