Question

我试图在一个GPU上并行安装多个小型Keras模型。由于我需要将它们从列表中删除并一次训练它们的原因。由于我不熟悉标准的多处理模块，我使用了pathos。

我试图做的是这样的事情：

from pathos.multiprocessing import ProcessPool as Pool
import tensorflow as tf
import keras.backend as K

def multiprocess_step(self, model):
    K.set_session(sess)
    with sess.graph.as_default():
        model = step(model, sess)
        return model

def step(model, sess):
    K.set_session(sess)
    with sess.graph.as_default():
        model.fit(x=data['X_train'], y=data['y_train'],
               batch_size=batch_size
               validation_data=(data['X_test'], data['y_test']), 
               verbose=verbose,
               shuffle=True,
               initial_epoch=self.step_num - 1)
        return model

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = "0"
sess = tf.Session(config=config)

K.set_session(sess)
with sess.graph.as_default():
    pool = Pool(8).map
    model_list = pool(multiprocess_step, model_list)

但无论我尝试什么，我都会收到一个错误，声称模型似乎不在同一个图表上......

ValueError: Tensor("training/RMSprop/Variable:0", shape=(25, 352), dtype=float32_ref) must be from the same graph as Tensor("RMSprop/rho/read:0", shape=(), dtype=float32).

异常来自于model.fit（）行，所以即使我试图在每个可能的位置设置它，我必须对会话图的分配做错了吗？

有没有人有过类似的经历？

Answer 1

Keras issue tracker建议如下。与使用多处理相比，我不确定该方法的相对优点。

in_1 = Input()
lstm_1 = LSTM(...)(in_1)
out_1 = Dense(...)(lstm_1)

in_2 = Input()
lstm_2 = LSTM(...)(in_2)
out_2 = Dense(...)(lstm_2)

model_1 = Model(input=in_1, output=out_1)
model_2 = Model(input=in_2, output=out_2)

model = Model(input = [in_1, in_2], output = [out_1, out_2])
model.compile(...)
model.fit(...)

model_1.predict(...)
model_2.predict(...)

Answer 2

考虑将后端设置为keras的tensorflow。您可以使用代码并对多个模型调用/多个模型加载进行并行处理。

def model1(dir_model):
    model = os.path.join(dir_model, 'model.json')
    dir_weights = os.path.join(dir_model, 'model.h5')
    graph1 = Graph()
    with graph1.as_default():
        session1 = Session(graph=graph1, config=config)
        with session1.as_default():
            with open(model, 'r') as data:
                model_json = data.read()
            model_1 = model_from_json(model_json)
            model_1.load_weights(dir_weights)
    return model_1,gap_weights,session1,graph1

def model_2(dir_model):
    model = os.path.join(dir_model, 'model.json')
    dir_weights = os.path.join(dir_model, 'model.h5')
    graph2 = Graph()
    with graph2.as_default():
        session2 = Session(graph=graph2, config=config)
        with session2.as_default():
            with open(model, 'r') as data:
                model_json = data.read()
            model_2 = model_from_json(model_json)
            model_2.load_weights(dir_weights)
    return model_2,session2,graph2

并且对于特定模型的调用，进行以下实验。对于模型1预测执行以下操作

K.set_session(session2)
with graph2.as_default():
     img_pred[img_name] = 
patch_dict[np.argmax(np.squeeze(model_2.predict(img_invoke)))

，对于模型2，它跟

一样

K.set_session(session2)
with graph2.as_default():
     img_pred[img_name] = 
patch_dict[np.argmax(np.squeeze(model_2.predict(img_invoke)))]

在单个GPU

2 个答案: