我正在使用GloVe嵌入查找创建LSTM +视频模型，但是失败了，因为我的占位符值显然与GloVe嵌入矩阵不在同一张图中。相关代码：
class BaseModel(object):
def __init__(self, batch_size, train_json_path, test_json_path):
# just sets some class variables
# ...
def build_retrieval_models(self, sentence_embed_train, visual_feature_train, p_time_stamp, sentence_embed_test, visual_feature_test, time_stamp_test):
with tf.variable_scope('Base_Model'):
print 'Building training network...............................\n'
bottom_positive = self.vid_model('v2s_lt', visual_feature_train, p_time_stamp)
query_text_train = self.text_model('t2s_lt', sentence_embed_train)
# reusing variables in model for test network
tf.get_variable_scope().reuse_variables()
print 'Building test network...............................\n'
transformed_video_test = vid_model('v2s_lt', visual_feature_test, time_stamp_test)
transformed_text_test = self.text_model('t2s_lt', sentence_embed_test)
return bottom_positive, query_text_train, transformed_video_test, transformed_text_test
def vid_model(self, name, bottom, time_stamp=None):
if time_stamp is not None:
bottom = tf.concat([bottom, time_stamp], axis=1)
with tf.variable_scope(name):
fc1 = fc_relu('vid_fc1', bottom, output_dim=self.visual_embedding_dim[0])
fc2 = fc('vid_fc2', fc1, output_dim=self.visual_embedding_dim[1])
return fc2
def text_model(self, name, text_seq_batch):
with tf.variable_scope(name):
embedding_mat = tf.get_variable('glove_embedding', initializer=self.glove_embedding, trainable=False)
embedded_seq = tf.nn.embedding_lookup(embedding_mat, text_seq_batch)
lstm_top = lstm('text_lstm1', embedded_seq, None, output_dim=self.language_embedding_dim[0],
num_layers=1, forget_bias=1.0, apply_dropout=False,
concat_output=False)[-1]
fc1 = fc('text_fc1', lstm_top, output_dim=self.language_embedding_dim[1])
return fc1
def init_placeholder(self):
# text_seq_batch is a set of langauge indices that mark index of vocab word in glove vocab list
text_seq_batch = tf.placeholder(tf.float32, shape=(self.batch_size, self.sentence_length))
video_seq_batch = tf.placeholder(tf.float32, shape=(self.batch_size, self.visual_feature_dim))
video_label_batch = tf.placeholder(tf.float32, shape=(self.batch_size, 2))
text_seq_batch_test = tf.placeholder(tf.float32, shape=(self.test_batch_size, self.sentence_length))
video_seq_batch_test = tf.placeholder(tf.float32, shape=(self.test_batch_size, self.visual_feature_dim))
video_label_batch_test = tf.placeholder(tf.float32, shape=(self.test_batch_size, 2))
return text_seq_batch, video_seq_batch, video_label_batch, text_seq_batch_test, video_seq_batch_test, video_label_batch_test
def training(self, loss):
v_dict = self.trainable_variables()
vs_optimizer = tf.train.AdamOptimizer(self.vs_lr, name='vs_adam')
vs_train_op = vs_optimizer.minimize(loss, var_list=v_dict)
return vs_train_op
def construct_model(self):
self.text_seq_batch, self.video_seq_batch, self.video_label_batch, self.text_seq_batch_test, self.video_seq_batch_test, self.video_label_batch_test = self.init_placeholder()
# build network
bottom_pos_train, query_text_train, bottom_pos_test, query_text_test = self.build_retrieval_models(self.text_seq_batch, self.video_seq_batch, self.video_label_batch, self.text_seq_batch_test, self.video_seq_batch_test, self.video_label_batch_test)
# compute loss, method not implemented for purposes of this question
self.loss = self.compute_ranking_loss(bottom_pos_train, query_text_train)
self.vs_train_op = self.training(self.loss)
return self.loss, self.vs_train_op, bottom_pos_test, query_text_test
fc，fc_relu和lstm之类的辅助函数是完全连接的层和LSTM的标准实现。我认为问题出在我在构建模型时如何确定模型的权重，但是我不太明白为什么。在训练模型的主要功能中，代码如下：
def train_model():
batch_size = 2
max_steps = 30000
train_json_path = 'data/train_data.json'
test_json_path = 'data/val_data.json'
model = base_model.BaseModel(batch_size, train_json_path, test_json_path)
with tf.Graph().as_default():
loss, vs_train_op, bottom_pos_test, query_text_test = model.construct_model()
# create a session for running Ops on the Graph
sess = tf.Session()
init = tf.initialize_all_variables()
sess.run(init)
for step in xrange(max_steps):
start_time = time.time()
feed_dict = model.fill_feed_dict_train_reg()
_, loss_value = sess.run([vs_train_op, loss], feed_dict=feed_dict)
duration = time.time() - start_time
print('Step %d: loss = %.3f (%.3f sec)' % (step, loss_value, duration))
def main(_):
train_model()
if __name__ == '__main__':
tf.app.run()
由于在Tensorflow图创建之后所有实例化了，所以不是所有的Tensor都位于同一图中吗？但是，我得到了错误：
ValueError: Tensor("Base_Model/Placeholder:0", shape=(2, 2), dtype=float32) must be from the same graph as Tensor("Base_Model/glove_embedding:0", shape=(400001, 50), dtype=float32_ref).
其源于embedded_seq = tf.nn.embedding_lookup(embedding_mat, text_seq_batch)的{{1}}函数中的行text_model。
似乎Tensorflow实际上在抱怨我的BaseModel变量，因为它的形状为video_label_batch，但在嵌入查找过程中没有涉及。那么为什么这是代码中的问题？