我目前正在努力改善张量流管道中最昂贵的操作的运行时间。
我正在尝试完成以下操作:我获得了包含一些患者数据的多个样本的3D张量,例如数据可能看起来像这样
n_hidden = 3 #number of elements per 1D tensor
batch_size = 3 #number of patients
n_mc_samples = 2 #number of samples per patient
rnn_grid_times = [2,3,1] #number of non zero 1D tensors per patient
all_outputs = tf.constant([[[0.15, 0.874, 0.2], [0.1,0.00878,0.58],[0.0,0.0,0.0]], #beginning of patient 1
[[0.456,0.454,0.003],[0.4564,0.4984,0.21], [0.0,0.0,0.0]],
[[0.121,0.22,0.45],[0.15,0.488,0.222], [0.11,0.849,0.45]], #beginning of patient 2
[[0.15, 0.5646, 0.15], [0.45,0.48949,0.56465], [0.4489,0.456,0.9]],
[[0.121, 0.22, 0.01], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]], #beginning of patient 3
[[0.15, 0.89, 0.42], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]])
该数据对应于3名患者,每位患者采样两次。如您所见,患者1和3的数据被填充以具有与患者2相同的大小。
我的目标是将每个非零1D张量馈送到具有一个隐藏层的单个输出神经网络,然后在零张量的位置处添加额外的填充,以便在患者之间保持统一的维度。所以这里有一个有效的结果
[[-0.11379365, -0.11188659, 0. ],
[-0.11379365, -0.11379365, 0. ],
[-0.1135166 , -0.11379365, -0.11379365],
[-0.11379365, -0.11359671, -0.11270589],
[-0.11379365, 0. , 0. ],
[-0.11379365, 0. , 0. ]]
重申一下,因为我意识到这有点复杂,第一个代码块中[0.15, 0.874, 0.2]
与-0.11379365
相关的输出在第二个代码块中为import tensorflow as tf
RANDOM_SEED = 42
tf.set_random_seed(RANDOM_SEED)
def code():
n_hidden = 3
batch_size = 3
n_mc_samples = 2
num_rnn_grid_times = tf.constant([2, 3, 1])
all_outputs = tf.constant([[[0.15, 0.874, 0.2], [0.1,0.00878,0.58],[0.0,0.0,0.0]], #beginning of patient 1
[[0.456,0.454,0.003],[0.4564,0.4984,0.21], [0.0,0.0,0.0]],
[[0.121,0.22,0.45],[0.15,0.488,0.222], [0.11,0.849,0.45]], #beginning of patient 2
[[0.15, 0.5646, 0.15], [0.45,0.48949,0.56465], [0.4489,0.456,0.9]],
[[0.121, 0.22, 0.01], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]], #beginning of patient 3
[[0.15, 0.89, 0.42], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]]])
n_extra_hidden_nodes = 2
extra_hidden_weights = tf.Variable(tf.random_normal([n_hidden, n_extra_hidden_nodes], stddev=0.1), name="HiddenSoftmax/W")
extra_hidden_biases = tf.Variable(tf.random_normal([n_extra_hidden_nodes], stddev=0.1), name="HiddenSoftmax/b")
out_weights = tf.Variable(tf.random_normal([n_extra_hidden_nodes, 1], stddev=0.1), name="Softmax/W")
out_biases = tf.Variable(tf.random_normal([1], stddev=0.1), name="Softmax/b")
nneth_array_total = tf.Variable([])
n = tf.constant(0)
inner_cond = lambda i, nneth_array, n: tf.less(i, num_rnn_grid_times[tf.floordiv(n,n_mc_samples)])
cond = lambda n, nneth_array_total: tf.less(n, batch_size*n_mc_samples)
def inner_body(i, nneth_array, n):
hidden = tf.nn.relu(tf.matmul(tf.expand_dims(all_outputs[n][i], 0), extra_hidden_weights) + extra_hidden_biases)
nneth = tf.matmul(hidden, out_weights) + out_biases
nneth = tf.reshape(nneth, [1]) #single output for the neural net
nneth_array = tf.concat([nneth_array, nneth], 0)
return i + 1, nneth_array, n
def body(n, nneth_array_total):
nneth_array = tf.Variable([])
i = tf.constant(0) #iterator over 1D tensors
i, nneth_array, n = tf.while_loop(inner_cond, inner_body, loop_vars=[i, nneth_array, n], shape_invariants=[i.get_shape(), tf.TensorShape([None]), n.get_shape()])
padding = tf.zeros([tf.reduce_max(num_rnn_grid_times) - num_rnn_grid_times[tf.floordiv(n,n_mc_samples)]],dtype=tf.float32)
nneth_array = tf.concat([nneth_array, padding],0) #add extra zeros so that all nneth_arrays have same shape
nneth_array_total= tf.concat([nneth_array_total, nneth_array], 0)
return n+1, nneth_array_total
n, nneth_array_total = tf.while_loop(cond, body, loop_vars=[n, nneth_array_total],
shape_invariants=[n.get_shape(), tf.TensorShape([None])])
nneth_array_total = tf.reshape(nneth_array_total, [batch_size*n_mc_samples, tf.reduce_max(num_rnn_grid_times)])
preds = nneth_array_total
return preds
if __name__ == '__main__':
pred = code()
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
sess.run(init)
print(sess.run([pred]))
。
这是隔离的代码,上面提供了玩具数据。如果你有一个工作的张量流环境
,这应该可以运行而没有问题import { App } from ../index
代码有效,但速度很慢。这是一个管道的一部分,需要大约1.25秒迭代一个病人,并且看起来大量的运行时间是由于上面的代码。这意味着我的数据集的一个纪元大约需要12个小时,与类似的方法相比有点太多了。
我已经google了一下,找到了将函数应用于多维张量的方法,但没有考虑填充的方法。任何见解?
答案 0 :(得分:1)
即使使用零向量也可以输入整个输入,从而获得最快的处理时间。但正如你所说,由于网络中的偏见,这将返回非零输出。由于当输入向量需要为零时你想要输出为零,一个简单的技巧就是应用一个掩码,如果输入向量为零,它将使预测为零。
当输入向量非零但在0时返回1的掩码可以通过一个简单的逻辑获得:
'インスタントグラム'
然后将预测与掩码相乘。