修改TensorFlow中的传入渐变

Question

Consider a neural network with 2 fully connected layers "l1_dense" and "l2_dense" in this order and some loss function. During backpropagation I want to compute the gradient of "l2_dense" wrt the loss function, do some manipulation of the gradient and use this manipulated gradient as the incoming gradient (in the chain rule) for the "l1_dense" layer. I know that I can use tf.train.Optimizer.compute_gradients() in order to compute the gradient wrt "l2_dense" and be able to manipulate it. What I do not know how to do is to feed the modified gradient for the computation of the "l1_dense" gradient.

As a very simplistic example, let's say that the way I want to manipulate the "l2_dense" gradient is to divide it by some number k . I know that all this is equivalent to just diving the loss by k , I am just giving this simple example for the purposes of the question. The code will be sth like:

import tensorflow as tf

i = tf.placeholder(tf.float32, shape=[None, 3])
y = tf.placeholder(tf.float32, shape=[None, 1])

x = tf.layers.dense(i, 4, tf.nn.relu, name="l1_dense")
x = tf.layers.dense(x, 1, tf.nn.relu, name="l2_dense")

loss = tf.losses.mean_squared_error(y, x)

opt = tf.train.AdamOptimizer()

gvars = tf.get_default_graph().get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
l1_dense_k = [v for v in gvars if v.name == "l1_dense/kernel:0"][0]
l1_dense_b = [v for v in gvars if v.name == "l1_dense/bias:0"][0]
l2_dense_k = [v for v in gvars if v.name == "l2_dense/kernel:0"][0]
l2_dense_b = [v for v in gvars if v.name == "l2_dense/bias:0"][0]

gvs = opt.compute_gradients(loss, var_list=[l2_dense_k, l2_dense_b])
# Manipulate gradients
gvs = [(g/10, v) for g,v in gvs]

# Compute gradients w.r.t. l1_dense_k and l1_dense_b using gvs ???

To be completely clear, my setting is far more complicated than that and I cannot automatically do the manipulation I need by changing the loss function. Furthermore, I need a solution in which the gradient wrt each variable is computed only once.

Answer 1

The answer is actually very simple - you need to use tf.gradients() . In case anyone gets stuck on this too, here is the solution:

import tensorflow as tf

i = tf.placeholder(tf.float32, shape=[None, 3])
y = tf.placeholder(tf.float32, shape=[None, 1])

x1 = tf.layers.dense(i, 4, tf.nn.relu, name="l1_dense")
x2 = tf.layers.dense(x1, 1, tf.nn.relu, name="l2_dense")

loss = tf.losses.mean_squared_error(y, x2)

gvars = tf.get_default_graph().get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
l1_k = [v for v in gvars if v.name == "l1_dense/kernel:0"][0]
l1_b = [v for v in gvars if v.name == "l1_dense/bias:0"][0]
l2_k = [v for v in gvars if v.name == "l2_dense/kernel:0"][0]
l2_b = [v for v in gvars if v.name == "l2_dense/bias:0"][0]

grads = tf.gradients(loss, [x1, l2_k, l2_b])
x1_grad, l2_k_grad, l2_b_grad = grads

# Manipulate the gradient
x1_grad = x1_grad / 10.0

# Backpropagate the gradient
grads = tf.gradients(x1, [l1_dense_k, l1_dense_b], x1_grad)
l1_k_grad, l1_b_grad = grads