How to add if condition in a TensorFlow graph?

ghz 1years ago ⋅ 3115 views

Question

Let's say I have following code:

x = tf.placeholder("float32", shape=[None, ins_size**2*3], name = "x_input")
condition = tf.placeholder("int32", shape=[1, 1], name = "condition")
W = tf.Variable(tf.zeros([ins_size**2*3,label_option]), name = "weights")
b = tf.Variable(tf.zeros([label_option]), name = "bias")

if condition > 0:
    y = tf.nn.softmax(tf.matmul(x, W) + b)
else:
    y = tf.nn.softmax(tf.matmul(x, W) - b)  

Would the if statement work in the calculation (I do not think so)? If not, how can I add an if statement into the TensorFlow calculation graph?


Answer

You're correct that the if statement doesn't work here, because the condition is evaluated at graph construction time, whereas presumably you want the condition to depend on the value fed to the placeholder at runtime. (In fact, it will always take the first branch, because condition > 0 evaluates to a Tensor, which is "truthy" in Python.)

To support conditional control flow, TensorFlow provides the tf.cond() operator, which evaluates one of two branches, depending on a boolean condition. To show you how to use it, I'll rewrite your program so that condition is a scalar tf.int32 value for simplicity:

x = tf.placeholder(tf.float32, shape=[None, ins_size**2*3], name="x_input")
condition = tf.placeholder(tf.int32, shape=[], name="condition")
W = tf.Variable(tf.zeros([ins_size**2 * 3, label_option]), name="weights")
b = tf.Variable(tf.zeros([label_option]), name="bias")

y = tf.cond(condition > 0, lambda: tf.matmul(x, W) + b, lambda: tf.matmul(x, W) - b)