Debug

Partial Runs

• avoid side-effects from fetching tensor
• can re-create and run subset of graph
• reuse results which have been generated previously

a = tf.placeholder(tf.float32, shape=[])
b = tf.placeholder(tf.float32, shape=[])
c = tf.placeholder(tf.float32, shape=[])

r1 = tf.add(a, b)
r2 = tf.multiply(r1, c)
r3 = tf.square(r2)

with tf.Session() as sess:
    h = sess.partial_run_setup([r1, r2, r3], [a, b, c])
    r1_eval = sess.partial_run(h, r1, feed_dict={a:1, b:2})
    r2_eval = sess.partial_run(h, r2, feed_dict={c:2})
    r3_eval = sess.partial_run(h, [r3])

tf.Print()

node1 = tf.Print(node1, [node1], "message: ")
out = tf.Print(out, [tf.argmax(out, 1)], "Predicted value = ", summarize=20. first_n=10)

tf.Assert()

r1 = tf.add(a, b)
r2 = tf.multiply(r1, c)
final = tf.square(r2)

#assert_op = tf.assert_positive(r2, [r2])
#assert_op = tf.assert_negative(r2, [r2])
#assert_op = tf.assert_equal(r2, r3)
#assert_op = tf.assert_type(r2, tf.float32)

assert_op = tf.Assert(tf.reduce_all(r2>0), [r2], name="assert_r2_positive")

with tf.control_dependencies([assert_op]):
    final = tf.identity(final)

with tf.Session() as sess:
    final_eval = sess.run(final, {a: 1, b:-10, c:10})
    print("Final result: ", final_eval)

#make a list of assertions ==================================
assert_op = tf.assert_positive(r2, [r2])
tf.add_to_collection('Asserts', assert_op);

assert_op = tf.assert_equal(r2, r3)
tf.add_to_collection('Asserts', assert_op);

assert_op = tf.assert_type(r2, tf.float32)
tf.add_to_collection('Asserts', assert_op);

with tf.Session() as sess:
    assert_op = tf.group(tf.get_collection('Asserts'))
    final_eval, _ = sess.run([final, assert_op], {a: 1, b:10, c:10})

python debugger

import ipdb; ipdb.set_trace() #add a break point

if (np.argmax(y_batch, axis=1)[:1] == [4]).all(): #conditional break point
    import ipdb; ipdb.set_trace()

TensorFlow tfdbg