Myujth

I need to get the gradient of weights and biases with tf.gradients():

 x = tf.placeholder(tf.float32, [batch_size, x_train.shape[1]])
 y = tf.placeholder(tf.float32, [batch_size, y_train.shape[1]])
 y_ = tf.placeholder(tf.float32, [batch_size, y_train.shape[1]])

 Wx=tf.Variable(tf.random_normal(stddev=0.1,shape=[x_train.shape[1],n_hidden]))
 Wy=tf.Variable(tf.random_normal(stddev=0.1,shape=[y_train.shape[1],n_hidden]))
 b=tf.Variable(tf.constant(0.1,shape=[n_hidden]))

 hidden_joint=tf.nn.relu((tf.matmul(x,Wx)+tf.matmul(y,Wy))+b)
 hidden_marg=tf.nn.relu(tf.matmul(x,Wx)+tf.matmul(y_,Wy)+b)

 Wout=tf.Variable(tf.random_normal(stddev=0.1,shape=[n_hidden, 1]))
 bout=tf.Variable(tf.constant(0.1,shape=[1]))

 out_joint=tf.matmul(hidden_joint,Wout)+bout
 out_marg=tf.matmul(hidden_marg,Wout)+bout

 optimizer = tf.train.AdamOptimizer(0.005)


 t = out_joint
 et = tf.exp(out_marg)

 ex_delta_t = tf.reduce_mean(tf.gradients(t, tf.trainable_variables()))
 ex_delta_et = tf.reduce_mean(tf.gradients(et, tf.trainable_variables()))

But I always get the following error:

 File "/home/ferdi/Documents/mine/mine.py", line 77, in get_mi_batched
 ex_delta_t = tf.reduce_mean(tf.gradients(t, tf.trainable_variables()))
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/util/deprecation.py", line 488, in new_func
 return func(*args, **kwargs)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/ops/math_ops.py", line 1490, in reduce_mean
 reduction_indices),
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/ops/math_ops.py", line 1272, in _ReductionDims
 return range(0, array_ops.rank(x))
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/ops/array_ops.py", line 368, in rank
 return rank_internal(input, name, optimize=True)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/ops/array_ops.py", line 388, in rank_internal
 input_tensor = ops.convert_to_tensor(input)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1048, in convert_to_tensor
 as_ref=False)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1144, in internal_convert_to_tensor
 ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/ops/array_ops.py", line 971, in _autopacking_conversion_function
 return _autopacking_helper(v, dtype, name or "packed")
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/ops/array_ops.py", line 923, in _autopacking_helper
 return gen_array_ops.pack(elems_as_tensors, name=scope)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/ops/gen_array_ops.py", line 4689, in pack
 "Pack", values=values, axis=axis, name=name)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 787, in _apply_op_helper
 op_def=op_def)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/util/deprecation.py", line 488, in new_func
 return func(*args, **kwargs)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 3272, in create_op
 op_def=op_def)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1790, in __init__
 control_input_ops)
 File "/home/ferdi/anaconda3/envs/ml_all/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1629, in _create_c_op
 raise ValueError(str(e))
ValueError: Shapes must be equal rank, but are 2 and 1
 From merging shape 3 with other shapes. for 'Rank/packed' (op: 'Pack') with input shapes: [512,20], [10,20], [20], [20,1], [1].

If I reshape or do similar things, there are other error occurring. I know that there are many similar question but I still couldn't figure it out. What am I doing wrong?

The solution:

ex_delta_t = tf.reduce_mean( tf.concat([tf.reshape(g, [-1]) for g in tf.gradients(t, tf.trainable_variables())], axis=0))
ex_delta_et = tf.reduce_mean( tf.concat([tf.reshape(g, [-1]) for g in tf.gradients(et, tf.trainable_variables())], axis=0))

Or the same code unfolded:

grads_t_0 = tf.gradients(t, tf.trainable_variables())
grads_et_0 = tf.gradients(t, tf.trainable_variables())

grads_t = 
grads_et = 
for gt,get in zip(grads_t_0, grads_et_0):
 grads_t.append(tf.reshape(gt, [-1]))
 grads_et.append(tf.reshape(get, [-1]))

grads_t_flatten = tf.concat(grads_t, axis=0)
grads_et_flatten = tf.concat(grads_et, axis=0)

ex_delta_t = tf.reduce_mean(grads_t_flatten)
ex_delta_et = tf.reduce_mean(grads_et_flatten)

Explanation:

You may get this error message, because your gradient functions

tf.gradients(t, tf.trainable_variables())
tf.gradients(et, tf.trainable_variables()

returns multiply shaped tensors.
As a result your tf.reduce_mean() operation complains, that it can't work with this multiply shaped tensors.

As a possibility to workaround this, you should first flatten than concatenate the gradient list, then pass it to the reduce_mean function.

Let's see a simple example code to simulate the error and its solution!

#You dummy gradients as the output of tf.gradients()
grad_wx = tf.constant(0.1, shape=[512, 20])
grad_wy = tf.constant(0.2, shape=[10, 20])
grad_b = tf.constant(0.3, shape=[20])
grad_wout = tf.constant(0.4, shape=[20, 1])
grad_bout = tf.constant(0.5, shape=[1])

grads_0 = [grad_wx, grad_wy, grad_b, grad_wout, grad_bout]

sess = tf.Session()

result = tf.reduce_mean(grads_0)
print(sess.run(result)

Out(error):

ValueError: Shapes must be equal rank, but are 2 and 1
 From merging shape 3 with other shapes. for 'Rank/packed' (op: 'Pack') with input shapes: [512,20], [10,20], [20], [20,1], [1].

Solution:

result = tf.reduce_mean( tf.concat([tf.reshape(g, [-1]) for g in grads_0], axis=0))
print(sess.run(result))

Out(fixed):

0.102899365