I have a tensorflow checkpoint that I'm able to load after redefining the graph corresponding to it with the regular routine tf.train.Saver() and saver.restore(session, 'my_checkpoint.ckpt').
However, now, I would like to modify the first layer of the network to accept an input of shape say [200, 200, 1] instead of [200, 200, 10].
To this end, I would like to modify the shape of the tensor corresponding to the first layer from [3, 3, 10, 32] (3x3 kernel, 10 input channels, 32 output channels) to [3, 3, 1, 32] by summing across the 3rd dimension.
How could I do that?
I found a way to do it, but in a not so straightforward way. Given a checkpoint, we can convert it to a serialized numpy array (or any other format that we might find suitable to save a dictionary of numpy arrays) as follow:
checkpoint = {}
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver.restore(sess, 'my_checkpoint.ckpt')
for x in tf.global_variables():
checkpoint[x.name] = x.eval()
np.save('checkpoint.npy', checkpoint)
There might be some exceptions to handle but let's keep the code simple.
Then, we can do whichever operations we like on the numpy arrays:
checkpoint = np.load('checkpoint.npy')
checkpoint = ...
np.save('checkpoint.npy', checkpoint)
Finally, we can load the weights manually as follow after having built the graph:
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
checkpoint = np.load('checkpoint.npy').item()
for key, data in checkpoint.iteritems():
var_scope = ... # to be extracted from key
var_name = ... #
with tf.variable_scope(var_scope, reuse=True):
var = tf.get_variable(var_name)
sess.run(var.assign(data))
If there is a more straightforward approach, I'm all ears!