TF1 code to TF2

General Discussion
#1

Hi,

I’m having trouble migrating this code to TF2. Could someone please help me?

input_image_size = args.img_resize
x = tf.placeholder(shape=input_shape, dtype=tf.float32, name="x")
y = tf.placeholder(dtype=tf.float32, name="y")
phase = tf.placeholder(tf.bool, name='phase')
global_step = tf.Variable(0, name='global_step', trainable=False)
out_seg, map_lambda1, map_lambda2 = architectures.ddunet(x,is_training)
y_out_dl = tf.round(out_seg)
x_acm = x[:, :, :, 0]
rounded_seg_acl = y_out_dl[:, :, :, 0]
dt_trans = tf.py_func(my_func, [rounded_seg_acl], tf.float32)
dt_trans.set_shape([args.batch_size, input_image_size, input_image_size])
phi_out,_, lambda1_tr, lambda2_tr = tf.map_fn(fn=active_contour_layer, elems=(x_acm, dt_trans, map_lambda1[:, :, :, 0], map_lambda2[:, :, :, 0]))
Dice = dice_soft(out_seg, y)
seg_loss = 1 - Dice
l2_loss = tf.losses.get_regularization_loss()
seg_loss += l2_loss
total_loss = seg_loss
rounded_seg = tf.round(out_seg)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
Dice_hard = dice_hard(out_seg, y)
adaptive_lr = tf.train.exponential_decay(args.lr, global_step, 100000, 0.96, staircase=False, name=None)

with tf.control_dependencies(update_ops):
    train_op = tf.train.AdamOptimizer(adaptive_lr).minimize(total_loss, global_step=global_step,
                                                            colocate_gradients_with_ops=True, name='train_op')

The problem is the last line, I know it should be changed to:

with tf.control_dependencies(update_ops):
    train_op = tf.optimizers.Adam().minimize(total_loss, var_list=???, tape=tf.GradientTape(persistent=False))

But I’m not sure what to put for var_list.
Thanks a lot for your help!

#3

Check this

#5

I have of course before posting here, and unfortunately it’s not helpful. I still get the same error.

#6

Sorry, but what error? You mention that you are not sure about the var_list. Could you please mention the error you’re facing (with reproducible code)? ( Also please ask your question in stack overflow? )

#7

So sorry. I get this error when I apply the changes suggested in that post:
ValueError: No gradients provided for any variable: ([‘x_5:0’],). Provided grads_and_vars is ((None, <tf.Variable ‘x_5:0’ shape=() dtype=float32>),).

#8

Some possible reason

Are you able to run your code with TF 1?
Also, the things you’re trying to do below don’t look like TF 2 best practices (IMO).

with tf.control_dependencies(update_ops):
    train_op = tf.train.AdamOptimizer(adaptive_lr).minimize(total_loss,
                      global_step=global_step, 
                      colocate_gradients_with_ops=True, 
                      name='train_op'
)

If possible, please share a reproducible code with some dummy data; it seems complicated to break down otherwise. Also, looking into your code, it looks like a segmentation-related task, in that case, you should try code in Keras.

#10

Thank you so much for your help!
I’m trying to run this repo in tf2: dals/main.py at master · ahatamiz/dals · GitHub , and specifically I have trouble over migrating the cost function.