Implementing Adaptive Loss Weights via Callback - tf.compat.v1

[PhilippBrendel]

Hi everyone,

I've read in other Issues (e.g. #215 and #908) that adaptive Loss-Weights are not high-priority for DeepXDE, but I still want to test some approaches for that, as I see quite some potential for my current use-case. However, implementing this via a Callback like the following does not really work for me so far (cf. Error-message below).

class LossWeightCallback(dde.callbacks.Callback):
    def __init__(self, model):
        super().__init__()
        self.model = model

    def on_epoch_end(self):
        ....
        self.model.compile("adam", lr=1e-3, decay=None,
                                       loss_weights=[1,1,1,1])

Click for Error Message

```log Training model... Step Train loss Test loss Test metric 0 [9.15e+04, 2.41e+00, 1.07e-07, 2.08e-06] [3.41e+04, 9.30e-01, 0.00e+00, 0.00e+00] [] Compiling model... 'compile' took 9.111588 s Traceback (most recent call last): File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\client\session.py", line 1375, in _do_call return fn(*args) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\client\session.py", line 1359, in _run_fn return self._call_tf_sessionrun(options, feed_dict, fetch_list, File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\client\session.py", line 1451, in _call_tf_sessionrun return tf_session.TF_SessionRun_wrapper(self._session, options, feed_dict, tensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value beta1_power_1 [[{{node beta1_power_1/read}}]] During handling of the above exception, another exception occurred: Traceback (most recent call last): File "maxwell_quasistatic.py", line 1689, in pinn.train() File "maxwell_quasistatic.py", line 1282, in train loss_hist, train_state = self.model.train( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\utils\internal.py", line 22, in wrapper result = f(*args, **kwargs) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\model.py", line 589, in train self._train_sgd(iterations, display_every) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\model.py", line 606, in _train_sgd self._train_step( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\model.py", line 505, in _train_step self.sess.run(self.train_step, feed_dict=feed_dict) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\client\session.py", line 967, in run result = self._run(None, fetches, feed_dict, options_ptr, File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\client\session.py", line 1190, in _run results = self._do_run(handle, final_targets, final_fetches, File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\client\session.py", line 1368, in _do_run return self._do_call(_run_fn, feeds, fetches, targets, options, File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\client\session.py", line 1394, in _do_call raise type(e)(node_def, op, message) tensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value beta1_power_1 [[node beta1_power_1/read (defined at D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\optimizers\tensorflow_compat_v1\optimizers.py:58) ]] Original stack trace for 'beta1_power_1/read': File "maxwell_quasistatic.py", line 1689, in pinn.train() File "maxwell_quasistatic.py", line 1282, in train loss_hist, train_state = self.model.train( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\utils\internal.py", line 22, in wrapper result = f(*args, **kwargs) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\model.py", line 589, in train self._train_sgd(iterations, display_every) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\model.py", line 618, in _train_sgd self.callbacks.on_epoch_end() File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\callbacks.py", line 78, in on_epoch_end callback.on_epoch_end() File "D:\pinns\src\deepxde\utils\callbacks.py", line 312, in on_epoch_end self.pinn_obj.model.compile("adam", lr=1e-3, decay=None, File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\utils\internal.py", line 22, in wrapper result = f(*args, **kwargs) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\model.py", line 124, in compile self._compile_tensorflow_compat_v1(lr, loss_fn, decay, loss_weights) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\model.py", line 177, in _compile_tensorflow_compat_v1 self.train_step = optimizers.get( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\deepxde\optimizers\tensorflow_compat_v1\optimizers.py", line 58, in get train_op = optim.minimize(loss, global_step=global_step) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\training\optimizer.py", line 412, in minimize return self.apply_gradients(grads_and_vars, global_step=global_step, File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\training\optimizer.py", line 597, in apply_gradients self._create_slots(var_list) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\training\adam.py", line 131, in _create_slots self._create_non_slot_variable( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\training\optimizer.py", line 830, in _create_non_slot_variable v = variable_scope.variable( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\ops\variables.py", line 260, in __call__ return cls._variable_v1_call(*args, **kwargs) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\ops\variables.py", line 206, in _variable_v1_call return previous_getter( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\ops\variables.py", line 199, in previous_getter = lambda **kwargs: default_variable_creator(None, **kwargs) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\ops\variable_scope.py", line 2620, in default_variable_creator return variables.RefVariable( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\ops\variables.py", line 264, in __call__ return super(VariableMetaclass, cls).__call__(*args, **kwargs) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\ops\variables.py", line 1656, in __init__ self._init_from_args( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\ops\variables.py", line 1861, in _init_from_args self._snapshot = array_ops.identity(self._variable, name="read") File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\util\dispatch.py", line 201, in wrapper return target(*args, **kwargs) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\ops\array_ops.py", line 287, in identity ret = gen_array_ops.identity(input, name=name) File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\ops\gen_array_ops.py", line 3941, in identity _, _, _op, _outputs = _op_def_library._apply_op_helper( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\framework\op_def_library.py", line 748, in _apply_op_helper op = g._create_op_internal(op_type_name, inputs, dtypes=None, File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\framework\ops.py", line 3528, in _create_op_internal ret = Operation( File "D:\Anaconda3\envs\deepxde_tf24\lib\site-packages\tensorflow\python\framework\ops.py", line 1990, in __init__ self._traceback = tf_stack.extract_stack() ```

I'm not an Expert on Core-Tensorflow (especially not TF1), so if anyone could give me an advice on what I'm doing wrong or how I can fix this, I'd really appreciate it!

Cheers, Philipp

Ps: In #331 are some more comments, but I don't think they apply to my problem, as I'm not interested in using gradients for the weights initially.

[PhilippBrendel]

• found the solution via this stackoverflow

        ...
        self.model.compile("adam", lr=1e-3, decay=None,
                                    loss_weights=[1, 1, 1, 1])
        self.model.sess.run(tf.global_variables_initializer())

Sorry for posting this a little early (in hindsight), but maybe this solution is helpful for others as well....

[sai-karthikeya-vemuri]

Hello Philipp,@PhilippBrendel and @lululxvi

Can you provide implementation of adaptive loss weights class?

I am also trying to apply the same thing for 2d wave equation. Since a FNN and MsFFN seem to fail, i would like to try adaptive weights.

Thanks in advance

[pescap]

Hi, I'll use this issue to centralize information regarding adaptive loss weighting:

• 215

• 331

• 787

• 908

• 910

• 1340

Other references of interest:

• Some exhaustive documentation in Modulus.
• Implementation of ReLoBRaLo
• Article on ReLoBRaLo
• TensorDiffEq model has some self-adaptive implementation.
• Reliable extrapolation of deep neural operators informed by physics or sparse observations:

Note: "lambda slighly improves the accuracy" (Fig 13). According to Fig. 13 it does not seem to be so efficient.

As stated by @lululxvi in #215, "based on my experience, fixed and adaptive weights have similar effects. As you can see in the papers you mentioned, the adaptive weights quickly converge to a fixed number, and thus fixed weights are basically sufficient. Also, it is recommended to use hard constraints for BC/IC, see FAQ".

I would definitely define adaptive weighting as callbacks. I'll try to figure out a structure for implementing a simple weighting scheme. I think that adaptive weighting can be useful for more involved loss terms (with e.g. 4-5 terms).

[pescap]

Hi! I have been exploring this issue.

For all these adaptive weighting techniques, we want to be able to update loss_weights during training with a callback, without having to compile the model again.

So loss_weights shall be initialized at first epoch.

The callback would adapt the weights as a on_epoch_end.

For example, in tensorflow.compat.v1 the multiplication is performed here:

https://github.com/lululxvi/deepxde/blob/683682c9baf876d58048160733d9897ad3090af3/deepxde/model.py#L181-L182

So, to begin with, we would comment out these two lines, and put them somewhere else at the beginning of train function.

Then, with a few changes, we could define and use self.loss_weights so that we can update the weights during training.

Do you agree @lululxvi ?

Also, do you prefer if I move this discussion to a new issue? Or could you please re-open this issue?

[lululxvi]

@pescap Yes, that sounds good.

[pescap]

• I'll start with a simple loss balancing algorithm (softadapt).
• Softadapt measures the ratio of the loss value at each iteration to its value at the previous iteration, making it simple to implement.
• This will allow me to set the infrastructure for loss balancing implementations.

As a first step, I am trying to facilitate the update of loss_weights during training, see #1511.

[jdellag]

Really looking forward to this implementation!

[pescap]

Really looking forward to this implementation!

Working on this in #1586

[haison19952013]

• I have experience with this type of work from my project, where I implemented my own PINN library in TensorFlow 2.0.
• If you're interested in implementing this feature in TensorFlow 2.0, I'm happy to lend a hand.

[pescap]

• I have experience with this type of work from my project, where I implemented my own PINN library in TensorFlow 2.0.
• If you're interested in implementing this feature in TensorFlow 2.0, I'm happy to lend a hand.

Hi! Thank you for proposing! We could start with the 'tensorflow' backend. How do you suggest we proceed? We could discuss via the Deepxde Slack workspace.

[haison19952013]

• I have experience with this type of work from my project, where I implemented my own PINN library in TensorFlow 2.0.
• If you're interested in implementing this feature in TensorFlow 2.0, I'm happy to lend a hand.

Hi! Thank you for proposing! We could start with the 'tensorflow' backend. How do you suggest we proceed? We could discuss via the Deepxde Slack workspace.

Yes, could you give me the Slack ID or URL?

[pescap]

• I have experience with this type of work from my project, where I implemented my own PINN library in TensorFlow 2.0.
• If you're interested in implementing this feature in TensorFlow 2.0, I'm happy to lend a hand.

Hi! Thank you for proposing! We could start with the 'tensorflow' backend. How do you suggest we proceed? We could discuss via the Deepxde Slack workspace.

Yes, could you give me the Slack ID or URL?

Can you please send an email to @lululxvi asking him to add you?