lucfra / far-ho Goto Github PK

I have emailed Luca Franceschi about some issues with this library, and he asked me to share it.
I've been working on a MLP and wanted to optimize the following parameters, but found some problems:

• Keep probability of a dropout layer: Luca explained to me that this is not posible since it has some non differentiable points.
• Regularization beta: We are using tf.nn.l2_loss, but can't optimize the beta.
• AdamOptimizer: When we tried to use far.AdamOptimizer() for the inner optimizer the code started crashing. Apparently there are some undefined variables: _beta1_power and _beta2_power. I think this is an error in the library.
  Until now we have been able to optimize only the learning rate. It would be great if there could be a list of the things you can and can't do with this library.

Best regards,
Nicolás Zorzano.

This is a nice package for HPO, but it seems that it is built only on TF.
Just wondering is there any version that supports pytorch?
I think that will make the package more popular for HPO reseachers!

I really don't get https://github.com/lucfra/FAR-HO/blob/master/far_ho/hyper_gradients.py#L351, why was it needed?

Aren't we already getting the hyper-gradient at https://github.com/lucfra/FAR-HO/blob/master/far_ho/hyper_gradients.py#L349

Also, please explain if this step is really needed as I don't see it explicitly stated in your paper, thanks.

When I run this notebook I get the following error:

First I am told that the param lambda is not even connected with the model, which doe not make any sense.

Then it fails at the assertion a t line 113 in file utils.py

Please help, and it would be really helpful, if you could also provide an example notebook with ForwardHG for MNIST (both) with and without online learning ...

Thanks in advance ...
Habib

//----------------------------------------------------------------------------------------------------------

Hyperparameter <tf.Variable 'lambda_components/0:0' shape=() dtype=float32_ref> is detached from this optimization dynamics.

AssertionError Traceback (most recent call last)
in ()
----> 1 ss, farho, cost, oo = _test(far.ForwardHG)
2
3 tf.global_variables_initializer().run()
4
5 # execution with gradient descent! (looks ol)

in _test(method)
25 optim_oo = tf.train.AdamOptimizer(.01)
26 farho = far.HyperOptimizer(hypergradient=method())
---> 27 farho.minimize(oo, optim_oo, cost, io_optim)
28 return ss, farho, cost, oo

/volume1/scratch/r0605927/backMeUpPlz/lib/python2.7/site-packages/far_ho/hyper_parameters.pyc in minimize(self, outer_objective, outer_objective_optimizer, inner_objective, inner_objective_optimizer, hyper_list, var_list, init_dynamics_dict, global_step, aggregation_fn, process_fn)
132 """
133 optim_dict = self.inner_problem(inner_objective, inner_objective_optimizer, var_list, init_dynamics_dict)
--> 134 self.outer_problem(outer_objective, optim_dict, outer_objective_optimizer, hyper_list, global_step)
135 return self.finalize(aggregation_fn=aggregation_fn, process_fn=process_fn)
136

/volume1/scratch/r0605927/backMeUpPlz/lib/python2.7/site-packages/far_ho/hyper_parameters.pyc in outer_problem(self, outer_objective, optim_dict, outer_objective_optimizer, hyper_list, global_step)
117 :return: itself
118 """
--> 119 hyper_list = self._hypergradient.compute_gradients(outer_objective, optim_dict, hyper_list=hyper_list)
120 self._h_optim_dict[outer_objective_optimizer].update(hyper_list)
121 self._global_step = global_step

/volume1/scratch/r0605927/backMeUpPlz/lib/python2.7/site-packages/far_ho/hyper_gradients.pyc in compute_gradients(self, outer_objective, optimizer_dict, hyper_list)
343 # d_E_T = dot(vectorize_all(d_oo_d_state), vectorize_all(zs))
344 d_E_T = [dot(d_oo_d_s, z) for d_oo_d_s, z in zip(d_oo_d_state, zs)
--> 345 if d_oo_d_s is not None and z is not None]
346 hg = maybe_add(tf.reduce_sum(d_E_T), d_oo_d_hyp) # this is right... the error is not here!
347 # hg = maybe_add(d_E_T, d_oo_d_hyp)

/volume1/scratch/r0605927/backMeUpPlz/lib/python2.7/site-packages/far_ho/utils.pyc in dot(a, b, name)
111 Dot product between vectors a and b with optional name
112 """
--> 113 assert a.shape.ndims == 1, '{} must be a vector'.format(a)
114 assert b.shape.ndims == 1, '{} must be a vector'.format(b)
115 with tf.name_scope(name, 'Dot', [a, b]):

AssertionError: Tensor("Mean_1_1/gradients/mul_5_grad/Reshape:0", shape=(2, 3), dtype=float32) must be a vector

Hi,

Thanks for providing the code. It is well written and very easy to use. I was wondering which version of Python and Tensorflow you used for testing?

Can you please provide support for python 2.7.x ?

Hi @lucfra ,
I ran the same code with CLASSES=20 with your h5 files. However, I get an error. Here is the code:

from far_ho.examples.hyper_representation import train, mini_imagenet_model

if __name__ == '__main__':
  CLASSES = 20
  SHOTS = 1
  META_BATCH_SIZE = 4
  from experiment_manager.datasets import load
  mini_imagenet = load.meta_mini_imagenet(std_num_classes=CLASSES,
                                          std_num_examples=(SHOTS*CLASSES, 15*CLASSES), h5=True)
  res = train(mini_imagenet, 'test', mini_imagenet_model, T=1, print_every=1000, MBS=META_BATCH_SIZE, n_episodes_testing=150, patience=20)

The error is related to dataset.

Traceback (most recent call last):
  File "run_hyper.py", line 11, in <module>
    res = train(mini_imagenet, 'maml', mini_imagenet_model, T=1, print_every=500, MBS=META_BATCH_SIZE, n_episodes_testing=150, patience=20)
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/far_ho/examples/hyper_representation.py", line 245, in train
    farho.run(T[0], trfd, vfd)  # one iteration of optimization of representation variables (hyperparameters)
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/savers/records.py", line 72, in _saver_wrapped
    self._execute_save(res, *args, **kwargs)
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/savers/records.py", line 125, in _execute_save
    super()._execute_save(res, *args, **kwargs)
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/savers/records.py", line 98, in _execute_save
    _res=res)
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/savers/save_and_load.py", line 535, in save
    rss = _compute_value(pt, save_dict)
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/savers/save_and_load.py", line 516, in _compute_value
    if callable(pt[1]) else _tf_run_catch_not_initialized(pt, _partial_save_dict)
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/savers/save_and_load.py", line 492, in _maybe_call
    _out = _method()
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/far_ho/examples/hyper_representation.py", line 157, in <lambda>
    'FLAT', lambda: accs_and_errs(metasets),
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/far_ho/examples/hyper_representation.py", line 133, in accs_and_errs
    for _d in meta_dataset.generate(n_episodes_testing, batch_size=MBS, rand=0):
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/datasets/structures.py", line 265, in generate
    yield self.generate_batch(batch_size, rand=rand, *args, **kwargs)
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/datasets/structures.py", line 271, in generate_batch
    return [self.generate_datasets(rand, *args, **kwargs) for _ in range(batch_size)]
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/datasets/structures.py", line 271, in <listcomp>
    return [self.generate_datasets(rand, *args, **kwargs) for _ in range(batch_size)]
  File "/home/amir/.conda/envs/farho/lib/python3.5/site-packages/experiment_manager/datasets/load.py", line 400, in generate_datasets
    random_classes = rand.choice(list(clss.keys()), size=(num_classes,), replace=False)
  File "mtrand.pyx", line 1437, in mtrand.RandomState.choice (numpy/random/mtrand/mtrand.c:17481)
ValueError: Cannot take a larger sample than population when 'replace=False'

Could you please let me know how to fix this?

Hi,

I wrote the following code to compare the hyper-gradient computed by ReverseHG and ForwardHG methods in the same file:

### ReverseHG
farho = far.HyperOptimizer()
hypergradient = farho.hypergradient
run = farho.minimize(val_loss, oo_optim, tr_loss, io_optim)
grads_hvars = [hypergradient.hgrads_hvars(hyper_list=hll)
    for opt, hll in farho._h_optim_dict.items()]
run(T, inner_objective_feed_dicts=tr_supplier, outer_objective_feed_dicts=val_supplier, _skip_hyper_ts=True)
grads_hvars_val = ss.run(grads_hvars, _opt_fd(farho._global_step, val_supplier))
print(grads_hvars_val)


### ForwardHG
hypergradient_fwd = far.ForwardHG()
farho_fwd = far.HyperOptimizer(hypergradient=hypergradient_fwd)
run_fwd = farho_fwd.minimize(val_loss, oo_optim, tr_loss, io_optim)
grads_hvars_fwd = [hypergradient_fwd.hgrads_hvars(hyper_list=hll)
    for opt, hll in farho_fwd._h_optim_dict.items()]
run_fwd(T, inner_objective_feed_dicts=tr_supplier, outer_objective_feed_dicts=val_supplier, _skip_hyper_ts=True)
grads_hvars_fwd_val = ss.run(grads_hvars_fwd, _opt_fd(farho_fwd._global_step, val_supplier))
print(grads_hvars_fwd_val)

They receive identical inputs and compute the hyper-gradient for the same hyper-variable (_skip_hyper_ts=True so the hyper-parameter remains unchanged) but for some reason their output is quiet different. If noticed that if I run them in separate files (with a fix random seed) or run ForwardHG block before ReverseHG block their outputs would be similar. I can not see how Reverse and Forward hyper-gradient computation can effect each other as they don't share any variable. Could you please explain how these two methods can be run in the same file?

I have also attached the complete python code for this experiment.

cp.py.zip

Hi, I was wondering if you could share a code that can reproduce the mini imagenet results in you workshop paper. I have tried couple of different learning rates and the best one-shot test accuracy I could get was around 43%. I used T=4 as it is mentioned in the paper.

Thanks,
Haamoon

We work on a sparse logistic regression task on data set 20newsgroups and want to find the best regularization lambda. When we tried to use far.AdamOptimizer() for the inner optimizer and ReverseHG() for hyper optimizer, lambda goes to NAN. We found that:

1. The hypergradient is an uninitialized value in this situation.
2. The same setting work perfect on ForwardHG() with far.AdamOptimizer().
3. A small learning rate for far.AdamOptimizer() will work, but it is not efficient.
4. A dense version fails too, the reason may be the large feature dimensions?

Best regards,
Xiang Geng

Hi @lucfra
While I wanted to try miniImageNet for meta batch size > 1, I got out of memory errors (tried on gtx 1080 and titax X gpus). Below is the code:

from hyper_representation import train, mini_imagenet_model

if __name__ == '__main__':
  CLASSES = 5
  SHOTS = 1
  META_BATCH_SIZE = 2
  from experiment_manager.datasets import load
  mini_imagenet = load.meta_mini_imagenet(std_num_classes=CLASSES,
                                          std_num_examples=(SHOTS*CLASSES, 15*CLASSES), h5=False, load_all_images=True)
  res = train(mini_imagenet, 'maml', mini_imagenet_model, T=1, print_every=500, MBS=META_BATCH_SIZE, n_episodes_testing=150, patience=20)

Could you please let me know how you ran the miniImageNet experiments?

Thanks