Removes unnecessary ViT-GP hyper-parameters.

PiperOrigin-RevId: 388484029

Author: jereliu

baselines/jft/experiments/jft300m_vit_base16_sngp.py

@@ -40,17 +40,17 @@ def get_config():
 
   pp_common = '|value_range(-1, 1)'
   pp_common += f'|onehot({config.num_classes})'
-  # To use ancestor "smearing", use this line instead:
-  # pp_common += f'|onehot({config.num_classes}, key="labels_extended", key_result="labels")  # pylint: disable=line-too-long
+  # To use ancestor 'smearing', use this line instead:
+  # pp_common += f'|onehot({config.num_classes}, key='labels_extended', key_result='labels')  # pylint: disable=line-too-long
   pp_common += '|keep("image", "labels")'
   config.pp_train = 'decode_jpeg_and_inception_crop(224)|flip_lr' + pp_common
   config.pp_eval = 'decode|resize_small(256)|central_crop(224)' + pp_common
   config.shuffle_buffer_size = 250_000  # Per host, so small-ish is ok.
 
   config.log_training_steps = 50
   config.log_eval_steps = 1000
-  # NOTE: eval is very fast O(seconds) so it's fine to run it often.
-  config.checkpoint_steps = 1000
+  # NOTE: For pretraining, save infrequently to prevent crowding diskspace.
+  config.checkpoint_steps = 517790
 
   # Model section
   config.model = ml_collections.ConfigDict()
@@ -66,11 +66,11 @@ def get_config():
   config.model.classifier = 'token'  # Or 'gap'
   config.model.representation_size = 768
 
-  # GP layer parameters.
+  # Gaussian process layer parameters.
   config.gp_layer = ml_collections.ConfigDict()
-  config.gp_layer.normalize_input = True
-  config.gp_layer.random_feature_scale = 1.  # 1. or None
-  config.gp_layer.random_feature_stddev = 0.025  # 1. or 0.025
+  # Use momentum for pre-training to prevent numeric error when inverting a
+  # precision matrix accumulated over 300M data.
+  config.gp_layer.covmat_momentum = .999
 
   # Optimizer section
   config.optim_name = 'Adam'
@@ -82,7 +82,8 @@ def get_config():
 
   # TODO(lbeyer): make a mini-language like preprocessings.
   config.lr = ml_collections.ConfigDict()
-  config.lr.base = 8e-4  # LR has to be lower for larger models!
+  # LR has to be lower for GP layer and on larger models.
+  config.lr.base = 4e-4
   config.lr.warmup_steps = 10_000
   config.lr.decay_type = 'linear'
   config.lr.linear_end = 1e-5
@@ -96,9 +97,4 @@ def get_config():
 
 
 def get_sweep(hyper):
-  # lr_grid = [3e-4, 4e-4, 5e-4, 6e-4]
-  # stddev_grid = [0.01, 0.02, 0.03, 0.04, 0.05]
-  return hyper.product([
-      # hyper.sweep('config.lr.base', lr_grid),
-      # hyper.sweep('config.gp_layer.random_feature_stddev', stddev_grid)
-  ])
+  return hyper.product([])

baselines/jft/sngp.py

@@ -65,7 +65,7 @@ def accumulate_gradient_with_states(
     accum_steps):
   """Improved version of `u.accumulate_gradient()` that allows for states."""
   # This function handles the `loss_and_grad_fn` function which takes a state
-  # arguement and returns ((losses, states), grads).
+  # argument and returns ((losses, states), grads).
   if accum_steps and accum_steps > 1:
     assert images.shape[0] % accum_steps == 0, (
         f'Bad accum_steps {accum_steps} for batch size {images.shape[0]}')
@@ -95,27 +95,16 @@ def acc_grad_and_loss(i, l_s_g):
 
 
 def get_gp_kwargs(gp_config):
-  """Extract keyword arguement parameters for the Gaussian process layer."""
-  normalize_input = gp_config.get('normalize_input', True)
-  kernel_stddev = gp_config.get('random_feature_stddev', 1.)
-  feature_scale = gp_config.get('random_feature_scale', -1.)
+  """Extract keyword argument parameters for the Gaussian process layer."""
   covmat_momentum = gp_config.get('covmat_momentum', 0.999)
 
-  logging.info('gp_config.normalize_input = %s', normalize_input)
-  logging.info('gp_config.random_feature_stddev = %s', kernel_stddev)
-  logging.info('gp_config.random_feature_scale = %s', feature_scale)
+  # Extracts model parameter.
   logging.info('gp_config.covmat_momentum = %s', covmat_momentum)
-
-  feature_scale = None if feature_scale < 0. else feature_scale
-  kernel_init = nn.initializers.normal(stddev=kernel_stddev)
-  hidden_kwargs = dict(feature_scale=feature_scale, kernel_init=kernel_init)
+  covmat_momentum = None if covmat_momentum < 0. else covmat_momentum
   covmat_kwargs = dict(momentum=covmat_momentum)
 
-  # Assemble into kwargs dictionary.
-  gp_layer_kwargs = dict(
-      normalize_input=normalize_input,
-      hidden_kwargs=hidden_kwargs,
-      covmat_kwargs=covmat_kwargs)
+  # Assembles into kwargs dictionary.
+  gp_layer_kwargs = dict(covmat_kwargs=covmat_kwargs)
 
   return gp_layer_kwargs
 
@@ -327,7 +316,7 @@ def representation_fn(params, images, labels, mask, states):
   @partial(jax.pmap, axis_name='batch', donate_argnums=(0,))
   def update_fn(opt, states, lr, images, labels, rng):
     """Update step."""
-
+    # TODO(jereliu): Expand to allow precision matrix resetting.
     measurements = {}
 
     if config.get('mixup') and config.mixup.p:
@@ -413,17 +402,17 @@ def decay_fn(v, wd):
                                              checkpoint['states'],
                                              checkpoint['extra'])
   elif config.get('model_init'):
-    write_note(f'Initialize model from {config.model_init}...')
-    raise ValueError(
-        'Load from `config.model_init` checkpoint is currently not supported.')
+    # Load trainable parameters from the checkpoint.
+    # This does not cause issue for SNGP since all non-trainable parameters
+    # (random feature, precision matrix, etc) are last-layer parameters that
+    # should be re-trained during fine-tuning.
+    write_note(f'Initialize trainable parameters from {config.model_init}...')
     # TODO(dusenberrymw): Replace and test load function.
-    # pylint:disable=unreachable
     loaded = resformer.load(params_cpu, config.model_init, config.get('model'))
     opt_cpu = opt_cpu.replace(target=loaded)
     if jax.host_id() == 0:
       logging.info('Restored parameter overview:')
       parameter_overview.log_parameter_overview(loaded)
-    # pylint:enable=unreachable
 
   write_note('Kicking off misc stuff...')
   first_step = int(opt_cpu.state.step)  # Might be a DeviceArray type.
@@ -472,6 +461,7 @@ def decay_fn(v, wd):
     mw.step_start(step)
 
     with jax.profiler.TraceContext('train_step', step_num=step, _r=1):
+      # TODO(jereliu): Expand to allow precision matrix resetting.
       (opt_repl, states_repl, loss_value, rngs_loop,
        extra_measurements) = update_fn(
            opt_repl,
@@ -495,8 +485,9 @@ def decay_fn(v, wd):
       # alive while they'll be updated in a future step, creating hard to debug
       # memory errors (see b/160593526). Also, takes device 0's params only.
       # We will also do the same for untrainable parameters (`states`). This is
-      # ok since both `random features` and `predictive covariance` are frozen
-      # or task-specific parameters that are not important for pre-training.
+      # ok since `random features` are frozen throughout pre-training, and
+      # `predictive covariance` are irrelevant for downstream finetuning and
+      # will be discarded anyway.
       opt_cpu = jax.tree_map(lambda x: np.array(x[0]), opt_repl)
       states_cpu = jax.tree_map(lambda x: np.array(x[0]), states_repl)
 

baselines/jft/sngp_test.py

@@ -116,10 +116,10 @@ def get_config(classifier, representation_size):
 class SNGPTest(parameterized.TestCase, tf.test.TestCase):
 
   @parameterized.parameters(
-      ('token', 2, 1111.4404296875, 16258.519965277777, 0.16999999806284904),
-      ('token', None, 13992.8515625, 3621.3713107638887, 0.20999999344348907),
-      ('gap', 2, 8779.61328125, 3998.798285590278, 0.12999999895691872),
-      ('gap', None, 11279.3515625, 3212.2536892361113, 0.2199999988079071),
+      ('token', 2, 916.2851, 1954.3369140625, 0.16999999806284904),
+      ('token', None, 290.0307, 915.987548828125, 0.20999999344348907),
+      ('gap', 2, 695.6460, 600.8613823784722, 0.12999999895691872),
+      ('gap', None, 192.9434, 341.7078450520833, 0.2199999988079071),
   )
   def test_sngp_script(self, classifier, representation_size,
                        correct_train_loss, correct_val_loss,