def test_initialize_rescale(self):
"""Test rescaling a single layer of a model."""
input_shape = (28, 28, 1)
output_shape = (10,)
model_str = 'fully_connected'
model_cls = models.get_model(model_str)
model_hps = models.get_model_hparams(model_str)
loss_name = 'cross_entropy'
metrics_name = 'classification_metrics'
hps = copy.copy(model_hps)
hps.update({'output_shape': output_shape})
rng = jax.random.PRNGKey(0)
model = model_cls(hps, {}, loss_name, metrics_name)
initializer = initializers.get_initializer('noop')
rng, init_rng = jax.random.split(rng)
# First initialize with no rescale.
flax_module, _ = trainer.initialize(
model.flax_module_def,
initializer,
model.loss_fn,
input_shape,
output_shape,
hps,
init_rng,
metrics_logger=None)
utils.log_pytree_shape_and_statistics(flax_module.params)
# Now rescale a layer by 100.
rescale_factor = 100
hps.layer_rescale_factors = {
'/Dense_1/kernel': rescale_factor,
}
rescaled_module, _ = trainer.initialize(
model.flax_module_def,
initializer,
model.loss_fn,
input_shape,
output_shape,
hps,
init_rng,
metrics_logger=None)
# Check the right variable is rescaled
v1 = flax_module.params['Dense_1']['kernel']
v2 = rescaled_module.params['Dense_1']['kernel']
diff = np.linalg.norm(v1.reshape(-1) * rescale_factor - v2.reshape(-1))
self.assertAlmostEqual(diff, 0.0)
# Check that other variables are the same
v1 = flax_module.params['Dense_2']['kernel']
v2 = rescaled_module.params['Dense_2']['kernel']
diff = np.linalg.norm(v1.reshape(-1) - v2.reshape(-1))
self.assertAlmostEqual(diff, 0.0)
def setUp(self):
super(CheckpointTest, self).setUp()
self.test_dir = tempfile.mkdtemp()
loss_name = 'cross_entropy'
metrics_name = 'classification_metrics'
model = models.get_model('fully_connected')
model_hps = models.get_model_hparams('fully_connected')
hps = copy.copy(model_hps)
hps.update({'output_shape': OUTPUT_SHAPE})
rng = jax.random.PRNGKey(0)
model = model(hps, {}, loss_name, metrics_name)
xs = jnp.array(np.random.normal(size=INPUT_SHAPE))
rng, params_rng = jax.random.split(rng)
_, self.flax_module = model.flax_module_def.create(params_rng, xs)
def setUp(self):
super(CheckpointTest, self).setUp()
self.test_dir = tempfile.mkdtemp()
loss_name = 'cross_entropy'
metrics_name = 'classification_metrics'
model = models.get_model('fully_connected')
model_hps = models.get_model_hparams('fully_connected')
hps = copy.copy(model_hps)
hps.update({'output_shape': OUTPUT_SHAPE})
rng = jax.random.PRNGKey(0)
model = model(hps, {}, loss_name, metrics_name)
xs = jnp.array(np.random.normal(size=INPUT_SHAPE))
rng, params_rng = jax.random.split(rng)
model_init_fn = jax.jit(
functools.partial(model.flax_module.init, train=False))
init_dict = model_init_fn({'params': params_rng}, xs)
self.params = init_dict['params']
def _load_model(model_name):
"""Load a test model."""
rng = jax.random.PRNGKey(0)
model_cls = models.get_model(model_name)
loss_name = 'cross_entropy'
metrics_name = 'classification_metrics'
model_hps = models.get_model_hparams(model_name)
hps = copy.copy(model_hps)
hps.update({'output_shape': OUTPUT_SHAPE})
model = model_cls(hps, {}, loss_name, metrics_name)
input_shape = (BATCH_SIZE, ) + MODEL_TO_INPUT_SHAPE[model_name]
_, flax_module = model.flax_module_def.create_by_shape(rng, [input_shape],
train=True)
utils.log_pytree_shape_and_statistics(flax_module.params)
return flax_module, input_shape
def test_graph_model(self):
"""Test forward pass of the GNN model."""
edge_input_shape = (5,)
node_input_shape = (5,)
output_shape = (5,)
model_str = 'gnn'
model_hps = models.get_model_hparams(model_str)
model_hps.update({'output_shape': output_shape,
'latent_dim': 10,
'hidden_dims': (10,),
'batch_size': 5,
'normalizer': 'batch_norm'})
model_cls = models.get_model(model_str)
rng = jax.random.PRNGKey(0)
dropout_rng, params_rng = jax.random.split(rng)
loss = 'sigmoid_binary_cross_entropy'
metrics = 'binary_classification_metrics'
model = model_cls(model_hps, {}, loss, metrics)
num_graphs = 5
node_per_graph = 3
edge_per_graph = 9
inputs = jraph.get_fully_connected_graph(
n_node_per_graph=node_per_graph,
n_graph=num_graphs,
node_features=np.ones((num_graphs * node_per_graph,) +
node_input_shape),
)
inputs = inputs._replace(
edges=np.ones((num_graphs * edge_per_graph,) + edge_input_shape))
padded_inputs = jraph.pad_with_graphs(inputs, 20, 50, 7)
model_init_fn = jax.jit(
functools.partial(model.flax_module.init, train=False))
init_dict = model_init_fn({'params': params_rng}, padded_inputs)
params = init_dict['params']
batch_stats = init_dict['batch_stats']
# Check that the forward pass works with mutated batch_stats.
outputs, _ = model.flax_module.apply(
{'params': params, 'batch_stats': batch_stats},
padded_inputs,
mutable=['batch_stats'],
rngs={'dropout': dropout_rng},
train=True)
self.assertEqual(outputs.shape, (7,) + output_shape)
def _load_model(model_name):
"""Load a test model."""
rng = jax.random.PRNGKey(0)
model_cls = models.get_model(model_name)
loss_name = 'cross_entropy'
metrics_name = 'classification_metrics'
model_hps = models.get_model_hparams(model_name)
hps = copy.copy(model_hps)
hps.update({'output_shape': OUTPUT_SHAPE})
model = model_cls(hps, {}, loss_name, metrics_name)
input_shape = (BATCH_SIZE, ) + MODEL_TO_INPUT_SHAPE[model_name]
model_init_fn = jax.jit(
functools.partial(model.flax_module.init, train=True))
init_dict = model_init_fn({'params': rng}, jnp.zeros(input_shape))
# Trainable model parameters.
params = init_dict['params']
utils.log_pytree_shape_and_statistics(params)
return model.flax_module, params, input_shape, hps
def test_classification_model(self, model_str):
"""Test forward pass of the image models."""
model_cls = models.get_model(model_str)
model_hps = models.get_model_hparams(model_str)
loss = 'cross_entropy'
metrics = 'classification_metrics'
hps = copy.copy(model_hps)
hps.update({'output_shape': OUTPUT_SHAPE['classification']})
rng = jax.random.PRNGKey(0)
dropout_rng, params_rng = jax.random.split(rng)
model = model_cls(hps, {}, loss, metrics)
xs = jnp.array(np.random.normal(size=INPUT_SHAPE['classification']))
model_init_fn = jax.jit(
functools.partial(model.flax_module.init, train=False))
init_dict = model_init_fn({'params': params_rng}, xs)
params = init_dict['params']
batch_stats = init_dict.get('batch_stats', {})
# Check that the forward pass works with mutated batch_stats.
outputs, new_batch_stats = model.flax_module.apply(
{'params': params, 'batch_stats': batch_stats},
xs,
mutable=['batch_stats'],
rngs={'dropout': dropout_rng},
train=True)
self.assertEqual(outputs.shape, (INPUT_SHAPE['classification'][0],
OUTPUT_SHAPE['classification'][-1]))
# If it's a batch norm model check the batch stats changed.
if batch_stats:
bflat, _ = ravel_pytree(batch_stats)
new_bflat, _ = ravel_pytree(new_batch_stats)
self.assertFalse(jnp.array_equal(bflat, new_bflat))
# Test batch_norm in inference mode.
outputs = model.flax_module.apply(
{'params': params, 'batch_stats': batch_stats}, xs, train=False)
self.assertEqual(
outputs.shape,
(INPUT_SHAPE['classification'][0], OUTPUT_SHAPE['classification'][-1]))
def test_autoencoder_model(self, model_str):
"""Test forward pass of the autoencoder models."""
model_cls = models.get_model(model_str)
model_hps = models.get_model_hparams(model_str)
loss = 'sigmoid_binary_cross_entropy'
metrics = 'binary_autoencoder_metrics'
hps = copy.copy(model_hps)
hps.update({'output_shape': OUTPUT_SHAPE[model_str]})
params_rng = jax.random.PRNGKey(0)
model = model_cls(hps, {}, loss, metrics)
xs = jnp.array(np.random.normal(size=INPUT_SHAPE[model_str]))
model_init_fn = jax.jit(
functools.partial(model.flax_module.init, train=False))
init_dict = model_init_fn({'params': params_rng}, xs)
params = init_dict['params']
batch_stats = init_dict.get('batch_stats', {})
# Check that the forward pass works with mutated batch_stats.
outputs, new_batch_stats = model.flax_module.apply(
{'params': params, 'batch_stats': batch_stats},
xs,
mutable=['batch_stats'],
train=True)
self.assertEqual(
outputs.shape,
tuple([INPUT_SHAPE[model_str][0]] + list(OUTPUT_SHAPE[model_str])))
# If it's a batch norm model check the batch stats changed.
if batch_stats:
bflat, _ = ravel_pytree(batch_stats)
new_bflat, _ = ravel_pytree(new_batch_stats)
self.assertFalse(jnp.array_equal(bflat, new_bflat))
# Test batch_norm in inference mode.
outputs = model.flax_module.apply(
{'params': params, 'batch_stats': batch_stats}, xs, train=False)
self.assertEqual(
outputs.shape,
tuple([INPUT_SHAPE[model_str][0]] + list(OUTPUT_SHAPE[model_str])))
def test_autoencoder_model(self, model_str):
"""Test forward pass of the autoencoder models."""
model_cls = models.get_model(model_str)
model_hps = models.get_model_hparams(model_str)
loss = 'sigmoid_binary_cross_entropy'
metrics = 'binary_autoencoder_metrics'
hps = copy.copy(model_hps)
hps.update({'output_shape': OUTPUT_SHAPE[model_str]})
rng = jax.random.PRNGKey(0)
model = model_cls(hps, {}, loss, metrics)
xs = jnp.array(np.random.normal(size=INPUT_SHAPE[model_str]))
rng, params_rng = jax.random.split(rng)
with nn.stateful() as batch_stats:
with nn.stochastic(params_rng):
_, flax_module = model.flax_module_def.create(params_rng, xs)
# Check that the forward pass works with mutated batch_stats.
with nn.stateful(batch_stats) as new_batch_stats:
with nn.stochastic(params_rng):
outputs = flax_module(xs)
self.assertEqual(
outputs.shape,
tuple([INPUT_SHAPE[model_str][0]] +
list(OUTPUT_SHAPE[model_str])))
# If it's a batch norm model check the batch stats changed.
if batch_stats.as_dict():
bflat, _ = ravel_pytree(batch_stats)
new_bflat, _ = ravel_pytree(new_batch_stats)
self.assertFalse(jnp.array_equal(bflat, new_bflat))
# Test batch_norm in inference mode.
with nn.stateful(batch_stats, mutable=False):
outputs = flax_module(xs, train=False)
self.assertEqual(
outputs.shape,
tuple([INPUT_SHAPE[model_str][0]] + list(OUTPUT_SHAPE[model_str])))
def test_dlrm_model_trainer(self):
"""Tests that dlrm model training decreases loss."""
rng = jax.random.PRNGKey(1337)
model_str = 'dlrm'
dataset_str = 'criteo1tb'
model_cls = models.get_model(model_str)
model_hps = models.get_model_hparams(model_str)
dataset_hps = datasets.get_dataset_hparams(dataset_str)
dataset_hps.update({
'batch_size': model_hps.batch_size,
'num_dense_features': model_hps.num_dense_features,
'vocab_sizes': model_hps.vocab_sizes,
})
eval_num_batches = 5
eval_batch_size = dataset_hps.batch_size
loss_name = 'sigmoid_binary_cross_entropy'
metrics_name = 'binary_classification_metrics'
dataset, dataset_meta_data = _get_fake_dlrm_dataset(
dataset_hps.batch_size, eval_num_batches, dataset_hps)
hps = copy.copy(model_hps)
hps.update({
'train_size':
15,
'valid_size':
10,
'test_size':
10,
'input_shape':
(model_hps.num_dense_features + len(model_hps.vocab_sizes), ),
'output_shape': (1, ),
'l2_decay_factor':
1e-4,
'l2_decay_rank_threshold':
2,
'num_device_prefetches':
0,
})
model = model_cls(hps, dataset_meta_data, loss_name, metrics_name)
initializer = initializers.get_initializer('noop')
metrics_logger, init_logger = utils.set_up_loggers(self.test_dir)
_ = list(
trainer.train(
train_dir=self.test_dir,
model=model,
dataset_builder=lambda *unused_args, **unused_kwargs: dataset,
initializer=initializer,
num_train_steps=10,
hps=hps,
rng=rng,
eval_batch_size=eval_batch_size,
eval_num_batches=eval_num_batches,
eval_train_num_batches=eval_num_batches,
eval_frequency=2,
checkpoint_steps=[],
metrics_logger=metrics_logger,
init_logger=init_logger))
with tf.io.gfile.GFile(os.path.join(self.test_dir,
'measurements.csv')) as f:
df = pandas.read_csv(f)
train_loss = df['train/ce_loss'].values
self.assertLess(train_loss[-1], train_loss[0])
def test_graph_model_trainer(self):
"""Tests that graph model training decreases loss."""
rng = jax.random.PRNGKey(1337)
model_str = 'gnn'
model_cls = models.get_model(model_str)
hps = models.get_model_hparams(model_str)
hps.update({
'batch_size': 2,
'input_edge_shape': (7, ),
'input_node_shape': (3, ),
'input_shape': (7, 3),
'output_shape': (5, ),
'model_dtype': 'float32',
'train_size': 15,
'valid_size': 10,
'test_size': 10,
'num_message_passing_steps': 1,
'normalizer': 'none',
'dropout_rate': 0.0,
'lr_hparams': {
'base_lr': 0.001,
'schedule': 'constant'
},
'num_device_prefetches': 0,
})
eval_num_batches = 5
eval_batch_size = hps.batch_size
loss_name = 'sigmoid_binary_cross_entropy'
metrics_name = 'binary_classification_metrics_ogbg_map'
dataset, dataset_meta_data = _get_fake_graph_dataset(
batch_size=hps.batch_size,
eval_num_batches=eval_num_batches,
hps=hps)
model = model_cls(hps, dataset_meta_data, loss_name, metrics_name)
initializer = initializers.get_initializer('noop')
metrics_logger, init_logger = utils.set_up_loggers(self.test_dir)
_ = list(
trainer.train(
train_dir=self.test_dir,
model=model,
dataset_builder=lambda *unused_args, **unused_kwargs: dataset,
initializer=initializer,
num_train_steps=10,
hps=hps,
rng=rng,
eval_batch_size=eval_batch_size,
eval_num_batches=eval_num_batches,
eval_train_num_batches=eval_num_batches,
# Note that for some reason, moving from the deprecated to linen
# Flax model API made training less stable so we need to eval more
# frequently in order to get a `train_loss[0]` that is earlier in
# training.
eval_frequency=2,
checkpoint_steps=[],
metrics_logger=metrics_logger,
init_logger=init_logger))
with tf.io.gfile.GFile(os.path.join(self.test_dir,
'measurements.csv')) as f:
df = pandas.read_csv(f)
train_loss = df['train/ce_loss'].values
self.assertLess(train_loss[-1], train_loss[0])
def build_hparams(model_name,
initializer_name,
dataset_name,
hparam_file,
hparam_overrides):
"""Build experiment hyperparameters.
Args:
model_name: the string model name.
initializer_name: the string initializer name.
dataset_name: the string dataset name.
hparam_file: the string to the hyperparameter override file (possibly on
CNS).
hparam_overrides: a dict of hyperparameter override names/values, or a JSON
string encoding of this hyperparameter override dict. Note that this is
applied after the hyperparameter file overrides.
Returns:
A ConfigDict of experiment hyperparameters.
"""
model_hps = models.get_model_hparams(model_name)
initializer_hps = initializers.get_initializer_hparams(initializer_name)
dataset_hps = datasets.get_dataset_hparams(dataset_name)
merged_dict = {}
hps_dicts = [
hps.to_dict() for hps in [model_hps, initializer_hps, dataset_hps]
]
total_hps = 0
for hps_dict in hps_dicts:
merged_dict.update(hps_dict)
total_hps += len(hps_dict.keys())
# Check that all provided have no overlap.
if total_hps != len(merged_dict.keys()):
raise ValueError('There is overlap in the provided hparams.')
# Convert to the Shallue and Lee label smoothing style.
if merged_dict.get('use_shallue_label_smoothing', False):
num_classes = merged_dict['output_shape'][-1]
merged_dict['label_smoothing'] *= num_classes / float(num_classes - 1)
merged = config_dict.ConfigDict(merged_dict)
merged.lock()
# Subconfig "opt_hparams" and "lr_hparams" are allowed to add new fields.
for key in ['opt_hparams', 'lr_hparams']:
if key not in merged:
with merged.unlocked():
merged[key] = config_dict.ConfigDict()
for key in ['opt_hparams', 'lr_hparams']:
merged[key].unlock()
if hparam_file:
logging.info('Loading hparams from %s', hparam_file)
with gfile.GFile(hparam_file, 'r') as f:
hparam_dict = json.load(f)
merged.update_from_flattened_dict(hparam_dict)
if hparam_overrides:
if isinstance(hparam_overrides, str):
hparam_overrides = json.loads(hparam_overrides)
# If the user is changing the learning rate schedule or optimizer. We must
# wipe all of the keys from the old dictionary.
if 'lr_hparams.schedule' in hparam_overrides and merged[
'lr_hparams']['schedule'] != hparam_overrides[
'lr_hparams.schedule']:
merged['lr_hparams'] = {}
if 'optimizer' in hparam_overrides and merged[
'optimizer'] != hparam_overrides['optimizer']:
merged['opt_hparams'] = {}
merged.update_from_flattened_dict(hparam_overrides)
return merged
def test_accumulation(self):
"""Test simple gradient accumulation."""
num_steps = 3
per_step_batch_size = 16
total_batch_size = 48
virtual_batch_size = 8
model_str = 'wide_resnet' # Pick a model with batch norm.
model_cls = models.get_model(model_str)
model_hps = models.get_model_hparams(model_str)
dataset_name = 'cifar10'
dataset_builder = datasets.get_dataset(dataset_name)
hps = copy.copy(model_hps)
hps.update(datasets.get_dataset_hparams(dataset_name))
# Compute updates using gradient accumulation.
hps.update({
'batch_size': per_step_batch_size,
'virtual_batch_size': virtual_batch_size,
'normalizer': 'virtual_batch_norm',
'total_accumulated_batch_size': total_batch_size,
})
grad_acc_params, grad_acc_batch_stats, grad_acc_training_cost = _init_model(
model_cls, hps)
total_dataset = dataset_builder(shuffle_rng=jax.random.PRNGKey(1),
batch_size=total_batch_size,
eval_batch_size=10,
hps=hps)
# Ensure we see the same exact batches.
train_iter = total_dataset.train_iterator_fn()
train_iter = itertools.islice(train_iter, 0, num_steps)
train_iter = itertools.cycle(train_iter)
def grad_acc_train_iter():
for _ in range(num_steps):
total_batch = next(train_iter)
# Split each total batch into sub batches.
num_sub_batches = total_batch_size // per_step_batch_size
start_index = 0
end_index = int(total_batch_size / num_sub_batches)
for bi in range(num_sub_batches):
yield jax.tree_map(lambda x: x[start_index:end_index],
total_batch) # pylint: disable=cell-var-from-loop
start_index = end_index
end_index = int(total_batch_size * (bi + 2) /
num_sub_batches)
lrs = jnp.array([1.0, 0.1, 1e-2])
sgd_opt_init, sgd_opt_update = optax.sgd(
learning_rate=lambda t: lrs.at[t].get())
opt_init, opt_update = gradient_accumulator.accumulate_gradients(
per_step_batch_size=per_step_batch_size,
total_batch_size=total_batch_size,
virtual_batch_size=virtual_batch_size,
base_opt_init_fn=sgd_opt_init,
base_opt_update_fn=sgd_opt_update)
grad_acc_params, grad_acc_batch_stats = _optimize(
# Run for 3x the number of steps to see the same number of examples.
num_steps=3 * num_steps,
params=grad_acc_params,
batch_stats=grad_acc_batch_stats,
training_cost=grad_acc_training_cost,
train_iter=grad_acc_train_iter(),
opt_init=opt_init,
opt_update=opt_update)
# Compute the same updates, but without gradient accumulation.
hps.update({
'batch_size': total_batch_size,
'total_accumulated_batch_size': None,
})
params, batch_stats, training_cost = _init_model(model_cls, hps)
params, batch_stats = _optimize(num_steps=num_steps,
params=params,
batch_stats=batch_stats,
training_cost=training_cost,
train_iter=train_iter,
opt_init=sgd_opt_init,
opt_update=sgd_opt_update)
diffs_params = jax.tree_multimap(lambda a, b: jnp.mean(jnp.abs(a - b)),
grad_acc_params, params)
def batch_stats_reduce(a, b):
if len(a.shape) > 0: # pylint: disable=g-explicit-length-test
return jnp.mean(
jnp.abs(jnp.mean(a, axis=0) - jnp.mean(b, axis=0)))
# The gradient accumulator counters are scalars.
return a - b
diffs_batch_stats = jax.tree_multimap(batch_stats_reduce,
grad_acc_batch_stats,
batch_stats)
# We sometimes get small floating point errors in the gradients, so we
# cannot test for the values being exactly the same.
acceptable_params_diff = 1e-4
acceptable_batch_stats_diff = 5e-3
def check_closeness(root_name, d, max_diff):
not_close_dict = {}
for name, dd in d.items():
new_name = root_name + '/' + name if root_name else name
if isinstance(dd, (dict, core.FrozenDict)):
not_close_dict.update(
check_closeness(new_name, dd, max_diff))
else:
if dd > max_diff:
not_close_dict[new_name] = dd
return not_close_dict
not_close_params = check_closeness('', diffs_params,
acceptable_params_diff)
self.assertEmpty(not_close_params)
not_close_batch_stats = check_closeness('', diffs_batch_stats,
acceptable_batch_stats_diff)
# Note that for the variance variables in the batch stats collection, they
# sometimes can start to diverge slightly over time (with a higher number of
# training steps), likely due to numerical issues.
self.assertEmpty(not_close_batch_stats)
def test_hessian_free_optimizer(self):
"""Tests the Hessian-free optimizer."""
model_str = 'autoencoder'
model_cls = models.get_model(model_str)
model_hps = models.get_model_hparams(model_str)
loss = 'sigmoid_binary_cross_entropy'
metrics = 'binary_autoencoder_metrics'
input_shape = (2, 2, 1)
output_shape = (4, )
hps = copy.copy(model_hps)
hps.update({
'optimizer': 'hessian_free',
'opt_hparams': {
'weight_decay': 0.0,
},
'hid_sizes': [2],
'activation_function': ['id'],
'input_shape': input_shape,
'output_shape': output_shape
})
model = model_cls(hps, {}, loss, metrics)
inputs = jnp.array([[[1, 0], [1, 1]], [[1, 0], [0, 1]]])
targets = inputs.reshape(tuple([inputs.shape[0]] + list(output_shape)))
batch = {'inputs': inputs, 'targets': targets}
def forward_fn(variables, inputs):
logits = model.flax_module.apply(variables, inputs, train=True)
return logits
def opt_cost(variables):
return model.loss_fn(forward_fn(variables, inputs), targets)
init_fn, update_fn = optimizers.get_optimizer(hps, model)
params = {
'Dense_0': {
'kernel': jnp.array([[-1., 2.], [2., 0.], [-1., 3.], [-2.,
2.]]),
'bias': jnp.array([0., 0.])
},
'Dense_1': {
'kernel': jnp.array([[4., 2., -2., 4.], [-3., 1., 2., -4.]]),
'bias': jnp.array([0., 0., 0., 0.])
}
}
variables = {'params': params}
grad_fn = jax.grad(opt_cost)
grads = grad_fn(variables)['params']
outputs = forward_fn(variables, batch['inputs'])
n = inputs.shape[0]
m = outputs.shape[-1]
d = ravel_pytree(params)[0].shape[0]
v = np.ones(d)
state = init_fn(params)
partial_forward_fn = partial(forward_fn, inputs=batch['inputs'])
partial_loss_fn = partial(model.loss_fn, targets=batch['targets'])
matmul_fn = partial(gvp, variables, outputs, state.inner_state.damping,
partial_forward_fn, partial_loss_fn)
jacobian = jax.jacfwd(partial_forward_fn)(variables)['params']
jacobian_tensor = np.concatenate(
(jacobian['Dense_0']['bias'].reshape(
n, m, -1), jacobian['Dense_0']['kernel'].reshape(
n, m, -1), jacobian['Dense_1']['bias'].reshape(n, m, -1),
jacobian['Dense_1']['kernel'].reshape(n, m, -1)),
axis=2)
ggn_matrix = 0
for i in range(n):
jacobian_matrix = jacobian_tensor[i]
hessian = jax.hessian(partial_loss_fn)(outputs[i, None])[0, :,
0, :]
ggn_matrix += np.transpose(
jacobian_matrix) @ hessian @ jacobian_matrix
ggn_matrix /= n
ggn_matrix += state.inner_state.damping * np.identity(d)
expected = ggn_matrix @ v
# Test the gvp function
self.assertAlmostEqual(jnp.linalg.norm(matmul_fn(v) - expected),
0,
places=4)
update_pmapped = jax.pmap(update_fn,
axis_name='batch',
in_axes=(None, None, None, 0, None))
batch_shard = data_utils.shard(batch)
state.hyperparams['learning_rate'] = 1.0
p, state = update_pmapped(grads, state, params, batch_shard, None)
# Test the damping parameter update
self.assertEqual(state.inner_state.damping, 3 / 2)
# Test the search direction
self.assertAlmostEqual(jnp.linalg.norm(
ravel_pytree(p)[0] +
jnp.linalg.inv(ggn_matrix) @ ravel_pytree(grads)[0]),
0,
places=4)
def test_cg_backtracking(self):
"""Tests CG backtracking."""
model_str = 'autoencoder'
model_cls = models.get_model(model_str)
model_hps = models.get_model_hparams(model_str)
loss = 'sigmoid_binary_cross_entropy'
metrics = 'binary_autoencoder_metrics'
input_shape = (2, 2, 1)
output_shape = (4, )
hps = copy.copy(model_hps)
hps.update({
'optimizer': 'hessian_free',
'opt_hparams': {
'weight_decay': 0.0,
},
'hid_sizes': [2],
'activation_function': ['id'],
'input_shape': input_shape,
'output_shape': output_shape
})
model = model_cls(hps, {}, loss, metrics)
inputs = jnp.array([[[1, 0], [1, 1]], [[1, 0], [0, 1]]])
targets = inputs.reshape(tuple([inputs.shape[0]] + list(output_shape)))
batch = {'inputs': inputs, 'targets': targets}
def forward_fn(variables, inputs):
return model.flax_module.apply(variables, inputs, train=False)
def opt_cost(params):
return model.loss_fn(forward_fn(params, inputs), targets)
params = {
'Dense_0': {
'kernel': jnp.array([[-1., 2.], [2., 0.], [-1., 3.], [-2.,
2.]]),
'bias': jnp.array([0., 0.])
},
'Dense_1': {
'kernel': jnp.array([[4., 2., -2., 4.], [-3., 1., 2., -4.]]),
'bias': jnp.array([0., 0., 0., 0.])
}
}
unravel_fn = ravel_pytree(params)[1]
p1 = np.array([
0.5, 0.2, 0.1, -0.4, -0.6, 0.4, 0.6, -0.7, 0.0, 0.5, -0.7, 0.2,
0.1, -0.2, 0.4, -0.6, -0.8, 0.7, 0.2, 0.9, -0.1, 0.5
])
p2 = np.array([
0.3, -0.1, -0.5, 0.2, -0.4, 0.8, -0.2, 0.0, 0.2, -0.4, 0.6, -0.2,
-0.4, 0.2, 0.3, 0.2, -0.2, -0.4, -0.5, 0.2, 0.2, -0.4
])
p_arr = jnp.array([p1, p2])
p_arr_idx = 1
partial_forward_fn = partial(forward_fn, inputs=batch['inputs'])
partial_loss_fn = partial(model.loss_fn, targets=batch['targets'])
def obj_fn(variables):
return partial_loss_fn(partial_forward_fn(variables))
flattened_p, obj_val = cg_backtracking(p_arr, p_arr_idx, obj_fn,
{'params': params}, unravel_fn)
# Test the backtracking function.
self.assertSameElements(flattened_p, p1)
updated_params = apply_updates(params, unravel_fn(p1))
self.assertEqual(opt_cost({'params': updated_params}), obj_val)
