def add_blobs(queue, num_samples):
blob = core.BlobReference("blob")
status = core.BlobReference("blob_status")
for i in range(num_samples):
data = self._create_test_tensor_protos(i)
data = np.array([data], dtype=str)
self._add_blob_to_queue(queue, data, blob, status)
def add_blobs(queue, num_samples):
blob = core.BlobReference("blob")
status = core.BlobReference("blob_status")
for i in range(num_samples):
self._add_blob_to_queue(queue,
self._create_test_tensor_protos(i),
blob, status)
def test_checkpoint_params(self):
def add_input_ops(model):
pass
def add_model_ops(model, loss_scale):
model.NHWC2NCHW("data", "data_nchw")
model.Conv("data_nchw",
'conv1',
3,
64,
weight_init=("MSRAFill", {}),
kernel=7,
stride=2,
pad=3,
no_bias=0)
model.SpatialBN('conv1', 'conv1_spatbn_relu', 64, epsilon=1e-3)
model.Relu('conv1_spatbn_relu', 'conv1_spatbn_relu')
model.MaxPool('conv1_spatbn_relu', 'pool1', kernel=3, stride=2)
model.FC('pool1', 'fc', dim_in=(64 * 56 * 56), dim_out=100)
model.Sigmoid('fc', 'fc_sigm')
model.Softmax('fc_sigm', 'softmax')
model.LabelCrossEntropy(['softmax', 'label'], 'xent')
loss = model.AveragedLoss('xent', 'loss')
# Add a duplicate param init to ensure it does not cause issues
model.param_init_net.ConstantFill([], ["fc_w"],
shape=((64 * 56 * 56), 1000))
return [loss]
def add_optimizer(model):
optimizer.build_sgd(model, 0.1, policy="fixed", momentum=0.9)
model = cnn.CNNModelHelper(
order="NHWC",
name="test",
)
data_parallel_model.Parallelize_CPU(
model,
input_builder_fun=add_input_ops,
forward_pass_builder_fun=add_model_ops,
optimizer_builder_fun=add_optimizer,
devices=[1, 2, 3],
)
# Only gpu_1 params should be returned (gpu_1 is the first gpu)
checkpoint_params = data_parallel_model.GetCheckpointParams(model)
for p in model.GetParams("cpu_1/"):
self.assertTrue(p in checkpoint_params)
self.assertTrue(p + "_momentum" in checkpoint_params)
for p in model.GetParams("cpu_2/"):
self.assertFalse(p in checkpoint_params)
self.assertTrue(
core.BlobReference("cpu_1/fc_w_momentum") in checkpoint_params)
for c in model.GetComputedParams("cpu_1/"):
self.assertTrue(c in checkpoint_params)
for c in model.GetComputedParams("cpu_2/"):
self.assertFalse(c in checkpoint_params)
self.assertFalse(core.BlobReference("cpu_1/data") in checkpoint_params)
self.assertTrue(
core.BlobReference("optimizer_iteration") in checkpoint_params)
def load_from_db(filename, db_type, device_option=None, *args, **kwargs):
# global_init_net in meta_net_def will load parameters from
# predictor_constants.PREDICTOR_DBREADER
create_db = core.CreateOperator(
'CreateDB', [],
[core.BlobReference(predictor_constants.PREDICTOR_DBREADER)],
db=filename, db_type=db_type)
assert workspace.RunOperatorOnce(create_db), (
'Failed to create db {}'.format(filename))
# predictor_constants.META_NET_DEF is always stored before the parameters
load_meta_net_def = core.CreateOperator(
'Load',
[core.BlobReference(predictor_constants.PREDICTOR_DBREADER)],
[core.BlobReference(predictor_constants.META_NET_DEF)])
assert workspace.RunOperatorOnce(load_meta_net_def)
blob = workspace.FetchBlob(predictor_constants.META_NET_DEF)
meta_net_def = serde.deserialize_protobuf_struct(
blob if isinstance(blob, bytes)
else str(blob).encode('utf-8'),
metanet_pb2.MetaNetDef)
if device_option is None:
device_option = scope.CurrentDeviceScope()
if device_option is not None:
# Set the device options of all loaded blobs
for kv in meta_net_def.nets:
net = kv.value
for op in net.op:
op.device_option.CopyFrom(device_option)
return meta_net_def
def testBlobReferenceIsIndependentFromNameScope(self):
blob_v = core.BlobReference("v")
with core.NameScope("foo"):
blob_w = core.BlobReference("w")
with core.NameScope("bar"):
blob_x = core.BlobReference("x")
self.assertEqual(str(blob_v), "v")
self.assertEqual(str(blob_w), "w")
self.assertEqual(str(blob_x), "x")
def _enqueue(self, blob_name, queue, data_arr):
'''
Enqueue the correctly sized batch arrays to Caffe2's queue.
'''
scratch_name = self._namescope + blob_name + \
"_scratch_" + self._input_source_name
blob = core.BlobReference(scratch_name)
status = core.BlobReference(scratch_name + "_status")
workspace.FeedBlob(blob, data_arr, device_option=self._device_option)
op = core.CreateOperator("SafeEnqueueBlobs", [queue, blob],
[blob, status],
device_option=self._device_option)
workspace.RunOperatorOnce(op)
def testAddLoss(self):
input_record_LR = self.new_record(
schema.Struct(('label', schema.Scalar((np.float64, (1, )))),
('prediction', schema.Scalar((np.float32, (2, )))),
('weight', schema.Scalar((np.float64, (1, ))))))
loss_LR = self.model.BatchLRLoss(input_record_LR)
self.model.add_loss(loss_LR)
assert 'unnamed' in self.model.loss
self.assertEqual(schema.Scalar((np.float32, tuple())),
self.model.loss.unnamed)
self.assertEqual(loss_LR, self.model.loss.unnamed)
self.model.add_loss(loss_LR, 'addLoss')
assert 'addLoss' in self.model.loss
self.assertEqual(schema.Scalar((np.float32, tuple())),
self.model.loss.addLoss)
self.assertEqual(loss_LR, self.model.loss.addLoss)
self.model.add_loss(
schema.Scalar(dtype=np.float32,
blob=core.BlobReference('loss_blob_1')), 'addLoss')
assert 'addLoss_auto_0' in self.model.loss
self.assertEqual(schema.Scalar((np.float32, tuple())),
self.model.loss.addLoss_auto_0)
assert core.BlobReference(
'loss_blob_1') in self.model.loss.field_blobs()
self.model.add_loss(
schema.Struct(
('structName',
schema.Scalar(dtype=np.float32,
blob=core.BlobReference('loss_blob_2')))),
'addLoss')
assert 'addLoss_auto_1' in self.model.loss
self.assertEqual(
schema.Struct(('structName', schema.Scalar(
(np.float32, tuple())))), self.model.loss.addLoss_auto_1)
assert core.BlobReference(
'loss_blob_2') in self.model.loss.field_blobs()
loss_in_tuple_0 = schema.Scalar(
dtype=np.float32, blob=core.BlobReference('loss_blob_in_tuple_0'))
loss_in_tuple_1 = schema.Scalar(
dtype=np.float32, blob=core.BlobReference('loss_blob_in_tuple_1'))
loss_tuple = schema.NamedTuple('loss_in_tuple',
*[loss_in_tuple_0, loss_in_tuple_1])
self.model.add_loss(loss_tuple, 'addLoss')
assert 'addLoss_auto_2' in self.model.loss
self.assertEqual(
schema.Struct(
('loss_in_tuple_0', schema.Scalar((np.float32, tuple()))),
('loss_in_tuple_1', schema.Scalar((np.float32, tuple())))),
self.model.loss.addLoss_auto_2)
assert core.BlobReference('loss_blob_in_tuple_0')\
in self.model.loss.field_blobs()
assert core.BlobReference('loss_blob_in_tuple_1')\
in self.model.loss.field_blobs()
def modify_net(self,
net,
init_net=None,
grad_map=None,
blob_to_device=None,
modify_output_record=False):
for blob_name in self._blobs:
blob = core.BlobReference(blob_name)
assert net.BlobIsDefined(
blob
), 'blob {} is not defined in net {} whose proto is {}'.format(
blob, net.Name(), net.Proto())
cast_blob = net.Cast(blob, to=core.DataType.FLOAT)
stats_name = net.NextScopedBlob(prefix=blob +
self._field_name_suffix)
stats = net.Summarize(cast_blob, stats_name, to_file=0)
net.Print(stats, [], every_n=self._logging_frequency)
if modify_output_record:
output_field_name = str(blob) + self._field_name_suffix
output_scalar = schema.Scalar((np.float, (1, )), stats)
if net.output_record() is None:
net.set_output_record(
schema.Struct((output_field_name, output_scalar)))
else:
net.AppendOutputRecordField(output_field_name,
output_scalar)
def __call__(self, rec):
assert not self._frozen
prefix = '/worker:%d/' % len(self._blob_maps)
blob_remap = {}
for net in self.thread_init_nets:
new_net, _ = core.clone_and_bind_net(net,
str(net) + prefix, prefix,
blob_remap)
self._cloned_init_nets.append(new_net)
new_net, remappings = core.clone_and_bind_net(self.net,
str(self.net) + prefix,
prefix, blob_remap, rec)
if self._stop_signal is None:
stop_signal = None
elif str(self._stop_signal) in remappings:
stop_signal = core.BlobReference(remappings[str(
self._stop_signal)],
net=new_net)
else:
stop_signal = self._stop_signal
self._blob_maps.append(remappings)
return Output([new_net], new_net.output_record(), stop_signal)
def test_bounded_grad_proj(self, X, left_open, right_open, eps, ub, lb, gc,
dc):
if ub - (eps if right_open else 0.) < lb + (eps if left_open else 0.):
return
param = core.BlobReference("X")
workspace.FeedBlob(param, X)
train_init_net, train_net = self.get_training_nets()
reg = regularizer.BoundedGradientProjection(lb=lb,
ub=ub,
left_open=left_open,
right_open=right_open,
epsilon=eps)
output = reg(train_net,
train_init_net,
param,
by=RegularizationBy.ON_LOSS)
reg(
train_net,
train_init_net,
param,
grad=None,
by=RegularizationBy.AFTER_OPTIMIZER,
)
workspace.RunNetOnce(train_init_net)
workspace.RunNetOnce(train_net)
def ref(X):
return np.clip(X, lb + (eps if left_open else 0.),
ub - (eps if right_open else 0.))
assert output is None
npt.assert_allclose(workspace.blobs[param], ref(X), atol=1e-7)
def _ComputeBlobsToSync(model):
'''
We sync all blobs that are generated by param init net and
are 'data parallel', i.e assigned to a gpu
'''
sync_names = set()
blobs_to_sync = []
for op in model.param_init_net.Proto().op:
dp_outputs = [
o for o in op.output
if o.startswith("{}_".format(model._device_prefix))
]
sync_names.update([stripParamName(o) for o in dp_outputs])
blobs_to_sync.extend(dp_outputs)
# Sanity check
diff = set(model._param_names) - sync_names
assert diff == set(), \
"Some params not instantiated in param init net: {}".format(diff)
# Remove duplicates and sort
blobs_to_sync = sorted(list(set(blobs_to_sync)))
blobs_to_sync = [core.BlobReference(b) for b in blobs_to_sync]
return (blobs_to_sync, sync_names)
def testNameScopeWithOp(self):
global_x = core.BlobReference("x")
global_y = core.BlobReference("y")
with core.NameScope("foo"):
# Raw strings should have namescope prepended.
op = core.CreateOperator("Relu", "x", "y")
self.assertEqual(len(op.input), 1)
self.assertEqual(op.input[0], "foo/x")
self.assertEqual(len(op.output), 1)
self.assertEqual(op.output[0], "foo/y")
# BlobReferences should not.
op = core.CreateOperator("Relu", global_x, global_y)
self.assertEqual(len(op.input), 1)
self.assertEqual(op.input[0], "x")
self.assertEqual(len(op.output), 1)
self.assertEqual(op.output[0], "y")
def modify_net(self,
net,
init_net=None,
grad_map=None,
blob_to_device=None):
for blob_name in self._blobs:
blob = core.BlobReference(blob_name)
if not net.BlobIsDefined(blob):
raise Exception('blob {0} is not defined in net {1}'.format(
blob, net.Name()))
cast_blob = net.Cast(blob, to=core.DataType.FLOAT)
stats_name = net.NextScopedBlob(prefix=blob + '_summary')
stats = net.Summarize(cast_blob, stats_name, to_file=0)
net.Print(stats, [], every_n=self._logging_frequency)
output_field_name = str(blob) + '_summary'
output_scalar = schema.Scalar((np.float, (1, )), stats)
if net.output_record() is None:
net.set_output_record(
schema.Struct((output_field_name, output_scalar)))
else:
net.AppendOutputRecordField(output_field_name, output_scalar)
def test_create_net(self):
action_blob = core.BlobReference("action")
N = 10
action_feature_ids = [100, 300, 200]
serving_max_scale = np.array(action_feature_ids) / 100.0
serving_min_scale = np.zeros(len(action_feature_ids)) - 5.0
actions = np.random.randn(N, len(action_feature_ids)).astype(np.float32)
workspace.FeedBlob(action_blob, actions)
ot = ActorOutputTransformer(
action_feature_ids, serving_max_scale, serving_min_scale
)
output_record = schema.Struct(("action", schema.Scalar(blob=action_blob)))
nets = ot.create_net(output_record)
workspace.RunNetOnce(nets.init_net)
workspace.RunNetOnce(nets.net)
external_outputs = {str(b) for b in nets.net.external_outputs}
def fetch_blob(b):
self.assertIn(b, external_outputs)
return workspace.FetchBlob(b)
lengths = fetch_blob("output/float_features.lengths")
keys = fetch_blob("output/float_features.keys")
values = fetch_blob("output/float_features.values")
scaled_actions = (actions + np.ones(len(action_feature_ids)) - 1e-6) / (
(1 - 1e-6) * 2
) * (serving_max_scale - serving_min_scale) + serving_min_scale
npt.assert_array_equal([len(action_feature_ids)] * N, lengths)
npt.assert_array_equal(action_feature_ids * N, keys)
npt.assert_array_almost_equal(scaled_actions.reshape(-1), values)
def create_param(self, param_name, shape, initializer, optimizer=None,
ps_param=None):
if isinstance(param_name, core.BlobReference):
param_name = str(param_name)
elif isinstance(param_name, six.string_types):
# Parameter name will be equal to current Namescope that got
# resolved with the respect of parameter sharing of the scopes.
param_name = parameter_sharing_context.get_parameter_name(
param_name)
else:
raise "Unsupported type for param_name"
param_blob = core.BlobReference(param_name)
if len(initializer) == 1:
init_op_args = {}
else:
assert len(initializer) == 2
init_op_args = initializer[1]
if shape is not None:
init_op_args.update({'shape': shape})
param = layers.LayerParameter(
parameter=param_blob,
initializer=core.CreateOperator(
initializer[0],
[],
param_blob,
**init_op_args
),
optimizer=optimizer,
ps_param=ps_param,
)
return param
def DoUntil(name, condition_blob_or_net, nets_or_steps):
"""
Similar to DoWhile() but execute nets_or_steps when
condition_blob_or_net returns false. It will execute
nets_or_steps before evaluating condition_blob_or_net.
Special case: if condition_blob_or_net is a blob and is pre-set to
true, then only the first net/step of nets_or_steps will be executed and
loop is exited. So you need to be careful about the initial value the
condition blob when using DoUntil(), esp when DoUntil() is called twice.
"""
if not isinstance(condition_blob_or_net, core.Net):
stop_blob = core.BlobReference(condition_blob_or_net)
return core.scoped_execution_step(_get_next_step_name('DoUntil', name),
nets_or_steps,
should_stop_blob=stop_blob)
nets_or_steps = _AppendNets(nets_or_steps, condition_blob_or_net)
stop_blob = GetConditionBlobFromNet(condition_blob_or_net)
# If stop_blob is pre-set to True (this may happen when DoWhile() is
# called twice), the loop will exit after executing the first net/step
# in nets_or_steps. This is not what we want. So we use BootNet to
# set stop_blob to False.
bool_net = BoolNet((stop_blob, False))
return Do(
name + '/DoUntil', bool_net,
core.scoped_execution_step(
_get_next_step_name('DoUntil-inner', name),
nets_or_steps,
should_stop_blob=stop_blob,
))
def modify_net(self,
net,
init_net=None,
grad_map=None,
blob_to_device=None):
p = self._p
for blob_name in self._blobs:
blob = core.BlobReference(blob_name)
if not net.BlobIsDefined(blob):
raise Exception('blob {0} is not defined in net {1}'.format(
blob, net.Name()))
norm_name = net.NextScopedBlob(prefix=blob + '_l{}_norm'.format(p))
norm = net.LpNorm(blob, norm_name, p=p)
if self._logging_frequency >= 1:
net.Print(norm, [], every_n=self._logging_frequency)
output_field_name = str(blob) + '_l{}_norm'.format(p)
output_scalar = schema.Scalar((np.float, (1, )), norm)
if net.output_record() is None:
net.set_output_record(
schema.Struct((output_field_name, output_scalar)))
else:
net.AppendOutputRecordField(output_field_name, output_scalar)
def test_log_barrier(self, X):
param = core.BlobReference("X")
workspace.FeedBlob(param, X)
train_init_net, train_net = self.get_training_nets()
reg = regularizer.LogBarrier(1.0)
output = reg(train_net,
train_init_net,
param,
by=RegularizationBy.ON_LOSS)
reg(
train_net,
train_init_net,
param,
grad=None,
by=RegularizationBy.AFTER_OPTIMIZER,
)
workspace.RunNetOnce(train_init_net)
workspace.RunNetOnce(train_net)
def ref(X):
return (
np.array(np.sum(-np.log(np.clip(X, 1e-9, None))) * 0.5).astype(
np.float32),
np.clip(X, 1e-9, None),
)
for x, y in zip(workspace.FetchBlobs([output, param]), ref(X)):
npt.assert_allclose(x, y, rtol=1e-3)
def testAddFieldByNestedName(self):
f_a = schema.Scalar(blob=core.BlobReference('blob1'))
f_b = schema.Struct(
('c', schema.Struct(
('d', schema.Scalar(blob=core.BlobReference('blob2'))),
)),
)
f_x = schema.Struct(
('x', schema.Scalar(blob=core.BlobReference('blob3'))),
)
with self.assertRaises(TypeError):
st = schema.Struct(
('a', f_a),
('b', f_b),
('b:c:d', f_x),
)
with self.assertRaises(TypeError):
st = schema.Struct(
('a', f_a),
('b', f_b),
('b:c:d:e', f_x),
)
st = schema.Struct(
('a', f_a),
('b', f_b),
('e:f', f_x),
)
self.assertEqual(['a', 'b:c:d', 'e:f:x'], st.field_names())
self.assertEqual(['blob1', 'blob2', 'blob3'], st.field_blobs())
st = schema.Struct(
('a', f_a),
('b:c:e', f_x),
('b', f_b),
)
self.assertEqual(['a', 'b:c:e:x', 'b:c:d'], st.field_names())
self.assertEqual(['blob1', 'blob3', 'blob2'], st.field_blobs())
st = schema.Struct(
('a:a1', f_a),
('b:b1', f_b),
('a', f_x),
)
self.assertEqual(['a:a1', 'a:x', 'b:b1:c:d'], st.field_names())
self.assertEqual(['blob1', 'blob3', 'blob2'], st.field_blobs())
def modify_net(self,
net,
init_net=None,
grad_map=None,
blob_to_device=None,
modify_output_record=False):
p = self._p
compute_averaged_norm = self._compute_averaged_norm
row_index = self.row_index
CPU = muji.OnCPU()
# if given, blob_to_device is a map from blob to device_option
blob_to_device = blob_to_device or {}
for blob_name in self._blobs:
blob = core.BlobReference(blob_name)
assert net.BlobIsDefined(
blob
), 'blob {} is not defined in net {} whose proto is {}'.format(
blob, net.Name(), net.Proto())
if blob in blob_to_device:
device = blob_to_device[blob]
else:
device = CPU
with core.DeviceScope(device):
if row_index and row_index >= 0:
blob = net.Slice(
[blob],
net.NextScopedBlob(prefix=blob +
'_row_{0}'.format(row_index)),
starts=[row_index, 0],
ends=[row_index + 1, -1])
cast_blob = net.Cast(blob,
net.NextScopedBlob(prefix=blob +
'_float'),
to=core.DataType.FLOAT)
norm_name = net.NextScopedBlob(prefix=blob +
self._field_name_suffix)
norm = net.LpNorm(cast_blob,
norm_name,
p=p,
average=compute_averaged_norm)
if self._logging_frequency >= 1:
net.Print(norm, [], every_n=self._logging_frequency)
if modify_output_record:
output_field_name = str(blob) + self._field_name_suffix
output_scalar = schema.Scalar((np.float, (1, )), norm)
if net.output_record() is None:
net.set_output_record(
schema.Struct((output_field_name, output_scalar)))
else:
net.AppendOutputRecordField(output_field_name,
output_scalar)
def _init_scratch(self):
self._scratch_blob = {}
self._scratch_status = {}
for blob_name in self._input_blob_names:
scratch_name = self._namescope + blob_name + \
"_scratch_" + self._input_source_name
self._scratch_blob[blob_name] = core.BlobReference(scratch_name)
self._scratch_status[blob_name] = core.BlobReference(scratch_name +
"_status")
# Feed empty arrays to the scratch blobs here, so that there won't be
# race conditions when calling FeedBlob (which calls wworkspace
# CreateBlob()) from enqueue threads
for b in self._scratch_blob.values() + self._scratch_status.values():
workspace.FeedBlob(
b,
np.array([]).astype(np.float32),
device_option=self._device_option,
)
def add_parameter_update_ops(model):
model.Iter("ITER")
LR = model.param_init_net.ConstantFill([],
'LR',
shape=[1],
value=0.1)
for param in model.GetParams():
param_grad = model.param_to_grad[param]
param_momentum = model.param_init_net.ConstantFill([param],
param +
'_momentum',
value=0.0)
model.net.MomentumSGDUpdate(
[param_grad, param_momentum, LR, param],
[param_grad, param_momentum, param],
)
model = cnn.CNNModelHelper(
order="NHWC",
name="test",
)
data_parallel_model.Parallelize_GPU(
model,
input_builder_fun=add_input_ops,
forward_pass_builder_fun=add_model_ops,
param_update_builder_fun=add_parameter_update_ops,
devices=[1, 2, 3],
)
# Only gpu_1 params should be returned (gpu_1 is the first gpu)
checkpoint_params = data_parallel_model.GetCheckpointParams(model)
for p in model.GetParams("gpu_1/"):
self.assertTrue(p in checkpoint_params)
self.assertTrue(p + "_momentum" in checkpoint_params)
for p in model.GetParams("gpu_2/"):
self.assertTrue(p in checkpoint_params)
for c in model.GetComputedParams("gpu_1/"):
self.assertFalse(c in checkpoint_params)
for c in model.GetComputedParams("gpu_2/"):
self.assertFalse(c in checkpoint_params)
self.assertFalse(
core.BlobReference("gpu_1/data") in checkpoint_params)
self.assertTrue(
core.BlobReference("gpu_1/ITER") in checkpoint_params)
def export_q_values(self, net, q_values, action_names, action_name_blob):
batch_size = self.get_batch_size_blob(net, q_values)
feature_lengths_blob = core.BlobReference(
"output/string_weighted_multi_categorical_features.lengths")
net.ConstantFill(
batch_size,
feature_lengths_blob,
value=1,
dtype=core.DataType.INT32,
input_as_shape=1,
)
feature_keys_blob = core.BlobReference(
"output/string_weighted_multi_categorical_features.keys")
net.ConstantFill(
batch_size,
feature_keys_blob,
value=0,
dtype=core.DataType.INT64,
input_as_shape=1,
)
values_lengths_blob = core.BlobReference(
"output/string_weighted_multi_categorical_features.values.lengths")
net.ConstantFill(
batch_size,
values_lengths_blob,
value=len(action_names),
dtype=core.DataType.INT32,
input_as_shape=1,
)
values_keys_blob = core.BlobReference(
"output/string_weighted_multi_categorical_features.values.keys")
net.Tile([action_name_blob, batch_size], values_keys_blob, axis=0)
values_values_blob = core.BlobReference(
"output/string_weighted_multi_categorical_features.values.values")
net.FlattenToVec(q_values, values_values_blob)
net.AddExternalOutput(
feature_lengths_blob,
feature_keys_blob,
values_lengths_blob,
values_keys_blob,
values_values_blob,
)
def _update_param_info(self):
assert len(self._param_info) <= len(self.params)
for param in self.params[len(self._param_info):]:
if not isinstance(param, core.BlobReference):
param = core.BlobReference(str(param), net=self._param_init_net)
self._param_info.append(ParameterInfo(
param_id=len(self._param_info),
param=param,
shape=self._infer_param_shape(param)))
for info in self._param_info:
info.grad = self.param_to_grad.get(info.name)
def __call__(self, net, param_init_net, param, grad=None):
if grad is None:
assert isinstance(param, parameter_info.ParameterInfo)
assert param.grad is not None
else:
if isinstance(param, basestring):
param = core.BlobReference(param)
param = parameter_info.ParameterInfo(
param_id=None, param=param, grad=grad)
self._run(net, param_init_net, param)
def test_create_blobs_queue_db(self):
num_samples = 1000
batch_size = 10
init_net = core.Net('init_net')
net = core.Net('test_create_blobs_queue_db')
queue = init_net.CreateBlobsQueue([], 'queue', capacity=num_samples)
reader = init_net.CreateBlobsQueueDB(
[queue],
'blobs_queue_db_reader',
value_blob_index=0,
timeout_secs=0.1,
)
workspace.RunNetOnce(init_net)
blob = core.BlobReference("blob")
status = core.BlobReference("blob_status")
for i in range(num_samples):
self._add_blob_to_queue(queue, self._create_test_tensor_protos(i),
blob, status)
net.TensorProtosDBInput([reader], ['image', 'label'],
batch_size=batch_size)
workspace.CreateNet(net)
close_net = core.Net('close_net')
close_net.CloseBlobsQueue([queue], [])
for i in range(int(num_samples / batch_size)):
print("Running net, iteration {}".format(i))
with timeout_guard.CompleteInTimeOrDie(2.0):
workspace.RunNet(net)
images = workspace.FetchBlob('image')
labels = workspace.FetchBlob('label')
self.assertEqual(batch_size, len(images))
self.assertEqual(batch_size, len(labels))
for idx, item in enumerate(images):
self.assertEqual(
"foo{}".format(i * batch_size + idx).encode('utf-8'), item)
for item in labels:
self.assertEqual(0, item)
workspace.RunNetOnce(close_net)
def create_net(self):
net = core.Net("feature_extractor")
init_net = core.Net("feature_extractor_init")
missing_scalar = self.create_const(init_net, "MISSING_SCALAR",
MISSING_VALUE)
input_schema = schema.Struct((
"float_features",
schema.Map(
keys=core.BlobReference("input/float_features.keys"),
values=core.BlobReference("input/float_features.values"),
lengths_blob=core.BlobReference(
"input/float_features.lengths"),
),
))
input_record = net.set_input_record(input_schema)
state = self.extract_float_features(
net,
"state",
input_record.float_features,
self.sorted_state_features,
missing_scalar,
)
output_record = schema.Struct(("state", state))
if self.sorted_action_features:
action = self.extract_float_features(
net,
"action",
input_record.float_features,
self.sorted_action_features,
missing_scalar,
)
output_record += schema.Struct(("action", action))
net.set_output_record(output_record)
return FeatureExtractorNet(net, init_net)
def add_fpn_onto_conv_body(model,
conv_body_func,
fpn_level_info_func,
P2only=False):
"""Add the specified conv body to the model and then add FPN levels to it.
"""
# Note: blobs_conv is in revsersed order: [fpn5, fpn4, fpn3, fpn2]
# similarly for dims_conv: [2048, 1024, 512, 256]
# similarly for spatial_scales_fpn: [1/32, 1/16, 1/8, 1/4]
conv_body_func(model)
blobs_fpn, dim_fpn, spatial_scales_fpn = add_fpn(model,
fpn_level_info_func())
if cfg.MODEL.SIBLING_BACKBONE_ON:
# Resnet stage to fork backbone weights
fork_at = cfg.SIBLING.FORK_AT
assert fork_at in [0, 2, 3, 4, 5]
prefix_len = 0
# No shared weights
if fork_at == 0:
fork_node = core.ScopedName("data".format(fork_at, fork_at))
# Fork weights at stage `fork_at`
else:
fork_node = core.ScopedName("res{}_{}_sum".format(
fork_at, fork_at))
ops = model.net.Proto().op
while (fork_node not in ops[prefix_len].output):
prefix_len += 1
prefix_len += 2
temp_net, _ = c2_utils.SuffixNet('temp_net', model.net, prefix_len,
blobs_fpn)
# Preffix sibling backbone
temp_net_preffixed, _ = c2_utils.RenameNet("temp_net_preffixed",
temp_net,
cfg.SIBLING.PREFFIX,
excluded_nodes=[fork_node])
model.AddParams([
core.BlobReference(input_name)
for op in temp_net_preffixed.Proto().op for input_name in op.input
if input_name[-2] == "_"
])
# Merge the backbones
model.net = c2_utils.MergeNets("net", [model.net, temp_net_preffixed])
del temp_net
del temp_net_preffixed
if P2only:
# use only the finest level
return blobs_fpn[-1], dim_fpn, spatial_scales_fpn[-1]
else:
# use all levels
return blobs_fpn, dim_fpn, spatial_scales_fpn
def GetConditionBlobFromNet(condition_net):
"""
The condition blob is the last external_output that must
be a single bool
"""
assert len(condition_net.Proto().external_output) > 0, (
"Condition net %s must has at least one external output" %
condition_net.Proto.name)
# we need to use a blob reference here instead of a string
# otherwise, it will add another name_scope to the input later
# when we create new ops (such as OR of two inputs)
return core.BlobReference(condition_net.Proto().external_output[-1])
def load_from_db(filename, db_type):
# global_init_net in meta_net_def will load parameters from
# predictor_constants.PREDICTOR_DBREADER
create_db = core.CreateOperator(
'CreateDB', [],
[core.BlobReference(predictor_constants.PREDICTOR_DBREADER)],
db=filename,
db_type=db_type)
assert workspace.RunOperatorOnce(create_db), (
'Failed to create db {}'.format(filename))
# predictor_constants.META_NET_DEF is always stored before the parameters
load_meta_net_def = core.CreateOperator(
'Load', [core.BlobReference(predictor_constants.PREDICTOR_DBREADER)],
[core.BlobReference(predictor_constants.META_NET_DEF)])
assert workspace.RunOperatorOnce(load_meta_net_def)
meta_net_def = serde.deserialize_protobuf_struct(
str(workspace.FetchBlob(predictor_constants.META_NET_DEF)),
metanet_pb2.MetaNetDef)
return meta_net_def