def __init__(self, solver_file, architecture_file=None, model_file=None):
solver = caffe.SGDSolver(solver_file)
self.net = solver.net
if model_file and architecture_file:
# Load the parameters from file
pretrained_net = caffe.Net(architecture_file, model_file)
self.net.copy_from(pretrained_net)
# Allocate shared memory for all memory data layers in the network
self.shared_data = []
self.shared_params = []
self.shared_grads = []
self.shmem = {}
for idx, layer in enumerate(self.net.layers):
if layer.type == barista.MEMORY_DATA_LAYER:
self.shared_data.append(SharedData(self.net, idx))
for param in self.net.params:
self.shared_params.append(SharedParameter(self.net, param))
self.shared_grads.append(SharedGradient(self.net, param))
self.batch_size = self.shared_data[0].data.shape[0]
# Allow convenient access to certain Caffe net properties
self.blobs = self.net.blobs
# Create semaphore for interprocess synchronization
self.compute_semaphore = posix_ipc.Semaphore(
"/"+prepend_pid("compute"),
flags=O_CREAT | O_EXCL)
self.model_semaphore = posix_ipc.Semaphore(
"/"+prepend_pid("model"),
flags=O_CREAT | O_EXCL)
def __init__(self, net, layer_idx):
self.data = None
self.label = None
self.shmem = []
self.name = net._layer_names[layer_idx]
handles = self.name.split('-', 1)
# print "Allocating shared memory for %s." % handles[0]
handle = prepend_pid(handles[0])
shmem, arr = create_shmem_ndarray('/'+handle,
net.blobs[handles[0]].data.shape,
np.float32,
flags=posix_ipc.O_CREAT)
self.data = arr
self.shmem.append(shmem)
if len(handles) == 2:
handle = prepend_pid(handles[1])
shmem, arr = create_shmem_ndarray('/'+handle,
net.blobs[handles[1]].data.shape,
np.float32,
flags=posix_ipc.O_CREAT)
self.label = arr
self.shmem.append(shmem)
net.set_input_arrays(self.data, self.label, layer_idx)
else:
if SharedData._null_array is None:
SharedData._null_array = np.empty(
(self.data.shape[0], 1, 1, 1),
dtype=np.float32)
print("[SharedData] Warning: didn't specify a handle for the "
"label in layer", layer_idx, ". Should not be used in net.",
file=sys.stderr)
net.set_input_arrays(self.data, SharedData._null_array, layer_idx)
def __init__(self, net, param_name):
self.caffe_grads = []
self.shared_grads = []
self.shmem = []
for i, param in enumerate(net.params[param_name]):
# Typically, we'll have two, a weight and bias.
handle = prepend_pid(param_name + '_' + SharedGradient.POSTFIX[i])
shmem, arr = create_shmem_ndarray('/' + handle,
param.diff.shape,
np.float32,
flags=posix_ipc.O_CREAT)
self.caffe_grads.append(param.diff)
self.shared_grads.append(arr)
self.shmem.append(shmem)
