def __init__(self,
prototxt,
weights,
inputs,
outputs,
volume_specs=None,
use_gpu=None):
for f in [prototxt, weights]:
if not os.path.isfile(f):
raise RuntimeError("%s does not exist" % f)
self.prototxt = prototxt
self.weights = weights
self.inputs = inputs
self.outputs = outputs
if use_gpu is not None:
logger.debug("Predict process: using GPU %d" % use_gpu)
caffe.enumerate_devices(False)
caffe.set_devices((use_gpu, ))
caffe.set_mode_gpu()
caffe.select_device(use_gpu, False)
self.net = caffe.Net(self.prototxt, self.weights, caffe.TEST)
self.net_io = NetIoWrapper(self.net, self.outputs.values())
def start(self):
logger.info("Initializing solver...")
if self.use_gpu is not None:
logger.debug("Predict process: using GPU %d" % self.use_gpu)
caffe.enumerate_devices(False)
caffe.set_devices((self.use_gpu, ))
caffe.set_mode_gpu()
caffe.select_device(self.use_gpu, False)
self.net = caffe.Net(self.prototxt, self.weights, caffe.TEST)
self.net_io = NetIoWrapper(self.net, self.outputs.keys())
def __predict(self, use_gpu):
if not self.net_initialized:
logger.info("Initializing solver...")
if use_gpu is not None:
logger.debug("Predict process: using GPU %d" % use_gpu)
caffe.enumerate_devices(False)
caffe.set_devices((use_gpu, ))
caffe.set_mode_gpu()
caffe.select_device(use_gpu, False)
self.net = caffe.Net(self.prototxt, self.weights, caffe.TEST)
self.net_io = NetIoWrapper(self.net)
self.net_initialized = True
start = time.time()
batch = self.batch_in.get()
fetch_time = time.time() - start
self.net_io.set_inputs({
'data':
batch.volumes[VolumeTypes.RAW].data[np.newaxis, np.newaxis, :],
})
self.net.forward()
output = self.net_io.get_outputs()
predict_time = time.time() - start
logger.info(
"Predict process: time=%f (including %f waiting for batch)" %
(predict_time, fetch_time))
assert len(
output['aff_pred'].shape
) == 5, "Got affinity prediction with unexpected number of dimensions, should be 1 (direction) + 3 (spatial) + 1 (batch, not used), but is %d" % len(
output['aff_pred'].shape)
batch.volumes[VolumeTypes.PRED_AFFINITIES] = Volume(
output['aff_pred'][0], Roi(), (1, 1, 1))
return batch
def start(self):
logger.info("Initializing solver...")
if self.use_gpu is not None:
logger.debug("Train process: using GPU %d", self.use_gpu)
caffe.enumerate_devices(False)
caffe.set_devices((self.use_gpu, ))
caffe.set_mode_gpu()
caffe.select_device(self.use_gpu, False)
self.solver = caffe.get_solver(self.solver_parameters)
if self.solver_parameters.resume_from is not None:
logger.debug("Train process: restoring solver state from %s",
self.solver_parameters.resume_from)
self.solver.restore(self.solver_parameters.resume_from)
names_net_outputs = self.outputs.keys() + self.gradients.keys()
self.net_io = NetIoWrapper(self.solver.net, names_net_outputs)
def __init__(self, prototxt, weights, input_key, output_key, gpu):
# TODO validate that gpu is available
assert os.path.exists(prototxt)
assert os.path.exists(weights)
for f in [prototxt, weights]:
if not os.path.isfile(f):
raise RuntimeError("%s does not exist" % f)
self.prototxt = prototxt
self.weights = weights
self.input_key = input_key
self.output_key = output_key
caffe.enumerate_devices(False)
caffe.set_devices((gpu, ))
caffe.set_mode_gpu()
caffe.select_device(gpu, False)
self.net = caffe.Net(self.prototxt, self.weights, caffe.TEST)
self.net_io = NetIoWrapper(self.net, [self.output_key])
def __train(self, use_gpu):
start = time.time()
if not self.solver_initialized:
logger.info("Initializing solver...")
if use_gpu is not None:
logger.debug("Train process: using GPU %d" % use_gpu)
caffe.enumerate_devices(False)
caffe.set_devices((use_gpu, ))
caffe.set_mode_gpu()
caffe.select_device(use_gpu, False)
self.solver = caffe.get_solver(self.solver_parameters)
if self.solver_parameters.resume_from is not None:
logger.debug("Train process: restoring solver state from " +
self.solver_parameters.resume_from)
self.solver.restore(self.solver_parameters.resume_from)
self.net_io = NetIoWrapper(self.solver.net)
self.solver_initialized = True
batch, request = self.batch_in.get()
data = {
'data':
batch.volumes[VolumeType.RAW].data[np.newaxis, np.newaxis, :],
'aff_label':
batch.volumes[VolumeType.GT_AFFINITIES].data[np.newaxis, :],
}
if self.solver_parameters.train_state.get_stage(0) == 'euclid':
logger.debug(
"Train process: preparing input data for Euclidean training")
self.__prepare_euclidean(batch, data)
else:
logger.debug(
"Train process: preparing input data for Malis training")
self.__prepare_malis(batch, data)
self.net_io.set_inputs(data)
loss = self.solver.step(1)
# self.__consistency_check()
output = self.net_io.get_outputs()
batch.volumes[VolumeType.PRED_AFFINITIES] = Volume(
output['aff_pred'][0],
batch.volumes[VolumeType.GT_AFFINITIES].roi,
batch.volumes[VolumeType.GT_AFFINITIES].resolution,
interpolate=True)
batch.loss = loss
batch.iteration = self.solver.iter
if VolumeType.LOSS_GRADIENT in request.volumes:
diffs = self.net_io.get_output_diffs()
batch.volumes[VolumeType.LOSS_GRADIENT] = Volume(
diffs['aff_pred'][0],
batch.volumes[VolumeType.GT_AFFINITIES].roi,
batch.volumes[VolumeType.GT_AFFINITIES].resolution,
interpolate=True)
time_of_iteration = time.time() - start
logger.info("Train process: iteration=%d loss=%f time=%f" %
(self.solver.iter, batch.loss, time_of_iteration))
return batch