def __init__(self):
self.jobs = []
self.active_jobs = []
self.apollo_net = ApolloNet()
self.solver = None
class Net:
def __init__(self):
self.jobs = []
self.active_jobs = []
self.apollo_net = ApolloNet()
self.solver = None
def add(self, job, states=None):
if states is not None:
job.states = lambda: states
self.jobs.append(job)
def forward(self, state, input_data={}):
for job in self.jobs:
if state in job.states():
if job.is_data_job():
job.run(self.apollo_net, input_data)
else:
job.run(self.apollo_net)
self.active_jobs.append(job)
def backward(self):
self.apollo_net.backward()
@property
def params(self):
return self.apollo_net.params
@property
def tops(self):
return self.apollo_net.tops
@property
def layers(self):
return self.apollo_net.layers
def load(self, weight_file):
self.apollo_net.load(weight_file)
def fit(self, train_data, val_data=None):
if self.solver is None:
raise AttributeError("Solver must be set in order to train.")
self.solver.solve(train_data, val_data)
def predict(self, data, tops=[], batch_size=1):
assert not isinstance(tops, str)
batch = None
outputs = None
for instance in data:
if batch == None:
batch = {k: [v] for k, v in instance.items()}
else:
for k, v in instance.iteritems():
batch[k].append(v)
if len(batch.values()[0]) == batch_size:
self.apollo_net.clear_forward()
self.forward('deploy', batch)
if outputs is None:
outputs = {k: np.copy(self.tops[k].data) for k in tops}
else:
for k in tops:
outputs[k] = np.vstack(
(outputs[k], np.copy(self.tops[k].data)))
batch = None
return outputs
def save(self, weight_file):
self.apollo_net.save(weight_file)
def draw_to_file(self, filename, rankdir='LR', require_nonempty=True):
self.apollo_net.draw_to_file(self, filename, rankdir='LR', require_nonempty=True)