|
from hypergan.gan_component import GANComponent
import tensorflow as tf
class MultiComponent():
"""
Used to combine multiple components into one. For example, `gan.loss = MultiComponent([loss1, loss2])`
"""
def __init__(self, components=[], combine='concat'):
self.components = components
self.gan = components[0].gan
self._combine = combine
self._cache = {}
def __getattr__(self, name):
if len(self.components) == 0:
return None
attributes = self.lookup(name)
self._cache[name] = self.combine(name, attributes)
return self._cache[name]
def lookup(self, name):
lookups = []
for component in self.components:
if hasattr(component, name):
lookups.append(getattr(component,name))
else:
print("Warning:Skipping lookup of "+name+" because None was returned")
return lookups
def combine(self, name, data):
if data == None or data == []:
return data
if isinstance(data[0], type({})):
full_dict = {}
for d in data:
full_dict.update(d)
return full_dict
# loss functions return [d_loss, g_loss]. We combine columnwise.
if isinstance(data, list) and isinstance(data[0], list) and isinstance(data[0][0], tf.Tensor):
if(name in self._cache):
return self._cache[name]
result = []
for j,_ in enumerate(data[0]):
column = []
for i,_ in enumerate(data):
column.append(data[i][j])
reduction = self.reduce(column)
result.append(reduction)
return result
if type(data[0]) == tf.Tensor:
if(name in self._cache):
return self._cache[name]
return self.reduce(data)
if callable(data[0]):
return self.call_each(data)
return data
def reduce(self, data):
data = [d for d in data if d is not None]
ops = self.gan.ops
if self._combine == 'concat':
return self.gan.ops.concat(values=data, axis=1)
elif self._combine == 'add':
data = [ops.reshape(d,ops.shape(data[0])) for d in data]
return self.gan.ops.add_n(data)
raise "Unknown combine"
def call_each(self, methods):
def do_call(*args, **kwargs):
results = []
for method in methods:
results.append(method(*args, **kwargs))
return self.combine(str(method), results)
return do_call
|
