def _train_cross_valid(self, _skiptrain=False):
df = DataFeeder(*(self.traindata + [self.traingold]))
splitter = SplitIdxIterator(df.size,
split=self.validsplits,
random=self.validrandom,
folds=self.validsplits)
err = []
verr = []
c = 0
for splitidxs in splitter:
trainf = self.buildtrainfun(self.model)
validf = self.getvalidfun(self.model)
tf, vf = df.isplit(splitidxs, df_randoms=(True, False))
tf.numbats(self.numbats)
vf.batsize = tf.batsize
serr, sverr = self.trainloop(trainf=self.getbatchloop(trainf, tf),
validf=self.getbatchloop(validf, vf),
_skiptrain=_skiptrain)
err.append(serr)
verr.append(sverr)
self.resetmodel(self.model)
err = np.asarray(err)
avgerr = np.mean(err, axis=0)
verr = np.asarray(verr)
avgverr = np.mean(verr, axis=0)
self.tt.tock("done")
return avgerr, avgverr, err, verr
def _train_split(self):
trainf = self.buildtrainfun(self.model)
validf = self.buildvalidfun(self.model)
df = DataFeeder(*(self.traindata + [self.traingold]))
dftrain, dfvalid = df.split(self.validsplits, self.validrandom)
err, verr = self.trainloop(
trainf=self.getbatchloop(trainf, dftrain.numbats(self.numbats)),
validf=self.getbatchloop(validf, dfvalid))
return err, verr, None, None
def _train_split(self):
trainf = self.buildtrainfun(self.model)
validf = self.buildvalidfun(self.model)
df = DataFeeder(*(self.traindata + [self.traingold]))
dftrain, dfvalid = df.split(self.validsplits, self.validrandom)
err, verr = self.trainloop(trainf=self.getbatchloop(
trainf, dftrain.numbats(self.numbats)),
validf=self.getbatchloop(validf, dfvalid))
return err, verr, None, None
def _train_validdata(self):
validf = self.buildvalidfun(self.model)
trainf = self.buildtrainfun(self.model)
df = DataFeeder(*(self.traindata + [self.traingold]))
vdf = DataFeeder(*(self.validdata + [self.validgold]))
#dfvalid = df.osplit(split=self.validsplits, random=self.validrandom)
err, verr = self.trainloop(
trainf=self.getbatchloop(trainf, df.numbats(self.numbats)),
validf=self.getbatchloop(validf, vdf))
return err, verr, None, None
def _train_validdata(self):
validf = self.buildvalidfun(self.model)
trainf = self.buildtrainfun(self.model)
df = DataFeeder(*(self.traindata + [self.traingold]))
vdf = DataFeeder(*(self.validdata + [self.validgold]))
#embed()
#dfvalid = df.osplit(split=self.validsplits, random=self.validrandom)
err, verr = self.trainloop(trainf=self.getbatchloop(
trainf, df.numbats(self.numbats)),
validf=self.getbatchloop(validf, vdf))
return err, verr, None, None
def _train_split(self, _skiptrain=False):
trainf = self.buildtrainfun(self.model)
validf = self.getvalidfun(self.model)
df = DataFeeder(*(self.traindata + [self.traingold]))
dftrain, dfvalid = df.split(self.validsplits,
self.validrandom,
df_randoms=(True, False))
dftrain.numbats(self.numbats)
dfvalid.batsize = dftrain.batsize
err, verr = self.trainloop(trainf=self.getbatchloop(trainf, dftrain),
validf=self.getbatchloop(validf, dfvalid),
_skiptrain=_skiptrain)
return err, verr, None, None
def _train_validdata(self, _skiptrain=False):
validf = self.getvalidfun(self.model)
trainf = self.buildtrainfun(self.model)
df = DataFeeder(*(self.traindata + [self.traingold])).numbats(
self.numbats)
vdf = DataFeeder(*(self.validdata + [self.validgold]), random=False)
vdf.batsize = df.batsize
#embed()
#dfvalid = df.osplit(split=self.validsplits, random=self.validrandom)
err, verr = self.trainloop(trainf=self.getbatchloop(trainf, df),
validf=self.getbatchloop(validf, vdf),
_skiptrain=_skiptrain)
return err, verr, None, None
def _train_full(self): # train on all data, no validation
trainf = self.buildtrainfun(self.model)
err, _ = self.trainloop(trainf=self.getbatchloop(
trainf,
DataFeeder(*(self.traindata +
[self.traingold])).numbats(self.numbats)))
return err, None, None, None
def test_fb_datafeed_validosplit(self):
gd, gmaxi = getglovedict(os.path.join(os.path.dirname(__file__), "../data/glove/miniglove.50d.txt"))
ed, emaxid = getentdict(os.path.join(os.path.dirname(__file__), "../data/freebase/entdic.small.map"), top=50)
dp = os.path.join(os.path.dirname(__file__), "../data/freebase/labelsrevlex.map.sample")
f = FreebaseEntFeedsMaker(dp, gd, ed, numwords=10, numchars=30)
self.assertEqual(f.worddic, gd)
dfeeder = DataFeeder(*([f.trainfeed] + [f.goldfeed]))
splits = 1
dfsplit = dfeeder.osplit(split=splits, random=False)
dfeeds = dfeeder.feeds
splitfeeds = dfsplit.feeds
for x, y in zip(dfeeds, splitfeeds):
self.assertEqual(x.__class__, y.__class__)
self.assertEqual(x.ndim, y.ndim)
self.assertEqual(y.shape[0], int(math.ceil(1.*x.shape[0]/splits)))
for dim in range(1, len(x.shape)):
self.assertEqual(x.shape[dim], y.shape[dim])
def _train_split(self, _lambda=False, _skiptrain=False):
df = DataFeeder(*(self.traindata + [self.traingold]))
dftrain, dfvalid = df.split(self.validsplits,
self.validrandom,
df_randoms=(True, False))
dftrain.numbats(self.numbats)
dfvalid.batsize = dftrain.batsize
trainf = self.buildtrainfun(self.model, dftrain.batsize)
validf = self.getvalidfun(self.model, dfvalid.batsize)
if _lambda:
return trainf, validf, dftrain, dfvalid
else:
err, verr = self.trainloop(trainf=self.getbatchloop(trainf,
dftrain,
phase="TRAIN"),
validf=self.getbatchloop(validf,
dfvalid,
phase="VALID"),
_skiptrain=_skiptrain)
return err, verr, None, None
def _train_validdata(self, _lambda=False, _skiptrain=False):
df = DataFeeder(*(self.traindata + [self.traingold])).numbats(
self.numbats)
vdf = DataFeeder(*(self.validdata + [self.validgold]), random=False)
vdf.batsize = df.batsize
trainf = self.buildtrainfun(self.model, df.batsize)
validf = self.getvalidfun(self.model, vdf.batsize)
#embed()
#dfvalid = df.osplit(split=self.validsplits, random=self.validrandom)
if _lambda:
return trainf, validf, df, vdf
else:
err, verr = self.trainloop(trainf=self.getbatchloop(trainf,
df,
phase="TRAIN"),
validf=self.getbatchloop(validf,
vdf,
phase="VALID"),
_skiptrain=_skiptrain)
return err, verr, None, None
def _train_cross_valid(self):
df = DataFeeder(*(self.traindata + [self.traingold]))
splitter = SplitIdxIterator(df.size, split=self.validsplits, random=self.validrandom, folds=self.validsplits)
err = []
verr = []
c = 0
for splitidxs in splitter:
trainf = self.buildtrainfun(self.model)
validf = self.buildvalidfun(self.model)
tf, vf = df.isplit(splitidxs)
serr, sverr = self.trainloop(
trainf=self.getbatchloop(trainf, tf.numbats(self.numbats)),
validf=self.getbatchloop(validf, vf))
err.append(serr)
verr.append(sverr)
self.resetmodel(self.model)
err = np.asarray(err)
avgerr = np.mean(err, axis=0)
verr = np.asarray(verr)
avgverr = np.mean(verr, axis=0)
self.tt.tock("done")
return avgerr, avgverr, err, verr
def _train_full(self,
_lambda=False,
_skiptrain=False): # on all data, no validation
df = DataFeeder(*(self.traindata + [self.traingold])).numbats(
self.numbats)
trainf = self.buildtrainfun(self.model, df.batsize)
if _lambda:
return trainf, None, df, None
else:
err, _ = self.trainloop(trainf=self.getbatchloop(trainf,
df,
phase="TRAIN"),
_skiptrain=_skiptrain)
return err, None, None, None
def test_fb_datafeed_validosplit(self):
gd, gmaxi = getglovedict(
os.path.join(os.path.dirname(__file__),
"../data/glove/miniglove.50d.txt"))
ed, emaxid = getentdict(os.path.join(
os.path.dirname(__file__), "../data/freebase/entdic.small.map"),
top=50)
dp = os.path.join(os.path.dirname(__file__),
"../data/freebase/labelsrevlex.map.sample")
f = FreebaseEntFeedsMaker(dp, gd, ed, numwords=10, numchars=30)
self.assertEqual(f.worddic, gd)
dfeeder = DataFeeder(*([f.trainfeed] + [f.goldfeed]))
splits = 1
dfsplit = dfeeder.osplit(split=splits, random=False)
dfeeds = dfeeder.feeds
splitfeeds = dfsplit.feeds
for x, y in zip(dfeeds, splitfeeds):
self.assertEqual(x.__class__, y.__class__)
self.assertEqual(x.ndim, y.ndim)
self.assertEqual(y.shape[0],
int(math.ceil(1. * x.shape[0] / splits)))
for dim in range(1, len(x.shape)):
self.assertEqual(x.shape[dim], y.shape[dim])
