def fprop(self,X):
"""
dimensions have to be in order (batch_size,48,48)
"""
# if X is a single image flattened or not
if tuple(X.shape) == (48,48) or len(X.shape)==1:
X = X.reshape((1,48,48,1))
else:
X = X.reshape((X.shape[0],48,48,1))
return utils.apply(self._fprop,X,self.batch_size)
def fprop(self, X):
"""
dimensions have to be in order (batch_size,48,48)
"""
# if X is a single image flattened or not
if tuple(X.shape) == (48, 48) or len(X.shape) == 1:
X = X.reshape((1, 48, 48, 1))
else:
X = X.reshape((X.shape[0], 48, 48, 1))
return utils.apply(self._fprop, X, self.batch_size)
def get_features(model_path, data, layer_idx=None):
model = serial.load(model_path)
batch_size = model.get_test_batch_size()
X = model.get_input_space().make_batch_theano()
rval = X
if layer_idx is None:
layers = model.layers
else:
layers = model.layers[:layer_idx]
for layer in layers:
rval = layer.fprop(rval)
f = function([X], rval)
return utils.apply(f, data, batch_size)
def get_features(model_path,data,layer_idx=None):
model = serial.load(model_path)
batch_size = model.get_test_batch_size()
X = model.get_input_space().make_batch_theano()
rval = X
if layer_idx is None:
layers = model.layers
else:
layers = model.layers[:layer_idx]
for layer in layers:
rval = layer.fprop(rval)
f = function([X], rval)
return utils.apply(f,data,batch_size)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--out", help="The directory where to extract predictions")
parser.add_argument("--data", action="append", required=True, help="Data path (.npy)")
parser.add_argument("--targets", action="append", help="Target path (.npy)")
parser.add_argument("--normalize", default=False)
parser.add_argument("model_paths", nargs="+", help="Pylearn2 models")
options = parser.parse_args()
out = options.out
data_paths = options.data
model_paths = options.model_paths
normalize = "True" == options.normalize
print model_paths
print data_paths
datas = [(data_path, np.load(data_path)) for data_path in data_paths]
from theano import config
from theano import function
from pylearn2.utils import serial
from emotiw.bouthilx.datasets import FeaturesDataset
print np.mean(
datas[0][1]
== FeaturesDataset(
features_paths=[
"audio/test26_best_audio_mlp_valid_feats_features.npy",
"pascal/afew2_valid_rmananinpic_emmanuel_features.npy",
],
targets_path="pascal/afew2_valid_rmananinpic_emmanuel_targets.npy",
base_path="/data/afew",
normalize=True,
shuffle=False,
).get_design_matrix()
)
predictions = []
if len(datas) < len(model_paths) and len(datas) == 1:
l = []
for model_path in model_paths:
l.append((model_path, datas[0]))
iterator = l
else:
iterator = zip(model_paths, datas)
for model_path, (data_path, data) in iterator:
model = serial.load(model_path)
batch_size = model.get_test_batch_size()
X = model.get_input_space().make_batch_theano()
f = function([X], model.fprop(X))
if normalize:
# set 0s to mean
tmp = data[:]
tmp -= tmp.min(0)
tmp /= (tmp == 0) + tmp.max(0)
# range is [-1,1]
tmp = tmp * 2.0 - 1.0
tmp = (data.min(0) != data.max(0)) * tmp # set empty dimensions to 0
data = tmp
# print data
y_hat = utils.apply(f, data, batch_size)
# print y_hat.sum(1)
if options.targets is not None:
targets = np.load(options.targets[0])
# print targets
# print y_hat.argmax(1)
# j = 0
# for i, target in enumerate(targets):
# j += classes[int(target)]==classes[y_hat[i].argmax()]
# print classes[int(target)], classes[y_hat[i].argmax()], j, len(targets)
#
# print np.mean(y_hat.argmax(1)==targets)
predictions.append(y_hat)
mean = np.array(predictions).mean(0)
if options.targets is not None:
targets = np.load(options.targets[0])
# print mean.argmax(1)
# print targets
print np.mean(mean.argmax(1) == targets)
j = 0
# for i, target in enumerate(targets):
# j += classes[int(target)]==classes[mean[i].argmax()]
# print classes[int(target)], classes[mean[i].argmax()], j, len(targets)
if options.out is not None:
np.save(options.out, mean)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--out",
help="The directory where to extract predictions")
parser.add_argument("--data",
action='append',
required=True,
help="Data path (.npy)")
parser.add_argument("--targets",
action='append',
help="Target path (.npy)")
parser.add_argument("--normalize", default=False)
parser.add_argument("model_paths", nargs='+', help="Pylearn2 models")
options = parser.parse_args()
out = options.out
data_paths = options.data
model_paths = options.model_paths
normalize = "True" == options.normalize
print model_paths
print data_paths
datas = [(data_path, np.load(data_path)) for data_path in data_paths]
from theano import config
from theano import function
from pylearn2.utils import serial
from emotiw.bouthilx.datasets import FeaturesDataset
print np.mean(datas[0][1] == FeaturesDataset(
features_paths=[
"audio/test26_best_audio_mlp_valid_feats_features.npy",
"pascal/afew2_valid_rmananinpic_emmanuel_features.npy"
],
targets_path="pascal/afew2_valid_rmananinpic_emmanuel_targets.npy",
base_path="/data/afew",
normalize=True,
shuffle=False).get_design_matrix())
predictions = []
if len(datas) < len(model_paths) and len(datas) == 1:
l = []
for model_path in model_paths:
l.append((model_path, datas[0]))
iterator = l
else:
iterator = zip(model_paths, datas)
for model_path, (data_path, data) in iterator:
model = serial.load(model_path)
batch_size = model.get_test_batch_size()
X = model.get_input_space().make_batch_theano()
f = function([X], model.fprop(X))
if normalize:
# set 0s to mean
tmp = data[:]
tmp -= tmp.min(0)
tmp /= (tmp == 0) + tmp.max(0)
# range is [-1,1]
tmp = tmp * 2.0 - 1.0
tmp = (data.min(0) !=
data.max(0)) * tmp # set empty dimensions to 0
data = tmp
# print data
y_hat = utils.apply(f, data, batch_size)
# print y_hat.sum(1)
if options.targets is not None:
targets = np.load(options.targets[0])
# print targets
# print y_hat.argmax(1)
# j = 0
# for i, target in enumerate(targets):
# j += classes[int(target)]==classes[y_hat[i].argmax()]
# print classes[int(target)], classes[y_hat[i].argmax()], j, len(targets)
#
# print np.mean(y_hat.argmax(1)==targets)
predictions.append(y_hat)
mean = np.array(predictions).mean(0)
if options.targets is not None:
targets = np.load(options.targets[0])
# print mean.argmax(1)
# print targets
print np.mean(mean.argmax(1) == targets)
j = 0
# for i, target in enumerate(targets):
# j += classes[int(target)]==classes[mean[i].argmax()]
# print classes[int(target)], classes[mean[i].argmax()], j, len(targets)
if options.out is not None:
np.save(options.out, mean)