def get_tagging_channels_from_state(self, state, target):
missingValuesFilter = T.neq(target, -1)
rval = OrderedDict()
y_hat = state > 0.5
y = target > 0.5
wrong_bit = T.cast(T.neq(y, y_hat), state.dtype) * missingValuesFilter
rval['mistagging'] = T.cast(wrong_bit.sum() / missingValuesFilter.sum(),
state.dtype)
y = T.cast(y, state.dtype)
y_hat = T.cast(y_hat, state.dtype)
tp = (y * y_hat * missingValuesFilter).sum()
fp = ((1-y) * y_hat * missingValuesFilter).sum()
precision = tp / T.maximum(1., tp + fp)
recall = tp / T.maximum(1., (y * missingValuesFilter).sum())
rval['precision'] = precision
rval['recall'] = recall
rval['f1'] = 2. * precision * recall / T.maximum(1, precision + recall)
tp = (y * y_hat * missingValuesFilter).sum(axis=0)
fp = ((1-y) * y_hat * missingValuesFilter).sum(axis=0)
precision = tp / T.maximum(1., tp + fp)
rval['per_output_precision.max'] = precision.max()
rval['per_output_precision.mean'] = precision.mean()
rval['per_output_precision.min'] = precision.min()
recall = tp / T.maximum(1., (y * missingValuesFilter).sum(axis=0))
rval['per_output_recall.max'] = recall.max()
rval['per_output_recall.mean'] = recall.mean()
rval['per_output_recall.min'] = recall.min()
f1 = 2. * precision * recall / T.maximum(1, precision + recall)
rval['per_output_f1.max'] = f1.max()
rval['per_output_f1.mean'] = f1.mean()
rval['per_output_f1.min'] = f1.min()
# Add computation of the mean average recision
from pylearn2_ECCV2014 import meanAvgPrec
(rval['min_avg_prec'],
rval['mean_avg_prec'],
rval['max_avg_prec'],
rval['mean_avg_prec_AnswerPhone'],
rval['mean_avg_prec_DriveCar'],
rval['mean_avg_prec_Eat'],
rval['mean_avg_prec_FightPerson'],
rval['mean_avg_prec_GetOutCar'],
rval['mean_avg_prec_HandShake'],
rval['mean_avg_prec_HugPerson'],
rval['mean_avg_prec_Kiss'],
rval['mean_avg_prec_Run'],
rval['mean_avg_prec_SitDown'],
rval['mean_avg_prec_SitUp'],
rval['mean_avg_prec_StandUp']) = meanAvgPrec.meanAveragePrecisionTheano(target, state)
return rval
def get_tagging_channels_from_state(self, state, target):
missingValuesFilter = T.neq(target, -1)
rval = OrderedDict()
y_hat = state > 0.5
y = target > 0.5
wrong_bit = T.cast(T.neq(y, y_hat), state.dtype) * missingValuesFilter
rval['mistagging'] = T.cast(
wrong_bit.sum() / missingValuesFilter.sum(), state.dtype)
y = T.cast(y, state.dtype)
y_hat = T.cast(y_hat, state.dtype)
tp = (y * y_hat * missingValuesFilter).sum()
fp = ((1 - y) * y_hat * missingValuesFilter).sum()
precision = tp / T.maximum(1., tp + fp)
recall = tp / T.maximum(1., (y * missingValuesFilter).sum())
rval['precision'] = precision
rval['recall'] = recall
rval['f1'] = 2. * precision * recall / T.maximum(1, precision + recall)
tp = (y * y_hat * missingValuesFilter).sum(axis=0)
fp = ((1 - y) * y_hat * missingValuesFilter).sum(axis=0)
precision = tp / T.maximum(1., tp + fp)
rval['per_output_precision.max'] = precision.max()
rval['per_output_precision.mean'] = precision.mean()
rval['per_output_precision.min'] = precision.min()
recall = tp / T.maximum(1., (y * missingValuesFilter).sum(axis=0))
rval['per_output_recall.max'] = recall.max()
rval['per_output_recall.mean'] = recall.mean()
rval['per_output_recall.min'] = recall.min()
f1 = 2. * precision * recall / T.maximum(1, precision + recall)
rval['per_output_f1.max'] = f1.max()
rval['per_output_f1.mean'] = f1.mean()
rval['per_output_f1.min'] = f1.min()
# Add computation of the mean average recision
from pylearn2_ECCV2014 import meanAvgPrec
(rval['min_avg_prec'], rval['mean_avg_prec'], rval['max_avg_prec'],
rval['mean_avg_prec_AnswerPhone'], rval['mean_avg_prec_DriveCar'],
rval['mean_avg_prec_Eat'], rval['mean_avg_prec_FightPerson'],
rval['mean_avg_prec_GetOutCar'], rval['mean_avg_prec_HandShake'],
rval['mean_avg_prec_HugPerson'], rval['mean_avg_prec_Kiss'],
rval['mean_avg_prec_Run'], rval['mean_avg_prec_SitDown'],
rval['mean_avg_prec_SitUp'], rval['mean_avg_prec_StandUp']
) = meanAvgPrec.meanAveragePrecisionTheano(target, state)
return rval
idx[which["test"]] = numpy.cast["int"](numpy.loadtxt(path + "test/Newtest.txt"))
num_feats = numpy.cast["int"](numpy.loadtxt("NumFeatures.txt"))
x = theano.tensor.tensor4()
y_hat = theano.tensor.fmatrix()
y = theano.tensor.fmatrix()
predict_fn = [0] * len(f2i)
for i in range(len(model)):
model[i] = pickle.load(file(path + "latest2class3_" + str(i2f[i]) + ".pkl"))
dset[which["train"]][i] = adn.AtousaDataset("train", num_feats=i2f[i])
dset[which["test"]][i] = adn.AtousaDataset("test", num_feats=i2f[i])
predict_fn[i] = theano.function([x], model[i].fprop(x))
theano_map = theano.function([y_hat, y], meanAveragePrecisionTheano(y_hat, y))
def predict(data, i):
transform = lambda x: x
if voting_mode:
transform = lambda x: onehot(x)
return transform(numpy.array(predict_fn[i](data)))
def mean_ap(target, pred):
return numpy.array(theano_map(target, pred))
def onehot(pred):
num_feats = numpy.cast['int'](numpy.loadtxt('NumFeatures.txt'))
x = theano.tensor.tensor4()
y_hat = theano.tensor.fmatrix()
y = theano.tensor.fmatrix()
predict_fn = [0] * len(f2i)
for i in range(len(model)):
model[i] = pickle.load(file(path + 'latest2class3_' + str(i2f[i]) +
'.pkl'))
dset[which['train']][i] = adn.AtousaDataset('train', num_feats=i2f[i])
dset[which['test']][i] = adn.AtousaDataset('test', num_feats=i2f[i])
predict_fn[i] = theano.function([x], model[i].fprop(x))
theano_map = theano.function([y_hat, y], meanAveragePrecisionTheano(y_hat, y))
def predict(data, i):
transform = lambda x: x
if voting_mode:
transform = lambda x: onehot(x)
return transform(numpy.array(predict_fn[i](data)))
def mean_ap(target, pred):
return numpy.array(theano_map(target, pred))
def onehot(pred):
