too lazy to fix merge conflict, fix it if you need this later
probably easier to just rewrite from scratch
<<<<<<< HEAD
import numpy as np
x = raw_input('use esp game dataset?')
if x == 'y':
from galatea.esp import Im2Word
dataset = Im2Word(start=99000, stop=100000
)
else:
from galatea.esp import FinalIm2Word
dataset = FinalIm2Word()
x = raw_input('use people?')
people_mask = np.zeros((dataset.y.shape[0],), dtype='bool')
for word in ['face', 'person', 'people', 'man', 'men', 'woman', 'women',
'child', 'children', 'kid', 'kids', 'guy', 'boy', 'boys', 'girl',
'girls', 'infant', 'infants', 'baby', 'babies']:
if word not in dataset.words:
continue
idx = dataset.words.index(word)
mask = dataset.y[:,idx].astype('bool')
people_mask = people_mask | mask
if x == 'y':
mask = people_mask
from galatea.esp import FinalIm2Word
import numpy as np
dataset = FinalIm2Word(start=0, stop=500
)
max_word = 50
dataset.y = dataset.y[:, 0:max_word]
from pylearn2.utils import serial
model = serial.load('rectifier_7.pkl')
last_layer = model.layers[-1]
last_layer.b.set_value(last_layer.b.get_value()[0:max_word])
W, = last_layer.transformer.get_params()
W.set_value(W.get_value()[:,0:max_word])
import theano.tensor as T
X = T.matrix()
X.tag.test_value = dataset.X
state = model.fprop(X)
target = T.matrix()
target.tag.test_value = dataset.y
wrong_target = T.matrix()
wrong_target.tag.test_value = dataset.y
right_cost = model.layers[-1].kl(Y=target, Y_hat=state)
wrong_cost = model.layers[-1].kl(Y=wrong_target, Y_hat=state)
from galatea.esp import FinalIm2Word
valid = FinalIm2Word()
from pylearn2.utils import serial
model = serial.load('rectifier_7.pkl')
import theano.tensor as T
X = T.matrix()
state = model.fprop(X)
target = T.matrix()
y_hat = state > 0.5
y = target > 0.5
y = T.cast(y, state.dtype)
y_hat = T.cast(y_hat, state.dtype)
tp = (y * y_hat).sum(axis=0)
fp = ((1 - y) * y_hat).sum(axis=0)
precision = tp / T.maximum(1., tp + fp)
recall = tp / T.maximum(1., y.sum(axis=0))
f1 = 2. * precision * recall / T.maximum(1, precision + recall)
from theano import function
ttp = tp.sum()
tpredp = y.sum()
from galatea.esp import FinalIm2Word
import numpy as np
dataset = FinalIm2Word(start=00, stop=1000)
from pylearn2.utils import serial
model = serial.load('rectifier_7_best.pkl')
import theano.tensor as T
X = T.matrix()
state = model.fprop(X)
target = T.matrix()
wrong_target = T.matrix()
right_cost = model.layers[-1].kl(Y=target, Y_hat=state)
wrong_cost = model.layers[-1].kl(Y=wrong_target, Y_hat=state)
acc = (wrong_cost > right_cost).mean()
from theano import function
f = function([X, target, wrong_target], acc)
wrong_target = dataset.y.copy()
used = np.zeros((dataset.y.shape[0], ), dtype='bool')
minmin = np.inf
for i in xrange(wrong_target.shape[0]):
dists = np.square(dataset.y - dataset.y[i, :]).sum(axis=1)
dists[i] = np.inf
dists[used] = np.inf