def plot_for_subtree2(nw, theta, n=0):
fig, ax = pt.new_fig()
pt.equal_aspects(ax)
M = theta[('composition', '#X#', '(#X#, #X#, #X#)', 'I', 'M')]
mbits = np.split(M, 3, axis=1)
B = theta[('composition', '#X#', '(#X#, #X#, #X#)', 'I', 'B')]
children = nw.inputs
reps = [c.a for c in children]
projs = [mbits[i].dot(reps[i]) for i in range(3)]
names = []
for c in children:
try:
names.append(c.key)
except:
names.append(str(c))
names.append('bias')
now = np.array([0, 0])
for i, proj in enumerate(projs + [B]):
nnow = now + proj
pt.scatter_and_annotate(ax, nnow, projColors[i], '', alpha=0)
if i == 0: postfix = 'L'
elif i == 1: postfix = 'M'
elif i == 2: postfix = 'R'
else: postfix = ''
pt.draw_arrow(ax,
now,
nnow,
label=names[i] + postfix,
color=projColors[i],
alpha=0.6,
head=0.1)
now = nnow
pt.scatter_and_annotate(ax, nnow, projColors[i], '')
# print nw.a-myActivation.activate(now,'tanh')[0], 'should be (close to) zero'
# pt.scatter_and_annotate(ax, now , 'black', 'sum')
pt.scatter_and_annotate(ax, nw.a, eC, str(nw), pos='bot')
pt.draw_arrow(ax,
now,
nw.a,
label='squash',
color='black',
alpha=0.6,
head=0.1)
pt.put_origin(ax)
pt.minimal_ticks(ax)
if outDir:
pt.save_figure(ax, False, outDir, 'example' + str(n) + 'All')
else:
pt.show(ax, False)
def plot_example(theta):
fig, ax = pt.new_fig(([-1.5, 3.5], [-3, 4.5]))
pt.minimal_ticks(ax)
me = arithmetics.mathExpression.fromstring('( 5 - ( 2 + 3 ) )')
ans = me.solve()
print str(me), ans
nw = data.NN.RNN(me)
nw.activate(theta)
# plot_for_subtree(ax, nw.root.inputs[2], theta,0)
# plot_for_subtree(ax, nw.root, theta, 10)
plot_for_subtree2(nw.root.inputs[2], theta, 0)
plot_for_subtree2(nw.root, theta, 10)
def plot_comparison(theta):
try:
left, right = np.split(theta[('comparison', 'M')], 2, axis=1)
except:
print 'no comparison matrix in theta'
return
mtb = arithmetics.training_treebank(
args['seed'],
languages={'L9': 25},
)
rnnTB = data.RNNTB(mtb.examples)
colors = pt.colorscale(120)
rnn, tar = rnnTB.examples[0]
rep = rnn.activate(theta)
d = len(rep)
for i in range(d):
for j in range(i + 1, d):
fig, ax = pt.new_fig()
for rnn, tar in rnnTB.get_examples():
rep = rnn.activate(theta)
leftie = left.dot(rep)
rightie = right.dot(rep)
# print leftie.shape, rightie.shape
# pt.scatter_and_annotate(ax, rep, color=colors(tar+60), label=tar, size=8, txtcolor='black')
pt.scatter_and_annotate(ax, (leftie[i], leftie[j]),
color=colors(tar + 60),
label=str(tar) + 'L',
size=8,
txtcolor='black')
pt.scatter_and_annotate(ax, (rightie[i], rightie[j]),
color=colors(tar + 60),
label=str(tar) + 'R',
size=8,
txtcolor='black')
pt.title('Effect of comparison layer on L9 expressions, dims ' +
str(i) + '-' + str(j))
if outDir:
pt.save_figure(ax, True, outDir,
'comparison' + str(i) + str(j))
else:
pt.show(ax, True)
def plot_embs(theta):
wm = theta[('word', )]
mbits = np.split(theta[('composition', '#X#', '(#X#, #X#, #X#)', 'I',
'M')],
3,
axis=1)
fig, ax = pt.new_fig()
ranges = [pt.colorscale(20, projColors[i]) for i in range(3)]
for w, v in wm.iteritems():
if w[-1].isdigit():
pt.scatter_and_annotate(ax,
point=v,
label=w,
color=eC,
pos='right')
for i in [0, 2]:
proj = mbits[i].dot(v)
pt.draw_arrow(ax,
v,
proj,
label='',
color=projColors[i],
head=0,
alpha=0.3)
if int(w) in [-10, -5, 0, 5, 10]:
label = w + ('L' if i == 0 else 'R')
else:
label = ''
pt.scatter_and_annotate(ax,
point=proj,
label=label,
color=ranges[i](int(w) + 10),
pos=('left' if i == 2 else 'right'))
else:
print w
pt.title('Effect of composition as left or right child on embeddings')
pt.improve_ticks(ax)
if outDir:
pt.save_figure(ax, True, outDir, 'embs')
else:
pt.show(ax, True)
def plot_answers(theta):
mtb = data.arithmetics.training_treebank(
args['seed'],
languages={'L9': 2000},
)
rnnTB = data.RNNTB(mtb.examples)
fig, ax = pt.new_fig()
colors = pt.colorscale(120, 'blue')
ranges = [pt.colorscale(120, projColors[i]) for i in range(3)]
mbits = np.split(theta[('composition', '#X#', '(#X#, #X#, #X#)', 'I',
'M')],
3,
axis=1)
tolabel = {
'+': {i: []
for i in range(-60, 61, 10)},
'-': {i: []
for i in range(-60, 61, 10)}
}
# tolabel=tolabel+tolabel[:]+tolabel[:]
doLater = []
for rnn, tar in rnnTB.get_examples():
rep = rnn.activate(theta)
for i in [0, 2]:
proj = mbits[i].dot(rep)
doLater.append((rep, proj, projColors[i], ranges[i](tar + 60)))
if tar in tolabel['-'].keys():
tolabel[rnn.root.inputs[1].key][tar].append(rep)
# pt.scatter_and_annotate(ax, rep, color=colors(tar+60), label=tar, size=10, txtcolor='black')
# del tolabel[tolabel.index(tar)]
else:
pt.scatter_and_annotate(ax,
rep,
color=colors(int(tar) + 60),
label='',
txtcolor='black')
stuff = []
for op, d in tolabel.iteritems():
print op
for tar, replist in d.iteritems():
if len(replist) == 0: continue
av = np.average(np.array(replist), axis=0)
stuff.append(
pt.scatter_and_annotate(ax,
av,
color=colors(tar + 60),
label=tar,
txtcolor='black',
pos=('right' if op == '+' else 'top')))
pt.title('Root representations of L9 expressions')
if outDir:
pt.save_figure(ax, True, outDir, 'answers')
else:
pt.show(ax, True)
[item.remove() for sublist in stuff for item in sublist]
for rep, proj, pColor, color in doLater[:200]:
pt.draw_arrow(ax, rep, proj, label='', color=pColor, head=0, alpha=0.3)
pt.scatter_and_annotate(ax, point=proj, label='', color=color)
pt.title('Root representations and Projections of L9 expressions')
if outDir:
pt.save_figure(ax, True, outDir, 'answersProj')
else:
pt.show(ax, True)