class UniformPrior(object):
'''A uniform color prior in RGB space.'''
def __init__(self, recurrent=False):
self.sampler = BucketsVectorizer([1], hsv=False)
self.recurrent = recurrent
def train(self, training_instances, listener_data='ignored'):
pass
def apply(self, input_vars):
c = input_vars[0]
if self.recurrent:
if c.ndim == 2:
ones = T.ones_like(c[:, 0])
elif c.ndim == 3:
ones = T.ones_like(c[:, 0, 0])
else:
assert False, 'need handling for higher rank color vectors (recurrent): %d' % c.ndim
else:
if c.ndim == 1:
ones = T.ones_like(c)
elif c.ndim == 2:
ones = T.ones_like(c[:, 0])
else:
assert False, 'need handling for higher rank color vectors (atomic): %d' % c.ndim
return -3.0 * np.log(256.0) * ones
def sample(self, num_samples):
'''
:return: a list of `num_samples` colors sampled uniformly in RGB space,
but expressed as HSV triples.
'''
colors = self.sampler.unvectorize_all(np.zeros(num_samples, dtype=np.int32),
random=True, hsv=True)
return [instance.Instance(c) for c in colors]
class ListenerLearner(NeuralLearner):
'''
An LSTM-based listener (guesses colors from descriptions).
'''
def __init__(self, id=None):
super(ListenerLearner, self).__init__(id=id)
self.word_counts = Counter()
self.seq_vec = SequenceVectorizer(
unk_threshold=self.options.listener_unk_threshold)
self.color_vec = BucketsVectorizer(
self.options.listener_color_resolution,
hsv=self.options.listener_hsv)
def predict_and_score(self, eval_instances, random=False, verbosity=0):
predictions = []
scores = []
batches = iterators.iter_batches(eval_instances,
self.options.listener_eval_batch_size)
num_batches = (len(eval_instances) -
1) // self.options.listener_eval_batch_size + 1
if self.options.verbosity + verbosity >= 2:
print('Testing')
progress.start_task('Eval batch', num_batches)
for batch_num, batch in enumerate(batches):
progress.progress(batch_num)
batch = list(batch)
xs, (y, ) = self._data_to_arrays(batch, test=True)
probs = self.model.predict(xs)
if random:
indices = sample(probs)
predictions.extend(self.unvectorize(indices, random=True))
else:
predictions.extend(self.unvectorize(probs.argmax(axis=1)))
scores_arr = np.log(probs[np.arange(len(batch)),
y]) + self.bucket_adjustment()
scores.extend(scores_arr.tolist())
progress.end_task()
if self.options.verbosity >= 9:
print('%s %ss:') % (self.id, 'sample' if random else 'prediction')
for inst, prediction in zip(eval_instances, predictions):
print('%s -> %s' % (repr(inst.input), repr(prediction)))
return predictions, scores
def unvectorize(self, indices, random=False):
return self.color_vec.unvectorize_all(indices, random=random, hsv=True)
def bucket_adjustment(self):
bucket_volume = (256.0**3) / self.color_vec.num_types
return -np.log(bucket_volume)
def on_iter_end(self, step, writer):
most_common = [
desc for desc, count in self.word_counts.most_common(10)
]
insts = [instance.Instance(input=desc) for desc in most_common]
xs, (y, ) = self._data_to_arrays(insts, test=True)
probs = self.model.predict(xs)
for i, desc in enumerate(most_common):
dist = probs[i, :]
for image, channel in zip(
self.color_vec.visualize_distribution(dist), '012'):
writer.log_image(step, '%s/%s/%s' % (self.id, desc, channel),
image)
super(ListenerLearner, self).on_iter_end(step, writer)
def _data_to_arrays(self,
training_instances,
init_vectorizer=False,
test=False,
inverted=False):
def get_multi(val):
if isinstance(val, tuple):
assert len(val) == 1
return val[0]
else:
return val
get_i, get_o = (lambda inst: inst.input), (lambda inst: inst.output)
get_desc, get_color = (get_o, get_i) if inverted else (get_i, get_o)
get_i_ind, get_o_ind = (
(lambda inst: inst.alt_inputs[get_multi(inst.input)]),
(lambda inst: inst.alt_outputs[get_multi(inst.output)]))
get_color_indexed = get_i_ind if inverted else get_o_ind
if hasattr(self.options, 'listener_tokenizer'):
tokenize = TOKENIZERS[self.options.listener_tokenizer]
else:
tokenize = TOKENIZERS['whitespace']
if init_vectorizer:
tokenized = [['<s>'] + tokenize(get_desc(inst)) + ['</s>']
for inst in training_instances]
self.seq_vec.add_all(tokenized)
unk_replaced = self.seq_vec.unk_replace_all(tokenized)
self.word_counts.update(
[get_desc(inst) for inst in training_instances])
config.dump(unk_replaced, 'unk_replaced.train.jsons', lines=True)
sentences = []
colors = []
if self.options.verbosity >= 9:
print('%s _data_to_arrays:' % self.id)
for i, inst in enumerate(training_instances):
desc = tokenize(get_desc(inst))
color = get_color(inst)
if isinstance(color, numbers.Number):
color = get_color_indexed(inst)
if not color:
assert test
color = (0.0, 0.0, 0.0)
s = ['<s>'] * (self.seq_vec.max_len - 1 - len(desc)) + desc
s.append('</s>')
if self.options.verbosity >= 9:
print('%s -> %s' % (repr(s), repr(color)))
sentences.append(s)
colors.append(color)
x = np.zeros((len(sentences), self.seq_vec.max_len), dtype=np.int32)
y = np.zeros((len(sentences), ), dtype=np.int32)
for i, sentence in enumerate(sentences):
if len(sentence) > x.shape[1]:
sentence = sentence[:x.shape[1]]
x[i, :] = self.seq_vec.vectorize(sentence)
y[i] = self.color_vec.vectorize(colors[i], hsv=True)
return [x], [y]
def _build_model(self, model_class=SimpleLasagneModel):
id_tag = (self.id + '/') if self.id else ''
input_var = T.imatrix(id_tag + 'inputs')
target_var = T.ivector(id_tag + 'targets')
self.l_out, self.input_layers = self._get_l_out([input_var])
self.loss = categorical_crossentropy
self.model = model_class(
[input_var], [target_var],
self.l_out,
loss=self.loss,
optimizer=OPTIMIZERS[self.options.listener_optimizer],
learning_rate=self.options.listener_learning_rate,
id=self.id)
def train_priors(self, training_instances, listener_data=False):
prior_class = PRIORS[self.options.listener_prior]
self.prior_emp = prior_class(
) # TODO: accurate values for empirical prior
self.prior_smooth = prior_class()
self.prior_emp.train(training_instances, listener_data=listener_data)
self.prior_smooth.train(training_instances,
listener_data=listener_data)
def _get_l_out(self, input_vars):
check_options(self.options)
id_tag = (self.id + '/') if self.id else ''
input_var = input_vars[0]
l_in = InputLayer(shape=(None, self.seq_vec.max_len),
input_var=input_var,
name=id_tag + 'desc_input')
l_in_embed = EmbeddingLayer(
l_in,
input_size=len(self.seq_vec.tokens),
output_size=self.options.listener_cell_size,
name=id_tag + 'desc_embed')
cell = CELLS[self.options.listener_cell]
cell_kwargs = {
'grad_clipping': self.options.listener_grad_clipping,
'num_units': self.options.listener_cell_size,
}
if self.options.listener_cell == 'LSTM':
cell_kwargs['forgetgate'] = Gate(
b=Constant(self.options.listener_forget_bias))
if self.options.listener_cell != 'GRU':
cell_kwargs['nonlinearity'] = NONLINEARITIES[
self.options.listener_nonlinearity]
l_rec1 = cell(l_in_embed, name=id_tag + 'rec1', **cell_kwargs)
if self.options.listener_dropout > 0.0:
l_rec1_drop = DropoutLayer(l_rec1,
p=self.options.listener_dropout,
name=id_tag + 'rec1_drop')
else:
l_rec1_drop = l_rec1
l_rec2 = cell(l_rec1_drop, name=id_tag + 'rec2', **cell_kwargs)
if self.options.listener_dropout > 0.0:
l_rec2_drop = DropoutLayer(l_rec2,
p=self.options.listener_dropout,
name=id_tag + 'rec2_drop')
else:
l_rec2_drop = l_rec2
l_hidden = DenseLayer(
l_rec2_drop,
num_units=self.options.listener_cell_size,
nonlinearity=NONLINEARITIES[self.options.listener_nonlinearity],
name=id_tag + 'hidden')
if self.options.listener_dropout > 0.0:
l_hidden_drop = DropoutLayer(l_hidden,
p=self.options.listener_dropout,
name=id_tag + 'hidden_drop')
else:
l_hidden_drop = l_hidden
l_scores = DenseLayer(l_hidden_drop,
num_units=self.color_vec.num_types,
nonlinearity=None,
name=id_tag + 'scores')
l_out = NonlinearityLayer(l_scores,
nonlinearity=softmax,
name=id_tag + 'out')
return l_out, [l_in]
def sample_prior_smooth(self, num_samples):
return self.prior_smooth.sample(num_samples)
class ListenerLearner(NeuralLearner):
'''
An LSTM-based listener (guesses colors from descriptions).
'''
def __init__(self, id=None):
super(ListenerLearner, self).__init__(id=id)
self.word_counts = Counter()
self.seq_vec = SequenceVectorizer(unk_threshold=self.options.listener_unk_threshold)
self.color_vec = BucketsVectorizer(self.options.listener_color_resolution,
hsv=self.options.listener_hsv)
def predict_and_score(self, eval_instances, random=False, verbosity=0):
predictions = []
scores = []
batches = iterators.iter_batches(eval_instances, self.options.listener_eval_batch_size)
num_batches = (len(eval_instances) - 1) // self.options.listener_eval_batch_size + 1
if self.options.verbosity + verbosity >= 2:
print('Testing')
progress.start_task('Eval batch', num_batches)
for batch_num, batch in enumerate(batches):
progress.progress(batch_num)
batch = list(batch)
xs, (y,) = self._data_to_arrays(batch, test=True)
probs = self.model.predict(xs)
self.on_predict(xs)
if random:
indices = sample(probs)
predictions.extend(self.unvectorize(indices, random=True))
else:
predictions.extend(self.unvectorize(probs.argmax(axis=1)))
scores_arr = np.log(probs[np.arange(len(batch)), y]) + self.bucket_adjustment()
scores.extend(scores_arr.tolist())
progress.end_task()
if self.options.verbosity >= 9:
print('%s %ss:') % (self.id, 'sample' if random else 'prediction')
for inst, prediction in zip(eval_instances, predictions):
print('%s -> %s' % (repr(inst.input), repr(prediction)))
return predictions, scores
def unvectorize(self, indices, random=False):
return self.color_vec.unvectorize_all(indices, random=random, hsv=True)
def bucket_adjustment(self):
bucket_volume = (256.0 ** 3) / self.color_vec.num_types
return -np.log(bucket_volume)
def on_predict(self, xs):
pass
def on_iter_end(self, step, writer):
most_common = [desc for desc, count in self.word_counts.most_common(10)]
insts = [instance.Instance(input=desc) for desc in most_common]
xs, (y,) = self._data_to_arrays(insts, test=True)
probs = self.model.predict(xs)
for i, desc in enumerate(most_common):
dist = probs[i, :]
for image, channel in zip(self.color_vec.visualize_distribution(dist), '012'):
writer.log_image(step, '%s/%s/%s' % (self.id, desc, channel), image)
super(ListenerLearner, self).on_iter_end(step, writer)
def _data_to_arrays(self, training_instances,
init_vectorizer=False, test=False, inverted=False):
def get_multi(val):
if isinstance(val, tuple):
assert len(val) == 1
return val[0]
else:
return val
get_i, get_o = (lambda inst: inst.input), (lambda inst: inst.output)
get_desc, get_color = (get_o, get_i) if inverted else (get_i, get_o)
get_i_ind, get_o_ind = ((lambda inst: inst.alt_inputs[get_multi(inst.input)]),
(lambda inst: inst.alt_outputs[get_multi(inst.output)]))
get_color_indexed = get_i_ind if inverted else get_o_ind
if hasattr(self.options, 'listener_tokenizer'):
tokenize = TOKENIZERS[self.options.listener_tokenizer]
else:
tokenize = TOKENIZERS['whitespace']
if init_vectorizer:
tokenized = [['<s>'] + tokenize(get_desc(inst)) + ['</s>']
for inst in training_instances]
self.seq_vec.add_all(tokenized)
unk_replaced = self.seq_vec.unk_replace_all(tokenized)
self.word_counts.update([get_desc(inst) for inst in training_instances])
config.dump(unk_replaced, 'unk_replaced.train.jsons', lines=True)
sentences = []
colors = []
if self.options.verbosity >= 9:
print('%s _data_to_arrays:' % self.id)
for i, inst in enumerate(training_instances):
desc = tokenize(get_desc(inst))
color = get_color(inst)
if isinstance(color, numbers.Number):
color = get_color_indexed(inst)
if not color:
assert test
color = (0.0, 0.0, 0.0)
s = ['<s>'] * (self.seq_vec.max_len - 1 - len(desc)) + desc
s.append('</s>')
if self.options.verbosity >= 9:
print('%s -> %s' % (repr(s), repr(color)))
sentences.append(s)
colors.append(color)
x = np.zeros((len(sentences), self.seq_vec.max_len), dtype=np.int32)
y = np.zeros((len(sentences),), dtype=np.int32)
for i, sentence in enumerate(sentences):
if len(sentence) > x.shape[1]:
sentence = sentence[:x.shape[1]]
x[i, :] = self.seq_vec.vectorize(sentence)
y[i] = self.color_vec.vectorize(colors[i], hsv=True)
return [x], [y]
def _build_model(self, model_class=SimpleLasagneModel):
id_tag = (self.id + '/') if self.id else ''
input_var = T.imatrix(id_tag + 'inputs')
target_var = T.ivector(id_tag + 'targets')
self.l_out, self.input_layers = self._get_l_out([input_var])
self.loss = categorical_crossentropy
self.model = model_class(
[input_var], [target_var], self.l_out,
loss=self.loss, optimizer=OPTIMIZERS[self.options.listener_optimizer],
learning_rate=self.options.listener_learning_rate,
id=self.id)
def train_priors(self, training_instances, listener_data=False):
prior_class = PRIORS[self.options.listener_prior]
self.prior_emp = prior_class() # TODO: accurate values for empirical prior
self.prior_smooth = prior_class()
self.prior_emp.train(training_instances, listener_data=listener_data)
self.prior_smooth.train(training_instances, listener_data=listener_data)
def _get_l_out(self, input_vars):
check_options(self.options)
id_tag = (self.id + '/') if self.id else ''
input_var = input_vars[0]
l_in = InputLayer(shape=(None, self.seq_vec.max_len), input_var=input_var,
name=id_tag + 'desc_input')
l_in_embed = EmbeddingLayer(l_in, input_size=len(self.seq_vec.tokens),
output_size=self.options.listener_cell_size,
name=id_tag + 'desc_embed')
cell = CELLS[self.options.listener_cell]
cell_kwargs = {
'grad_clipping': self.options.listener_grad_clipping,
'num_units': self.options.listener_cell_size,
}
if self.options.listener_cell == 'LSTM':
cell_kwargs['forgetgate'] = Gate(b=Constant(self.options.listener_forget_bias))
if self.options.listener_cell != 'GRU':
cell_kwargs['nonlinearity'] = NONLINEARITIES[self.options.listener_nonlinearity]
l_rec1 = cell(l_in_embed, name=id_tag + 'rec1', **cell_kwargs)
if self.options.listener_dropout > 0.0:
l_rec1_drop = DropoutLayer(l_rec1, p=self.options.listener_dropout,
name=id_tag + 'rec1_drop')
else:
l_rec1_drop = l_rec1
l_rec2 = cell(l_rec1_drop, name=id_tag + 'rec2', **cell_kwargs)
if self.options.listener_dropout > 0.0:
l_rec2_drop = DropoutLayer(l_rec2, p=self.options.listener_dropout,
name=id_tag + 'rec2_drop')
else:
l_rec2_drop = l_rec2
l_hidden = DenseLayer(l_rec2_drop, num_units=self.options.listener_cell_size,
nonlinearity=NONLINEARITIES[self.options.listener_nonlinearity],
name=id_tag + 'hidden')
if self.options.listener_dropout > 0.0:
l_hidden_drop = DropoutLayer(l_hidden, p=self.options.listener_dropout,
name=id_tag + 'hidden_drop')
else:
l_hidden_drop = l_hidden
l_scores = DenseLayer(l_hidden_drop, num_units=self.color_vec.num_types, nonlinearity=None,
name=id_tag + 'scores')
l_out = NonlinearityLayer(l_scores, nonlinearity=softmax, name=id_tag + 'out')
return l_out, [l_in]
def sample_prior_smooth(self, num_samples):
return self.prior_smooth.sample(num_samples)