class SCSTLayer(caffe.Layer):
"""
Self-Critical Sequence Training (SCST) layer. Takes beam search and
outputs weights for training.
"""
def setup(self, bottom, top):
if len(bottom) != 2:
raise Exception("Inputs 2 bottom blobs - image_ids and captions.")
if len(top) != 4:
raise Exception("Outputs 3 top blobs - score_weights, input_sentence, target_sentence, mean_score.")
params = ast.literal_eval(self.param_str)
self._end_of_sequence = params['end_of_sequence']
self._ignore_label = params['ignore_label']
# Load vocab
self._vocab = []
with open(params['vocab_path']) as vocab_file:
for word in vocab_file:
self._vocab.append(word.lower().strip())
self._cider = CiderScorer(params['gt_caption_paths'])
def _translate(self, blob):
# Results will be lower case, tokenized, without full stop
# (to match reference tokenization)
caption = [];
for ix in blob:
next_word = self._vocab[int(ix)]
if next_word == '.':
break
caption.append(next_word)
return caption
def reshape(self, bottom, top):
self._batch_size = bottom[1].shape[0]
self._beam_size = bottom[1].shape[2]
self._sequence_length = bottom[1].shape[3]
top[0].reshape(self._batch_size*self._beam_size, self._sequence_length)
top[1].reshape(self._batch_size*self._beam_size, self._sequence_length)
top[2].reshape(self._batch_size*self._beam_size, self._sequence_length)
top[3].reshape(1)
def forward(self, bottom, top):
top[1].data[...] = self._end_of_sequence
top[2].data[...] = self._ignore_label
# Score captions and generate training input and target output
image_ids = []
captions = []
for n in range(self._batch_size):
for b in range(self._beam_size):
image_ids.append(int(bottom[0].data[n][0]))
seq = bottom[1].data[n][0][b]
captions.append(self._translate(seq))
caption = seq[:len(captions[-1])].tolist()
top[1].data[n*self._beam_size+b,1:min(self._sequence_length,len(caption)+1)] = \
caption[:self._sequence_length-1] # input_sentence
caption.append(self._end_of_sequence)
top[2].data[n*self._beam_size+b,:min(self._sequence_length,len(caption))] = \
caption[:self._sequence_length] # target_sentence
raw_scores = np.array(self._cider.compute_scores(image_ids,captions))
# Generate score output
for n in range(self._batch_size):
baseline = np.mean(raw_scores[n*self._beam_size:(n+1)*self._beam_size])
for b in range(self._beam_size):
score = raw_scores[n*self._beam_size+b]
top[0].data[n*self._beam_size+b,:] = score - baseline
top[3].data[0] = np.mean(raw_scores)
def backward(self, top, propagate_down, bottom):
"""This layer does not propagate gradients."""
pass
class SCSTSamplingLayer(caffe.Layer):
"""
Self-Critical Sequence Training (SCST) layer. Takes argmax and sampled captions and
outputs weights for training.
"""
def setup(self, bottom, top):
if len(bottom) != 2:
raise Exception("Inputs 2 bottom blobs - image_ids and captions.")
if len(top) != 4:
raise Exception("Outputs 4 top blobs - score_weights, target_sentence, mean_score, scores.")
params = ast.literal_eval(self.param_str)
self._end_of_sequence = params['end_of_sequence']
self._ignore_label = params['ignore_label']
# Load vocab
self._vocab = []
with open(params['vocab_path']) as vocab_file:
for word in vocab_file:
self._vocab.append(word.lower().strip())
self._cider = CiderScorer(params['gt_caption_paths'], include_eos=True)
def _translate(self, blob):
# Results will be lower case, tokenized, without full stop
# (to match reference tokenization)
caption = [];
for ix in blob:
next_word = self._vocab[int(ix)]
if next_word == '.':
caption.append(next_word) # Include EOS
break
caption.append(next_word)
return caption
def reshape(self, bottom, top):
self._batch_size = bottom[1].shape[0]
self._sequence_length = bottom[1].shape[1]
top[0].reshape(self._batch_size, self._sequence_length)
top[1].reshape(self._batch_size, self._sequence_length)
top[2].reshape(1)
top[3].reshape(self._batch_size)
def forward(self, bottom, top):
top[0].data[...] = 0
top[1].data[...] = self._ignore_label
# Score captions and generate target output
image_ids = []
captions = []
for n in range(self._batch_size):
image_ids.append(int(bottom[0].data[n/2][0]))
seq = bottom[1].data[n]
captions.append(self._translate(seq))
if n % 2 == 1: # Generate targets
caption = seq[:len(captions[-1])].tolist()
top[1].data[n,:min(self._sequence_length,len(caption))] = \
caption[:self._sequence_length] # target_sentence
raw_scores = self._cider.compute_scores(image_ids,captions)
# Generate score weights
for n in range(self._batch_size/2):
baseline_score = raw_scores[n*2]
sample_score = raw_scores[n*2+1]
top[3].data[n*2] = baseline_score
top[3].data[n*2+1] = sample_score
if sample_score > 0:
sample_score = math.log(sample_score)
if baseline_score > 0:
baseline_score = math.log(baseline_score)
top[0].data[n*2+1] = max(0.0, sample_score - baseline_score)
top[2].data[0] = np.mean(raw_scores[::2])
def backward(self, top, propagate_down, bottom):
"""This layer does not propagate gradients."""
pass
