def _forward(self,
sg_data,
fc_feats,
att_feats,
seq,
weak_rela,
att_masks=None):
core_args = self.prepare_core_args(sg_data, fc_feats, att_feats,
att_masks)
# make seq_per_img copies of the encoded inputs: shape: (B, ...) => (B*seq_per_image, ...)
core_args = expand_feats(core_args, self.seq_per_img)
weak_rela = expand_feats([weak_rela], self.seq_per_img)[0]
batch_size = fc_feats.size(0) * self.seq_per_img
state = self.init_hidden(batch_size, weak_rela)
outputs = fc_feats.new_zeros(batch_size,
seq.size(1) - 1, self.vocab_size + 1)
outputs_tag = fc_feats.new_zeros(batch_size, seq.size(1) - 1, 3)
# teacher forcing
for i in range(seq.size(1) - 1):
# scheduled sampling
if self.training and i >= 1 and self.ss_prob > 0.0:
sample_prob = fc_feats.new(batch_size).uniform_(0, 1)
sample_mask = sample_prob < self.ss_prob
if sample_mask.sum() == 0:
it = seq[:, i].clone()
else:
sample_ind = sample_mask.nonzero().view(-1)
it = seq[:, i].data.clone()
prob_prev = torch.exp(outputs[:, i - 1].detach(
)) # fetch prev distribution: shape Nx(M+1)
it.index_copy_(
0, sample_ind,
torch.multinomial(prob_prev, 1).view(-1).index_select(
0, sample_ind))
else:
it = seq[:, i].clone()
# break if all the sequences end
if i >= 1 and seq[:, i].sum() == 0:
break
# output, state = self.get_logprobs_state(it, state, core_args)
output, output_tag, state = self.get_logprobs_state(
it, state, core_args)
outputs[:, i] = output
outputs_tag[:, i] = output_tag
return outputs, outputs_tag
def _sample_beam(self, sg_data, fc_feats, att_feats, att_masks=None, opt={}):
beam_size = opt.get('beam_size', 10)
batch_size = fc_feats.size(0)
core_args = self.prepare_core_args(sg_data, fc_feats, att_feats, att_masks)
assert beam_size <= self.vocab_size + 1, 'lets assume this for now, otherwise this corner case causes a few headaches down the road. can be dealt with in future if needed'
seq = torch.LongTensor(self.seq_length, batch_size).zero_()
seqLogprobs = torch.FloatTensor(self.seq_length, batch_size)
# lets process every image independently for now, for simplicity
self.done_beams = [[] for _ in range(batch_size)]
for k in range(batch_size):
state = self.init_hidden(beam_size)
sample_core_args = []
for item in core_args:
if type(item) is list or item is None:
sample_core_args.append(item)
continue
else:
sample_core_args.append(item[k:k+1])
sample_core_args = expand_feats(sample_core_args, beam_size)
for t in range(1):
if t == 0: # input <bos>
it = fc_feats.new_zeros([beam_size], dtype=torch.long)
logprobs, state = self.get_logprobs_state(it, state, sample_core_args)
self.done_beams[k] = self.beam_search(state, logprobs, sample_core_args, opt=opt)
seq[:, k] = self.done_beams[k][0]['seq'] # the first beam has highest cumulative score
seqLogprobs[:, k] = self.done_beams[k][0]['logps']
# return the samples and their log likelihoods
return seq.transpose(0, 1), seqLogprobs.transpose(0, 1)
def get_self_critical_reward(model, core_args, sg_data, fc_feats, att_feats,
weak_relas, att_masks, data, gen_result, opt):
batch_size = gen_result.size(0)
seq_per_img = batch_size // len(data['gts'])
# get greedy decoding baseline
model.eval()
with torch.no_grad():
greedy_res, _ = model(sg_data,
fc_feats,
att_feats,
weak_relas,
att_masks=att_masks,
_core_args=core_args,
opt={'expand_features': False},
mode='sample')
model.train()
greedy_res = expand_feats([greedy_res], seq_per_img)[0]
res = OrderedDict()
gen_result = gen_result.data.cpu().numpy()
greedy_res = greedy_res.data.cpu().numpy()
for i in range(batch_size):
res[i] = [array_to_str(gen_result[i])]
for i in range(batch_size):
res[batch_size + i] = [array_to_str(greedy_res[i])]
gts = OrderedDict()
for i in range(len(data['gts'])):
gts[i] = [
array_to_str(data['gts'][i][j]) for j in range(len(data['gts'][i]))
]
res_ = [{'image_id': i, 'caption': res[i]} for i in range(2 * batch_size)]
res__ = {i: res[i] for i in range(2 * batch_size)}
gts = {
i: gts[i % batch_size // seq_per_img]
for i in range(2 * batch_size)
}
if opt.cider_reward_weight > 0:
_, cider_scores = CiderD_scorer.compute_score(gts, res_)
print('Cider scores:', _)
else:
cider_scores = 0
if opt.bleu_reward_weight > 0:
_, bleu_scores = Bleu_scorer.compute_score(gts, res__)
bleu_scores = np.array(bleu_scores[3])
print('Bleu scores:', _[3])
else:
bleu_scores = 0
scores = opt.cider_reward_weight * cider_scores + opt.bleu_reward_weight * bleu_scores
scores = scores[:batch_size] - scores[batch_size:]
# batch_size * seq_length
rewards = np.repeat(scores[:, np.newaxis], gen_result.shape[1], 1)
return rewards
def _sample(self, sg_data, fc_feats, att_feats, att_masks=None, opt={}, _core_args=None):
sample_max = opt.get('sample_max', 1)
beam_size = opt.get('beam_size', 1)
temperature = opt.get('temperature', 1.0)
decoding_constraint = opt.get('decoding_constraint', 0)
return_core_args = opt.get('return_core_args', False)
expand_features = opt.get('expand_features', True)
if beam_size > 1:
return self._sample_beam(sg_data, fc_feats, att_feats, att_masks, opt)
if _core_args is not None:
# reuse the core_args calculated during generating sampled captions
# when generating greedy captions for SCST,
core_args = _core_args
else:
core_args = self.prepare_core_args(sg_data, fc_feats, att_feats, att_masks)
# make seq_per_img copies of the encoded inputs: shape: (B, ...) => (B*seq_per_image, ...)
# should be True when training (xe or scst), False when evaluation
if expand_features:
if return_core_args:
_core_args = core_args
core_args = expand_feats(core_args, self.seq_per_img)
batch_size = fc_feats.size(0)*self.opt.seq_per_img
else:
batch_size = fc_feats.size(0)
state = self.init_hidden(batch_size)
seq = fc_feats.new_zeros((batch_size, self.seq_length), dtype=torch.long)
seqLogprobs = fc_feats.new_zeros(batch_size, self.seq_length)
for t in range(self.seq_length + 1):
if t == 0: # input <bos>
it = fc_feats.new_zeros(batch_size, dtype=torch.long)
logprobs, state = self.get_logprobs_state(it, state, core_args)
if decoding_constraint and t > 0:
tmp = logprobs.new_zeros(logprobs.size())
tmp.scatter_(1, seq[:,t-1].data.unsqueeze(1), float('-inf'))
logprobs = logprobs + tmp
# sample the next word
if t == self.seq_length: # skip if we achieve maximum length
break
if sample_max:
sampleLogprobs, it = torch.max(logprobs.data, 1)
it = it.view(-1).long()
else:
if temperature == 1.0:
prob_prev = torch.exp(logprobs.data) # fetch prev distribution: shape Nx(M+1)
else:
# scale logprobs by temperature
prob_prev = torch.exp(torch.div(logprobs.data, temperature))
it = torch.multinomial(prob_prev, 1)
sampleLogprobs = logprobs.gather(1, it) # gather the logprobs at sampled positions
it = it.view(-1).long() # and flatten indices for downstream processing
# stop when all finished
if t == 0:
unfinished = it > 0
else:
unfinished = unfinished * (it > 0)
it = it * unfinished.type_as(it)
seq[:,t] = it
seqLogprobs[:,t] = sampleLogprobs.view(-1)
# quit loop if all sequences have finished
if unfinished.sum() == 0:
break
returns = [seq, seqLogprobs]
if return_core_args:
returns.append(_core_args)
return returns
