def evaluate(model, data, indices, parallel=False):
start_time = time.time()
eval_loss = 0.
eval_num_words = 0
model.eval()
with torch.no_grad():
# fetch the first batch
batch = [dh.make_batch(data, indices[0], separate_caption=args.separate_caption, pretrained_elmo=args.pretrained_elmo, pretrained_bert=args.pretrained_bert, bert_tokenizer=bert_tokenizer, pretrained_all=args.pretrained_all, bert_model=args.bert_model, concat_his=args.concat_his)]
# evaluation loop
it = tqdm(six.moves.range(len(indices)), desc="evaluation", ncols=0)
for j in it:
#if args.separate_caption:
# x_batch, h_batch, q_batch, a_batch_in, a_batch_out, c_batch = batch.pop()
#else:
# x_batch, h_batch, q_batch, a_batch_in, a_batch_out = batch.pop()
b = batch.pop()
if j < len(indices)-1:
prefetch = threading.Thread(target=fetch_batch,
args=([dh, data, indices[j+1], args.separate_caption, batch]))
prefetch.start()
# propagate for training
x = [torch.from_numpy(x) for x in b[0]]
if args.concat_his:
h = [torch.from_numpy(h_i) for h_i in b[1]]
else:
h = [[torch.from_numpy(h) for h in hb] for hb in b[1]]
q = [torch.from_numpy(q) for q in b[2]]
ai = [torch.from_numpy(ai) for ai in b[3]]
ao = [torch.from_numpy(ao) for ao in b[4]]
if args.separate_caption:
c = [torch.from_numpy(c) for c in b[5]]
else:
c = None
if args.pretrained_elmo or args.pretrained_bert:
if args.pretrained_all:
context_q, context_h, context_ai = b[-3:]
else:
context_q = b[-1]
context_h = None
context_ai = None
else:
context_q = None
context_h = None
context_ai = None
if args.exclude_video:
x = None
if parallel:
_, _, loss = model.module.loss(x, h, q, ai, ao, c, context_q, context_h, context_ai)
else:
_, _, loss = model.loss(x, h, q, ai, ao, c, context_q, context_h, context_ai)
num_words = sum([len(s) for s in ao])
eval_loss += loss.cpu().data.numpy() * num_words
eval_num_words += num_words
prefetch.join()
model.train()
wall_time = time.time() - start_time
return math.exp(eval_loss/eval_num_words), wall_time
def run_epoch(data, indices, vocab, epoch, model, loss_compute, eval=False):
"Standard Training and Logging Function"
start = time.time()
total_tokens = 0
total_loss = 0
tokens = 0
it = tqdm(range(len(indices)), desc="epoch {}/{}".format(epoch+1, args.num_epochs), ncols=0)
for j in it:
batch = dh.make_batch(data, indices[j], vocab, separate_caption=args.separate_caption, cut_a=args.cut_a)
b = batch
if True:
out, ae_out = model.forward(b)
if args.auto_encoder_ft == 'caption' or args.auto_encoder_ft == 'summary':
ntokens_cap = (b.cap != vocab['<blank>']).data.sum()
loss = loss_compute(out, b.trg_y, b.ntokens, ae_out, b.cap, ntokens_cap)
elif args.auto_encoder_ft == 'query':
ntokens_query = (b.query != vocab['<blank>']).data.sum()
loss = loss_compute(out, b.trg_y, b.ntokens, ae_out, b.query, ntokens_query)
total_loss += loss
total_tokens += b.ntokens
tokens += b.ntokens
if (j+1) % args.report_interval == 0 and not eval:
elapsed = time.time() - start
print("Epoch: %d Step: %d Loss: %f Tokens per Sec: %f" %
(epoch+1,j+1, loss / b.ntokens.float(), float(tokens) / elapsed))
with open(train_log_path, "a") as f:
f.write("{},{},{:e},{}\n".format(epoch+1,j+1,loss/b.ntokens.float(),float(tokens)/elapsed))
start = time.time()
tokens = 0
#prefetch.join()
return total_loss / total_tokens.float()
def generate_caption(model,
data,
batch_indices,
vocab,
dim,
stride=1,
maxlen=20,
beam=5,
penalty=2.0,
nbest=1):
vocablist = sorted(vocab.keys(), key=lambda s: vocab[s])
result = []
c = 0
for j in six.moves.range(len(batch_indices[0])):
start_time = time.time()
x_batch = [None] * len(data)
for n in six.moves.range(len(data)):
x_batch[n], Q_batch, A_batch = dh.make_batch(data[n],
batch_indices[n][j],
dim=dim[n],
stride=stride)
pred_out, logp = model.generate(x_batch,
Q_batch,
A_batch,
maxlen=maxlen,
beam=beam,
penalty=penalty)
for i in six.moves.range(Q_batch.shape[1]):
c = c + 1
print c, batch_indices[0][j][0][0] + '_' + str(i + 1)
print 'REF:',
for n in six.moves.range(A_batch.shape[0]):
number = A_batch[n][i][0]
if number == 3:
continue
else:
print vocablist[A_batch[n][i][0]],
print
for n in six.moves.range(min(nbest, len(pred_out[i]))):
# pdb.set_trace()
print 'HYP[%d]:' % (n + 1),
pred = pred_out[i][n]
if (isinstance(pred, list)):
# The format used in run_youtube_mm.sh.
for w in pred:
print vocablist[w],
print '( {:f} )'.format(logp[i][n])
else:
# The format used in run_youtube.sh.
print(vocablist[pred_out[i][n]])
print('( {:f} )'.format(logp[i]))
print 'ElapsedTime:', time.time() - start_time
print '-----------------------'
## print sentence
result.append(pred_out)
return result
def generate_response(model,
data,
batch_indices,
vocab,
maxlen=20,
beam=5,
penalty=2.0,
nbest=1):
vocablist = sorted(vocab.keys(), key=lambda s: vocab[s])
result_dialogs = []
model.eval()
with torch.no_grad():
qa_id = 0
for dialog in data['original']['dialogs']:
vid = dialog['image_id']
pred_dialog = {
'image_id': vid,
'dialog': copy.deepcopy(dialog['dialog'])
}
result_dialogs.append(pred_dialog)
for t, qa in enumerate(dialog['dialog']):
logging.info('%d %s_%d' % (qa_id, vid, t))
logging.info('QS: ' + qa['question'])
logging.info('REF: ' + qa['answer'])
# prepare input data
start_time = time.time()
x_batch, h_batch, q_batch, a_batch_in, a_batch_out, s_batch = \
dh.make_batch(data, batch_indices[qa_id])
qa_id += 1
x = [torch.from_numpy(x) for x in x_batch]
h = [[torch.from_numpy(h) for h in hb] for hb in h_batch]
q = [torch.from_numpy(q) for q in q_batch]
s = torch.from_numpy(s_batch).cuda().float()
# generate sequences
pred_out, _ = model.generate(x,
h,
q,
s,
maxlen=maxlen,
beam=beam,
penalty=penalty,
nbest=nbest)
for n in six.moves.range(min(nbest, len(pred_out))):
pred = pred_out[n]
hypstr = ' '.join([vocablist[w] for w in pred[0]])
logging.info('HYP[%d]: %s ( %f )' %
(n + 1, hypstr, pred[1]))
if n == 0:
pred_dialog['dialog'][t]['answer'] = hypstr
logging.info('ElapsedTime: %f' % (time.time() - start_time))
logging.info('-----------------------')
return {'dialogs': result_dialogs}
def evaluate(model, data, indices):
start_time = time.time()
eval_loss = 0.
eval_num_words = 0
model.eval()
with torch.no_grad():
# fetch the first batch
batch = [dh.make_batch(data, indices[0])]
# evaluation loop
for j in six.moves.range(len(indices)):
# get a fetched batch
x_batch, h_batch, q_batch, a_batch_in, a_batch_out, s_batch = batch.pop(
)
# fetch the next batch in parallel
if j < len(indices) - 1:
prefetch = threading.Thread(
target=fetch_batch,
args=([dh, data, indices[j + 1], batch]))
prefetch.start()
# propagate for training
x = [torch.from_numpy(x) for x in x_batch]
h = [[torch.from_numpy(h) for h in hb] for hb in h_batch]
q = [torch.from_numpy(q) for q in q_batch]
ai = [torch.from_numpy(ai) for ai in a_batch_in]
ao = [torch.from_numpy(ao) for ao in a_batch_out]
s = torch.from_numpy(s_batch).cuda().float()
_, _, loss = model.loss(x, h, q, ai, ao, s)
num_words = sum([len(s) for s in ao])
eval_loss += loss.cpu().data.numpy() * num_words
eval_num_words += num_words
# wait prefetch completion
prefetch.join()
model.train()
wall_time = time.time() - start_time
return math.exp(eval_loss / eval_num_words), wall_time
def evaluate(model, data, batch_indices, dim, stride=1):
start_time = time.time()
eval_loss = 0.
eval_hit = 0
num_tokens = 0
for j in six.moves.range(len(batch_indices)):
x_batch = [None] * len(data)
for m in six.moves.range(len(data)):
x_batch[m], Q_batch, A_batch = dh.make_batch(data[m],
batch_indices[m][j],
dim=dim[m],
stride=stride)
loss, hit, num = model(x_batch,
Q_batch,
A_batch,
predicted_context=False,
istraining=False)
eval_loss += loss
eval_hit += hit
num_tokens += num
wall_time = time.time() - start_time
return math.exp(eval_loss /
num_tokens), float(eval_hit) / num_tokens, wall_time
# do training iterations
for i in six.moves.range(args.num_epochs):
if is_sgd == True:
print('Epoch %d : SGD learning rate = %g' % (i + 1, optimizer.lr))
else:
print('Epoch %d : %s' % (i + 1, args.optimizer))
train_loss = 0.
for j in six.moves.range(len(ids)):
# prepare input data
k = ids[j]
x_batch = [None] * len(data)
for m in six.moves.range(len(data)):
x_batch[m], Q_batch, A_batch = dh.make_batch(
data[m],
train_indices[m][k],
dim=args.in_size[m],
stride=args.frame_stride)
# propagate for training
loss = model(x_batch,
Q_batch,
A_batch,
predicted_context=False,
istraining=True)
if multiGPU:
loss = loss.sum()
wj = loss.cpu()
cur_log_perp += wj.data.numpy()
num_words += A_batch.shape[0] * A_batch.shape[1]
if (n + 1) % report_interval == 0:
now = time.time()
def fetch_batch(dh, data, index, result):
result.append(dh.make_batch(data, index))
min_valid_ppl = 1.0e+10
n = 0
report_interval = 1000 / args.batch_size
bestmodel_num = 0
random.shuffle(train_indices)
# do training iterations
for i in six.moves.range(args.num_epochs):
logging.info('Epoch %d : %s' % (i + 1, args.optimizer))
train_loss = 0.
train_num_words = 0
batch_time = AverageMeter()
data_time = AverageMeter()
end = time.time()
# fetch the first batch
batch = [dh.make_batch(train_data, train_indices[0])]
# train iterations
for j in six.moves.range(len(train_indices)):
data_time.update(time.time() - end)
# get fetched batch
x_batch, h_batch, q_batch, a_batch_in, a_batch_out, s_batch = batch.pop(
)
# fetch the next batch in parallel
if j < len(train_indices) - 1:
prefetch = threading.Thread(
target=fetch_batch,
args=([dh, train_data, train_indices[j + 1], batch]))
prefetch.start()
# propagate for training
x = [torch.from_numpy(x) for x in x_batch]
def generate_response(model,
data,
batch_indices,
vocab,
maxlen=20,
beam=5,
penalty=2.0,
nbest=1,
ref_data=None):
vocablist = sorted(vocab.keys(), key=lambda s: vocab[s])
result_dialogs = []
model.eval()
with torch.no_grad():
qa_id = 0
for idx, dialog in enumerate(data['original']['dialogs']):
vid = dialog['image_id']
if args.undisclosed_only:
out_dialog = dialog['dialog'][-1:]
if ref_data is not None:
ref_dialog = ref_data['dialogs'][idx]
assert ref_dialog['image_id'] == vid
ref_dialog = ref_dialog['dialog'][-1:]
else:
out_dialog = dialog['dialog']
pred_dialog = {
'image_id': vid,
'dialog': copy.deepcopy(out_dialog)
}
result_dialogs.append(pred_dialog)
for t, qa in enumerate(out_dialog):
if args.undisclosed_only:
assert qa['answer'] == '__UNDISCLOSED__'
logging.info('%d %s_%d' % (qa_id, vid, t))
logging.info('QS: ' + qa['question'])
if args.undisclosed_only and ref_data is not None:
logging.info('REF: ' + ref_dialog[t]['answer'])
else:
logging.info('REF: ' + qa['answer'])
# prepare input data
start_time = time.time()
batch = dh.make_batch(
data,
batch_indices[qa_id],
vocab,
separate_caption=train_args.separate_caption)
qa_id += 1
if args.decode_style == 'beam_search':
pred_out, _ = beam_search_decode(
model,
batch,
maxlen,
start_symbol=vocab['<sos>'],
unk_symbol=vocab['<unk>'],
end_symbol=vocab['<eos>'],
pad_symbol=vocab['<blank>'])
for n in range(min(nbest, len(pred_out))):
pred = pred_out[n]
hypstr = []
for w in pred[0]:
if w == vocab['<eos>']:
break
hypstr.append(vocablist[w])
hypstr = " ".join(hypstr)
#hypstr = " ".join([vocablist[w] for w in pred[0]])
logging.info('HYP[%d]: %s ( %f )' %
(n + 1, hypstr, pred[1]))
if n == 0:
pred_dialog['dialog'][t]['answer'] = hypstr
elif args.decode_style == 'greedy':
output = greedy_decode(model,
batch,
maxlen,
start_symbol=vocab['<sos>'],
pad_symbol=vocab['<blank>'])
output = [i for i in output[0].cpu().numpy()]
hypstr = []
for i in output[1:]:
if i == vocab['<eos>']:
break
hypstr.append(vocablist[i])
hypstr = ' '.join(hypstr)
logging.info('HYP: {}'.format(hypstr))
pred_dialog['dialog'][t]['answer'] = hypstr
logging.info('ElapsedTime: %f' % (time.time() - start_time))
logging.info('-----------------------')
return {'dialogs': result_dialogs}
def generate_response(model,
data,
batch_indices,
vocab,
maxlen=20,
beam=5,
penalty=2.0,
nbest=1):
vocablist = sorted(vocab.keys(), key=lambda s: vocab[s])
result_dialogs = []
model.eval()
with torch.no_grad():
qa_id = 0
for dialog in data['original']['dialogs']:
vid = dialog['image_id']
pred_dialog = {
'image_id': vid,
'dialog': copy.deepcopy(dialog['dialog'])
}
result_dialogs.append(pred_dialog)
for t, qa in enumerate(dialog['dialog']):
logging.info('%d %s_%d' % (qa_id, vid, t))
logging.info('QS: ' + qa['question'])
logging.info('REF: ' + qa['answer'])
# prepare input data
start_time = time.time()
b = dh.make_batch(data,
batch_indices[qa_id],
separate_caption=train_args.separate_caption,
pretrained_elmo=train_args.pretrained_elmo,
pretrained_bert=train_args.pretrained_bert,
bert_tokenizer=bert_tokenizer,
pretrained_all=train_args.pretrained_all,
bert_model=train_args.bert_model,
concat_his=train_args.concat_his)
qa_id += 1
x = [torch.from_numpy(x) for x in b[0]]
if train_args.concat_his:
h = [torch.from_numpy(h_i) for h_i in b[1]]
else:
h = [[torch.from_numpy(h) for h in hb] for hb in b[1]]
q = [torch.from_numpy(q) for q in b[2]]
# generate sequences
if train_args.separate_caption:
c = [torch.from_numpy(c) for c in b[5]]
else:
c = None
if train_args.pretrained_elmo or train_args.pretrained_bert:
if train_args.pretrained_all:
context_q, context_h, context_ai = b[-3:]
else:
context_q = b[-1]
context_h = None
context_ai = None
else:
context_q = None
context_h = None
context_ai = None
if train_args.exclude_video:
x = None
if hasattr(train_args,
"i3d_two_stream") and train_args.i3d_two_stream:
min_num_frames = min(x[0].shape[0], x[1].shape[0])
x = [
torch.cat([
x[0][:min_num_frames, :, :],
x[1][:min_num_frames, :, :]
],
dim=2), x[2]
]
if hasattr(model, 'generate'):
pred_out, _ = model.generate(x,
h,
q,
context_q,
context_h,
context_ai,
maxlen=maxlen,
beam=beam,
penalty=penalty,
nbest=nbest,
c=c)
else:
pred_out, _ = model.module.generate(x,
h,
q,
context_q,
context_ai,
maxlen=maxlen,
beam=beam,
penalty=penalty,
nbest=nbest,
c=c)
for n in six.moves.range(min(nbest, len(pred_out))):
pred = pred_out[n]
if 'openai-gpt' in train_args.bert_model:
hypstr = ' '.join([
vocablist[w].replace('</w>', '') for w in pred[0]
])
elif 'gpt2' in train_args.bert_model:
out_tokens = []
for w in pred[0]:
token = vocablist[w]
if token[0].isupper():
out_tokens.append(token[1:])
else:
out_tokens.append(token)
hypstr = ' '.join(out_tokens)
else:
hypstr = ' '.join([vocablist[w] for w in pred[0]])
logging.info('HYP[%d]: %s ( %f )' %
(n + 1, hypstr, pred[1]))
if n == 0:
pred_dialog['dialog'][t]['answer'] = hypstr
logging.info('ElapsedTime: %f' % (time.time() - start_time))
logging.info('-----------------------')
return {'dialogs': result_dialogs}
def fetch_batch(dh, data, index, separate_caption, result):
result.append(dh.make_batch(data, index, separate_caption=separate_caption, pretrained_elmo=args.pretrained_elmo, pretrained_bert=args.pretrained_bert, bert_tokenizer=bert_tokenizer, pretrained_all=args.pretrained_all, bert_model=args.bert_model, concat_his=args.concat_his))
train_log_path = args.model+'_train.csv'
with open(train_log_path, "w") as f:
f.write('epoch,step,perplexity\n')
print("Saving training results to {}".format(train_log_path))
print("Saving val results to {}".format(trace_log_path))
# do training iterations
for i in six.moves.range(args.num_epochs):
if args.lr_scheduler:
scheduler.step()
logging.info('-------------------------Epoch %d : %s-----------------------' % (i+1, args.optimizer))
train_loss = 0.
train_num_words = 0
# fetch the first batch
batch = [dh.make_batch(train_data, train_indices[0], separate_caption=args.separate_caption, pretrained_elmo=args.pretrained_elmo, pretrained_bert=args.pretrained_bert, bert_tokenizer=bert_tokenizer,
pretrained_all=args.pretrained_all, bert_model=args.bert_model,
concat_his=args.concat_his)]
# train iterations
if args.n_batches > 0:
n_batches = args.n_batches
else:
n_batches = len(train_indices)
it = tqdm(six.moves.range(n_batches), desc="epoch {}/{}".format(i, args.num_epochs), ncols=0)
for j in it:
b = batch.pop()
# fetch the next batch in parallel
if j < len(train_indices)-1:
prefetch = threading.Thread(target=fetch_batch,
args=([dh, train_data, train_indices[j+1], args.separate_caption, batch]))
prefetch.start()
