def beam_search_exe_test(sess,
data,
cate_info,
batch_size,
v2i,
i2v,
hf,
feature_shape,
predict_words,
input_video,
input_captions,
input_categories,
y,
finished_beam,
logprobs_finished_beams,
past_logprobs,
capl=16):
caption_output = []
total_data = len(data)
num_batch = int(round(total_data * 1.0 / batch_size))
for batch_idx in xrange(num_batch):
batch_caption = data[batch_idx *
batch_size:min((batch_idx + 1) *
batch_size, total_data)]
data_v = MsrDataUtil.getBatchVideoFeature(batch_caption, hf,
feature_shape)
data_c, data_y = MsrDataUtil.getBatchTestCaptionWithSparseLabel(
batch_caption, v2i, capl=capl)
data_cate = MsrDataUtil.getBatchVideoCategoriesInfo(
batch_caption, cate_info, feature_shape)
[gw, tw, gp, gl] = sess.run(
[
finished_beam, predict_words, logprobs_finished_beams,
past_logprobs
],
feed_dict={
input_video: data_v,
input_captions: data_c,
input_categories: data_cate,
y: data_y
})
generated_captions = MsrDataUtil.convertCaptionI2V(
batch_caption, gw, i2v)
for idx, sen in enumerate(generated_captions):
print('%s : %s' % (batch_caption[idx].keys()[0], sen))
caption_output.append({
'image_id': batch_caption[idx].keys()[0],
'caption': sen
})
js = {}
js['val_predictions'] = caption_output
return js
def exe_train(sess,
data,
cate_info,
batch_size,
v2i,
hf1,
hf2,
feature_shape1,
feature_shape2,
train,
loss,
input_video1,
input_video2,
input_captions,
input_categories,
y,
capl=16):
np.random.shuffle(data)
total_data = len(data)
num_batch = int(round(total_data * 1.0 / batch_size))
total_loss = 0.0
for batch_idx in xrange(num_batch):
# for batch_idx in xrange(500):
# if batch_idx < 100:
batch_caption = data[batch_idx *
batch_size:min((batch_idx + 1) *
batch_size, total_data)]
data_v1 = MsrDataUtil.getBatchVideoFeature(batch_caption, hf1,
feature_shape1)
data_v2 = MsrDataUtil.getBatchC3DVideoFeature(batch_caption, hf2,
feature_shape2)
flag = np.random.randint(0, 2)
if flag == 1:
data_v1 = data_v1[:, ::-1, :]
data_v2 = data_v2[:, ::-1, :]
data_c, data_y = MsrDataUtil.getBatchTrainCaptionWithSparseLabel(
batch_caption, v2i, capl=capl)
data_cate = MsrDataUtil.getBatchVideoCategoriesInfo(
batch_caption, cate_info, feature_shape1)
_, l = sess.run(
[train, loss],
feed_dict={
input_video1: data_v1,
input_video2: data_v2,
input_captions: data_c,
input_categories: data_cate,
y: data_y
})
total_loss += l
print(' batch_idx:%d/%d, loss:%.5f' % (batch_idx + 1, num_batch, l))
total_loss = total_loss / num_batch
return total_loss
def exe_test(sess, data, cate_info, batch_size, v2i, i2v, hf1, hf2, feature_shape1, feature_shape2,
predict_words, input_video1, input_video2, input_captions, input_categories, y, capl=16):
caption_output = []
total_data = len(data)
num_batch = int(round(total_data*1.0/batch_size))+1
for batch_idx in xrange(num_batch):
batch_caption = data[batch_idx*batch_size:min((batch_idx+1)*batch_size,total_data)]
data_v1 = MsrDataUtil.getBatchVideoFeature(batch_caption,hf1,feature_shape1)
data_v2 = MsrDataUtil.getBatchC3DVideoFeature(batch_caption,hf2,feature_shape2)
data_c, data_y = MsrDataUtil.getBatchTestCaptionWithSparseLabel(batch_caption, v2i, capl=capl)
data_cate = MsrDataUtil.getBatchVideoCategoriesInfo(batch_caption, cate_info, feature_shape1)
[gw] = sess.run([predict_words],feed_dict={input_video1:data_v1, input_video2:data_v2, input_captions:data_c, input_categories:data_cate, y:data_y})
generated_captions = MsrDataUtil.convertCaptionI2V(batch_caption, gw, i2v)
for idx, sen in enumerate(generated_captions):
print('%s : %s' %(batch_caption[idx].keys()[0],sen))
caption_output.append({'image_id':batch_caption[idx].keys()[0],'caption':sen})
js = {}
js['val_predictions'] = caption_output
return js
def exe_train(sess,
data,
audio_info,
cate_info,
batch_size,
v2i,
hf,
feature_shape,
train,
loss,
input_video,
input_captions,
input_categories,
input_audio,
y,
capl=16):
np.random.shuffle(data)
total_data = len(data)
num_batch = int(round(total_data * 1.0 / batch_size))
total_loss = 0.0
for batch_idx in xrange(num_batch):
# for batch_idx in xrange(500):
# if batch_idx < 100:
batch_caption = data[batch_idx *
batch_size:min((batch_idx + 1) *
batch_size, total_data)]
data_v = MsrDataUtil.getBatchVideoFeature(batch_caption, hf,
feature_shape)
data_c, data_y = MsrDataUtil.getBatchTrainCaptionWithSparseLabel(
batch_caption, v2i, capl=capl)
data_cate = MsrDataUtil.getBatchVideoCategoriesInfo(
batch_caption, cate_info, feature_shape)
data_audio = MsrDataUtil.getBatchVideoAudioInfo(
batch_caption, audio_info, feature_shape)
_, l = sess.run(
[train, loss],
feed_dict={
input_video: data_v,
input_captions: data_c,
input_categories: data_cate,
input_audio: data_audio,
y: data_y
})
total_loss += l
print(' batch_idx:%d/%d, loss:%.5f' % (batch_idx + 1, num_batch, l))
total_loss = total_loss / num_batch
return total_loss
def beamsearch_exe_test(sess,
data,
cate_info,
batch_size,
v2i,
i2v,
hf1,
hf2,
feature_shape1,
feature_shape2,
predict_words,
input_video1,
input_video2,
input_captions,
input_categories,
y,
finished_beam,
logprobs_finished_beams,
capl=16):
caption_output = []
total_data = len(data)
num_batch = int(round(total_data * 1.0 / batch_size))
for batch_idx in xrange(num_batch):
batch_caption = data[batch_idx *
batch_size:min((batch_idx + 1) *
batch_size, total_data)]
data_v1 = MsrFinalDataUtil.getBatchVideoFeature(
batch_caption, hf1, feature_shape1)
data_v2 = MsrFinalDataUtil.getBatchC3DVideoFeature(
batch_caption, hf2, feature_shape2)
data_c, data_y = MsrDataUtil.getBatchTestCaptionWithSparseLabel(
batch_caption, v2i, capl=capl)
data_cate = MsrDataUtil.getBatchVideoCategoriesInfo(
batch_caption, cate_info, feature_shape1)
[tw, gw, gp] = sess.run(
[predict_words, finished_beam, logprobs_finished_beams],
feed_dict={
input_video1: data_v1,
input_video2: data_v2,
input_captions: data_c,
input_categories: data_cate,
y: data_y
})
generated_captions = MsrDataUtil.convertCaptionI2V(
batch_caption, gw, i2v)
for idx, sen in enumerate(generated_captions):
print('%s : %s' % (batch_caption[idx].keys()[0], sen))
caption_output.append({
'video_id': batch_caption[idx].keys()[0],
'caption': sen.strip()
})
js = {}
js['result'] = caption_output
js['version'] = '3'
js['external_data'] = {
'used':
'true',
'details':
'the features of video frames are extracted by ResNet152 and C3D'
}
return js
