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_train(sess, data, batch_size, v2i, hf1, hf2, feature_shape, train, loss, input_video, input_captions, 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_shape[0],2048)) data_v2 = MsrDataUtil.getBatchC3DVideoFeature(batch_caption,hf2,(feature_shape[0],4096)) # data_v1 = data_v1/(np.linalg.norm(data_v1, ord=None, axis=-1, keepdims=True)+sys.float_info.epsilon) # data_v2 = data_v2/(np.linalg.norm(data_v2, ord=None, axis=-1, keepdims=True)+sys.float_info.epsilon) data_v = np.concatenate((data_v1,data_v2),axis=-1) data_c, data_y = MsrDataUtil.getBatchTrainCaptionWithSparseLabel(batch_caption, v2i, capl=capl) _, l = sess.run([train,loss],feed_dict={input_video:data_v, input_captions:data_c, 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_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 exe_train(sess, data, batch_size, v2i, hf, feature_shape, train, loss, input_video, input_captions, 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) flag = np.random.randint(0, 2) if flag == 1: data_v = data_v[:, ::-1, :] data_c, data_y = MsrDataUtil.getBatchTrainCaptionWithSparseLabel( batch_caption, v2i, capl=capl) _, l = sess.run([train, loss], feed_dict={ input_video: data_v, input_captions: data_c, 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