caffe---之如何实现多任务学习

一.制作数据集：

首先制作lmdb的多任务标签，使用python来做代码如下（做两个lmdb一个用于存储data,一个用于存储label,这是训练集的做法，验证集同理）：

    X = np.arange(4356).reshape(4,33,33)
    map_size = X.nbytes*100000*2
    env_data = lmdb.open('/yourlmdbpath/33-multilabel-train', map_size=map_size)
    env_label = lmdb.open('yourlmdbpath/33-multilabel-val', map_size=map_size)
    data = np.zeros((4,33,33))
    q = 0
    env_data_txn = env_data.begin(write=True)
    env_label_txn = env_label.begin(write=True)
    for i in index_sam:
        for j in range(4):
            patch_data = get_patch(four_modalities[i[0],:,:,j],(i[1],i[2]))
            data[j,:,:] = patch_data
            #data = data.astype('int16')
        datum = caffe.proto.caffe_pb2.Datum()
        datum.channels = data.shape[0]
        datum.height = data.shape[1]
        datum.width = data.shape[2]
        #datum.data = data.tobytes() # or .tostring() if numpy < 1.9
        datum.float_data.extend(data.flat)
        str_id = '{:08}'.format(q)
        q = q+1
        print q
        # The encode is only essential in Python 3
        env_data_txn.put(str_id.encode('ascii'), datum.SerializeToString())
        print str_id,'data_label:',brain_label[i[0],i[1],i[2]]

	#制作label lmdb
	multi_label = np.zeros((3,1,1),np.int8)
	multi_label[0,0,0] = int(brain_label[i[0],i[1],i[2]])
	if brain_label[i[0],i[1],i[2]] == 0:
		multi_label[1,0,0] == int(0)
	else: 
		multi_label[1,0,0] == int(1)
	if brain_label[i[0],i[1],i[2]] == 0 or brain_label[i[0],i[1],i[2]] == 2:
		multi_label[2,0,0] == int(0)
	else:
		multi_label[2,0,0] == int(1)
	datum_label = caffe.io.array_to_datum(multi_label)
	env_label_txn.put('{:0>10d}'.format(q), datum_label.SerializeToString())

    env_data_txn.commit()
    env_label_txn.commit()
    env_data.close()
    env_label.close()
                    
    print 'dnoe'

做完数据集后，使用4个data层读入数据,分别两个data用于训练，两个data用于测试，要用slice层来切分标签，

slice层和4个data层的使用方法如下：

#训练数据载入
layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TRAIN
  }
  #transform_param {
  #  scale: 0.00390625
  #}
  data_param {
    source: "/media/czn/e04e3ecf-cf63-416c-afd7-6d737e09968a/BrainTuomr_new_experiment/HGG_Test/data/multi_label_lmdb/train/train_data_lmdb"
    batch_size: 128
    backend: LMDB
  }
}

layer {
  name: "label"
  type: "Data"
  top: "label"
  include {
    phase: TRAIN
  }
  #transform_param {
  #  scale: 0.00390625
  #}
  data_param {
    # 这里定义了验证集的数据来源
    source: "/media/czn/e04e3ecf-cf63-416c-afd7-6d737e09968a/BrainTuomr_new_experiment/HGG_Test/data/multi_label_lmdb/train/train_label_lmdb"
    batch_size: 128
    backend: LMDB
  }
}
#测试数据载入
layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TEST
  }
  #transform_param {
  #  scale: 0.00390625
  #}
  data_param {
    source: "/media/czn/e04e3ecf-cf63-416c-afd7-6d737e09968a/BrainTuomr_new_experiment/HGG_Test/data/multi_label_lmdb/val/33-multilabel-train"
    batch_size: 128
    backend: LMDB
  }
}

layer {
  name: "label"
  type: "Data"
  top: "label"
  include {
    phase: TEST
  }
  #transform_param {
  #  scale: 0.00390625
  #}
  data_param {
    # 这里定义了验证集的数据来源
    source: "/media/czn/e04e3ecf-cf63-416c-afd7-6d737e09968a/BrainTuomr_new_experiment/HGG_Test/data/multi_label_lmdb/val/33-multilabel-val"
    batch_size: 128
    backend: LMDB
  }
}
#第一个Slice层
layer {
  name: "slice"
  type: "Slice"
  bottom: "label"
  top: "label_1"
  top: "label_2"
  top: "label_3"
  slice_param {
    axis: 1
    slice_point: 1
    slice_point: 2
  }
}