class multilabel(caffe.Layer):
'''
This is a simple datalayer for train multilabel model on clothes datasets
'''
def setup(self, bottom, top):
print "my datalayer is inilializing...."
label_numbers = len(par.top_names)
layer_params = eval(self.param_str)
self.phase = layer_params['phase']
top[0].reshape(par.batchsize_dic[self.phase], par.channel, par.height,
par.width)
for i in range(1, label_numbers):
top[i].reshape(par.batchsize_dic[self.phase], 1)
self.batchloader = Batchloader(par.data_label_dic[self.phase][0],
par.data_label_dic[self.phase][1])
self.set_queue()
print "---------------------inilialized %s----------------" % (
self.phase, )
def set_queue(self):
self.blob_queue = Queue(par.queue_num)
self.prefetch_process = BlobFetcher(par.data_label_dic[self.phase][0],
par.data_label_dic[self.phase][1],
self.blob_queue, self.phase)
self.prefetch_process.start()
def clean():
print "Terminating BlobFetcher"
self.prefetch_process.terminate()
self.prefetch_process.join()
import atexit
atexit.register(clean)
def reshape(self, bottom, top):
'''do nothing'''
pass
def get_batch(self):
'''interface control if use multi-process'''
if par.Using_Process_dic[self.phase]:
return self.blob_queue.get()
else:
return self.batchloader.get_minibatch(self.phase)
def forward(self, bottom, top):
blob = self.get_batch(
) #blob=[[image,[color,long,sleeve]],[image,[color,long,sleeve]],....[]]
num_labels = len(blob[0][1])
for i in xrange(par.batchsize_dic[self.phase]):
top[0].data[i, ...] = blob[i][0] #data
for j in range(num_labels):
top[j + 1].data[i, ...] = int(blob[i][1][j]) #label
def backward(self, bottom, top):
'''do nothing'''
pass
class BlobFetcher(Process):
'''using multi-process prepare blobs'''
def __init__(self,imagepath,annotion,queue,phase):
super(BlobFetcher,self).__init__()
self.batch=Batchloader(imagepath,annotion)
self.queue=queue
self.phase=phase
def run(self):
print "---------------BlobFetcher start------------------------"
while True:
blobs=self.batch.get_minibatch(self.phase)
self.queue.put(blobs)