def main():
args = xdnn_io.processCommandLine()
runner = Runner(args['vitis_rundir'])
inTensors = runner.get_input_tensors()
outTensors = runner.get_output_tensors()
batch_sz = args['batch_sz']
if batch_sz == -1:
# use Runner's suggested batch size
batch_sz = inTensors[0].dims[0]
if args['golden']:
goldenMap = xdnn_io.getGoldenMap(args['golden'])
top5Count = 0
top1Count = 0
fpgaBlobs = []
for io in [inTensors, outTensors]:
blobs = []
for t in io:
shape = (batch_sz,) + tuple([t.dims[i] for i in range(t.ndims)][1:])
blobs.append(np.empty((shape), dtype=np.float32, order='C'))
fpgaBlobs.append(blobs)
img_paths = xdnn_io.getFilePaths(args['images'])
labels = xdnn_io.get_labels(args['labels'])
xdnnCPUOp = xdnn.XDNNCPUOp("%s/weights.h5" % args['vitis_rundir'])
fcOutput = np.empty((batch_sz, args['outsz'],), dtype=np.float32, order='C')
fpgaInput = fpgaBlobs[0][0]
for i in range(0, len(img_paths), batch_sz):
pl = []
# fill tensor input data from image file
for j, p in enumerate(img_paths[i:i + batch_sz]):
img, _ = xdnn_io.loadImageBlobFromFile(p,
args['img_raw_scale'], args['img_mean'], args['img_input_scale'],
fpgaInput.shape[2], fpgaInput.shape[3])
pl.append(p)
np.copyto(fpgaInput[j], img)
jid = runner.execute_async(fpgaBlobs[0], fpgaBlobs[1])
runner.wait(jid)
xdnnCPUOp.computeFC(fpgaBlobs[1][0], fcOutput)
softmaxOut = xdnnCPUOp.computeSoftmax(fcOutput)
if args['golden']:
for j,p in enumerate(img_paths[i:i + batch_sz]):
top1Count += xdnn_io.isTopK(softmaxOut[j], goldenMap, p, labels, 1)
top5Count += xdnn_io.isTopK(softmaxOut[j], goldenMap, p, labels, 5)
else:
xdnn_io.printClassification(softmaxOut, pl, labels)
if args['golden']:
print ( ("\nAverage accuracy (n=%d) Top-1: %.1f%%, Top-5: %.1f%%\n") % (len(img_paths), float(top1Count)/float(len(img_paths))*100., float(top5Count)/float(len(img_paths))*100.) )
def run(self, imgList, fpgaOutput_list, fpgaOutputShape_list, shape_list):
fpgaOutput = fpgaOutput_list[0]
fpgaOutputShape = fpgaOutputShape_list[0]
if self.numProcessed == 0:
self.startTime = timeit.default_timer()
self.labels = xdnn_io.get_labels(self.args['labels'])
self.zmqPub = None
if self.args['zmqpub']:
self.zmqPub = ZmqResultPublisher(self.args['deviceID'])
self.goldenMap = None
if self.args['golden']:
self.goldenMap = xdnn_io.getGoldenMap(self.args['golden'])
self.top5Count = 0
self.top1Count = 0
self.fcOutput = np.empty((
self.args['batch_sz'],
self.args['outsz'],
),
dtype=np.float32,
order='C')
self.numProcessed += len(imgList)
npout_view = fpgaOutput
if self.cpuOp._weight is not None:
self.cpuOp.computeFC(npout_view, self.fcOutput)
else:
self.fcOutput = npout_view
smaxOutput = self.cpuOp.computeSoftmax(self.fcOutput)
if self.args['golden']:
for i, p in enumerate(imgList):
#topk = xdnn_io.getTopK( smaxOutput[i], self.labels, 1)
#print(imgList[i], topk)
self.top1Count += xdnn_io.isTopK(\
smaxOutput[i], self.goldenMap, p, self.labels, 1)
self.top5Count += xdnn_io.isTopK(\
smaxOutput[i], self.goldenMap, p, self.labels, 5)
if self.zmqPub is not None:
predictMsg = xdnn_io.getClassification(\
smaxOutput, imgList, self.labels, zmqPub=True)
self.zmqPub.send(predictMsg)