def InferImage(net, image, labels):
import numpy as np
import xdnn_io
global board_avail
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2, 0, 1))
transformer.set_mean('data', np.array([104, 117, 123]))
transformer.set_raw_scale('data', 255)
transformer.set_channel_swap('data',
(2, 1, 0)) # if using RGB instead if BGR
img = caffe.io.load_image(image)
net.blobs['data'].data[...] = transformer.preprocess('data', img)
ptxtShape = net.blobs["data"].data.shape
print("Running with shape of: ", ptxtShape)
with board_avail:
out = net.forward()
for key in out:
try:
if out[key].shape[1] == 1000:
softmax = out[key]
except:
pass
Labels = xdnn_io.get_labels(labels)
xdnn_io.printClassification(softmax, [image], Labels)
result = xdnn_io.getClassification(softmax, [image], Labels)
return result
def post_process(qFrom, args, img_paths, streamQ, fpgaOutputs):
numProcessed = 0
labels = xdnn_io.get_labels(args['labels'])
zmqPub = None
if args['zmqpub']:
zmqPub = ZmqResultPublisher(args['deviceID'])
goldenMap = None
if args['golden']:
goldenMap = xdnn_io.getGoldenMap(args['golden'])
top5Count = 0
top1Count = 0
(fcWeight, fcBias) = xdnn_io.loadFCWeightsBias(args)
bsz = args['batch_sz']
fcOutput = np.empty((
bsz,
args['outsz'],
), dtype=np.float32, order='C')
start = 0
while True:
(sId, img_idx) = qFrom.get()
if numProcessed == 0:
start = timeit.default_timer()
if sId is None or img_idx is None:
break
imgList = []
for x in np.nditer(img_idx):
if x >= 0:
imgList.append(img_paths[x])
numProcessed += 1
npout_view = np.frombuffer(fpgaOutputs[sId].get_obj(),
dtype=np.float32)
xdnn.computeFC(fcWeight, fcBias, npout_view, bsz, args['outsz'],
args['fpgaoutsz'], fcOutput)
streamQ.put(sId)
smaxOutput = xdnn.computeSoftmax(fcOutput)
if args['golden']:
for i, p in enumerate(imgList):
top1Count += xdnn_io.isTopK(smaxOutput[i], goldenMap, p,
labels, 1)
top5Count += xdnn_io.isTopK(smaxOutput[i], goldenMap, p,
labels, 5)
if zmqPub is not None:
predictMsg = xdnn_io.getClassification(smaxOutput,
imgList,
labels,
zmqPub=True)
zmqPub.send(predictMsg)
print("%g images/s" % (float(numProcessed) / (time.time() - start)))
if args['golden']:
print ("\nAverage accuracy (n=%d) Top-1: %.1f%%, Top-5: %.1f%%\n") \
% (numProcessed,
float(top1Count)/float(numProcessed)*100.,
float(top5Count)/float(numProcessed)*100.)
def main(argv):
args = xdnn_io.processCommandLine(argv)
ret, handles = xdnn.createHandle(args['xclbin'], "kernelSxdnn_0")
# ret = xdnn.createHandle(g_xclbin, "kernelSxdnn_0", g_xdnnLib)
if ret != 0:
sys.exit(1)
labels = xdnn_io.get_labels(args['labels'])
# TODO dict of tuples instead?
fpgaRT = {}
fpgaOutputs = {}
fcWeights = {}
fcBiases = {}
netFiles = {}
confNames = []
args = args['jsoncfg'] # we do not use other args' keys
for netconf_args in args:
confName = str(netconf_args['name'])
confNames += [confName]
# netconf_args['netcfg'] = './data/{}_{}.json'.format(netconf_args['net'], netconf_args['dsp'])
fpgaRT[confName] = xdnn.XDNNFPGAOp(handles, netconf_args)
netconf_args['in_shape'] = tuple((netconf_args['batch_sz'],) + tuple(fpgaRT[confName].getInputDescriptors().itervalues().next()[1:] ))
(fcWeights[confName],
fcBiases[confName]) = xdnn_io.loadFCWeightsBias(netconf_args)
fpgaOutputs[confName] = np.empty ((netconf_args['batch_sz'], int(netconf_args['fpgaoutsz']),), dtype=np.float32, order='C')
netFiles[confName] = str(netconf_args['netcfg'])
batchArrays = []
for streamId, netconf_args in enumerate(args):
batchArrays.append(np.empty(netconf_args['in_shape'], dtype=np.float32, order='C'))
pl = []
img_paths = xdnn_io.getFilePaths(netconf_args['images'])
for j, p in enumerate(img_paths[:netconf_args['batch_sz']]):
batchArrays[-1][j, ...], _ = xdnn_io.loadImageBlobFromFile(p, netconf_args['img_raw_scale'],
netconf_args['img_mean'],
netconf_args['img_input_scale'],
netconf_args['in_shape'][2],
netconf_args['in_shape'][3])
pl.append(p)
confName = str(netconf_args['name'])
firstInputName = fpgaRT[confName].getInputs().iterkeys().next()
firstOutputName = fpgaRT[confName].getOutputs().iterkeys().next()
fpgaRT[confName].exec_async({ firstInputName : batchArrays[-1] }, { firstOutputName : fpgaOutputs[confName] }, streamId)
for streamId, confName in enumerate(confNames):
fpgaRT[confName].get_result (streamId)
for netconf_args in args:
confName = str(netconf_args['name'])
fcOut = np.empty( (netconf_args['batch_sz'], netconf_args['outsz']), dtype=np.float32, order = 'C')
xdnn.computeFC (fcWeights[confName], fcBiases[confName], fpgaOutputs[confName], fcOut)
softmaxOut = xdnn.computeSoftmax(fcOut)
xdnn_io.printClassification(softmaxOut, netconf_args['images'], labels);
xdnn.closeHandle()
def init_fpga():
# Instead of using command line, we hard code it here.
# Typing correct args is almost impossible so either do it in .sh or .py
#
global g_args
global g_ctxt
print(" --- INIT FPGA --- \n")
xdnnArgs = build_xdnn_args()
print(xdnnArgs)
g_args = xdnn_io.processCommandLine(xdnnArgs)
print(" --- After parsing --- \n")
print(g_args)
print(" --- Create handle --- \n")
ret, handles = xdnn.createHandle(g_args['xclbin'], "kernelSxdnn_0")
if ret != 0:
print(" --- !!! FAILED: Cannot create handle. --- \n")
sys.exit(1)
print(" --- Create fpgaRT --- \n")
fpgaRT = xdnn.XDNNFPGAOp(handles, g_args)
g_ctxt["fpgaRT"] = fpgaRT
print(" --- Weight and Bias --- \n")
fcWeight, fcBias = xdnn_io.loadFCWeightsBias(g_args)
g_ctxt["fcWeight"] = fcWeight
g_ctxt["fcBias"] = fcBias
print(" --- Init input input/output area --- \n")
if is_deploymode():
g_ctxt['fpgaOutput'] = fpgaRT.getOutputs()
g_ctxt['fpgaInput'] = fpgaRT.getInputs()
g_ctxt['inShape'] = (g_args['batch_sz'], ) + tuple(
fpgaRT.getInputDescriptors().itervalues().next()[1:])
else:
g_ctxt['fpgaOutput'] = np.empty((
g_args['batch_sz'],
g_args['fpgaoutsz'],
),
dtype=np.float32,
order='C')
g_ctxt['batch_array'] = np.empty(
((g_args['batch_sz'], ) + g_args['in_shape']),
dtype=np.float32,
order='C')
g_ctxt['fcOutput'] = np.empty((
g_args['batch_sz'],
g_args['outsz'],
),
dtype=np.float32,
order='C')
print(" --- Get lables --- \n")
g_ctxt['labels'] = xdnn_io.get_labels(g_args['labels'])
# golden? What is that?
# Seems we are done.
print(" --- FPGA INITIALIZED! ---\n")
def main():
args = xdnn_io.processCommandLine()
ret, handles = xdnn.createHandle(args['xclbin'], "kernelSxdnn_0")
if ret != 0:
sys.exit(1)
fpgaRT = xdnn.XDNNFPGAOp(handles, args)
fcWeight, fcBias = xdnn_io.loadFCWeightsBias(args)
img_paths = xdnn_io.getFilePaths(args['images'])
fpgaOutput = np.empty((
args['batch_sz'],
args['fpgaoutsz'],
),
dtype=np.float32,
order='C')
fcOutput = np.empty((
args['batch_sz'],
args['outsz'],
),
dtype=np.float32,
order='C')
batch_array = np.empty(((args['batch_sz'], ) + args['in_shape']),
dtype=np.float32,
order='C')
labels = xdnn_io.get_labels(args['labels'])
if args['golden']:
goldenMap = xdnn_io.getGoldenMap(args['golden'])
top5Count = 0
top1Count = 0
for i in xrange(0, len(img_paths), args['batch_sz']):
pl = []
for j, p in enumerate(img_paths[i:i + args['batch_sz']]):
batch_array[j, ...], _ = xdnn_io.loadImageBlobFromFile(
p, args['img_raw_scale'], args['img_mean'],
args['img_input_scale'], args['in_shape'][2],
args['in_shape'][1])
pl.append(p)
fpgaRT.execute(batch_array, fpgaOutput)
xdnn.computeFC(fcWeight, fcBias, fpgaOutput, args['batch_sz'],
args['outsz'], args['fpgaoutsz'], fcOutput)
softmaxOut = xdnn.computeSoftmax(fcOutput)
xdnn_io.printClassification(softmaxOut, pl, labels)
if args['golden']:
for j, p in enumerate(img_paths[i:i + args['batch_sz']]):
top1Count += xdnn_io.isTopK(softmaxOut[j], goldenMap, p,
labels, 1)
top5Count += xdnn_io.isTopK(softmaxOut[j], goldenMap, p,
labels, 5)
xdnn.closeHandle()
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 __init__(self, maxNumStreams):
self._maxNumStreams = maxNumStreams
self._streamsAvailable = []
self._streamInputs = []
self._streamOutputs = []
self._config = xdnn_io.processCommandLine()
ret, handles = xdnn.createHandle(self._config['xclbin'])
if ret != 0:
sys.exit(1)
self._fpgaRT = xdnn.XDNNFPGAOp(handles, self._config)
self._fcWeight, self._fcBias = xdnn_io.loadFCWeightsBias(self._config)
self._labels = xdnn_io.get_labels(self._config['labels'])
for i in range(maxNumStreams):
self._streamsAvailable.append(i)
self._streamInputs.append(None)
self._streamOutputs.append(None)