def main(args=None):
os.environ['LS_BIND_NOW'] = "1"
args = xdnn_io.processCommandLine()
images = xdnn_io.getFilePaths(args['images'])
# spawn dispatcher
dispatcher = Dispatcher(args['vitis_rundir'], g_nFPGA, g_nDispatchers,
args['batch_sz'])
inshape = dispatcher.inshape
# send work to system
work = []
for qIdx in range(g_nQueries):
idx = qIdx * inshape[0]
workBatch = [
images[(idx + i) % len(images)] for i in range(inshape[0])
]
work.append((qIdx, workBatch, (args['img_raw_scale'], args['img_mean'],
args['img_input_scale'])))
startTime = timeit.default_timer()
dispatcher.run(work)
del dispatcher
t = timeit.default_timer() - startTime
print("Queries: %d, Elapsed: %.2fs, QPS: %.2f, FPS: %.2f" \
% (g_nQueries, t, g_nQueries / t, g_nQueries * inshape[0] / t))
sys.stdout.flush()
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 __init__(self):
global pool
global pFpgaRT
self.args = xdnn_io.processCommandLine() # Get command line args
pool = Pool(
self.args['numproc']
) # Depends on new switch, new switch is not added to xdnn_io.py as of yet
# Split Images into batches - list of lists
self.batches = []
self.all_image_paths = xdnn_io.getFilePaths(
self.args['images']) #[0:10000]
for i in range(0, len(self.all_image_paths), self.args['batch_sz']):
self.batches.append(self.all_image_paths[i:i +
self.args['batch_sz']])
pFpgaRT = xdnn.XDNNFPGAOp(self.args) # Parent process gets handle
self.args['inShape'] = (self.args['batch_sz'], ) + tuple(
tuple(pFpgaRT.getInputDescriptors().values())[0]
[1:]) # Save input shape for children
self.mpid = pFpgaRT.getMPID() # Save handle to use in child processes
def main():
args = xdnn_io.processCommandLine()
images = xdnn_io.getFilePaths(args['images'])
# start comms
xserver = xstream.Server()
# acquire resources
fmaster = FpgaMaster(args['vitis_rundir'])
inshape = list(fmaster.inshape)
if args['batch_sz'] != -1:
inshape[0] = args['batch_sz'] # update batch size
# spawn dispatchers
dispatcher = Dispatcher(g_nDispatchers, g_nWorkers, inshape)
# spawn workers
workers = WorkerPool(args['vitis_rundir'] + "_worker", g_nWorkers, args)
# send work to system
work = []
for qIdx in range(g_nQueries):
idx = qIdx * inshape[0]
workBatch = [
images[(idx + i) % len(images)] for i in range(inshape[0])
]
work.append((qIdx, workBatch, (args['img_raw_scale'], args['img_mean'],
args['img_input_scale'])))
startTime = timeit.default_timer()
dispatcher.run(work)
del dispatcher
t = timeit.default_timer() - startTime
print("Queries: %d, Elapsed: %.2fs, QPS: %.2f, FPS: %.2f" \
% (g_nQueries, t, g_nQueries / t, g_nQueries * inshape[0] / t))
# cleanup
del workers
del fmaster
del xserver
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.))
if __name__ == '__main__':
print("\n\n\n\n\n\n\n\n" + '\33[32m' +
"Running Inference with HW Pre-processing with JPEG decoder" +
'\33[0m')
args = xdnn_io.processCommandLine()
#Create a queue for passing the pre-processed data
q = mp.Queue()
#Creating a process to run HW pre-processing kernel
p_preprocess = mp.Process(target=pre_process, args=(q, args))
#Process to run XDNN
p_xdnn = mp.Process(target=process_xdnn, args=(q, args))
p_preprocess.start()
p_xdnn.start()
p_preprocess.join()
p_xdnn.join()
