def _run(work):
try:
(idx, images, args) = work
chanIdx = idx % Dispatcher.nWorkers
token = pid2TokenStr()
shape = Dispatcher.inshape
for i, img in enumerate(images):
cached = Dispatcher.inBlobCache.get(img)
if cached is None:
Dispatcher.inBlob[token][
i, ...], _ = xdnn_io.loadImageBlobFromFile(
img, args[0], args[1], args[2], shape[2], shape[3])
Dispatcher.inBlobCache.set(
img, np.copy(Dispatcher.inBlob[token][i]))
else:
Dispatcher.inBlob[token][i, ...] = cached
meta = {'id': idx, 'from': token, 'shape': shape, 'images': images}
if idx % 1000 == 0:
print("Put query %d to objstore" % idx)
sys.stdout.flush()
Dispatcher.xspub[token].put_blob(chanIdx2Str(chanIdx),
Dispatcher.inBlob[token], meta)
Dispatcher.xstoken[token].get_msg()
except Exception as e:
logging.error("Producer exception " + str(e))
def pre_process(self, image_paths, fpgaInput):
input_buffer = list(fpgaInput.values(
))[0] # Assume the first network input needs to have image data loaded
for index, image_path in enumerate(image_paths):
input_buffer[index, ...], _ = xdnn_io.loadImageBlobFromFile(
image_path, self.args['img_raw_scale'], self.args['img_mean'],
self.args['img_input_scale'], self.args['inShape'][2],
self.args['inShape'][3])
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, inum_chunk):
write_slot = self._shared_trans_arrs.openWriteId()
write_arrs = self._shared_trans_arrs.accessNumpyBuffer(write_slot)
if not self._args['benchmarkmode']:
for i, inum in enumerate(inum_chunk):
write_arrs[0][i][:], shape = xdnn_io.loadImageBlobFromFile(self._imgpaths[inum], self._args['img_raw_scale'], self._meanarr,
self._args['img_input_scale'], self._firstInputShape[2], self._firstInputShape[3])
write_arrs[-1][i][0] = inum
write_arrs[-1][i][1:4] = shape
# Fill -1 for unfilled image slots in whole batch
write_arrs[-1][len(inum_chunk):][:] = -1
self._shared_trans_arrs.closeWriteId(write_slot)