def request_process(args, img_paths, inchan, outchan):
# print("Starting REQ PROC")
q = Queue(100) # only allow this many requests in flight at once
def throttler():
xsResult = xstream.Subscribe(outchan)
while True:
msg = q.get()
if msg == None:
break
xsResult.get()
t = Thread(target=throttler)
t.start()
# pad num images to be multiple of batch_sz
if len(img_paths) % args['batch_sz']:
npad = args['batch_sz'] - (len(img_paths) % args['batch_sz'])
for i in range(npad):
img_paths.append(img_paths[-1])
xs = xstream.Publisher()
for inum, fpath in enumerate(img_paths):
(buf, objId) = xs.new(1,
obj_id="img_%d" % inum,
meta={
'id': inum,
'path': fpath
})
q.put('+')
xs.put(inchan, objId)
# print("REQ put data : {} ".format(fpath))
q.put(None)
t.join()
def __init__(self):
signal.signal(signal.SIGINT, self.signal_handler)
# signal.signal(signal.SIGQUIT, self.sigquit_handler)
self.do_exit = False
self.hbeat_id = 0
self.hbeat = 0
# self.do_restart = False
self.xserver = xstream.Server()
parse_command_line()
self.web_app = ServerWebApplication()
# Add handlers for default services here so they can be destroyed if
# needed. Handlers installed in the web_app __init__ cannot be destroyed.
recipes = self.get_recipes()
for recipe in recipes:
tornado_handler = construct.construct(recipe)
self.web_app.add_handlers(
r".*", # match any host
tornado_handler)
self.web_server = self.web_app.listen(options.port)
self.ws_app = tornado.web.Application([(r"/", WebSocketHandler)])
self.ws_server = self.ws_app.listen(options.wsport)
self.xspub = xstream.Publisher()
self.xs2server = threading.Thread(target=ServerApp.xstream2server)
self.xs2server.start()
self.heartbeat_thread = threading.Thread(target=ServerApp.heartbeat,
args=(lambda: self.do_exit, ))
self.heartbeat_thread.start()
def run(rundir, chanIdx, q, args):
xspub = xstream.Publisher()
xssub = xstream.Subscribe(chanIdx2Str(chanIdx))
runner = Runner(rundir)
inTensors = runner.get_input_tensors()
outTensors = runner.get_output_tensors()
q.put(1) # ready for work
fpgaBlobs = None
fcOutput = None
labels = xdnn_io.get_labels(args['labels'])
xdnnCPUOp = xdnn.XDNNCPUOp("%s/weights.h5" % rundir)
while True:
try:
payload = xssub.get()
if not payload:
break
(meta, buf) = payload
if fpgaBlobs == None:
# allocate buffers
fpgaBlobs = []
batchsz = meta['shape'][0] # inTensors[0].dims[0]
for io in [inTensors, outTensors]:
blobs = []
for t in io:
shape = (batchsz,) + tuple([t.dims[i] for i in range(t.ndims)][1:])
blobs.append(np.empty((shape), dtype=np.float32, order='C'))
fpgaBlobs.append(blobs)
fcOutput = np.empty((batchsz, args['outsz'],), dtype=np.float32, order='C')
fpgaInput = fpgaBlobs[0][0]
assert(tuple(meta['shape']) == fpgaInput.shape)
data = np.frombuffer(buf, dtype=np.float32).reshape(fpgaInput.shape)
np.copyto(fpgaInput, data)
jid = runner.execute_async(fpgaBlobs[0], fpgaBlobs[1])
runner.wait(jid)
xdnnCPUOp.computeFC(fpgaBlobs[1][0], fcOutput)
softmaxOut = xdnnCPUOp.computeSoftmax(fcOutput)
xdnn_io.printClassification(softmaxOut, meta['images'], labels)
sys.stdout.flush()
if meta['id'] % 1000 == 0:
print("Recvd query %d" % meta['id'])
sys.stdout.flush()
del data
del buf
del payload
xspub.send(meta['from'], "success")
except Exception as e:
logging.error("Worker exception " + str(e))
def _init(nWorkers, inshape):
token = pid2TokenStr()
if token not in Dispatcher.xspub:
Dispatcher.xspub[token] = xstream.Publisher()
Dispatcher.xstoken[token] = xstream.Subscribe(token)
Dispatcher.inshape = inshape
Dispatcher.nWorkers = nWorkers
Dispatcher.inBlob[token] = np.zeros(tuple(inshape), dtype=np.float32, order='C')
def __init__(self, rundir, nWorkers, workerArgs):
self.xspub = xstream.Publisher()
self.workers = []
self.wq = mp.Queue()
for wi in range(nWorkers):
w = mp.Process(target=WorkerPool.run, args=(rundir, wi, self.wq, workerArgs,))
w.start()
self.workers.append(w)
# wait for worker to be ready before starting next worker
# because the last worker overwrites the IP programming
# (Note: this assumes all workers must have the same IP instructions)
self.wq.get()
def run(rundir, chanIdx, q, args):
xspub = xstream.Publisher()
xssub = xstream.Subscribe(chanIdx2Str(chanIdx))
runner = Runner(rundir)
inTensors = runner.get_input_tensors()
outTensors = runner.get_output_tensors()
q.put(1) # ready for work
fpgaBlobs = None
labels = xdnn_io.get_labels(args['labels'])
if args['yolo_version'] == 'v2': yolo_postproc = yolo.yolov2_postproc
elif args['yolo_version'] == 'v3': yolo_postproc = yolo.yolov3_postproc
else:
assert args['yolo_version'] in (
'v2', 'v3'), "--yolo_version should be <v2|v3>"
biases = bias_selector(args)
if (args['visualize']): colors = generate_colors(len(labels))
while True:
try:
payload = xssub.get()
if not payload:
break
(meta, buf) = payload
if fpgaBlobs == None:
# allocate buffers
fpgaBlobs = []
batchsz = meta['shape'][0] # inTensors[0].dims[0]
for io in [inTensors, outTensors]:
blobs = []
for t in io:
shape = (batchsz, ) + tuple(
[t.dims[i] for i in range(t.ndims)][1:])
blobs.append(
np.empty((shape), dtype=np.float32, order='C'))
fpgaBlobs.append(blobs)
fcOutput = np.empty((
batchsz,
args['outsz'],
),
dtype=np.float32,
order='C')
fpgaInput = fpgaBlobs[0][0]
assert (tuple(meta['shape']) == fpgaInput.shape)
data = np.frombuffer(buf,
dtype=np.float32).reshape(fpgaInput.shape)
np.copyto(fpgaInput, data)
jid = runner.execute_async(fpgaBlobs[0], fpgaBlobs[1])
runner.wait(jid)
boxes = yolo_postproc(fpgaBlobs[1],
args,
meta['image_shapes'],
biases=biases)
if (not args['profile']):
for i in range(min(batchsz, len(meta['image_shapes']))):
print("Detected {} boxes in {}".format(
len(boxes[i]), meta['images'][i]),
flush=True)
# Save the result
if (args['results_dir']):
for i in range(min(batchsz, len(meta['image_shapes']))):
fname = meta['images'][i]
filename = os.path.splitext(os.path.basename(fname))[0]
out_file_txt = os.path.join(args['results_dir'],
filename + '.txt')
print("Saving {} boxes to {}".format(
len(boxes[i]), out_file_txt))
sys.stdout.flush()
saveDetectionDarknetStyle(out_file_txt, boxes[i],
meta['image_shapes'][i])
if (args['visualize']):
out_file_png = os.path.join(
args['results_dir'], filename + '.png')
print("Saving result to {}".format(out_file_png))
sys.stdout.flush()
draw_boxes(fname, boxes[i], labels, colors,
out_file_png)
if meta['id'] % 1000 == 0:
print("Recvd query %d" % meta['id'])
sys.stdout.flush()
del data
del buf
del payload
xspub.send(meta['from'], "success")
except Exception as e:
logging.error("Worker exception " + str(e))
