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 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 loop(self): fpgaOutputShapes = [] for idx in range(len(self.output_shapes)): fpgaOutputShape_l = self.output_shapes[idx] fpgaOutputShape_l[0] = self.args['batch_sz'] fpgaOutputShapes.append(fpgaOutputShape_l) if self.args['yolo_version'] == 'v2': self.yolo_postproc = yolo.yolov2_postproc elif self.args['yolo_version'] == 'v3': self.yolo_postproc = yolo.yolov3_postproc self.biases = bias_selector(self.args) self.labels = xdnn_io.get_labels(self.args['labels']) self.colors = generate_colors(len(self.labels)) while True: read_slot = self._shared_output_arrs.openReadId() if read_slot is None: break read_slot_arrs = self._shared_output_arrs.accessNumpyBuffer( read_slot) imgList = [] shape_list = [] #image_id = self._qFrom.get() num_images = (read_slot_arrs[-1].shape)[0] for image_num in range(num_images): image_id = read_slot_arrs[-1][image_num][0] if image_id == -1: break imgList.append(self.img_paths[int(image_id)]) shape_list.append(read_slot_arrs[-1][image_num][1:4]) if self.args["benchmarkmode"]: self.numProcessed += len(imgList) #self.streamQ.put(sId) self._shared_output_arrs.closeReadId(read_slot) continue self.run(imgList, read_slot_arrs[0:-1], fpgaOutputShapes, shape_list) self._shared_output_arrs.closeReadId(read_slot) self.finish()
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)
def post_process(self, image_paths, fpgaOutput): global labels global fcOutput global cCpuRT if cCpuRT is None: cCpuRT = xdnn.XDNNCPUOp(self.args['weights']) fcOutput = np.empty(( self.args['batch_sz'], self.args['outsz'], ), dtype=np.float32, order='C') labels = xdnn_io.get_labels(self.args['labels']) input_buffer = list(fpgaOutput.values())[ 0] # Assume the first network output will feed CPU layers cCpuRT.computeFC(input_buffer, fcOutput) softmaxOut = cCpuRT.computeSoftmax(fcOutput) if self.args['golden']: xdnn_io.printClassification(softmaxOut, image_paths, labels)
def _run(self, imgList, imgShape, fpgaOutput): if self.numProcessed == 0: self.startTime = timeit.default_timer() self.labels = xdnn_io.get_labels(self._args['labels']) self.colors = generate_colors(len(self.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 bboxes = yolo.yolov2_postproc([fpgaOutput], self._args, imgShape, biases=self.biases) #if self._args['servermode']: return bboxes
def main(): parser = xdnn_io.default_parser_args() parser = yolo_parser_args(parser) args = parser.parse_args() args = xdnn_io.make_dict_args(args) g_nDispatchers = args['numprepproc'] g_nWorkers = args['numworkers'] # Setup the environment images = xdnn_io.getFilePaths(args['images']) if (args['golden'] or args['visualize']): assert args['labels'], "Provide --labels to compute mAP." assert args[ 'results_dir'], "For accuracy measurements, provide --results_dir to save the detections." # start comms xserver = xstream.Server() # acquire resources fmaster = FpgaMaster(args['vitis_rundir']) # update batch size inshape = list(fmaster.inshape) if args['batch_sz'] != -1: inshape[0] = args['batch_sz'] args['net_h'] = inshape[2] args['net_w'] = inshape[3] # spawn dispatchers dispatcher = yoloDispatcher(g_nDispatchers, g_nWorkers, inshape) # spawn workers workers = yoloWorkerPool(args['vitis_rundir'] + "_worker", g_nWorkers, args) # send work to system g_nQueries = int(np.ceil(len(images) / inshape[0])) 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() # cleanup del workers del fmaster del xserver # mAP calculation if (args['golden']): print() print("Computing mAP score : ") labels = xdnn_io.get_labels(args['labels']) print("Class names are : {} ".format(labels)) mAP = calc_detector_mAP(args['results_dir'], args['golden'], len(labels), labels, args['prob_threshold'], args['mapiouthresh'], args['points']) sys.stdout.flush()
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))
def run(args=None): if not args: parser = xdnn_io.default_parser_args() parser = yolo_parser_args(parser) parser.add_argument('--startxstream', default=True, action='store_true', help='automatically start obj store server') parser.add_argument('--servermode', default=False, action='store_true', help='accept images from another process') parser.add_argument("--deploymodel", type=str, default='', help='Original prototxt') parser.add_argument("--caffemodel", type=str, default='', help='Original caffemodel') args = parser.parse_args() args = xdnn_io.make_dict_args(args) args['preprocseq'] = [('resize', (224, 224)), ('meansub', [104.007, 116.669, 122.679]), ('chtranspose', (2, 0, 1))] if (args['golden'] or args['visualize']): assert args['labels'], "Provide --labels to compute mAP." assert args[ 'results_dir'], "For accuracy measurements, provide --results_dir to save the detections." labels = xdnn_io.get_labels(args['labels']) colors = generate_colors(len(labels)) args['startxstream'] = True args['servermode'] = False timerQ = Queue() args['timerQ'] = timerQ compJson = xdnn.CompilerJsonParser(args['netcfg']) firstInputShape = next(itervalues(compJson.getInputs())) args['net_h'] = firstInputShape[2] args['net_w'] = firstInputShape[3] # start object store # (make sure to 'pip install pyarrow') xserver = None if args['startxstream']: xserver = xstream.Server() graph = grapher.Graph("yolo_v2") graph.node("prep", yolov2_pre.Node, args) graph.node("fpga", yolov2_fpga.Node, args) graph.node("post", yolov2_post.Node, args) graph.edge("START", None, "prep") graph.edge("prep", "prep", "fpga") graph.edge("fpga", "fpga", "post") graph.edge("DONE", "post", "fpga") graph.edge("DONE", "post", None) if not args['servermode']: graph.serve(background=True) img_paths = xdnn_io.getFilePaths(args['images']) reqProc = mp.Process(target=request_process, args=( args, img_paths, graph._in[0], graph._out[0], )) t = timeit.default_timer() reqProc.start() reqProc.join() graph.stop(kill=False) t2 = args['timerQ'].get() full_time = t2 - t args['timerQ'].close() print("Total time : {}s for {} images".format(full_time, len(img_paths))) print("Average FPS : {} imgs/sec".format(len(img_paths) / full_time)) else: print("Serving %s -> %s" % (graph._in[0], graph._out[0])) graph.serve() # mAP calculation if (args['golden']): print(flush=True) print("Computing mAP score : ", flush=True) print("Class names are : {} ".format(labels), flush=True) mAP = calc_detector_mAP(args['results_dir'], args['golden'], len(labels), labels,\ args['prob_threshold'], args['mapiouthresh'], args['points']) sys.stdout.flush()
def main(): parser = xdnn_io.default_parser_args() parser = yolo_parser_args(parser) args = parser.parse_args() args = xdnn_io.make_dict_args(args) # Setup the environment img_paths = xdnn_io.getFilePaths(args['images']) if (args['golden'] or args['visualize']): assert args['labels'], "Provide --labels to compute mAP." assert args[ 'results_dir'], "For accuracy measurements, provide --results_dir to save the detections." labels = xdnn_io.get_labels(args['labels']) colors = generate_colors(len(labels)) if args['yolo_version'] == 'v2': yolo_postproc = yolo.yolov2_postproc elif args['yolo_version'] == 'v3': yolo_postproc = yolo.yolov3_postproc runner = Runner(args['vitis_rundir']) # Setup the blobs inTensors = runner.get_input_tensors() outTensors = runner.get_output_tensors() batch_sz = args['batch_sz'] if batch_sz == -1: batch_sz = inTensors[0].dims[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) fpgaInput = fpgaBlobs[0][0] # Setup the YOLO config net_h, net_w = fpgaInput.shape[-2:] args['net_h'] = net_h args['net_w'] = net_w biases = bias_selector(args) # Setup profiling env prep_time = 0 exec_time = 0 post_time = 0 # Start the execution for i in range(0, len(img_paths), batch_sz): pl = [] img_shapes = [] # Prep images t1 = timeit.default_timer() for j, p in enumerate(img_paths[i:i + batch_sz]): fpgaInput[j, ...], img_shape = xdnn_io.loadYoloImageBlobFromFile( p, net_h, net_w) pl.append(p) img_shapes.append(img_shape) t2 = timeit.default_timer() # Execute jid = runner.execute_async(fpgaBlobs[0], fpgaBlobs[1]) runner.wait(jid) # Post Proc t3 = timeit.default_timer() boxes = yolo_postproc(fpgaBlobs[1], args, img_shapes, biases=biases) t4 = timeit.default_timer() prep_time += (t2 - t1) exec_time += (t3 - t2) post_time += (t4 - t3) for i in range(min(batch_sz, len(img_shapes))): print("Detected {} boxes in {}".format(len(boxes[i]), pl[i])) # Save the result if (args['results_dir']): for i in range(min(batch_sz, len(img_shapes))): filename = os.path.splitext(os.path.basename(pl[i]))[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], img_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(pl[i], boxes[i], labels, colors, out_file_png) # Profiling results if (args['profile']): print("\nAverage Latency in ms:") print(" Image Prep: {0:3f}".format(prep_time * 1000.0 / len(img_paths))) print(" Exec: {0:3f}".format(exec_time * 1000.0 / len(img_paths))) print(" Post Proc: {0:3f}".format(post_time * 1000.0 / len(img_paths))) sys.stdout.flush() # mAP calculation if (args['golden']): print() print("Computing mAP score : ") print("Class names are : {} ".format(labels)) mAP = calc_detector_mAP(args['results_dir'], args['golden'], len(labels), labels, args['prob_threshold'], args['mapiouthresh'], args['points']) sys.stdout.flush()
def yolo_gpu_inference(backend_path, image_dir, deploy_model, weights, out_labels, IOU_threshold, scorethresh, mean_value, pxscale, transpose, channel_swap, yolo_model, num_classes, args): # Setup the environment images = xdnn_io.getFilePaths(args['images']) if (args['golden'] or args['visualize']): assert args['labels'], "Provide --labels to compute mAP." assert args[ 'results_dir'], "For accuracy measurements, provide --results_dir to save the detections." labels = xdnn_io.get_labels(args['labels']) colors = generate_colors(len(labels)) # Select postproc and biases if args['yolo_version'] == 'v2': yolo_postproc = yolo.yolov2_postproc elif args['yolo_version'] == 'v3': yolo_postproc = yolo.yolov3_postproc biases = bias_selector(args) import caffe caffe.set_mode_cpu() print(args) if (args['gpu'] is not None): caffe.set_mode_gpu() caffe.set_device(args['gpu']) net = caffe.Net(deploy_model, weights, caffe.TEST) net_h, net_w = net.blobs['data'].data.shape[-2:] args['net_h'] = net_h args['net_w'] = net_w for i, img in enumerate(images): if ((i + 1) % 100 == 0): print(i + 1, "images processed") raw_img, img_shape = xdnn_io.loadYoloImageBlobFromFile( img, net_h, net_w) net.blobs['data'].data[...] = raw_img out = net.forward() caffeOutput = sorted(out.values(), key=lambda item: item.shape[-1]) boxes = yolo_postproc(caffeOutput, args, [img_shape], biases=biases) print("{}. Detected {} boxes in {}".format(i, len(boxes[0]), img)) # Save the result boxes = boxes[0] if (args['results_dir']): filename = os.path.splitext(os.path.basename(img))[0] out_file_txt = os.path.join(args['results_dir'], filename + '.txt') print("Saving {} boxes to {}".format(len(boxes), out_file_txt)) sys.stdout.flush() saveDetectionDarknetStyle(out_file_txt, boxes, img_shape) 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(img, boxes, labels, colors, out_file_png) # draw_boxes(images[i],bboxes,class_names,colors=[(0,0,0)]*num_classes) return len(images)
def main(): parser = argparse.ArgumentParser() parser = yolo_parser_args(parser) parser.add_argument( '--deploymodel', help="network definition prototxt file in case of caffe", required=True, type=extant_file, metavar="FILE") parser.add_argument( '--caffemodel', help="network weights caffe model file in case of caffe", required=True, type=extant_file, metavar="FILE") parser.add_argument('--images', nargs='*', help='directory or raw image files to use as input', required=True, type=extant_file, metavar="FILE") parser.add_argument('--labels', help='label ID', type=extant_file, metavar="FILE") parser.add_argument('--golden', help='Ground truth directory', type=extant_file, metavar="FILE") parser.add_argument( '--mean_value', type=int, nargs=3, default=[0, 0, 0], # BGR for Caffe help='image mean values ') parser.add_argument('--pxscale', type=float, default=(1.0 / 255.0), help='pix cale value') parser.add_argument( '--transpose', type=int, default=[2, 0, 1], nargs=3, help= "Passed to caffe.io.Transformer function set_transpose, default 2,0,1") parser.add_argument( '--channel_swap', type=int, default=[2, 1, 0], nargs=3, help= "Passed to caffe.io.Transformer function set_channel_swap, default 2,1,0" ) parser.add_argument('--caffe_backend_path', help='caffe backend') parser.add_argument('--gpu', type=int, default=None, help='GPU-ID to run Caffe inference on GPU') args = parser.parse_args() args = xdnn_io.make_dict_args(args) num_images_processed = yolo_gpu_inference( args['caffe_backend_path'], args['images'], args['deploymodel'], args['caffemodel'], args['results_dir'], args['iouthresh'], args['scorethresh'], args['mean_value'], args['pxscale'], args['transpose'], args['channel_swap'], args['yolo_model'], args['classes'], args) print('num images processed : ', num_images_processed) # mAP calculation if (args['golden']): labels = xdnn_io.get_labels(args['labels']) print() print("Computing mAP score : ") print("Class names are : {} ".format(labels)) mAP = calc_detector_mAP(args['results_dir'], args['golden'], len(labels), labels, args['prob_threshold'], args['iouthresh']) sys.stdout.flush()