def infer(): data_path = DatasetPath('val') test_list = data_path.get_file_list() cocoGt = COCO(test_list) num_id_to_cat_id_map = {i + 1: v for i, v in enumerate(cocoGt.getCatIds())} category_ids = cocoGt.getCatIds() label_list = { item['id']: item['name'] for item in cocoGt.loadCats(category_ids) } label_list[0] = ['background'] image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size] class_nums = cfg.class_num model = model_builder.RCNN( add_conv_body_func=resnet.add_ResNet50_conv4_body, add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head, use_pyreader=False, is_train=False) model.build_model(image_shape) pred_boxes = model.eval_bbox_out() if cfg.MASK_ON: masks = model.eval_mask_out() place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) # yapf: disable if cfg.pretrained_model: def if_exist(var): return os.path.exists(os.path.join(cfg.pretrained_model, var.name)) fluid.io.load_vars(exe, cfg.pretrained_model, predicate=if_exist) # yapf: enable infer_reader = reader.infer() feeder = fluid.DataFeeder(place=place, feed_list=model.feeds()) dts_res = [] segms_res = [] if cfg.MASK_ON: fetch_list = [pred_boxes, masks] else: fetch_list = [pred_boxes] data = next(infer_reader()) im_info = [data[0][1]] result = exe.run(fetch_list=[v.name for v in fetch_list], feed=feeder.feed(data), return_numpy=False) pred_boxes_v = result[0] if cfg.MASK_ON: masks_v = result[1] new_lod = pred_boxes_v.lod() nmsed_out = pred_boxes_v path = os.path.join(cfg.image_path, cfg.image_name) image = None if cfg.MASK_ON: segms_out = segm_results(nmsed_out, masks_v, im_info) image = draw_mask_on_image(path, segms_out, cfg.draw_threshold) draw_bounding_box_on_image(path, nmsed_out, cfg.draw_threshold, label_list, num_id_to_cat_id_map, image)
def train(): learning_rate = cfg.learning_rate image_shape = [3, cfg.TRAIN.max_size, cfg.TRAIN.max_size] devices_num = get_device_num() total_batch_size = devices_num * cfg.TRAIN.im_per_batch use_random = True model = model_builder.RCNN( add_conv_body_func=resnet. add_ResNet50_conv4_body, # res4: [-1, 1024, 84, 84] add_roi_box_head_func=resnet. add_ResNet_roi_conv5_head, # res5: [-1, 2048, 7, 7] use_pyreader=cfg.use_pyreader, use_random=use_random) model.build_model(image_shape) losses, keys = model.loss() loss = losses[0] fetch_list = losses boundaries = cfg.lr_steps gamma = cfg.lr_gamma step_num = len(cfg.lr_steps) values = [learning_rate * (gamma**i) for i in range(step_num + 1)] lr = exponential_with_warmup_decay(learning_rate=learning_rate, boundaries=boundaries, values=values, warmup_iter=cfg.warm_up_iter, warmup_factor=cfg.warm_up_factor) optimizer = fluid.optimizer.Momentum( learning_rate=lr, regularization=fluid.regularizer.L2Decay(cfg.weight_decay), momentum=cfg.momentum) optimizer.minimize(loss) fetch_list = fetch_list + [lr] for var in fetch_list: var.persistable = True gpu_id = int(os.environ.get('FLAGS_selected_gpus', 0)) place = fluid.CUDAPlace(gpu_id) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(fluid.default_startup_program()) if cfg.pretrained_model: def if_exist(var): return os.path.exists(os.path.join(cfg.pretrained_model, var.name)) fluid.io.load_vars(exe, cfg.pretrained_model, predicate=if_exist) if cfg.parallel: build_strategy = fluid.BuildStrategy() build_strategy.memory_optimize = False build_strategy.enable_inplace = True exec_strategy = fluid.ExecutionStrategy() exec_strategy.num_iteration_per_drop_scope = 10 if num_trainers > 1 and cfg.use_gpu: dist_utils.prepare_for_multi_process(exe, build_strategy, fluid.default_main_program()) # the process is fast when num_threads is 1 for multi-process training exec_strategy.num_threads = 1 train_exe = fluid.ParallelExecutor(use_cuda=bool(cfg.use_gpu), loss_name=loss.name, build_strategy=build_strategy, exec_strategy=exec_strategy) else: train_exe = exe shuffle = True # NOTE: do not shuffle dataset when using multi-process training shuffle_seed = None if num_trainers > 1: shuffle_seed = 1 if cfg.use_pyreader: train_reader = reader.train(batch_size=cfg.TRAIN.im_per_batch, total_batch_size=total_batch_size, padding_total=cfg.TRAIN.padding_minibatch, shuffle=shuffle, shuffle_seed=shuffle_seed) if num_trainers > 1: assert shuffle_seed is not None, "If num_trainers > 1, the shuffle_seed must be set, because the order of batch data generated by reader must be the same in the respective processes" train_reader = fluid.contrib.reader.distributed_batch_reader( train_reader) py_reader = model.py_reader py_reader.decorate_paddle_reader(train_reader) else: if num_trainers > 1: shuffle = False train_reader = reader.train(batch_size=total_batch_size, shuffle=shuffle) feeder = fluid.DataFeeder(place=place, feed_list=model.feeds()) def save_model(postfix): model_path = os.path.join(cfg.model_save_dir, postfix) if os.path.isdir(model_path): shutil.rmtree(model_path) fluid.io.save_persistables(exe, model_path) def train_loop_pyreader(): py_reader.start() train_stats = TrainingStats(cfg.log_window, keys) try: start_time = time.time() for iter_id in range(cfg.max_iter): prev_start_time = start_time start_time = time.time() outs = train_exe.run(fetch_list=[v.name for v in fetch_list]) stats = { k: np.array(v).mean() for k, v in zip(keys, outs[:-1]) } train_stats.update(stats) logs = train_stats.log() strs = '{}, iter: {}, lr: {:.5f}, {}, time: {:.3f}'.format( now_time(), iter_id, np.mean(outs[-1]), logs, start_time - prev_start_time) print(strs) sys.stdout.flush() if (iter_id + 1) % cfg.TRAIN.snapshot_iter == 0: save_model("model_iter{}".format(iter_id)) end_time = time.time() total_time = end_time - first_start_time last_loss = np.array(outs[0]).mean() except (StopIteration, fluid.core.EOFException): py_reader.reset() def train_loop(): train_stats = TrainingStats(cfg.log_window, keys) start_time = time.time() for iter_id, data in enumerate(train_reader()): prev_start_time = start_time start_time = time.time() outs = train_exe.run(fetch_list=[v.name for v in fetch_list], feed=feeder.feed(data)) stats = {k: np.array(v).mean() for k, v in zip(keys, outs[:-1])} train_stats.update(stats) logs = train_stats.log() stats = '{}, iter: {}, lr: {:.5f}, {}, time: {:.3f}'.format( now_time(), iter_id, np.mean(outs[-1]), logs, start_time - prev_start_time) print(stats) sys.stdout.flush() if (iter_id + 1) % cfg.TRAIN.snapshot_iter == 0: save_model("model_iter{}".format(iter_id)) if (iter_id + 1) == cfg.max_iter: break end_time = time.time() total_time = end_time - start_time last_loss = np.array(outs[0]).mean() if cfg.use_pyreader: train_loop_pyreader() else: train_loop() save_model('model_final')
def infer(): try: from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval, Params data_path = DatasetPath('val') test_list = data_path.get_file_list() coco_api = COCO(test_list) cid = coco_api.getCatIds() cat_id_to_num_id_map = { v: i + 1 for i, v in enumerate(coco_api.getCatIds()) } category_ids = coco_api.getCatIds() labels_map = { cat_id_to_num_id_map[item['id']]: item['name'] for item in coco_api.loadCats(category_ids) } labels_map[0] = 'background' except: print("The COCO dataset or COCO API is not exist, use the default " "mapping of class index and real category name on COCO17.") assert cfg.dataset == 'coco2017' labels_map = coco17_labels() image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size] class_nums = cfg.class_num model = model_builder.RCNN( add_conv_body_func=resnet.add_ResNet50_conv4_body, add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head, use_pyreader=False, mode='infer') model.build_model(image_shape) pred_boxes = model.eval_bbox_out() if cfg.MASK_ON: masks = model.eval_mask_out() place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) # yapf: disable if not os.path.exists(cfg.pretrained_model): raise ValueError("Model path [%s] does not exist." % (cfg.pretrained_model)) def if_exist(var): return os.path.exists(os.path.join(cfg.pretrained_model, var.name)) fluid.io.load_vars(exe, cfg.pretrained_model, predicate=if_exist) # yapf: enable infer_reader = reader.infer(cfg.image_path) feeder = fluid.DataFeeder(place=place, feed_list=model.feeds()) dts_res = [] segms_res = [] if cfg.MASK_ON: fetch_list = [pred_boxes, masks] else: fetch_list = [pred_boxes] data = next(infer_reader()) im_info = [data[0][1]] result = exe.run(fetch_list=[v.name for v in fetch_list], feed=feeder.feed(data), return_numpy=False) pred_boxes_v = result[0] if cfg.MASK_ON: masks_v = result[1] new_lod = pred_boxes_v.lod() nmsed_out = pred_boxes_v image = None if cfg.MASK_ON: segms_out = segm_results(nmsed_out, masks_v, im_info) image = draw_mask_on_image(cfg.image_path, segms_out, cfg.draw_threshold) draw_bounding_box_on_image(cfg.image_path, nmsed_out, cfg.draw_threshold, labels_map, image)
def train(): learning_rate = cfg.learning_rate image_shape = [3, cfg.TRAIN.max_size, cfg.TRAIN.max_size] if cfg.enable_ce: fluid.default_startup_program().random_seed = 1000 fluid.default_main_program().random_seed = 1000 import random random.seed(0) np.random.seed(0) devices = os.getenv("CUDA_VISIBLE_DEVICES") or "" devices_num = len(devices.split(",")) total_batch_size = devices_num * cfg.TRAIN.im_per_batch use_random = True if cfg.enable_ce: use_random = False model = model_builder.RCNN( add_conv_body_func=resnet.add_ResNet50_conv4_body, add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head, use_pyreader=cfg.use_pyreader, use_random=use_random) model.build_model(image_shape) losses, keys = model.loss() loss = losses[0] fetch_list = losses boundaries = cfg.lr_steps gamma = cfg.lr_gamma step_num = len(cfg.lr_steps) values = [learning_rate * (gamma**i) for i in range(step_num + 1)] lr = exponential_with_warmup_decay(learning_rate=learning_rate, boundaries=boundaries, values=values, warmup_iter=cfg.warm_up_iter, warmup_factor=cfg.warm_up_factor) optimizer = fluid.optimizer.Momentum( learning_rate=lr, regularization=fluid.regularizer.L2Decay(cfg.weight_decay), momentum=cfg.momentum) optimizer.minimize(loss) fetch_list = fetch_list + [lr] for var in fetch_list: var.persistable = True #fluid.memory_optimize(fluid.default_main_program(), skip_opt_set=set(fetch_list)) place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(fluid.default_startup_program()) if cfg.pretrained_model: def if_exist(var): return os.path.exists(os.path.join(cfg.pretrained_model, var.name)) fluid.io.load_vars(exe, cfg.pretrained_model, predicate=if_exist) if cfg.parallel: build_strategy = fluid.BuildStrategy() build_strategy.memory_optimize = False build_strategy.enable_inplace = False exec_strategy = fluid.ExecutionStrategy() exec_strategy.use_experimental_executor = True train_exe = fluid.ParallelExecutor(use_cuda=bool(cfg.use_gpu), loss_name=loss.name, build_strategy=build_strategy, exec_strategy=exec_strategy) else: train_exe = exe shuffle = True if cfg.enable_ce: shuffle = False if cfg.use_pyreader: train_reader = reader.train(batch_size=cfg.TRAIN.im_per_batch, total_batch_size=total_batch_size, padding_total=cfg.TRAIN.padding_minibatch, shuffle=shuffle) py_reader = model.py_reader py_reader.decorate_paddle_reader(train_reader) else: train_reader = reader.train(batch_size=total_batch_size, shuffle=shuffle) feeder = fluid.DataFeeder(place=place, feed_list=model.feeds()) def save_model(postfix): model_path = os.path.join(cfg.model_save_dir, postfix) if os.path.isdir(model_path): shutil.rmtree(model_path) fluid.io.save_persistables(exe, model_path) def train_loop_pyreader(): py_reader.start() train_stats = TrainingStats(cfg.log_window, keys) try: start_time = time.time() prev_start_time = start_time for iter_id in range(cfg.max_iter): prev_start_time = start_time start_time = time.time() outs = train_exe.run(fetch_list=[v.name for v in fetch_list]) stats = { k: np.array(v).mean() for k, v in zip(keys, outs[:-1]) } train_stats.update(stats) logs = train_stats.log() strs = '{}, iter: {}, lr: {:.5f}, {}, time: {:.3f}'.format( now_time(), iter_id, np.mean(outs[-1]), logs, start_time - prev_start_time) print(strs) sys.stdout.flush() if (iter_id + 1) % cfg.TRAIN.snapshot_iter == 0: save_model("model_iter{}".format(iter_id)) end_time = time.time() total_time = end_time - start_time last_loss = np.array(outs[0]).mean() if cfg.enable_ce: gpu_num = devices_num epoch_idx = iter_id + 1 loss = last_loss print("kpis\teach_pass_duration_card%s\t%s" % (gpu_num, total_time / epoch_idx)) print("kpis\ttrain_loss_card%s\t%s" % (gpu_num, loss)) except (StopIteration, fluid.core.EOFException): py_reader.reset() def train_loop(): start_time = time.time() prev_start_time = start_time start = start_time train_stats = TrainingStats(cfg.log_window, keys) for iter_id, data in enumerate(train_reader()): prev_start_time = start_time start_time = time.time() outs = train_exe.run(fetch_list=[v.name for v in fetch_list], feed=feeder.feed(data)) stats = {k: np.array(v).mean() for k, v in zip(keys, outs[:-1])} train_stats.update(stats) logs = train_stats.log() strs = '{}, iter: {}, lr: {:.5f}, {}, time: {:.3f}'.format( now_time(), iter_id, np.mean(outs[-1]), logs, start_time - prev_start_time) print(strs) sys.stdout.flush() if (iter_id + 1) % cfg.TRAIN.snapshot_iter == 0: save_model("model_iter{}".format(iter_id)) if (iter_id + 1) == cfg.max_iter: break end_time = time.time() total_time = end_time - start_time last_loss = np.array(outs[0]).mean() # only for ce if cfg.enable_ce: gpu_num = devices_num epoch_idx = iter_id + 1 loss = last_loss print("kpis\teach_pass_duration_card%s\t%s" % (gpu_num, total_time / epoch_idx)) print("kpis\ttrain_loss_card%s\t%s" % (gpu_num, loss)) return np.mean(every_pass_loss) if cfg.use_pyreader: train_loop_pyreader() else: train_loop() save_model('model_final')
def eval(): data_path = DatasetPath('val') test_list = data_path.get_file_list() image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size] class_nums = cfg.class_num devices = os.getenv("CUDA_VISIBLE_DEVICES") or "" devices_num = len(devices.split(",")) total_batch_size = devices_num * cfg.TRAIN.im_per_batch cocoGt = COCO(test_list) num_id_to_cat_id_map = {i + 1: v for i, v in enumerate(cocoGt.getCatIds())} category_ids = cocoGt.getCatIds() label_list = { item['id']: item['name'] for item in cocoGt.loadCats(category_ids) } label_list[0] = ['background'] model = model_builder.RCNN( add_conv_body_func=resnet.add_ResNet50_conv4_body, add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head, use_pyreader=False, mode='val') model.build_model(image_shape) pred_boxes = model.eval_bbox_out() if cfg.MASK_ON: masks = model.eval_mask_out() place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(fluid.default_startup_program()) # yapf: disable if cfg.pretrained_model: def if_exist(var): return os.path.exists(os.path.join(cfg.pretrained_model, var.name)) fluid.io.load_vars(exe, cfg.pretrained_model, predicate=if_exist) # yapf: enable test_reader = reader.test(total_batch_size) feeder = fluid.DataFeeder(place=place, feed_list=model.feeds()) dts_res = [] segms_res = [] if cfg.MASK_ON: fetch_list = [pred_boxes, masks] else: fetch_list = [pred_boxes] eval_start = time.time() for batch_id, batch_data in enumerate(test_reader()): start = time.time() im_info = [] for data in batch_data: im_info.append(data[1]) results = exe.run(fetch_list=[v.name for v in fetch_list], feed=feeder.feed(batch_data), return_numpy=False) pred_boxes_v = results[0] if cfg.MASK_ON: masks_v = results[1] new_lod = pred_boxes_v.lod() nmsed_out = pred_boxes_v dts_res += get_dt_res(total_batch_size, new_lod[0], nmsed_out, batch_data, num_id_to_cat_id_map) if cfg.MASK_ON and np.array(masks_v).shape != (1, 1): segms_out = segm_results(nmsed_out, masks_v, im_info) segms_res += get_segms_res(total_batch_size, new_lod[0], segms_out, batch_data, num_id_to_cat_id_map) end = time.time() print('batch id: {}, time: {}'.format(batch_id, end - start)) eval_end = time.time() total_time = eval_end - eval_start print('average time of eval is: {}'.format(total_time / (batch_id + 1))) assert len(dts_res) > 0, "The number of valid bbox detected is zero.\n \ Please use reasonable model and check input data." if cfg.MASK_ON: assert len( segms_res) > 0, "The number of valid mask detected is zero.\n \ Please use reasonable model and check input data." with io.open("detection_bbox_result.json", 'w') as outfile: encode_func = unicode if six.PY2 else str outfile.write(encode_func(json.dumps(dts_res))) print("start evaluate bbox using coco api") cocoDt = cocoGt.loadRes("detection_bbox_result.json") cocoEval = COCOeval(cocoGt, cocoDt, 'bbox') cocoEval.evaluate() cocoEval.accumulate() cocoEval.summarize() if cfg.MASK_ON: with io.open("detection_segms_result.json", 'w') as outfile: encode_func = unicode if six.PY2 else str outfile.write(encode_func(json.dumps(segms_res))) print("start evaluate mask using coco api") cocoDt = cocoGt.loadRes("detection_segms_result.json") cocoEval = COCOeval(cocoGt, cocoDt, 'segm') cocoEval.evaluate() cocoEval.accumulate() cocoEval.summarize()
'hair drier', 'toothbrush' ] #80 #config.py & (main in utility.py) args = parse_args() print_arguments(args) #Step 1 image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size] class_nums = cfg.class_num #Step 2 model = model_builder.RCNN( add_conv_body_func=resnet.add_ResNet50_conv4_body, add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head, use_pyreader=False, mode='infer') model.build_model(image_shape) model_pred_boxes = model.eval_bbox_out() model_masks = model.eval_mask_out() place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) #Step 3 def if_exist(var): return os.path.exists(os.path.join(cfg.pretrained_model, var.name)) fluid.io.load_vars(exe, cfg.pretrained_model, predicate=if_exist)
def train(): learning_rate = cfg.learning_rate image_shape = [3, cfg.TRAIN.max_size, cfg.TRAIN.max_size] num_iterations = cfg.max_iter devices = os.getenv("CUDA_VISIBLE_DEVICES") or "" devices_num = len(devices.split(",")) total_batch_size = devices_num * cfg.TRAIN.im_per_batch model = model_builder.RCNN( add_conv_body_func=resnet.add_ResNet50_conv4_body, add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head, use_pyreader=cfg.use_pyreader, use_random=False) model.build_model(image_shape) losses, keys = model.loss() loss = losses[0] fetch_list = [loss] boundaries = cfg.lr_steps gamma = cfg.lr_gamma step_num = len(cfg.lr_steps) values = [learning_rate * (gamma**i) for i in range(step_num + 1)] optimizer = fluid.optimizer.Momentum( learning_rate=exponential_with_warmup_decay( learning_rate=learning_rate, boundaries=boundaries, values=values, warmup_iter=500, warmup_factor=1.0 / 3.0), regularization=fluid.regularizer.L2Decay(0.0001), momentum=0.9) optimizer.minimize(loss) fluid.memory_optimize(fluid.default_main_program()) place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(fluid.default_startup_program()) if cfg.pretrained_model: def if_exist(var): return os.path.exists(os.path.join(cfg.pretrained_model, var.name)) fluid.io.load_vars(exe, cfg.pretrained_model, predicate=if_exist) if cfg.parallel: train_exe = fluid.ParallelExecutor(use_cuda=bool(cfg.use_gpu), loss_name=loss.name) if cfg.use_pyreader: train_reader = reader.train(batch_size=cfg.TRAIN.im_per_batch, total_batch_size=total_batch_size, padding_total=cfg.TRAIN.padding_minibatch, shuffle=False) py_reader = model.py_reader py_reader.decorate_paddle_reader(train_reader) else: train_reader = reader.train(batch_size=total_batch_size, shuffle=False) feeder = fluid.DataFeeder(place=place, feed_list=model.feeds()) def run(iterations): reader_time = [] run_time = [] total_images = 0 for batch_id in range(iterations): start_time = time.time() data = next(train_reader()) end_time = time.time() reader_time.append(end_time - start_time) start_time = time.time() if cfg.parallel: outs = train_exe.run(fetch_list=[v.name for v in fetch_list], feed=feeder.feed(data)) else: outs = exe.run(fluid.default_main_program(), fetch_list=[v.name for v in fetch_list], feed=feeder.feed(data)) end_time = time.time() run_time.append(end_time - start_time) total_images += len(data) print("Batch {:d}, loss {:.6f} ".format(batch_id, np.mean(outs[0]))) return reader_time, run_time, total_images def run_pyreader(iterations): reader_time = [0] run_time = [] total_images = 0 py_reader.start() try: for batch_id in range(iterations): start_time = time.time() if cfg.parallel: outs = train_exe.run( fetch_list=[v.name for v in fetch_list]) else: outs = exe.run(fluid.default_main_program(), fetch_list=[v.name for v in fetch_list]) end_time = time.time() run_time.append(end_time - start_time) total_images += devices_num print("Batch {:d}, loss {:.6f} ".format( batch_id, np.mean(outs[0]))) except fluid.core.EOFException: py_reader.reset() return reader_time, run_time, total_images run_func = run if not cfg.use_pyreader else run_pyreader # warm-up run_func(2) # profiling start = time.time() if cfg.use_profile: with profiler.profiler('GPU', 'total', '/tmp/profile_file'): reader_time, run_time, total_images = run_func(num_iterations) else: reader_time, run_time, total_images = run_func(num_iterations) end = time.time() total_time = end - start print( "Total time: {0}, reader time: {1} s, run time: {2} s, images/s: {3}". format(total_time, np.sum(reader_time), np.sum(run_time), total_images / total_time))