def main(): cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # disable npu in config by default and check use_npu if 'use_npu' not in cfg: cfg.use_npu = False check_npu(cfg.use_npu) # check if paddlepaddle version is satisfied check_version() main_arch = cfg.architecture dataset = cfg.TestReader['dataset'] test_images = get_test_images(FLAGS.infer_dir, FLAGS.infer_img) dataset.set_images(test_images) if cfg.use_gpu: place = fluid.CUDAPlace(0) elif cfg.use_npu: place = fluid.NPUPlace(0) else: place = fluid.CPUPlace() exe = fluid.Executor(place) model = create(main_arch) startup_prog = fluid.Program() infer_prog = fluid.Program() with fluid.program_guard(infer_prog, startup_prog): with fluid.unique_name.guard(): inputs_def = cfg['TestReader']['inputs_def'] inputs_def['iterable'] = True feed_vars, loader = model.build_inputs(**inputs_def) test_fetches = model.test(feed_vars) infer_prog = infer_prog.clone(True) reader = create_reader(cfg.TestReader, devices_num=1) loader.set_sample_list_generator(reader, place) exe.run(startup_prog) if cfg.weights: checkpoint.load_params(exe, infer_prog, cfg.weights) # parse infer fetches assert cfg.metric in ['COCO', 'VOC', 'OID', 'WIDERFACE'], \ "unknown metric type {}".format(cfg.metric) extra_keys = [] if cfg['metric'] in ['COCO', 'OID']: extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg['metric'] == 'VOC' or cfg['metric'] == 'WIDERFACE': extra_keys = ['im_id', 'im_shape'] keys, values, _ = parse_fetches(test_fetches, infer_prog, extra_keys) # parse dataset category if cfg.metric == 'COCO': from ppdet.utils.coco_eval import bbox2out, mask2out, segm2out, get_category_info if cfg.metric == 'OID': from ppdet.utils.oid_eval import bbox2out, get_category_info if cfg.metric == "VOC": from ppdet.utils.voc_eval import bbox2out, get_category_info if cfg.metric == "WIDERFACE": from ppdet.utils.widerface_eval_utils import bbox2out, lmk2out, get_category_info anno_file = dataset.get_anno() with_background = dataset.with_background use_default_label = dataset.use_default_label clsid2catid, catid2name = get_category_info(anno_file, with_background, use_default_label) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() # use VisualDL to log image if FLAGS.use_vdl: assert six.PY3, "VisualDL requires Python >= 3.5" from visualdl import LogWriter vdl_writer = LogWriter(FLAGS.vdl_log_dir) vdl_image_step = 0 vdl_image_frame = 0 # each frame can display ten pictures at most. imid2path = dataset.get_imid2path() for iter_id, data in enumerate(loader()): outs = exe.run(infer_prog, feed=data, fetch_list=values, return_numpy=False) res = { k: (np.array(v), v.recursive_sequence_lengths()) for k, v in zip(keys, outs) } logger.info('Infer iter {}'.format(iter_id)) if 'TTFNet' in cfg.architecture: res['bbox'][1].append([len(res['bbox'][0])]) if 'CornerNet' in cfg.architecture: from ppdet.utils.post_process import corner_post_process post_config = getattr(cfg, 'PostProcess', None) corner_post_process(res, post_config, cfg.num_classes) bbox_results = None mask_results = None segm_results = None lmk_results = None if 'bbox' in res: bbox_results = bbox2out([res], clsid2catid, is_bbox_normalized) if 'mask' in res: mask_results = mask2out([res], clsid2catid, model.mask_head.resolution) if 'segm' in res: segm_results = segm2out([res], clsid2catid) if 'landmark' in res: lmk_results = lmk2out([res], is_bbox_normalized) # visualize result im_ids = res['im_id'][0] for im_id in im_ids: image_path = imid2path[int(im_id)] image = Image.open(image_path).convert('RGB') image = ImageOps.exif_transpose(image) # use VisualDL to log original image if FLAGS.use_vdl: original_image_np = np.array(image) vdl_writer.add_image( "original/frame_{}".format(vdl_image_frame), original_image_np, vdl_image_step) image = visualize_results(image, int(im_id), catid2name, FLAGS.draw_threshold, bbox_results, mask_results, segm_results, lmk_results) # use VisualDL to log image with bbox if FLAGS.use_vdl: infer_image_np = np.array(image) vdl_writer.add_image("bbox/frame_{}".format(vdl_image_frame), infer_image_np, vdl_image_step) vdl_image_step += 1 if vdl_image_step % 10 == 0: vdl_image_step = 0 vdl_image_frame += 1 save_name = get_save_image_name(FLAGS.output_dir, image_path) logger.info("Detection bbox results save in {}".format(save_name)) image.save(save_name, quality=95)
def main(): env = os.environ cfg = load_config(FLAGS.config) if 'architecture' in cfg: main_arch = cfg.architecture else: raise ValueError("'architecture' not specified in config file.") merge_config(FLAGS.opt) if 'log_iter' not in cfg: cfg.log_iter = 20 # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) if cfg.use_gpu: devices_num = fluid.core.get_cuda_device_count() else: devices_num = int(os.environ.get('CPU_NUM', 1)) if 'FLAGS_selected_gpus' in env: device_id = int(env['FLAGS_selected_gpus']) else: device_id = 0 place = fluid.CUDAPlace(device_id) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) # build program model = create(main_arch) inputs_def = cfg['TrainReader']['inputs_def'] train_feed_vars, train_loader = model.build_inputs(**inputs_def) train_fetches = model.train(train_feed_vars) loss = train_fetches['loss'] start_iter = 0 train_reader = create_reader(cfg.TrainReader, (cfg.max_iters - start_iter) * devices_num, cfg) train_loader.set_sample_list_generator(train_reader, place) eval_prog = fluid.Program() with fluid.program_guard(eval_prog, fluid.default_startup_program()): with fluid.unique_name.guard(): model = create(main_arch) inputs_def = cfg['EvalReader']['inputs_def'] test_feed_vars, eval_loader = model.build_inputs(**inputs_def) fetches = model.eval(test_feed_vars) eval_prog = eval_prog.clone(True) eval_reader = create_reader(cfg.EvalReader) eval_loader.set_sample_list_generator(eval_reader, place) teacher_cfg = load_config(FLAGS.teacher_config) merge_config(FLAGS.opt) teacher_arch = teacher_cfg.architecture teacher_program = fluid.Program() teacher_startup_program = fluid.Program() with fluid.program_guard(teacher_program, teacher_startup_program): with fluid.unique_name.guard(): teacher_feed_vars = OrderedDict() for name, var in train_feed_vars.items(): teacher_feed_vars[name] = teacher_program.global_block( )._clone_variable(var, force_persistable=False) model = create(teacher_arch) train_fetches = model.train(teacher_feed_vars) teacher_loss = train_fetches['loss'] exe.run(teacher_startup_program) assert FLAGS.teacher_pretrained, "teacher_pretrained should be set" checkpoint.load_params(exe, teacher_program, FLAGS.teacher_pretrained) teacher_program = teacher_program.clone(for_test=True) data_name_map = { 'target0': 'target0', 'target1': 'target1', 'target2': 'target2', 'image': 'image', 'gt_bbox': 'gt_bbox', 'gt_class': 'gt_class', 'gt_score': 'gt_score' } merge(teacher_program, fluid.default_main_program(), data_name_map, place) yolo_output_names = [ 'strided_slice_0.tmp_0', 'strided_slice_1.tmp_0', 'strided_slice_2.tmp_0', 'strided_slice_3.tmp_0', 'strided_slice_4.tmp_0', 'transpose_0.tmp_0', 'strided_slice_5.tmp_0', 'strided_slice_6.tmp_0', 'strided_slice_7.tmp_0', 'strided_slice_8.tmp_0', 'strided_slice_9.tmp_0', 'transpose_2.tmp_0', 'strided_slice_10.tmp_0', 'strided_slice_11.tmp_0', 'strided_slice_12.tmp_0', 'strided_slice_13.tmp_0', 'strided_slice_14.tmp_0', 'transpose_4.tmp_0' ] assert cfg.use_fine_grained_loss, \ "Only support use_fine_grained_loss=True, Please set it in config file or '-o use_fine_grained_loss=true'" distill_loss = split_distill(yolo_output_names, 1000) loss = distill_loss + loss lr_builder = create('LearningRate') optim_builder = create('OptimizerBuilder') lr = lr_builder() opt = optim_builder(lr) opt.minimize(loss) exe.run(fluid.default_startup_program()) checkpoint.load_params(exe, fluid.default_main_program(), cfg.pretrain_weights) assert FLAGS.pruned_params is not None, \ "FLAGS.pruned_params is empty!!! Please set it by '--pruned_params' option." pruned_params = FLAGS.pruned_params.strip().split(",") logger.info("pruned params: {}".format(pruned_params)) pruned_ratios = [float(n) for n in FLAGS.pruned_ratios.strip().split(",")] logger.info("pruned ratios: {}".format(pruned_ratios)) assert len(pruned_params) == len(pruned_ratios), \ "The length of pruned params and pruned ratios should be equal." assert pruned_ratios > [0] * len(pruned_ratios) and pruned_ratios < [1] * len(pruned_ratios), \ "The elements of pruned ratios should be in range (0, 1)." pruner = Pruner() distill_prog = pruner.prune(fluid.default_main_program(), fluid.global_scope(), params=pruned_params, ratios=pruned_ratios, place=place, only_graph=False)[0] base_flops = flops(eval_prog) eval_prog = pruner.prune(eval_prog, fluid.global_scope(), params=pruned_params, ratios=pruned_ratios, place=place, only_graph=True)[0] pruned_flops = flops(eval_prog) logger.info("FLOPs -{}; total FLOPs: {}; pruned FLOPs: {}".format( float(base_flops - pruned_flops) / base_flops, base_flops, pruned_flops)) build_strategy = fluid.BuildStrategy() build_strategy.fuse_all_reduce_ops = False build_strategy.fuse_all_optimizer_ops = False build_strategy.fuse_elewise_add_act_ops = True # only enable sync_bn in multi GPU devices sync_bn = getattr(model.backbone, 'norm_type', None) == 'sync_bn' build_strategy.sync_batch_norm = sync_bn and devices_num > 1 \ and cfg.use_gpu exec_strategy = fluid.ExecutionStrategy() # iteration number when CompiledProgram tries to drop local execution scopes. # Set it to be 1 to save memory usages, so that unused variables in # local execution scopes can be deleted after each iteration. exec_strategy.num_iteration_per_drop_scope = 1 parallel_main = fluid.CompiledProgram(distill_prog).with_data_parallel( loss_name=loss.name, build_strategy=build_strategy, exec_strategy=exec_strategy) compiled_eval_prog = fluid.compiler.CompiledProgram(eval_prog) # parse eval fetches extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_bbox', 'gt_class', 'is_difficult'] eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog, extra_keys) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() map_type = cfg.map_type if 'map_type' in cfg else '11point' best_box_ap_list = [0.0, 0] #[map, iter] cfg_name = os.path.basename(FLAGS.config).split('.')[0] save_dir = os.path.join(cfg.save_dir, cfg_name) train_loader.start() for step_id in range(start_iter, cfg.max_iters): teacher_loss_np, distill_loss_np, loss_np, lr_np = exe.run( parallel_main, fetch_list=[ 'teacher_' + teacher_loss.name, distill_loss.name, loss.name, lr.name ]) if step_id % cfg.log_iter == 0: logger.info( "step {} lr {:.6f}, loss {:.6f}, distill_loss {:.6f}, teacher_loss {:.6f}" .format(step_id, lr_np[0], loss_np[0], distill_loss_np[0], teacher_loss_np[0])) if step_id % cfg.snapshot_iter == 0 and step_id != 0 or step_id == cfg.max_iters - 1: save_name = str( step_id) if step_id != cfg.max_iters - 1 else "model_final" checkpoint.save(exe, distill_prog, os.path.join(save_dir, save_name)) # eval results = eval_run(exe, compiled_eval_prog, eval_loader, eval_keys, eval_values, eval_cls) resolution = None box_ap_stats = eval_results(results, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_eval, map_type, cfg['EvalReader']['dataset']) if box_ap_stats[0] > best_box_ap_list[0]: best_box_ap_list[0] = box_ap_stats[0] best_box_ap_list[1] = step_id checkpoint.save(exe, distill_prog, os.path.join("./", "best_model")) logger.info("Best test box ap: {}, in step: {}".format( best_box_ap_list[0], best_box_ap_list[1])) train_loader.reset()
def main(): parser = ArgsParser() parser.add_argument('--n', '-n', default=9, type=int, help='num of clusters') parser.add_argument('--iters', '-i', default=1000, type=int, help='num of iterations for kmeans') parser.add_argument('--gen_iters', '-gi', default=1000, type=int, help='num of iterations for genetic algorithm') parser.add_argument('--thresh', '-t', default=0.25, type=float, help='anchor scale threshold') parser.add_argument('--verbose', '-v', default=True, type=bool, help='whether print result') parser.add_argument('--size', '-s', default=None, type=str, help='image size: w,h, using comma as delimiter') parser.add_argument('--method', '-m', default='v2', type=str, help='cluster method, [v2, v5] are supported now') parser.add_argument('--cache_path', default='cache', type=str, help='cache path') parser.add_argument('--cache', action='store_true', help='whether use cache') FLAGS = parser.parse_args() cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # check if paddlepaddle version is satisfied check_version() # get dataset dataset = cfg['TrainReader']['dataset'] if FLAGS.size: if ',' in FLAGS.size: size = list(map(int, FLAGS.size.split(','))) assert len(size) == 2, "the format of size is incorrect" else: size = int(FLAGS.size) size = [size, size] elif 'image_shape' in cfg['TrainReader']['inputs_def']: size = cfg['TrainReader']['inputs_def']['image_shape'][1:] else: raise ValueError('size is not specified') if FLAGS.method == 'v2': cluster = YOLOv2AnchorCluster(FLAGS.n, dataset, size, FLAGS.cache_path, FLAGS.cache, FLAGS.iters, FLAGS.verbose) elif FLAGS.method == 'v5': cluster = YOLOv5AnchorCluster(FLAGS.n, dataset, size, FLAGS.cache_path, FLAGS.cache, FLAGS.iters, FLAGS.gen_iters, FLAGS.thresh, FLAGS.verbose) else: raise ValueError('cluster method: %s is not supported' % FLAGS.method) anchors = cluster()
def main(): env = os.environ FLAGS.dist = 'PADDLE_TRAINER_ID' in env and 'PADDLE_TRAINERS_NUM' in env if FLAGS.dist: trainer_id = int(env['PADDLE_TRAINER_ID']) import random local_seed = (99 + trainer_id) random.seed(local_seed) np.random.seed(local_seed) cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # check if paddlepaddle version is satisfied check_version() main_arch = cfg.architecture if cfg.use_gpu: devices_num = fluid.core.get_cuda_device_count() else: devices_num = int(os.environ.get('CPU_NUM', 1)) if 'FLAGS_selected_gpus' in env: device_id = int(env['FLAGS_selected_gpus']) else: device_id = 0 place = fluid.CUDAPlace(device_id) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) lr_builder = create('LearningRate') optim_builder = create('OptimizerBuilder') # add NAS config = ([(cfg.search_space)]) server_address = (cfg.server_ip, cfg.server_port) load_checkpoint = FLAGS.resume_checkpoint if FLAGS.resume_checkpoint else None sa_nas = SANAS(config, server_addr=server_address, init_temperature=cfg.init_temperature, reduce_rate=cfg.reduce_rate, search_steps=cfg.search_steps, save_checkpoint=cfg.save_dir, load_checkpoint=load_checkpoint, is_server=cfg.is_server) start_iter = 0 train_reader = create_reader(cfg.TrainReader, (cfg.max_iters - start_iter) * devices_num, cfg) eval_reader = create_reader(cfg.EvalReader) constraint = create('Constraint') for step in range(cfg.search_steps): logger.info('----->>> search step: {} <<<------'.format(step)) archs = sa_nas.next_archs()[0] # build program startup_prog = fluid.Program() train_prog = fluid.Program() with fluid.program_guard(train_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) if FLAGS.fp16: assert (getattr(model.backbone, 'norm_type', None) != 'affine_channel'), \ '--fp16 currently does not support affine channel, ' \ ' please modify backbone settings to use batch norm' with mixed_precision_context(FLAGS.loss_scale, FLAGS.fp16) as ctx: inputs_def = cfg['TrainReader']['inputs_def'] feed_vars, train_loader = model.build_inputs(**inputs_def) train_fetches = archs(feed_vars, 'train', cfg) loss = train_fetches['loss'] if FLAGS.fp16: loss *= ctx.get_loss_scale_var() lr = lr_builder() optimizer = optim_builder(lr) optimizer.minimize(loss) if FLAGS.fp16: loss /= ctx.get_loss_scale_var() current_constraint = constraint.compute_constraint(train_prog) logger.info('current steps: {}, constraint {}'.format( step, current_constraint)) if (constraint.max_constraint != None and current_constraint > constraint.max_constraint) or ( constraint.min_constraint != None and current_constraint < constraint.min_constraint): continue # parse train fetches train_keys, train_values, _ = parse_fetches(train_fetches) train_values.append(lr) if FLAGS.eval: eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) inputs_def = cfg['EvalReader']['inputs_def'] feed_vars, eval_loader = model.build_inputs(**inputs_def) fetches = archs(feed_vars, 'eval', cfg) eval_prog = eval_prog.clone(True) # When iterable mode, set set_sample_list_generator(eval_reader, place) eval_loader.set_sample_list_generator(eval_reader) extra_keys = ['im_id', 'im_shape', 'gt_bbox'] eval_keys, eval_values, eval_cls = parse_fetches( fetches, eval_prog, extra_keys) # compile program for multi-devices build_strategy = fluid.BuildStrategy() build_strategy.fuse_all_optimizer_ops = False build_strategy.fuse_elewise_add_act_ops = True exec_strategy = fluid.ExecutionStrategy() # iteration number when CompiledProgram tries to drop local execution scopes. # Set it to be 1 to save memory usages, so that unused variables in # local execution scopes can be deleted after each iteration. exec_strategy.num_iteration_per_drop_scope = 1 if FLAGS.dist: dist_utils.prepare_for_multi_process(exe, build_strategy, startup_prog, train_prog) exec_strategy.num_threads = 1 exe.run(startup_prog) compiled_train_prog = fluid.CompiledProgram( train_prog).with_data_parallel(loss_name=loss.name, build_strategy=build_strategy, exec_strategy=exec_strategy) if FLAGS.eval: compiled_eval_prog = fluid.CompiledProgram(eval_prog) # When iterable mode, set set_sample_list_generator(train_reader, place) train_loader.set_sample_list_generator(train_reader) train_stats = TrainingStats(cfg.log_iter, train_keys) train_loader.start() end_time = time.time() cfg_name = os.path.basename(FLAGS.config).split('.')[0] save_dir = os.path.join(cfg.save_dir, cfg_name) time_stat = deque(maxlen=cfg.log_iter) ap = 0 for it in range(start_iter, cfg.max_iters): start_time = end_time end_time = time.time() time_stat.append(end_time - start_time) time_cost = np.mean(time_stat) eta_sec = (cfg.max_iters - it) * time_cost eta = str(datetime.timedelta(seconds=int(eta_sec))) outs = exe.run(compiled_train_prog, fetch_list=train_values) stats = { k: np.array(v).mean() for k, v in zip(train_keys, outs[:-1]) } train_stats.update(stats) logs = train_stats.log() if it % cfg.log_iter == 0 and (not FLAGS.dist or trainer_id == 0): strs = 'iter: {}, lr: {:.6f}, {}, time: {:.3f}, eta: {}'.format( it, np.mean(outs[-1]), logs, time_cost, eta) logger.info(strs) if (it > 0 and it == cfg.max_iters - 1) and (not FLAGS.dist or trainer_id == 0): save_name = str( it) if it != cfg.max_iters - 1 else "model_final" checkpoint.save(exe, train_prog, os.path.join(save_dir, save_name)) if FLAGS.eval: # evaluation results = eval_run(exe, compiled_eval_prog, eval_loader, eval_keys, eval_values, eval_cls) ap = calculate_ap_py(results) train_loader.reset() eval_loader.reset() logger.info('rewards: ap is {}'.format(ap)) sa_nas.reward(float(ap)) current_best_tokens = sa_nas.current_info()['best_tokens'] logger.info("All steps end, the best BlazeFace-NAS structure is: ") sa_nas.tokens2arch(current_best_tokens)
def main(): env = os.environ FLAGS.dist = 'PADDLE_TRAINER_ID' in env and 'PADDLE_TRAINERS_NUM' in env if FLAGS.dist: trainer_id = int(env['PADDLE_TRAINER_ID']) import random local_seed = (99 + trainer_id) random.seed(local_seed) np.random.seed(local_seed) cfg = load_config(FLAGS.config) if 'architecture' in cfg: main_arch = cfg.architecture else: raise ValueError("'architecture' not specified in config file.") merge_config(FLAGS.opt) if 'log_iter' not in cfg: cfg.log_iter = 20 # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # check if paddlepaddle version is satisfied check_version() if not FLAGS.dist or trainer_id == 0: print_total_cfg(cfg) if cfg.use_gpu: devices_num = fluid.core.get_cuda_device_count() else: devices_num = int(os.environ.get('CPU_NUM', 1)) if 'FLAGS_selected_gpus' in env: device_id = int(env['FLAGS_selected_gpus']) else: device_id = 0 place = fluid.CUDAPlace(device_id) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) lr_builder = create('LearningRate') optim_builder = create('OptimizerBuilder') # build program startup_prog = fluid.Program() train_prog = fluid.Program() with fluid.program_guard(train_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) if FLAGS.fp16: assert (getattr(model.backbone, 'norm_type', None) != 'affine_channel'), \ '--fp16 currently does not support affine channel, ' \ ' please modify backbone settings to use batch norm' with mixed_precision_context(FLAGS.loss_scale, FLAGS.fp16) as ctx: inputs_def = cfg['TrainReader']['inputs_def'] feed_vars, train_loader = model.build_inputs(**inputs_def) train_fetches = model.train(feed_vars) loss = train_fetches['loss'] if FLAGS.fp16: loss *= ctx.get_loss_scale_var() lr = lr_builder() optimizer = optim_builder(lr) optimizer.minimize(loss) if FLAGS.fp16: loss /= ctx.get_loss_scale_var() # parse train fetches train_keys, train_values, _ = parse_fetches(train_fetches) train_values.append(lr) if FLAGS.print_params: param_delimit_str = '-' * 20 + "All parameters in current graph" + '-' * 20 print(param_delimit_str) for block in train_prog.blocks: for param in block.all_parameters(): print("parameter name: {}\tshape: {}".format( param.name, param.shape)) print('-' * len(param_delimit_str)) return if FLAGS.eval: eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) inputs_def = cfg['EvalReader']['inputs_def'] feed_vars, eval_loader = model.build_inputs(**inputs_def) fetches = model.eval(feed_vars) eval_prog = eval_prog.clone(True) eval_reader = create_reader(cfg.EvalReader) eval_loader.set_sample_list_generator(eval_reader, place) # parse eval fetches extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_bbox', 'gt_class', 'is_difficult'] if cfg.metric == 'WIDERFACE': extra_keys = ['im_id', 'im_shape', 'gt_bbox'] eval_keys, eval_values, eval_cls = parse_fetches( fetches, eval_prog, extra_keys) # compile program for multi-devices build_strategy = fluid.BuildStrategy() build_strategy.fuse_all_optimizer_ops = False build_strategy.fuse_elewise_add_act_ops = True # only enable sync_bn in multi GPU devices sync_bn = getattr(model.backbone, 'norm_type', None) == 'sync_bn' build_strategy.sync_batch_norm = sync_bn and devices_num > 1 \ and cfg.use_gpu exec_strategy = fluid.ExecutionStrategy() # iteration number when CompiledProgram tries to drop local execution scopes. # Set it to be 1 to save memory usages, so that unused variables in # local execution scopes can be deleted after each iteration. exec_strategy.num_iteration_per_drop_scope = 1 if FLAGS.dist: dist_utils.prepare_for_multi_process(exe, build_strategy, startup_prog, train_prog) exec_strategy.num_threads = 1 exe.run(startup_prog) fuse_bn = getattr(model.backbone, 'norm_type', None) == 'affine_channel' start_iter = 0 if cfg.pretrain_weights: checkpoint.load_params(exe, train_prog, cfg.pretrain_weights) pruned_params = FLAGS.pruned_params assert FLAGS.pruned_params is not None, \ "FLAGS.pruned_params is empty!!! Please set it by '--pruned_params' option." pruned_params = FLAGS.pruned_params.strip().split(",") logger.info("pruned params: {}".format(pruned_params)) pruned_ratios = [float(n) for n in FLAGS.pruned_ratios.strip().split(",")] logger.info("pruned ratios: {}".format(pruned_ratios)) assert len(pruned_params) == len(pruned_ratios), \ "The length of pruned params and pruned ratios should be equal." assert (pruned_ratios > [0] * len(pruned_ratios) and pruned_ratios < [1] * len(pruned_ratios) ), "The elements of pruned ratios should be in range (0, 1)." pruner = Pruner() train_prog = pruner.prune(train_prog, fluid.global_scope(), params=pruned_params, ratios=pruned_ratios, place=place, only_graph=False)[0] compiled_train_prog = fluid.CompiledProgram(train_prog).with_data_parallel( loss_name=loss.name, build_strategy=build_strategy, exec_strategy=exec_strategy) if FLAGS.eval: base_flops = flops(eval_prog) eval_prog = pruner.prune(eval_prog, fluid.global_scope(), params=pruned_params, ratios=pruned_ratios, place=place, only_graph=True)[0] pruned_flops = flops(eval_prog) logger.info("FLOPs -{}; total FLOPs: {}; pruned FLOPs: {}".format( float(base_flops - pruned_flops) / base_flops, base_flops, pruned_flops)) compiled_eval_prog = fluid.compiler.CompiledProgram(eval_prog) if FLAGS.resume_checkpoint: checkpoint.load_checkpoint(exe, train_prog, FLAGS.resume_checkpoint) start_iter = checkpoint.global_step() train_reader = create_reader(cfg.TrainReader, (cfg.max_iters - start_iter) * devices_num, cfg) train_loader.set_sample_list_generator(train_reader, place) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() # if map_type not set, use default 11point, only use in VOC eval map_type = cfg.map_type if 'map_type' in cfg else '11point' train_stats = TrainingStats(cfg.log_smooth_window, train_keys) train_loader.start() start_time = time.time() end_time = time.time() cfg_name = os.path.basename(FLAGS.config).split('.')[0] save_dir = os.path.join(cfg.save_dir, cfg_name) time_stat = deque(maxlen=cfg.log_smooth_window) best_box_ap_list = [0.0, 0] #[map, iter] # use tb-paddle to log data if FLAGS.use_tb: from tb_paddle import SummaryWriter tb_writer = SummaryWriter(FLAGS.tb_log_dir) tb_loss_step = 0 tb_mAP_step = 0 if FLAGS.eval: # evaluation results = eval_run(exe, compiled_eval_prog, eval_loader, eval_keys, eval_values, eval_cls) resolution = None if 'mask' in results[0]: resolution = model.mask_head.resolution dataset = cfg['EvalReader']['dataset'] box_ap_stats = eval_results(results, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_eval, map_type, dataset=dataset) for it in range(start_iter, cfg.max_iters): start_time = end_time end_time = time.time() time_stat.append(end_time - start_time) time_cost = np.mean(time_stat) eta_sec = (cfg.max_iters - it) * time_cost eta = str(datetime.timedelta(seconds=int(eta_sec))) outs = exe.run(compiled_train_prog, fetch_list=train_values) stats = {k: np.array(v).mean() for k, v in zip(train_keys, outs[:-1])} # use tb-paddle to log loss if FLAGS.use_tb: if it % cfg.log_iter == 0: for loss_name, loss_value in stats.items(): tb_writer.add_scalar(loss_name, loss_value, tb_loss_step) tb_loss_step += 1 train_stats.update(stats) logs = train_stats.log() if it % cfg.log_iter == 0 and (not FLAGS.dist or trainer_id == 0): strs = 'iter: {}, lr: {:.6f}, {}, time: {:.3f}, eta: {}'.format( it, np.mean(outs[-1]), logs, time_cost, eta) logger.info(strs) if (it > 0 and it % cfg.snapshot_iter == 0 or it == cfg.max_iters - 1) \ and (not FLAGS.dist or trainer_id == 0): save_name = str(it) if it != cfg.max_iters - 1 else "model_final" checkpoint.save(exe, train_prog, os.path.join(save_dir, save_name)) if FLAGS.eval: # evaluation results = eval_run(exe, compiled_eval_prog, eval_loader, eval_keys, eval_values, eval_cls) resolution = None if 'mask' in results[0]: resolution = model.mask_head.resolution box_ap_stats = eval_results(results, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_eval, map_type, dataset=dataset) # use tb_paddle to log mAP if FLAGS.use_tb: tb_writer.add_scalar("mAP", box_ap_stats[0], tb_mAP_step) tb_mAP_step += 1 if box_ap_stats[0] > best_box_ap_list[0]: best_box_ap_list[0] = box_ap_stats[0] best_box_ap_list[1] = it checkpoint.save(exe, train_prog, os.path.join(save_dir, "best_model")) logger.info("Best test box ap: {}, in iter: {}".format( best_box_ap_list[0], best_box_ap_list[1])) train_loader.reset()
def main(): """ Main evaluate function """ cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # check if paddlepaddle version is satisfied check_version() main_arch = cfg.architecture multi_scale_test = getattr(cfg, 'MultiScaleTEST', None) # define executor place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) # build program model = create(main_arch) startup_prog = fluid.Program() eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): inputs_def = cfg['EvalReader']['inputs_def'] feed_vars, loader = model.build_inputs(**inputs_def) if multi_scale_test is None: fetches = model.eval(feed_vars) else: fetches = model.eval(feed_vars, multi_scale_test) eval_prog = eval_prog.clone(True) reader = create_reader(cfg.EvalReader, devices_num=1) loader.set_sample_list_generator(reader, place) dataset = cfg['EvalReader']['dataset'] # eval already exists json file if FLAGS.json_eval: logger.info( "In json_eval mode, PaddleDetection will evaluate json files in " "output_eval directly. And proposal.json, bbox.json and mask.json " "will be detected by default.") json_eval_results(cfg.metric, json_directory=FLAGS.output_eval, dataset=dataset) return compile_program = fluid.compiler.CompiledProgram( eval_prog).with_data_parallel() assert cfg.metric != 'OID', "eval process of OID dataset \ is not supported." if cfg.metric == "WIDERFACE": raise ValueError("metric type {} does not support in tools/eval.py, " "please use tools/face_eval.py".format(cfg.metric)) assert cfg.metric in ['COCO', 'VOC'], \ "unknown metric type {}".format(cfg.metric) extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_bbox', 'gt_class', 'is_difficult'] keys, values, cls = parse_fetches(fetches, eval_prog, extra_keys) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() sub_eval_prog = None sub_keys = None sub_values = None # build sub-program if 'Mask' in main_arch and multi_scale_test: sub_eval_prog = fluid.Program() with fluid.program_guard(sub_eval_prog, startup_prog): with fluid.unique_name.guard(): inputs_def = cfg['EvalReader']['inputs_def'] inputs_def['mask_branch'] = True feed_vars, eval_loader = model.build_inputs(**inputs_def) sub_fetches = model.eval(feed_vars, multi_scale_test, mask_branch=True) assert cfg.metric == 'COCO' extra_keys = ['im_id', 'im_shape'] sub_keys, sub_values, _ = parse_fetches(sub_fetches, sub_eval_prog, extra_keys) sub_eval_prog = sub_eval_prog.clone(True) #if 'weights' in cfg: # checkpoint.load_params(exe, sub_eval_prog, cfg.weights) # load model exe.run(startup_prog) if 'weights' in cfg: checkpoint.load_params(exe, startup_prog, cfg.weights) resolution = None if 'Mask' in cfg.architecture: resolution = model.mask_head.resolution results = eval_run(exe, compile_program, loader, keys, values, cls, cfg, sub_eval_prog, sub_keys, sub_values, resolution) #print(cfg['EvalReader']['dataset'].__dict__) # evaluation # if map_type not set, use default 11point, only use in VOC eval map_type = cfg.map_type if 'map_type' in cfg else '11point' save_only = getattr(cfg, 'save_prediction_only', False) eval_results(results, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_eval, map_type, dataset=dataset, save_only=save_only)
def main(): """ Main evaluate function """ cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # check if paddlepaddle version is satisfied check_version() main_arch = cfg.architecture # define executor place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) # build program model = create(main_arch) startup_prog = fluid.Program() eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): inputs_def = cfg['EvalReader']['inputs_def'] test_feed_vars, loader = model.build_inputs(**inputs_def) test_fetches = model.eval(test_feed_vars) eval_prog = eval_prog.clone(True) reader = create_reader(cfg.EvalReader) # When iterable mode, set set_sample_list_generator(reader, place) loader.set_sample_list_generator(reader) # eval already exists json file if FLAGS.json_eval: logger.info( "In json_eval mode, PaddleDetection will evaluate json files in " "output_eval directly. And proposal.json, bbox.json and mask.json " "will be detected by default.") json_eval_results( cfg.metric, json_directory=FLAGS.output_eval, dataset=dataset) return assert cfg.metric != 'OID', "eval process of OID dataset \ is not supported." if cfg.metric == "WIDERFACE": raise ValueError("metric type {} does not support in tools/eval.py, " "please use tools/face_eval.py".format(cfg.metric)) assert cfg.metric in ['COCO', 'VOC'], \ "unknown metric type {}".format(cfg.metric) extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_bbox', 'gt_class', 'is_difficult'] keys, values, cls = parse_fetches(test_fetches, eval_prog, extra_keys) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() dataset = cfg['EvalReader']['dataset'] sub_eval_prog = None sub_keys = None sub_values = None not_quant_pattern = [] if FLAGS.not_quant_pattern: not_quant_pattern = FLAGS.not_quant_pattern config = { 'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max', 'quantize_op_types': ['depthwise_conv2d', 'mul', 'conv2d'], 'not_quant_pattern': not_quant_pattern } eval_prog = quant_aware(eval_prog, place, config, for_test=True) # load model exe.run(startup_prog) if 'weights' in cfg: checkpoint.load_params(exe, eval_prog, cfg.weights) eval_prog = convert(eval_prog, place, config, save_int8=False) compile_program = fluid.CompiledProgram(eval_prog).with_data_parallel() results = eval_run(exe, compile_program, loader, keys, values, cls, cfg, sub_eval_prog, sub_keys, sub_values) # evaluation resolution = None if 'mask' in results[0]: resolution = model.mask_head.resolution # if map_type not set, use default 11point, only use in VOC eval map_type = cfg.map_type if 'map_type' in cfg else '11point' eval_results( results, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_eval, map_type, dataset=dataset)
def main(): cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # check if paddlepaddle version is satisfied check_version() main_arch = cfg.architecture dataset = cfg.TestReader['dataset'] test_images = get_test_images(FLAGS.infer_dir, FLAGS.infer_img) dataset.set_images(test_images) place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) model = create(main_arch) startup_prog = fluid.Program() infer_prog = fluid.Program() with fluid.program_guard(infer_prog, startup_prog): with fluid.unique_name.guard(): inputs_def = cfg['TestReader']['inputs_def'] inputs_def['iterable'] = True feed_vars, loader = model.build_inputs(**inputs_def) test_fetches = model.test(feed_vars) infer_prog = infer_prog.clone(True) reader = create_reader(cfg.TestReader, devices_num=1) loader.set_sample_list_generator(reader, place) exe.run(startup_prog) if cfg.weights: checkpoint.load_params(exe, infer_prog, cfg.weights) # parse infer fetches assert cfg.metric in ['COCO', 'VOC', 'OID', 'WIDERFACE','traffic'], \ "unknown metric type {}".format(cfg.metric) extra_keys = [] if cfg['metric'] in ['COCO', 'OID']: extra_keys = ['im_info', 'im_id', 'im_shape'] elif cfg['metric'] == 'VOC' or cfg['metric'] == 'WIDERFACE': extra_keys = ['im_id', 'im_shape'] else: extra_keys = ['im_id', 'im_shape'] keys, values, _ = parse_fetches(test_fetches, infer_prog, extra_keys) # parse dataset category if cfg.metric == 'COCO': from ppdet.utils.coco_eval import bbox2out, mask2out, get_category_info elif cfg.metric == 'OID': from ppdet.utils.oid_eval import bbox2out, get_category_info elif cfg.metric == "VOC": from ppdet.utils.voc_eval import bbox2out, get_category_info elif cfg.metric == "WIDERFACE": from ppdet.utils.widerface_eval_utils import bbox2out, get_category_info else: from ppdet.utils.generic_eval import bbox2out, get_category_info anno_file = dataset.get_anno() with_background = dataset.with_background use_default_label = dataset.use_default_label clsid2catid, catid2name = get_category_info(anno_file, with_background, use_default_label) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() # use tb-paddle to log image if FLAGS.use_tb: from tb_paddle import SummaryWriter tb_writer = SummaryWriter(FLAGS.tb_log_dir) tb_image_step = 0 tb_image_frame = 0 # each frame can display ten pictures at most. imid2path = dataset.get_imid2path() for iter_id, data in enumerate(loader()): outs = exe.run(infer_prog, feed=data, fetch_list=values, return_numpy=False) res = { k: (np.array(v), v.recursive_sequence_lengths()) for k, v in zip(keys, outs) } logger.info('Infer iter {}'.format(iter_id)) bbox_results = None mask_results = None if 'bbox' in res: bbox_results = bbox2out([res], clsid2catid, is_bbox_normalized) if 'mask' in res: mask_results = mask2out([res], clsid2catid, model.mask_head.resolution) # visualize result im_ids = res['im_id'][0] for im_id in im_ids: image_path = imid2path[int(im_id)] image = Image.open(image_path).convert('RGB') # use tb-paddle to log original image if FLAGS.use_tb: original_image_np = np.array(image) tb_writer.add_image( "original/frame_{}".format(tb_image_frame), original_image_np, tb_image_step, dataformats='HWC') image = visualize_results(image, int(im_id), catid2name, FLAGS.draw_threshold, bbox_results, mask_results) # use tb-paddle to log image with bbox if FLAGS.use_tb: infer_image_np = np.array(image) tb_writer.add_image( "bbox/frame_{}".format(tb_image_frame), infer_image_np, tb_image_step, dataformats='HWC') tb_image_step += 1 if tb_image_step % 10 == 0: tb_image_step = 0 tb_image_frame += 1 save_name = get_save_image_name(FLAGS.output_dir, image_path) logger.info("Detection bbox results save in {}".format(save_name)) image.save(save_name, quality=95)
def main(): """ Main evaluate function """ cfg = load_config(FLAGS.config) if 'architecture' in cfg: main_arch = cfg.architecture else: raise ValueError("'architecture' not specified in config file.") merge_config(FLAGS.opt) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) if cfg.use_gpu: devices_num = fluid.core.get_cuda_device_count() else: devices_num = int( os.environ.get('CPU_NUM', multiprocessing.cpu_count())) if 'eval_feed' not in cfg: eval_feed = create(main_arch + 'EvalFeed') else: eval_feed = create(cfg.eval_feed) # define executor place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) # build program model = create(main_arch) startup_prog = fluid.Program() eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): pyreader, feed_vars = create_feed(eval_feed) fetches = model.eval(feed_vars) eval_prog = eval_prog.clone(True) reader = create_reader(eval_feed, args_path=FLAGS.dataset_dir) pyreader.decorate_sample_list_generator(reader, place) # compile program for multi-devices if devices_num <= 1: compile_program = fluid.compiler.CompiledProgram(eval_prog) else: build_strategy = fluid.BuildStrategy() build_strategy.memory_optimize = False build_strategy.enable_inplace = False compile_program = fluid.compiler.CompiledProgram( eval_prog).with_data_parallel(build_strategy=build_strategy) # load model exe.run(startup_prog) if 'weights' in cfg: checkpoint.load_pretrain(exe, eval_prog, cfg.weights) assert cfg.metric in ['COCO', 'VOC'], \ "unknown metric type {}".format(cfg.metric) extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_box', 'gt_label', 'is_difficult'] keys, values, cls = parse_fetches(fetches, eval_prog, extra_keys) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() results = eval_run(exe, compile_program, pyreader, keys, values, cls) # evaluation resolution = None if 'mask' in results[0]: resolution = model.mask_head.resolution eval_results(results, eval_feed, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_file)
def main(): env = os.environ FLAGS.dist = 'PADDLE_TRAINER_ID' in env \ and 'PADDLE_TRAINERS_NUM' in env \ and int(env['PADDLE_TRAINERS_NUM']) > 1 num_trainers = int(env.get('PADDLE_TRAINERS_NUM', 1)) if FLAGS.dist: trainer_id = int(env['PADDLE_TRAINER_ID']) local_seed = (99 + trainer_id) random.seed(local_seed) np.random.seed(local_seed) if FLAGS.enable_ce: random.seed(0) np.random.seed(0) cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # check if paddlepaddle version is satisfied check_version() save_only = getattr(cfg, 'save_prediction_only', False) if save_only: raise NotImplementedError('The config file only support prediction,' ' training stage is not implemented now') main_arch = cfg.architecture if cfg.use_gpu: devices_num = fluid.core.get_cuda_device_count() else: devices_num = int(os.environ.get('CPU_NUM', 1)) if 'FLAGS_selected_gpus' in env: device_id = int(env['FLAGS_selected_gpus']) else: device_id = 0 place = fluid.CUDAPlace(device_id) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) lr_builder = create('LearningRate') optim_builder = create('OptimizerBuilder') # build program startup_prog = fluid.Program() train_prog = fluid.Program() if FLAGS.enable_ce: startup_prog.random_seed = 1000 train_prog.random_seed = 1000 with fluid.program_guard(train_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) if FLAGS.fp16: assert (getattr(model.backbone, 'norm_type', None) != 'affine_channel'), \ '--fp16 currently does not support affine channel, ' \ ' please modify backbone settings to use batch norm' with mixed_precision_context(FLAGS.loss_scale, FLAGS.fp16) as ctx: inputs_def = cfg['TrainReader']['inputs_def'] feed_vars, train_loader = model.build_inputs(**inputs_def) train_fetches = model.train(feed_vars) loss = train_fetches['loss'] if FLAGS.fp16: loss *= ctx.get_loss_scale_var() lr = lr_builder() optimizer = optim_builder(lr) optimizer.minimize(loss) if FLAGS.fp16: loss /= ctx.get_loss_scale_var() if 'use_ema' in cfg and cfg['use_ema']: global_steps = _decay_step_counter() ema = ExponentialMovingAverage(cfg['ema_decay'], thres_steps=global_steps) ema.update() # parse train fetches train_keys, train_values, _ = parse_fetches(train_fetches) train_values.append(lr) if FLAGS.eval: eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) inputs_def = cfg['EvalReader']['inputs_def'] feed_vars, eval_loader = model.build_inputs(**inputs_def) fetches = model.eval(feed_vars) eval_prog = eval_prog.clone(True) eval_reader = create_reader(cfg.EvalReader, devices_num=1) # When iterable mode, set set_sample_list_generator(eval_reader, place) eval_loader.set_sample_list_generator(eval_reader) # parse eval fetches extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_bbox', 'gt_class', 'is_difficult'] if cfg.metric == 'WIDERFACE': extra_keys = ['im_id', 'im_shape', 'gt_bbox'] eval_keys, eval_values, eval_cls = parse_fetches( fetches, eval_prog, extra_keys) # compile program for multi-devices build_strategy = fluid.BuildStrategy() build_strategy.fuse_all_optimizer_ops = False # only enable sync_bn in multi GPU devices sync_bn = getattr(model.backbone, 'norm_type', None) == 'sync_bn' build_strategy.sync_batch_norm = sync_bn and devices_num > 1 \ and cfg.use_gpu exec_strategy = fluid.ExecutionStrategy() # iteration number when CompiledProgram tries to drop local execution scopes. # Set it to be 1 to save memory usages, so that unused variables in # local execution scopes can be deleted after each iteration. exec_strategy.num_iteration_per_drop_scope = 1 if FLAGS.dist: dist_utils.prepare_for_multi_process(exe, build_strategy, startup_prog, train_prog) exec_strategy.num_threads = 1 exe.run(startup_prog) compiled_train_prog = fluid.CompiledProgram(train_prog).with_data_parallel( loss_name=loss.name, build_strategy=build_strategy, exec_strategy=exec_strategy) if FLAGS.eval: compiled_eval_prog = fluid.CompiledProgram(eval_prog) fuse_bn = getattr(model.backbone, 'norm_type', None) == 'affine_channel' ignore_params = cfg.finetune_exclude_pretrained_params \ if 'finetune_exclude_pretrained_params' in cfg else [] start_iter = 0 if FLAGS.resume_checkpoint: checkpoint.load_checkpoint(exe, train_prog, FLAGS.resume_checkpoint) start_iter = checkpoint.global_step() elif cfg.pretrain_weights and fuse_bn and not ignore_params: checkpoint.load_and_fusebn(exe, train_prog, cfg.pretrain_weights) elif cfg.pretrain_weights: checkpoint.load_params(exe, train_prog, cfg.pretrain_weights, ignore_params=ignore_params) train_reader = create_reader(cfg.TrainReader, (cfg.max_iters - start_iter) * devices_num, cfg, devices_num=devices_num, num_trainers=num_trainers) # When iterable mode, set set_sample_list_generator(train_reader, place) train_loader.set_sample_list_generator(train_reader) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() # if map_type not set, use default 11point, only use in VOC eval map_type = cfg.map_type if 'map_type' in cfg else '11point' train_stats = TrainingStats(cfg.log_smooth_window, train_keys) train_loader.start() start_time = time.time() end_time = time.time() cfg_name = os.path.basename(FLAGS.config).split('.')[0] save_dir = os.path.join(cfg.save_dir, cfg_name) time_stat = deque(maxlen=cfg.log_smooth_window) best_box_ap_list = [0.0, 0] #[map, iter] # use VisualDL to log data if FLAGS.use_vdl: assert six.PY3, "VisualDL requires Python >= 3.5" from visualdl import LogWriter vdl_writer = LogWriter(FLAGS.vdl_log_dir) vdl_loss_step = 0 vdl_mAP_step = 0 for it in range(start_iter, cfg.max_iters): start_time = end_time end_time = time.time() time_stat.append(end_time - start_time) time_cost = np.mean(time_stat) eta_sec = (cfg.max_iters - it) * time_cost eta = str(datetime.timedelta(seconds=int(eta_sec))) outs = exe.run(compiled_train_prog, fetch_list=train_values) stats = {k: np.array(v).mean() for k, v in zip(train_keys, outs[:-1])} # use vdl-paddle to log loss if FLAGS.use_vdl: if it % cfg.log_iter == 0: for loss_name, loss_value in stats.items(): vdl_writer.add_scalar(loss_name, loss_value, vdl_loss_step) vdl_loss_step += 1 train_stats.update(stats) logs = train_stats.log() if it % cfg.log_iter == 0 and (not FLAGS.dist or trainer_id == 0): ips = float(cfg['TrainReader']['batch_size']) / time_cost strs = 'iter: {}, lr: {:.6f}, {}, eta: {}, batch_cost: {:.5f} sec, ips: {:.5f} images/sec'.format( it, np.mean(outs[-1]), logs, eta, time_cost, ips) logger.info(strs) # NOTE : profiler tools, used for benchmark if FLAGS.is_profiler and it == 5: profiler.start_profiler("All") elif FLAGS.is_profiler and it == 10: profiler.stop_profiler("total", FLAGS.profiler_path) return if (it > 0 and it % cfg.snapshot_iter == 0 or it == cfg.max_iters - 1) \ and (not FLAGS.dist or trainer_id == 0): save_name = str(it) if it != cfg.max_iters - 1 else "model_final" if 'use_ema' in cfg and cfg['use_ema']: exe.run(ema.apply_program) checkpoint.save(exe, train_prog, os.path.join(save_dir, save_name)) if FLAGS.eval: # evaluation resolution = None if 'Mask' in cfg.architecture: resolution = model.mask_head.resolution results = eval_run(exe, compiled_eval_prog, eval_loader, eval_keys, eval_values, eval_cls, cfg, resolution=resolution) box_ap_stats = eval_results(results, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_eval, map_type, cfg['EvalReader']['dataset']) # use vdl_paddle to log mAP if FLAGS.use_vdl: vdl_writer.add_scalar("mAP", box_ap_stats[0], vdl_mAP_step) vdl_mAP_step += 1 if box_ap_stats[0] > best_box_ap_list[0]: best_box_ap_list[0] = box_ap_stats[0] best_box_ap_list[1] = it checkpoint.save(exe, train_prog, os.path.join(save_dir, "best_model")) logger.info("Best test box ap: {}, in iter: {}".format( best_box_ap_list[0], best_box_ap_list[1])) if 'use_ema' in cfg and cfg['use_ema']: exe.run(ema.restore_program) train_loader.reset()
def main(): cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # check if paddlepaddle version is satisfied check_version() main_arch = cfg.architecture dataset = cfg.TestReader['dataset'] test_images = get_test_images(FLAGS.infer_dir, FLAGS.infer_img) dataset.set_images(test_images) place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) model = create(main_arch) startup_prog = fluid.Program() infer_prog = fluid.Program() with fluid.program_guard(infer_prog, startup_prog): with fluid.unique_name.guard(): inputs_def = cfg['TestReader']['inputs_def'] feed_vars, loader = model.build_inputs(**inputs_def) test_fetches = model.test(feed_vars) infer_prog = infer_prog.clone(True) reader = create_reader(cfg.TestReader) # When iterable mode, set set_sample_list_generator(reader, place) loader.set_sample_list_generator(reader) not_quant_pattern = [] if FLAGS.not_quant_pattern: not_quant_pattern = FLAGS.not_quant_pattern config = { 'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max', 'quantize_op_types': ['depthwise_conv2d', 'mul', 'conv2d'], 'not_quant_pattern': not_quant_pattern } infer_prog = quant_aware(infer_prog, place, config, for_test=True) exe.run(startup_prog) if cfg.weights: checkpoint.load_params(exe, infer_prog, cfg.weights) infer_prog = convert(infer_prog, place, config, save_int8=False) # parse infer fetches assert cfg.metric in ['COCO', 'VOC', 'OID', 'WIDERFACE'], \ "unknown metric type {}".format(cfg.metric) extra_keys = [] if cfg['metric'] in ['COCO', 'OID']: extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg['metric'] == 'VOC' or cfg['metric'] == 'WIDERFACE': extra_keys = ['im_id', 'im_shape'] keys, values, _ = parse_fetches(test_fetches, infer_prog, extra_keys) # parse dataset category if cfg.metric == 'COCO': from ppdet.utils.coco_eval import bbox2out, mask2out, get_category_info if cfg.metric == 'OID': from ppdet.utils.oid_eval import bbox2out, get_category_info if cfg.metric == "VOC": from ppdet.utils.voc_eval import bbox2out, get_category_info if cfg.metric == "WIDERFACE": from ppdet.utils.widerface_eval_utils import bbox2out, get_category_info anno_file = dataset.get_anno() with_background = dataset.with_background use_default_label = dataset.use_default_label clsid2catid, catid2name = get_category_info(anno_file, with_background, use_default_label) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() imid2path = dataset.get_imid2path() iter_id = 0 try: loader.start() while True: outs = exe.run(infer_prog, fetch_list=values, return_numpy=False) res = { k: (np.array(v), v.recursive_sequence_lengths()) for k, v in zip(keys, outs) } logger.info('Infer iter {}'.format(iter_id)) iter_id += 1 bbox_results = None mask_results = None if 'bbox' in res: bbox_results = bbox2out([res], clsid2catid, is_bbox_normalized) if 'mask' in res: mask_results = mask2out([res], clsid2catid, model.mask_head.resolution) # visualize result im_ids = res['im_id'][0] for im_id in im_ids: image_path = imid2path[int(im_id)] image = Image.open(image_path).convert('RGB') image = visualize_results(image, int(im_id), catid2name, FLAGS.draw_threshold, bbox_results, mask_results) save_name = get_save_image_name(FLAGS.output_dir, image_path) logger.info( "Detection bbox results save in {}".format(save_name)) image.save(save_name, quality=95) except (StopIteration, fluid.core.EOFException): loader.reset()
def main(): cfg = load_config(FLAGS.config) if 'architecture' in cfg: main_arch = cfg.architecture else: raise ValueError("'architecture' not specified in config file.") merge_config(FLAGS.opt) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) if 'test_feed' not in cfg: test_feed = create(main_arch + 'TestFeed') else: test_feed = create(cfg.test_feed) test_images = get_test_images(FLAGS.infer_dir, FLAGS.infer_img) test_feed.dataset.add_images(test_images) place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) model = create(main_arch) startup_prog = fluid.Program() infer_prog = fluid.Program() with fluid.program_guard(infer_prog, startup_prog): with fluid.unique_name.guard(): _, feed_vars = create_feed(test_feed, use_pyreader=False) test_fetches = model.test(feed_vars) infer_prog = infer_prog.clone(True) reader = create_reader(test_feed) feeder = fluid.DataFeeder(place=place, feed_list=feed_vars.values()) exe.run(startup_prog) if cfg.weights: checkpoint.load_checkpoint(exe, infer_prog, cfg.weights) if FLAGS.save_inference_model: save_infer_model(FLAGS, exe, feed_vars, test_fetches, infer_prog) # parse infer fetches assert cfg.metric in ['COCO', 'VOC'], \ "unknown metric type {}".format(cfg.metric) extra_keys = [] if cfg['metric'] == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg['metric'] == 'VOC': extra_keys = ['im_id', 'im_shape'] keys, values, _ = parse_fetches(test_fetches, infer_prog, extra_keys) # parse dataset category if cfg.metric == 'COCO': from ppdet.utils.coco_eval import bbox2out, mask2out, get_category_info if cfg.metric == "VOC": from ppdet.utils.voc_eval import bbox2out, get_category_info anno_file = getattr(test_feed.dataset, 'annotation', None) with_background = getattr(test_feed, 'with_background', True) use_default_label = getattr(test_feed, 'use_default_label', False) clsid2catid, catid2name = get_category_info(anno_file, with_background, use_default_label) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() imid2path = reader.imid2path for iter_id, data in enumerate(reader()): outs = exe.run(infer_prog, feed=feeder.feed(data), fetch_list=values, return_numpy=False) res = { k: (np.array(v), v.recursive_sequence_lengths()) for k, v in zip(keys, outs) } logger.info('Infer iter {}'.format(iter_id)) bbox_results = None mask_results = None if 'bbox' in res: bbox_results = bbox2out([res], clsid2catid, is_bbox_normalized) if 'mask' in res: mask_results = mask2out([res], clsid2catid, model.mask_head.resolution) # visualize result im_ids = res['im_id'][0] for im_id in im_ids: image_path = imid2path[int(im_id)] image = Image.open(image_path).convert('RGB') image = visualize_results(image, int(im_id), catid2name, FLAGS.draw_threshold, bbox_results, mask_results) save_name = get_save_image_name(FLAGS.output_dir, image_path) logger.info("Detection bbox results save in {}".format(save_name)) image.save(save_name, quality=95)
def main(): cfg = load_config(FLAGS.config) if 'architecture' in cfg: main_arch = cfg.architecture else: raise ValueError("'architecture' not specified in config file.") merge_config(FLAGS.opt) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # print_total_cfg(cfg) if 'test_feed' not in cfg: test_feed = create(main_arch + 'TestFeed') else: test_feed = create(cfg.test_feed) test_images = get_test_images(FLAGS.infer_dir, FLAGS.infer_img) test_feed.dataset.add_images(test_images) place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) infer_prog, feed_var_names, fetch_list = fluid.io.load_inference_model( dirname=FLAGS.model_path, model_filename=FLAGS.model_name, params_filename=FLAGS.params_name, executor=exe) reader = create_reader(test_feed) feeder = fluid.DataFeeder( place=place, feed_list=feed_var_names, program=infer_prog) # parse infer fetches assert cfg.metric in ['COCO', 'VOC'], \ "unknown metric type {}".format(cfg.metric) extra_keys = [] if cfg['metric'] == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg['metric'] == 'VOC': extra_keys = ['im_id', 'im_shape'] keys, values, _ = parse_fetches({ 'bbox': fetch_list }, infer_prog, extra_keys) # parse dataset category if cfg.metric == 'COCO': from ppdet.utils.coco_eval import bbox2out, mask2out, get_category_info if cfg.metric == "VOC": from ppdet.utils.voc_eval import bbox2out, get_category_info anno_file = getattr(test_feed.dataset, 'annotation', None) with_background = getattr(test_feed, 'with_background', True) use_default_label = getattr(test_feed, 'use_default_label', False) clsid2catid, catid2name = get_category_info(anno_file, with_background, use_default_label) # whether output bbox is normalized in model output layer is_bbox_normalized = False # use tb-paddle to log image if FLAGS.use_tb: from tb_paddle import SummaryWriter tb_writer = SummaryWriter(FLAGS.tb_log_dir) tb_image_step = 0 tb_image_frame = 0 # each frame can display ten pictures at most. imid2path = reader.imid2path keys = ['bbox'] infer_time = True compile_prog = fluid.compiler.CompiledProgram(infer_prog) for iter_id, data in enumerate(reader()): feed_data = [[d[0], d[1]] for d in data] # for infer time if infer_time: warmup_times = 10 repeats_time = 100 feed_data_dict = feeder.feed(feed_data) for i in range(warmup_times): exe.run(compile_prog, feed=feed_data_dict, fetch_list=fetch_list, return_numpy=False) start_time = time.time() for i in range(repeats_time): exe.run(compile_prog, feed=feed_data_dict, fetch_list=fetch_list, return_numpy=False) print("infer time: {} ms/sample".format((time.time() - start_time) * 1000 / repeats_time)) infer_time = False outs = exe.run(compile_prog, feed=feeder.feed(feed_data), fetch_list=fetch_list, return_numpy=False) res = { k: (np.array(v), v.recursive_sequence_lengths()) for k, v in zip(keys, outs) } res['im_id'] = [[d[2] for d in data]] logger.info('Infer iter {}'.format(iter_id)) bbox_results = None mask_results = None if 'bbox' in res: bbox_results = bbox2out([res], clsid2catid, is_bbox_normalized) if 'mask' in res: mask_results = mask2out([res], clsid2catid, model.mask_head.resolution) # visualize result im_ids = res['im_id'][0] for im_id in im_ids: image_path = imid2path[int(im_id)] image = Image.open(image_path).convert('RGB') # use tb-paddle to log original image if FLAGS.use_tb: original_image_np = np.array(image) tb_writer.add_image( "original/frame_{}".format(tb_image_frame), original_image_np, tb_image_step, dataformats='HWC') image = visualize_results(image, int(im_id), catid2name, FLAGS.draw_threshold, bbox_results, mask_results) # use tb-paddle to log image with bbox if FLAGS.use_tb: infer_image_np = np.array(image) tb_writer.add_image( "bbox/frame_{}".format(tb_image_frame), infer_image_np, tb_image_step, dataformats='HWC') tb_image_step += 1 if tb_image_step % 10 == 0: tb_image_step = 0 tb_image_frame += 1 save_name = get_save_image_name(FLAGS.output_dir, image_path) logger.info("Detection bbox results save in {}".format(save_name)) image.save(save_name, quality=95)
def main(): env = os.environ FLAGS.dist = 'PADDLE_TRAINER_ID' in env and 'PADDLE_TRAINERS_NUM' in env if FLAGS.dist: trainer_id = int(env['PADDLE_TRAINER_ID']) local_seed = (99 + trainer_id) random.seed(local_seed) np.random.seed(local_seed) if FLAGS.enable_ce: random.seed(0) np.random.seed(0) cfg = load_config(FLAGS.config) if 'architecture' in cfg: main_arch = cfg.architecture else: raise ValueError("'architecture' not specified in config file.") merge_config(FLAGS.opt) if 'log_iter' not in cfg: cfg.log_iter = 20 # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) if not FLAGS.dist or trainer_id == 0: print_total_cfg(cfg) if cfg.use_gpu: devices_num = fluid.core.get_cuda_device_count() else: devices_num = int(os.environ.get('CPU_NUM', 1)) if 'train_feed' not in cfg: train_feed = create(main_arch + 'TrainFeed') else: train_feed = create(cfg.train_feed) if FLAGS.eval: if 'eval_feed' not in cfg: eval_feed = create(main_arch + 'EvalFeed') else: eval_feed = create(cfg.eval_feed) if 'FLAGS_selected_gpus' in env: device_id = int(env['FLAGS_selected_gpus']) else: device_id = 0 place = fluid.CUDAPlace(device_id) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) lr_builder = create('LearningRate') optim_builder = create('OptimizerBuilder') # build program startup_prog = fluid.Program() train_prog = fluid.Program() if FLAGS.enable_ce: startup_prog.random_seed = 1000 train_prog.random_seed = 1000 with fluid.program_guard(train_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) train_pyreader, feed_vars = create_feed(train_feed) if FLAGS.fp16: assert (getattr(model.backbone, 'norm_type', None) != 'affine_channel'), \ '--fp16 currently does not support affine channel, ' \ ' please modify backbone settings to use batch norm' with mixed_precision_context(FLAGS.loss_scale, FLAGS.fp16) as ctx: train_fetches = model.train(feed_vars) loss = train_fetches['loss'] if FLAGS.fp16: loss *= ctx.get_loss_scale_var() lr = lr_builder() optimizer = optim_builder(lr) optimizer.minimize(loss) if FLAGS.fp16: loss /= ctx.get_loss_scale_var() # parse train fetches train_keys, train_values, _ = parse_fetches(train_fetches) train_values.append(lr) if FLAGS.eval: eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) eval_pyreader, feed_vars = create_feed(eval_feed) fetches = model.eval(feed_vars) eval_prog = eval_prog.clone(True) eval_reader = create_reader(eval_feed, args_path=FLAGS.dataset_dir) eval_pyreader.decorate_sample_list_generator(eval_reader, place) # parse eval fetches extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_box', 'gt_label', 'is_difficult'] if cfg.metric == 'WIDERFACE': extra_keys = ['im_id', 'im_shape', 'gt_box'] eval_keys, eval_values, eval_cls = parse_fetches( fetches, eval_prog, extra_keys) # compile program for multi-devices build_strategy = fluid.BuildStrategy() build_strategy.fuse_all_optimizer_ops = False build_strategy.fuse_elewise_add_act_ops = True # only enable sync_bn in multi GPU devices sync_bn = getattr(model.backbone, 'norm_type', None) == 'sync_bn' build_strategy.sync_batch_norm = sync_bn and devices_num > 1 \ and cfg.use_gpu exec_strategy = fluid.ExecutionStrategy() # iteration number when CompiledProgram tries to drop local execution scopes. # Set it to be 1 to save memory usages, so that unused variables in # local execution scopes can be deleted after each iteration. exec_strategy.num_iteration_per_drop_scope = 1 if FLAGS.dist: dist_utils.prepare_for_multi_process(exe, build_strategy, startup_prog, train_prog) exec_strategy.num_threads = 1 exe.run(startup_prog) compiled_train_prog = fluid.CompiledProgram(train_prog).with_data_parallel( loss_name=loss.name, build_strategy=build_strategy, exec_strategy=exec_strategy) if FLAGS.eval: compiled_eval_prog = fluid.compiler.CompiledProgram(eval_prog) fuse_bn = getattr(model.backbone, 'norm_type', None) == 'affine_channel' ignore_params = cfg.finetune_exclude_pretrained_params \ if 'finetune_exclude_pretrained_params' in cfg else [] start_iter = 0 if FLAGS.resume_checkpoint: checkpoint.load_checkpoint(exe, train_prog, FLAGS.resume_checkpoint) start_iter = checkpoint.global_step() elif cfg.pretrain_weights and fuse_bn and not ignore_params: checkpoint.load_and_fusebn(exe, train_prog, cfg.pretrain_weights) elif cfg.pretrain_weights: checkpoint.load_params(exe, train_prog, cfg.pretrain_weights, ignore_params=ignore_params) train_reader = create_reader(train_feed, (cfg.max_iters - start_iter) * devices_num, FLAGS.dataset_dir) train_pyreader.decorate_sample_list_generator(train_reader, place) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() # if map_type not set, use default 11point, only use in VOC eval map_type = cfg.map_type if 'map_type' in cfg else '11point' train_stats = TrainingStats(cfg.log_smooth_window, train_keys) train_pyreader.start() start_time = time.time() end_time = time.time() cfg_name = os.path.basename(FLAGS.config).split('.')[0] save_dir = os.path.join(cfg.save_dir, cfg_name) time_stat = deque(maxlen=cfg.log_smooth_window) best_box_ap_list = [0.0, 0] #[map, iter] # use tb-paddle to log data if FLAGS.use_tb: from tb_paddle import SummaryWriter tb_writer = SummaryWriter(FLAGS.tb_log_dir) tb_loss_step = 0 tb_mAP_step = 0 for it in range(start_iter, cfg.max_iters): start_time = end_time end_time = time.time() time_stat.append(end_time - start_time) time_cost = np.mean(time_stat) eta_sec = (cfg.max_iters - it) * time_cost eta = str(datetime.timedelta(seconds=int(eta_sec))) outs = exe.run(compiled_train_prog, fetch_list=train_values) stats = {k: np.array(v).mean() for k, v in zip(train_keys, outs[:-1])} # use tb-paddle to log loss if FLAGS.use_tb: if it % cfg.log_iter == 0: for loss_name, loss_value in stats.items(): tb_writer.add_scalar(loss_name, loss_value, tb_loss_step) tb_loss_step += 1 train_stats.update(stats) logs = train_stats.log() if it % cfg.log_iter == 0 and (not FLAGS.dist or trainer_id == 0): strs = 'iter: {}, lr: {:.6f}, {}, time: {:.3f}, eta: {}'.format( it, np.mean(outs[-1]), logs, time_cost, eta) logger.info(strs) #only for continuous evaluation if FLAGS.enable_ce and it == cfg.max_iters - 1: print("kpis\t{}_train_loss\t{}".format(cfg.architecture, stats['loss'])) print("kpis\t{}_train_time\t{}".format(cfg.architecture, time_cost)) # profiler tools, used for benchmark if FLAGS.is_profiler and it == 5: profiler.start_profiler("All") elif FLAGS.is_profiler and it == 10: profiler.stop_profiler("total", FLAGS.profiler_path) return if (it > 0 and it % cfg.snapshot_iter == 0 or it == cfg.max_iters - 1) \ and (not FLAGS.dist or trainer_id == 0): save_name = str(it) if it != cfg.max_iters - 1 else "model_final" checkpoint.save(exe, train_prog, os.path.join(save_dir, save_name)) if FLAGS.eval: # evaluation results = eval_run(exe, compiled_eval_prog, eval_pyreader, eval_keys, eval_values, eval_cls) resolution = None if 'mask' in results[0]: resolution = model.mask_head.resolution box_ap_stats = eval_results(results, eval_feed, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_eval, map_type) # use tb_paddle to log mAP if FLAGS.use_tb: tb_writer.add_scalar("mAP", box_ap_stats[0], tb_mAP_step) tb_mAP_step += 1 if box_ap_stats[0] > best_box_ap_list[0]: best_box_ap_list[0] = box_ap_stats[0] best_box_ap_list[1] = it checkpoint.save(exe, train_prog, os.path.join(save_dir, "best_model")) logger.info("Best test box ap: {}, in iter: {}".format( best_box_ap_list[0], best_box_ap_list[1])) train_pyreader.reset()
def main(): cfg = load_config(FLAGS.config) if 'architecture' in cfg: main_arch = cfg.architecture else: raise ValueError("'architecture' not specified in config file.") merge_config(FLAGS.opt) if 'log_iter' not in cfg: cfg.log_iter = 20 # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) if cfg.use_gpu: devices_num = fluid.core.get_cuda_device_count() else: devices_num = int( os.environ.get('CPU_NUM', multiprocessing.cpu_count())) if 'train_feed' not in cfg: train_feed = create(main_arch + 'TrainFeed') else: train_feed = create(cfg.train_feed) if 'eval_feed' not in cfg: eval_feed = create(main_arch + 'EvalFeed') else: eval_feed = create(cfg.eval_feed) place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) lr_builder = create('LearningRate') optim_builder = create('OptimizerBuilder') # build program startup_prog = fluid.Program() train_prog = fluid.Program() with fluid.program_guard(train_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) _, feed_vars = create_feed(train_feed, True) train_fetches = model.train(feed_vars) loss = train_fetches['loss'] lr = lr_builder() optimizer = optim_builder(lr) optimizer.minimize(loss) train_reader = create_reader(train_feed, cfg.max_iters, FLAGS.dataset_dir) # parse train fetches train_keys, train_values, _ = parse_fetches(train_fetches) train_keys.append("lr") train_values.append(lr.name) train_fetch_list = [] for k, v in zip(train_keys, train_values): train_fetch_list.append((k, v)) eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) _, test_feed_vars = create_feed(eval_feed, True) fetches = model.eval(test_feed_vars) eval_prog = eval_prog.clone(True) eval_reader = create_reader(eval_feed, args_path=FLAGS.dataset_dir) test_data_feed = fluid.DataFeeder(test_feed_vars.values(), place) # parse eval fetches extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_box', 'gt_label', 'is_difficult'] eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog, extra_keys) eval_fetch_list = [] for k, v in zip(eval_keys, eval_values): eval_fetch_list.append((k, v)) exe.run(startup_prog) checkpoint.load_params(exe, train_prog, cfg.pretrain_weights) best_box_ap_list = [] def eval_func(program, scope): #place = fluid.CPUPlace() #exe = fluid.Executor(place) results = eval_run(exe, program, eval_reader, eval_keys, eval_values, eval_cls, test_data_feed, cfg) resolution = None if 'mask' in results[0]: resolution = model.mask_head.resolution box_ap_stats = eval_results(results, eval_feed, cfg.metric, cfg.num_classes, resolution, False, FLAGS.output_eval) if len(best_box_ap_list) == 0: best_box_ap_list.append(box_ap_stats[0]) elif box_ap_stats[0] > best_box_ap_list[0]: best_box_ap_list[0] = box_ap_stats[0] logger.info("Best test box ap: {}".format(best_box_ap_list[0])) return best_box_ap_list[0] test_feed = [('image', test_feed_vars['image'].name), ('im_size', test_feed_vars['im_size'].name)] com = Compressor(place, fluid.global_scope(), train_prog, train_reader=train_reader, train_feed_list=[(key, value.name) for key, value in feed_vars.items()], train_fetch_list=train_fetch_list, eval_program=eval_prog, eval_reader=eval_reader, eval_feed_list=test_feed, eval_func={'map': eval_func}, eval_fetch_list=[eval_fetch_list[0]], save_eval_model=True, prune_infer_model=[["image", "im_size"], ["multiclass_nms_0.tmp_0"]], train_optimizer=None) com.config(FLAGS.slim_file) com.run()
def main(): if FLAGS.eval is False: raise ValueError( "Currently only supports `--eval==True` while training in `quantization`." ) env = os.environ FLAGS.dist = 'PADDLE_TRAINER_ID' in env and 'PADDLE_TRAINERS_NUM' in env if FLAGS.dist: trainer_id = int(env['PADDLE_TRAINER_ID']) import random local_seed = (99 + trainer_id) random.seed(local_seed) np.random.seed(local_seed) cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) # check if set use_gpu=True in paddlepaddle cpu version check_gpu(cfg.use_gpu) # check if paddlepaddle version is satisfied check_version() main_arch = cfg.architecture if cfg.use_gpu: devices_num = fluid.core.get_cuda_device_count() else: devices_num = int(os.environ.get('CPU_NUM', 1)) if 'FLAGS_selected_gpus' in env: device_id = int(env['FLAGS_selected_gpus']) else: device_id = 0 place = fluid.CUDAPlace(device_id) if cfg.use_gpu else fluid.CPUPlace() exe = fluid.Executor(place) lr_builder = create('LearningRate') optim_builder = create('OptimizerBuilder') # build program startup_prog = fluid.Program() train_prog = fluid.Program() with fluid.program_guard(train_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) inputs_def = cfg['TrainReader']['inputs_def'] feed_vars, train_loader = model.build_inputs(**inputs_def) train_fetches = model.train(feed_vars) loss = train_fetches['loss'] lr = lr_builder() optimizer = optim_builder(lr) optimizer.minimize(loss) # parse train fetches train_keys, train_values, _ = parse_fetches(train_fetches) train_values.append(lr) if FLAGS.eval: eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) inputs_def = cfg['EvalReader']['inputs_def'] feed_vars, eval_loader = model.build_inputs(**inputs_def) fetches = model.eval(feed_vars) eval_prog = eval_prog.clone(True) eval_reader = create_reader(cfg.EvalReader) eval_loader.set_sample_list_generator(eval_reader, place) # parse eval fetches extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_bbox', 'gt_class', 'is_difficult'] if cfg.metric == 'WIDERFACE': extra_keys = ['im_id', 'im_shape', 'gt_bbox'] eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog, extra_keys) # compile program for multi-devices build_strategy = fluid.BuildStrategy() build_strategy.fuse_all_optimizer_ops = False build_strategy.fuse_elewise_add_act_ops = True build_strategy.fuse_all_reduce_ops = False # only enable sync_bn in multi GPU devices sync_bn = getattr(model.backbone, 'norm_type', None) == 'sync_bn' sync_bn = False build_strategy.sync_batch_norm = sync_bn and devices_num > 1 \ and cfg.use_gpu exec_strategy = fluid.ExecutionStrategy() # iteration number when CompiledProgram tries to drop local execution scopes. # Set it to be 1 to save memory usages, so that unused variables in # local execution scopes can be deleted after each iteration. exec_strategy.num_iteration_per_drop_scope = 1 if FLAGS.dist: dist_utils.prepare_for_multi_process(exe, build_strategy, startup_prog, train_prog) exec_strategy.num_threads = 1 exe.run(startup_prog) not_quant_pattern = [] if FLAGS.not_quant_pattern: not_quant_pattern = FLAGS.not_quant_pattern config = { 'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max', 'quantize_op_types': ['depthwise_conv2d', 'mul', 'conv2d'], 'not_quant_pattern': not_quant_pattern } ignore_params = cfg.finetune_exclude_pretrained_params \ if 'finetune_exclude_pretrained_params' in cfg else [] fuse_bn = getattr(model.backbone, 'norm_type', None) == 'affine_channel' if not FLAGS.resume_checkpoint: if cfg.pretrain_weights and fuse_bn and not ignore_params: checkpoint.load_and_fusebn(exe, train_prog, cfg.pretrain_weights) elif cfg.pretrain_weights: checkpoint.load_params( exe, train_prog, cfg.pretrain_weights, ignore_params=ignore_params) # insert quantize op in train_prog, return type is CompiledProgram train_prog_quant = quant_aware(train_prog, place, config, for_test=False) compiled_train_prog = train_prog_quant.with_data_parallel( loss_name=loss.name, build_strategy=build_strategy, exec_strategy=exec_strategy) if FLAGS.eval: # insert quantize op in eval_prog eval_prog = quant_aware(eval_prog, place, config, for_test=True) compiled_eval_prog = fluid.CompiledProgram(eval_prog) start_iter = 0 if FLAGS.resume_checkpoint: checkpoint.load_checkpoint(exe, eval_prog, FLAGS.resume_checkpoint) load_global_step(exe, train_prog, FLAGS.resume_checkpoint) start_iter = checkpoint.global_step() train_reader = create_reader(cfg.TrainReader, (cfg.max_iters - start_iter) * devices_num) train_loader.set_sample_list_generator(train_reader, place) # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() # if map_type not set, use default 11point, only use in VOC eval map_type = cfg.map_type if 'map_type' in cfg else '11point' train_stats = TrainingStats(cfg.log_smooth_window, train_keys) train_loader.start() start_time = time.time() end_time = time.time() cfg_name = os.path.basename(FLAGS.config).split('.')[0] save_dir = os.path.join(cfg.save_dir, cfg_name) time_stat = deque(maxlen=cfg.log_smooth_window) best_box_ap_list = [0.0, 0] #[map, iter] for it in range(start_iter, cfg.max_iters): start_time = end_time end_time = time.time() time_stat.append(end_time - start_time) time_cost = np.mean(time_stat) eta_sec = (cfg.max_iters - it) * time_cost eta = str(datetime.timedelta(seconds=int(eta_sec))) outs = exe.run(compiled_train_prog, fetch_list=train_values) stats = {k: np.array(v).mean() for k, v in zip(train_keys, outs[:-1])} train_stats.update(stats) logs = train_stats.log() if it % cfg.log_iter == 0 and (not FLAGS.dist or trainer_id == 0): strs = 'iter: {}, lr: {:.6f}, {}, time: {:.3f}, eta: {}'.format( it, np.mean(outs[-1]), logs, time_cost, eta) logger.info(strs) if (it > 0 and it % cfg.snapshot_iter == 0 or it == cfg.max_iters - 1) \ and (not FLAGS.dist or trainer_id == 0): save_name = str(it) if it != cfg.max_iters - 1 else "model_final" save_checkpoint(exe, eval_prog, os.path.join(save_dir, save_name), train_prog) if FLAGS.eval: # evaluation results = eval_run( exe, compiled_eval_prog, eval_loader, eval_keys, eval_values, eval_cls, cfg=cfg) resolution = None if 'mask' in results[0]: resolution = model.mask_head.resolution box_ap_stats = eval_results( results, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_eval, map_type, cfg['EvalReader']['dataset']) if box_ap_stats[0] > best_box_ap_list[0]: best_box_ap_list[0] = box_ap_stats[0] best_box_ap_list[1] = it save_checkpoint(exe, eval_prog, os.path.join(save_dir, "best_model"), train_prog) logger.info("Best test box ap: {}, in iter: {}".format( best_box_ap_list[0], best_box_ap_list[1])) train_loader.reset()
def main(): env = os.environ print("FLAGS.config: {}".format(FLAGS.config)) cfg = load_config(FLAGS.config) merge_config(FLAGS.opt) check_config(cfg) check_version() main_arch = cfg.architecture place = fluid.CUDAPlace(0) exe = fluid.Executor(place) # build program startup_prog = fluid.Program() eval_prog = fluid.Program() with fluid.program_guard(eval_prog, startup_prog): with fluid.unique_name.guard(): model = create(main_arch) inputs_def = cfg['EvalReader']['inputs_def'] feed_vars, eval_loader = model.build_inputs(**inputs_def) fetches = model.eval(feed_vars) eval_prog = eval_prog.clone(True) if FLAGS.print_params: print( "-------------------------All parameters in current graph----------------------" ) for block in eval_prog.blocks: for param in block.all_parameters(): print("parameter name: {}\tshape: {}".format( param.name, param.shape)) print( "------------------------------------------------------------------------------" ) return eval_reader = create_reader(cfg.EvalReader) eval_loader.set_sample_list_generator(eval_reader, place) # parse eval fetches extra_keys = [] if cfg.metric == 'COCO': extra_keys = ['im_info', 'im_id', 'im_shape'] if cfg.metric == 'VOC': extra_keys = ['gt_bbox', 'gt_class', 'is_difficult'] if cfg.metric == 'WIDERFACE': extra_keys = ['im_id', 'im_shape', 'gt_box'] eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog, extra_keys) exe.run(startup_prog) fuse_bn = getattr(model.backbone, 'norm_type', None) == 'affine_channel' ignore_params = cfg.finetune_exclude_pretrained_params \ if 'finetune_exclude_pretrained_params' in cfg else [] start_iter = 0 if cfg.weights: checkpoint.load_params(exe, eval_prog, cfg.weights) else: logger.warn("Please set cfg.weights to load trained model.") # whether output bbox is normalized in model output layer is_bbox_normalized = False if hasattr(model, 'is_bbox_normalized') and \ callable(model.is_bbox_normalized): is_bbox_normalized = model.is_bbox_normalized() # if map_type not set, use default 11point, only use in VOC eval map_type = cfg.map_type if 'map_type' in cfg else '11point' def test(program): compiled_eval_prog = fluid.CompiledProgram(program) results = eval_run(exe, compiled_eval_prog, eval_loader, eval_keys, eval_values, eval_cls, cfg=cfg) resolution = None if 'mask' in results[0]: resolution = model.mask_head.resolution dataset = cfg['EvalReader']['dataset'] box_ap_stats = eval_results(results, cfg.metric, cfg.num_classes, resolution, is_bbox_normalized, FLAGS.output_eval, map_type, dataset=dataset) return box_ap_stats[0] pruned_params = FLAGS.pruned_params assert ( FLAGS.pruned_params is not None ), "FLAGS.pruned_params is empty!!! Please set it by '--pruned_params' option." pruned_params = FLAGS.pruned_params.strip().split(",") logger.info("pruned params: {}".format(pruned_params)) pruned_ratios = [float(n) for n in FLAGS.pruned_ratios.strip().split(" ")] logger.info("pruned ratios: {}".format(pruned_ratios)) sensitivity(eval_prog, place, pruned_params, test, sensitivities_file=FLAGS.sensitivities_file, pruned_ratios=pruned_ratios)