From d51e5fdc7f9125d73616a4260b1bc17d8c533843 Mon Sep 17 00:00:00 2001 From: D-X-Y <280835372@qq.com> Date: Wed, 12 May 2021 15:45:45 +0800 Subject: [PATCH] Update LFNA ablation codes --- exps/LFNA/backup/lfna-debug.py | 280 ---------------------- exps/LFNA/lfna-test-hpnet.py | 176 ++++++++++++++ exps/LFNA/lfna-ttss-hpnet.py | 134 +++++++++++ exps/LFNA/lfna_models.py | 97 ++++++++ lib/xlayers/super_core.py | 1 + lib/xlayers/super_positional_embedding.py | 35 +++ 6 files changed, 443 insertions(+), 280 deletions(-) delete mode 100644 exps/LFNA/backup/lfna-debug.py create mode 100644 exps/LFNA/lfna-test-hpnet.py create mode 100644 exps/LFNA/lfna-ttss-hpnet.py create mode 100644 exps/LFNA/lfna_models.py diff --git a/exps/LFNA/backup/lfna-debug.py b/exps/LFNA/backup/lfna-debug.py deleted file mode 100644 index 969d9d1..0000000 --- a/exps/LFNA/backup/lfna-debug.py +++ /dev/null @@ -1,280 +0,0 @@ -##################################################### -# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 # -##################################################### -# python exps/LFNA/lfna-debug.py --env_version v1 --hidden_dim 16 -##################################################### -import sys, time, copy, torch, random, argparse -from tqdm import tqdm -from copy import deepcopy -from pathlib import Path - -lib_dir = (Path(__file__).parent / ".." / ".." / "lib").resolve() -if str(lib_dir) not in sys.path: - sys.path.insert(0, str(lib_dir)) -from procedures import prepare_seed, prepare_logger, save_checkpoint, copy_checkpoint -from log_utils import time_string -from log_utils import AverageMeter, convert_secs2time - -from utils import split_str2indexes - -from procedures.advanced_main import basic_train_fn, basic_eval_fn -from procedures.metric_utils import SaveMetric, MSEMetric, ComposeMetric -from datasets.synthetic_core import get_synthetic_env -from models.xcore import get_model -from xlayers import super_core - - -from lfna_utils import lfna_setup, train_model, TimeData -from lfna_models import HyperNet - - -class LFNAmlp: - """A LFNA meta-model that uses the MLP as delta-net.""" - - def __init__(self, obs_dim, hidden_sizes, act_name, criterion): - self.delta_net = super_core.SuperSequential( - super_core.SuperLinear(obs_dim, hidden_sizes[0]), - super_core.super_name2activation[act_name](), - super_core.SuperLinear(hidden_sizes[0], hidden_sizes[1]), - super_core.super_name2activation[act_name](), - super_core.SuperLinear(hidden_sizes[1], 1), - ) - self.meta_optimizer = torch.optim.Adam( - self.delta_net.parameters(), lr=0.01, amsgrad=True - ) - self.criterion = criterion - - def adapt(self, model, seq_flatten_w): - delta_inputs = torch.stack(seq_flatten_w, dim=-1) - delta = self.delta_net(delta_inputs) - container = model.get_w_container() - unflatten_delta = container.unflatten(delta) - future_container = container.create_container(unflatten_delta) - return future_container - - def step(self): - torch.nn.utils.clip_grad_norm_(self.delta_net.parameters(), 1.0) - self.meta_optimizer.step() - - def zero_grad(self): - self.meta_optimizer.zero_grad() - self.delta_net.zero_grad() - - def state_dict(self): - return dict( - delta_net=self.delta_net.state_dict(), - meta_optimizer=self.meta_optimizer.state_dict(), - ) - - -def main(args): - logger, env_info, model_kwargs = lfna_setup(args) - dynamic_env = env_info["dynamic_env"] - model = get_model(dict(model_type="simple_mlp"), **model_kwargs) - - total_time = env_info["total"] - for i in range(total_time): - for xkey in ("timestamp", "x", "y"): - nkey = "{:}-{:}".format(i, xkey) - assert nkey in env_info, "{:} no in {:}".format(nkey, list(env_info.keys())) - train_time_bar = total_time // 2 - - criterion = torch.nn.MSELoss() - logger.log("There are {:} weights.".format(model.get_w_container().numel())) - - adaptor = LFNAmlp(args.meta_seq, (200, 200), "leaky_relu", criterion) - - # pre-train the model - dataset = init_dataset = TimeData(0, env_info["0-x"], env_info["0-y"]) - - shape_container = model.get_w_container().to_shape_container() - hypernet = HyperNet(shape_container, 16) - - optimizer = torch.optim.Adam(hypernet.parameters(), lr=args.init_lr, amsgrad=True) - - best_loss, best_param = None, None - for _iepoch in range(args.epochs): - container = hypernet(None) - - preds = model.forward_with_container(dataset.x, container) - optimizer.zero_grad() - loss = criterion(preds, dataset.y) - loss.backward() - optimizer.step() - # save best - if best_loss is None or best_loss > loss.item(): - best_loss = loss.item() - best_param = copy.deepcopy(model.state_dict()) - print("hyper-net : best={:.4f}".format(best_loss)) - - init_loss = train_model(model, init_dataset, args.init_lr, args.epochs) - logger.log("The pre-training loss is {:.4f}".format(init_loss)) - import pdb - - pdb.set_trace() - - all_past_containers = [] - ground_truth_path = ( - logger.path(None) / ".." / "use-same-timestamp-v1-d16" / "final-ckp.pth" - ) - ground_truth_data = torch.load(ground_truth_path) - all_gt_containers = ground_truth_data["w_container_per_epoch"] - all_gt_flattens = dict() - for idx, container in all_gt_containers.items(): - all_gt_flattens[idx] = container.no_grad_clone().flatten() - - # LFNA meta-training - meta_loss_meter = AverageMeter() - per_epoch_time, start_time = AverageMeter(), time.time() - for iepoch in range(args.epochs): - - need_time = "Time Left: {:}".format( - convert_secs2time(per_epoch_time.avg * (args.epochs - iepoch), True) - ) - logger.log( - "[{:}] [{:04d}/{:04d}] ".format(time_string(), iepoch, args.epochs) - + need_time - ) - - adaptor.zero_grad() - - meta_losses = [] - for ibatch in range(args.meta_batch): - future_timestamp = random.randint(args.meta_seq, train_time_bar) - future_dataset = TimeData( - future_timestamp, - env_info["{:}-x".format(future_timestamp)], - env_info["{:}-y".format(future_timestamp)], - ) - seq_datasets = [] - for iseq in range(args.meta_seq): - cur_time = future_timestamp - iseq - 1 - cur_x = env_info["{:}-x".format(cur_time)] - cur_y = env_info["{:}-y".format(cur_time)] - seq_datasets.append(TimeData(cur_time, cur_x, cur_y)) - seq_datasets.reverse() - seq_flatten_w = [ - all_gt_flattens[dataset.timestamp] for dataset in seq_datasets - ] - future_container = adaptor.adapt(network, seq_flatten_w) - """ - future_y_hat = network.forward_with_container( - future_dataset.x, future_container - ) - future_loss = adaptor.criterion(future_y_hat, future_dataset.y) - """ - future_loss = adaptor.criterion( - future_container.flatten(), all_gt_flattens[future_timestamp] - ) - # import pdb; pdb.set_trace() - meta_losses.append(future_loss) - meta_loss = torch.stack(meta_losses).mean() - meta_loss.backward() - adaptor.step() - - meta_loss_meter.update(meta_loss.item()) - - logger.log( - "meta-loss: {:.4f} ({:.4f}) ".format( - meta_loss_meter.avg, meta_loss_meter.val - ) - ) - if iepoch % 200 == 0: - save_checkpoint( - {"adaptor": adaptor.state_dict(), "iepoch": iepoch}, - logger.path("model"), - logger, - ) - per_epoch_time.update(time.time() - start_time) - start_time = time.time() - - w_container_per_epoch = dict() - # import pdb; pdb.set_trace() - for idx in range(1, env_info["total"]): - future_time = env_info["{:}-timestamp".format(idx)] - future_x = env_info["{:}-x".format(idx)] - future_y = env_info["{:}-y".format(idx)] - seq_datasets = [] - for iseq in range(1, args.meta_seq + 1): - cur_time = future_timestamp - iseq - 1 - if cur_time < 0: - cur_time = 0 - cur_x = env_info["{:}-x".format(cur_time)] - cur_y = env_info["{:}-y".format(cur_time)] - seq_datasets.append(TimeData(cur_time, cur_x, cur_y)) - seq_datasets.reverse() - seq_flatten_w = [all_gt_flattens[dataset.timestamp] for dataset in seq_datasets] - future_container = adaptor.adapt(network, seq_flatten_w) - w_container_per_epoch[idx] = future_container.no_grad_clone() - with torch.no_grad(): - future_y_hat = network.forward_with_container( - future_x, w_container_per_epoch[idx] - ) - future_loss = adaptor.criterion(future_y_hat, future_y) - logger.log("meta-test: [{:03d}] -> loss={:.4f}".format(idx, future_loss.item())) - - save_checkpoint( - {"w_container_per_epoch": w_container_per_epoch}, - logger.path(None) / "final-ckp.pth", - logger, - ) - - logger.log("-" * 200 + "\n") - logger.close() - - -if __name__ == "__main__": - parser = argparse.ArgumentParser("Use the data in the past.") - parser.add_argument( - "--save_dir", - type=str, - default="./outputs/lfna-synthetic/lfna-debug", - help="The checkpoint directory.", - ) - parser.add_argument( - "--env_version", - type=str, - required=True, - help="The synthetic enviornment version.", - ) - parser.add_argument( - "--hidden_dim", - type=int, - required=True, - help="The hidden dimension.", - ) - ##### - parser.add_argument( - "--init_lr", - type=float, - default=0.1, - help="The initial learning rate for the optimizer (default is Adam)", - ) - parser.add_argument( - "--meta_batch", - type=int, - default=32, - help="The batch size for the meta-model", - ) - parser.add_argument( - "--meta_seq", - type=int, - default=10, - help="The length of the sequence for meta-model.", - ) - parser.add_argument( - "--epochs", - type=int, - default=2000, - help="The total number of epochs.", - ) - # Random Seed - parser.add_argument("--rand_seed", type=int, default=-1, help="manual seed") - args = parser.parse_args() - if args.rand_seed is None or args.rand_seed < 0: - args.rand_seed = random.randint(1, 100000) - assert args.save_dir is not None, "The save dir argument can not be None" - args.save_dir = "{:}-{:}-d{:}".format( - args.save_dir, args.env_version, args.hidden_dim - ) - main(args) diff --git a/exps/LFNA/lfna-test-hpnet.py b/exps/LFNA/lfna-test-hpnet.py new file mode 100644 index 0000000..3cf5552 --- /dev/null +++ b/exps/LFNA/lfna-test-hpnet.py @@ -0,0 +1,176 @@ +##################################################### +# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 # +##################################################### +# python exps/LFNA/lfna-test-hpnet.py --env_version v1 --hidden_dim 16 +##################################################### +import sys, time, copy, torch, random, argparse +from tqdm import tqdm +from copy import deepcopy +from pathlib import Path + +lib_dir = (Path(__file__).parent / ".." / ".." / "lib").resolve() +if str(lib_dir) not in sys.path: + sys.path.insert(0, str(lib_dir)) +from procedures import prepare_seed, prepare_logger, save_checkpoint, copy_checkpoint +from log_utils import time_string +from log_utils import AverageMeter, convert_secs2time + +from utils import split_str2indexes + +from procedures.advanced_main import basic_train_fn, basic_eval_fn +from procedures.metric_utils import SaveMetric, MSEMetric, ComposeMetric +from datasets.synthetic_core import get_synthetic_env +from models.xcore import get_model +from xlayers import super_core, trunc_normal_ + + +from lfna_utils import lfna_setup, train_model, TimeData + +# from lfna_models import HyperNet_VX as HyperNet +from lfna_models import HyperNet + + +def main(args): + logger, env_info, model_kwargs = lfna_setup(args) + dynamic_env = env_info["dynamic_env"] + model = get_model(dict(model_type="simple_mlp"), **model_kwargs) + criterion = torch.nn.MSELoss() + + logger.log("There are {:} weights.".format(model.get_w_container().numel())) + + shape_container = model.get_w_container().to_shape_container() + hypernet = HyperNet(shape_container, args.hidden_dim, args.task_dim) + # task_embed = torch.nn.Parameter(torch.Tensor(env_info["total"], args.task_dim)) + task_embed = torch.nn.Parameter(torch.Tensor(1, args.task_dim)) + trunc_normal_(task_embed, std=0.02) + + parameters = list(hypernet.parameters()) + [task_embed] + optimizer = torch.optim.Adam(parameters, lr=args.init_lr, amsgrad=True) + lr_scheduler = torch.optim.lr_scheduler.MultiStepLR( + optimizer, + milestones=[ + int(args.epochs * 0.8), + int(args.epochs * 0.9), + ], + gamma=0.1, + ) + + # total_bar = env_info["total"] - 1 + total_bar = 1 + # LFNA meta-training + loss_meter = AverageMeter() + per_epoch_time, start_time = AverageMeter(), time.time() + for iepoch in range(args.epochs): + + need_time = "Time Left: {:}".format( + convert_secs2time(per_epoch_time.avg * (args.epochs - iepoch), True) + ) + head_str = ( + "[{:}] [{:04d}/{:04d}] ".format(time_string(), iepoch, args.epochs) + + need_time + ) + + losses = [] + # for ibatch in range(args.meta_batch): + for cur_time in range(total_bar): + # cur_time = random.randint(0, total_bar) + cur_task_embed = task_embed + cur_container = hypernet(cur_task_embed) + cur_x = env_info["{:}-x".format(cur_time)] + cur_y = env_info["{:}-y".format(cur_time)] + cur_dataset = TimeData(cur_time, cur_x, cur_y) + + preds = model.forward_with_container(cur_dataset.x, cur_container) + optimizer.zero_grad() + loss = criterion(preds, cur_dataset.y) + + losses.append(loss) + + final_loss = torch.stack(losses).mean() + final_loss.backward() + optimizer.step() + lr_scheduler.step() + + loss_meter.update(final_loss.item()) + if iepoch % 200 == 0: + logger.log( + head_str + + "meta-loss: {:.4f} ({:.4f}) :: lr={:.5f}, batch={:}".format( + loss_meter.avg, + loss_meter.val, + min(lr_scheduler.get_lr()), + len(losses), + ) + ) + + save_checkpoint( + { + "hypernet": hypernet.state_dict(), + "task_embed": task_embed, + "lr_scheduler": lr_scheduler.state_dict(), + "iepoch": iepoch, + }, + logger.path("model"), + logger, + ) + loss_meter.reset() + per_epoch_time.update(time.time() - start_time) + start_time = time.time() + + print(model) + print(hypernet) + + logger.log("-" * 200 + "\n") + logger.close() + + +if __name__ == "__main__": + parser = argparse.ArgumentParser("Use the data in the past.") + parser.add_argument( + "--save_dir", + type=str, + default="./outputs/lfna-synthetic/lfna-test-hpnet", + help="The checkpoint directory.", + ) + parser.add_argument( + "--env_version", + type=str, + required=True, + help="The synthetic enviornment version.", + ) + parser.add_argument( + "--hidden_dim", + type=int, + required=True, + help="The hidden dimension.", + ) + ##### + parser.add_argument( + "--init_lr", + type=float, + default=0.1, + help="The initial learning rate for the optimizer (default is Adam)", + ) + parser.add_argument( + "--meta_batch", + type=int, + default=64, + help="The batch size for the meta-model", + ) + parser.add_argument( + "--epochs", + type=int, + default=2000, + help="The total number of epochs.", + ) + # Random Seed + parser.add_argument("--rand_seed", type=int, default=-1, help="manual seed") + args = parser.parse_args() + if args.rand_seed is None or args.rand_seed < 0: + args.rand_seed = random.randint(1, 100000) + assert args.save_dir is not None, "The save dir argument can not be None" + args.task_dim = args.hidden_dim + args.save_dir = "{:}-{:}-d{:}".format( + args.save_dir, args.env_version, args.hidden_dim + ) + main(args) diff --git a/exps/LFNA/lfna-ttss-hpnet.py b/exps/LFNA/lfna-ttss-hpnet.py new file mode 100644 index 0000000..1f3bbde --- /dev/null +++ b/exps/LFNA/lfna-ttss-hpnet.py @@ -0,0 +1,134 @@ +##################################################### +# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 # +##################################################### +# python exps/LFNA/lfna-ttss-hpnet.py --env_version v1 --hidden_dim 16 +##################################################### +import sys, time, copy, torch, random, argparse +from tqdm import tqdm +from copy import deepcopy +from pathlib import Path + +lib_dir = (Path(__file__).parent / ".." / ".." / "lib").resolve() +if str(lib_dir) not in sys.path: + sys.path.insert(0, str(lib_dir)) +from procedures import prepare_seed, prepare_logger, save_checkpoint, copy_checkpoint +from log_utils import time_string +from log_utils import AverageMeter, convert_secs2time + +from utils import split_str2indexes + +from procedures.advanced_main import basic_train_fn, basic_eval_fn +from procedures.metric_utils import SaveMetric, MSEMetric, ComposeMetric +from datasets.synthetic_core import get_synthetic_env +from models.xcore import get_model +from xlayers import super_core + + +from lfna_utils import lfna_setup, train_model, TimeData +from lfna_models import HyperNet_VX as HyperNet + + +def main(args): + logger, env_info, model_kwargs = lfna_setup(args) + dynamic_env = env_info["dynamic_env"] + model = get_model(dict(model_type="simple_mlp"), **model_kwargs) + + total_time = env_info["total"] + for i in range(total_time): + for xkey in ("timestamp", "x", "y"): + nkey = "{:}-{:}".format(i, xkey) + assert nkey in env_info, "{:} no in {:}".format(nkey, list(env_info.keys())) + train_time_bar = total_time // 2 + + criterion = torch.nn.MSELoss() + logger.log("There are {:} weights.".format(model.get_w_container().numel())) + + # pre-train the model + dataset = init_dataset = TimeData(0, env_info["0-x"], env_info["0-y"]) + + shape_container = model.get_w_container().to_shape_container() + hypernet = HyperNet(shape_container, 16) + print(hypernet) + + optimizer = torch.optim.Adam(hypernet.parameters(), lr=args.init_lr, amsgrad=True) + + best_loss, best_param = None, None + for _iepoch in range(args.epochs): + container = hypernet(None) + + preds = model.forward_with_container(dataset.x, container) + optimizer.zero_grad() + loss = criterion(preds, dataset.y) + loss.backward() + optimizer.step() + # save best + if best_loss is None or best_loss > loss.item(): + best_loss = loss.item() + best_param = copy.deepcopy(model.state_dict()) + print("hyper-net : best={:.4f}".format(best_loss)) + + init_loss = train_model(model, init_dataset, args.init_lr, args.epochs) + logger.log("The pre-training loss is {:.4f}".format(init_loss)) + + print(model) + print(hypernet) + + logger.log("-" * 200 + "\n") + logger.close() + + +if __name__ == "__main__": + parser = argparse.ArgumentParser("Use the data in the past.") + parser.add_argument( + "--save_dir", + type=str, + default="./outputs/lfna-synthetic/lfna-debug", + help="The checkpoint directory.", + ) + parser.add_argument( + "--env_version", + type=str, + required=True, + help="The synthetic enviornment version.", + ) + parser.add_argument( + "--hidden_dim", + type=int, + required=True, + help="The hidden dimension.", + ) + ##### + parser.add_argument( + "--init_lr", + type=float, + default=0.1, + help="The initial learning rate for the optimizer (default is Adam)", + ) + parser.add_argument( + "--meta_batch", + type=int, + default=32, + help="The batch size for the meta-model", + ) + parser.add_argument( + "--meta_seq", + type=int, + default=10, + help="The length of the sequence for meta-model.", + ) + parser.add_argument( + "--epochs", + type=int, + default=2000, + help="The total number of epochs.", + ) + # Random Seed + parser.add_argument("--rand_seed", type=int, default=-1, help="manual seed") + args = parser.parse_args() + if args.rand_seed is None or args.rand_seed < 0: + args.rand_seed = random.randint(1, 100000) + assert args.save_dir is not None, "The save dir argument can not be None" + args.save_dir = "{:}-{:}-d{:}".format( + args.save_dir, args.env_version, args.hidden_dim + ) + main(args) diff --git a/exps/LFNA/lfna_models.py b/exps/LFNA/lfna_models.py new file mode 100644 index 0000000..c85f4fa --- /dev/null +++ b/exps/LFNA/lfna_models.py @@ -0,0 +1,97 @@ +##################################################### +# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 # +##################################################### +import copy +import torch + +from xlayers import super_core +from xlayers import trunc_normal_ +from models.xcore import get_model + + +class HyperNet(super_core.SuperModule): + """The hyper-network.""" + + def __init__( + self, shape_container, layer_embeding, task_embedding, return_container=True + ): + super(HyperNet, self).__init__() + self._shape_container = shape_container + self._num_layers = len(shape_container) + self._numel_per_layer = [] + for ilayer in range(self._num_layers): + self._numel_per_layer.append(shape_container[ilayer].numel()) + + self.register_parameter( + "_super_layer_embed", + torch.nn.Parameter(torch.Tensor(self._num_layers, layer_embeding)), + ) + trunc_normal_(self._super_layer_embed, std=0.02) + + model_kwargs = dict( + input_dim=layer_embeding + task_embedding, + output_dim=max(self._numel_per_layer), + hidden_dim=layer_embeding * 4, + act_cls="sigmoid", + norm_cls="identity", + ) + self._generator = get_model(dict(model_type="simple_mlp"), **model_kwargs) + self._return_container = return_container + print("generator: {:}".format(self._generator)) + + def forward_raw(self, task_embed): + task_embed = task_embed.view(1, -1).expand(self._num_layers, -1) + joint_embed = torch.cat((task_embed, self._super_layer_embed), dim=-1) + weights = self._generator(joint_embed) + if self._return_container: + weights = torch.split(weights, 1) + return self._shape_container.translate(weights) + else: + return weights + + def forward_candidate(self, input): + raise NotImplementedError + + def extra_repr(self) -> str: + return "(_super_layer_embed): {:}".format(list(self._super_layer_embed.shape)) + + +class HyperNet_VX(super_core.SuperModule): + def __init__(self, shape_container, input_embeding, return_container=True): + super(HyperNet_VX, self).__init__() + self._shape_container = shape_container + self._num_layers = len(shape_container) + self._numel_per_layer = [] + for ilayer in range(self._num_layers): + self._numel_per_layer.append(shape_container[ilayer].numel()) + + self.register_parameter( + "_super_layer_embed", + torch.nn.Parameter(torch.Tensor(self._num_layers, input_embeding)), + ) + trunc_normal_(self._super_layer_embed, std=0.02) + + model_kwargs = dict( + input_dim=input_embeding, + output_dim=max(self._numel_per_layer), + hidden_dim=input_embeding * 4, + act_cls="sigmoid", + norm_cls="identity", + ) + self._generator = get_model(dict(model_type="simple_mlp"), **model_kwargs) + self._return_container = return_container + print("generator: {:}".format(self._generator)) + + def forward_raw(self, input): + weights = self._generator(self._super_layer_embed) + if self._return_container: + weights = torch.split(weights, 1) + return self._shape_container.translate(weights) + else: + return weights + + def forward_candidate(self, input): + raise NotImplementedError + + def extra_repr(self) -> str: + return "(_super_layer_embed): {:}".format(list(self._super_layer_embed.shape)) diff --git a/lib/xlayers/super_core.py b/lib/xlayers/super_core.py index 8b9d3fd..a6564d4 100644 --- a/lib/xlayers/super_core.py +++ b/lib/xlayers/super_core.py @@ -41,4 +41,5 @@ super_name2activation = { from .super_trade_stem import SuperAlphaEBDv1 +from .super_positional_embedding import SuperDynamicPositionE from .super_positional_embedding import SuperPositionalEncoder diff --git a/lib/xlayers/super_positional_embedding.py b/lib/xlayers/super_positional_embedding.py index 9205990..c69ed86 100644 --- a/lib/xlayers/super_positional_embedding.py +++ b/lib/xlayers/super_positional_embedding.py @@ -10,6 +10,41 @@ from .super_module import SuperModule from .super_module import IntSpaceType +class SuperDynamicPositionE(SuperModule): + """Applies a positional encoding to the input positions.""" + + def __init__(self, dimension: int, scale: float = 1.0) -> None: + super(SuperDynamicPositionE, self).__init__() + + self._scale = scale + self._dimension = dimension + # weights to be optimized + self.register_buffer( + "_div_term", + torch.exp( + torch.arange(0, dimension, 2).float() * (-math.log(10000.0) / dimension) + ), + ) + + @property + def abstract_search_space(self): + root_node = spaces.VirtualNode(id(self)) + return root_node + + def forward_candidate(self, input: torch.Tensor) -> torch.Tensor: + return self.forward_raw(input) + + def forward_raw(self, input: torch.Tensor) -> torch.Tensor: + import pdb + + pdb.set_trace() + print("---") + return F.linear(input, self._super_weight, self._super_bias) + + def extra_repr(self) -> str: + return "scale={:}, dim={:}".format(self._scale, self._dimension) + + class SuperPositionalEncoder(SuperModule): """Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf https://github.com/pytorch/examples/blob/master/word_language_model/model.py#L65