From a2b1d0d2272a1b9ffa9a8db6a378f69da0a549d5 Mon Sep 17 00:00:00 2001 From: D-X-Y <280835372@qq.com> Date: Thu, 13 May 2021 15:32:44 +0800 Subject: [PATCH] Add more functions for synthetic env --- exps/LFNA/{ => backup}/lfna-fix-init.py | 0 exps/LFNA/{ => backup}/lfna-ttss-hpnet.py | 0 exps/LFNA/{ => backup}/lfna-v1.py | 0 exps/LFNA/lfna.py | 230 ++++++++++++++++++++++ exps/LFNA/lfna_models_v2.py | 72 +++++++ lib/datasets/synthetic_env.py | 4 + lib/datasets/synthetic_utils.py | 2 +- 7 files changed, 307 insertions(+), 1 deletion(-) rename exps/LFNA/{ => backup}/lfna-fix-init.py (100%) rename exps/LFNA/{ => backup}/lfna-ttss-hpnet.py (100%) rename exps/LFNA/{ => backup}/lfna-v1.py (100%) create mode 100644 exps/LFNA/lfna.py create mode 100644 exps/LFNA/lfna_models_v2.py diff --git a/exps/LFNA/lfna-fix-init.py b/exps/LFNA/backup/lfna-fix-init.py similarity index 100% rename from exps/LFNA/lfna-fix-init.py rename to exps/LFNA/backup/lfna-fix-init.py diff --git a/exps/LFNA/lfna-ttss-hpnet.py b/exps/LFNA/backup/lfna-ttss-hpnet.py similarity index 100% rename from exps/LFNA/lfna-ttss-hpnet.py rename to exps/LFNA/backup/lfna-ttss-hpnet.py diff --git a/exps/LFNA/lfna-v1.py b/exps/LFNA/backup/lfna-v1.py similarity index 100% rename from exps/LFNA/lfna-v1.py rename to exps/LFNA/backup/lfna-v1.py diff --git a/exps/LFNA/lfna.py b/exps/LFNA/lfna.py new file mode 100644 index 0000000..8e9cfae --- /dev/null +++ b/exps/LFNA/lfna.py @@ -0,0 +1,230 @@ +##################################################### +# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 # +##################################################### +# python exps/LFNA/lfna.py --env_version v1 --hidden_dim 16 --layer_dim 32 --epochs 50000 +##################################################### +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_v2 import HyperNet + + +def main(args): + logger, env_info, model_kwargs = lfna_setup(args) + dynamic_env = env_info["dynamic_env"] + model = get_model(**model_kwargs) + model = model.to(args.device) + criterion = torch.nn.MSELoss() + + logger.log("There are {:} weights.".format(model.get_w_container().numel())) + # meta_train_range = (dynamic_env.min_timestamp, (dynamic_env.min_timestamp + dynamic_env.max_timestamp) / 2) + # meta_train_interval = dynamic_env.timestamp_interval + + shape_container = model.get_w_container().to_shape_container() + + # pre-train the hypernetwork + timestamps = list( + dynamic_env.get_timestamp(index) for index in range(len(dynamic_env) // 2) + ) + + hypernet = HyperNet(shape_container, args.layer_dim, args.task_dim, timestamps) + hypernet = hypernet.to(args.device) + + import pdb + + pdb.set_trace() + + # task_embed = torch.nn.Parameter(torch.Tensor(env_info["total"], args.task_dim)) + total_bar = 16 + task_embeds = [] + for i in range(total_bar): + tensor = torch.Tensor(1, args.task_dim).to(args.device) + task_embeds.append(torch.nn.Parameter(tensor)) + for task_embed in task_embeds: + trunc_normal_(task_embed, std=0.02) + + model.train() + hypernet.train() + + parameters = list(hypernet.parameters()) + task_embeds + # optimizer = torch.optim.Adam(parameters, lr=args.init_lr, amsgrad=True) + optimizer = torch.optim.Adam(parameters, lr=args.init_lr, weight_decay=1e-5) + 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 + # 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_embeds[cur_time] + cur_container = hypernet(cur_task_embed) + cur_x = env_info["{:}-x".format(cur_time)].to(args.device) + cur_y = env_info["{:}-y".format(cur_time)].to(args.device) + 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 % 100 == 0: + logger.log( + head_str + + " meta-loss: {:.4f} ({:.4f}) :: lr={:.5f}, batch={:}".format( + loss_meter.avg, + loss_meter.val, + min(lr_scheduler.get_last_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) + + w_container_per_epoch = dict() + for idx in range(0, total_bar): + future_time = env_info["{:}-timestamp".format(idx)] + future_x = env_info["{:}-x".format(idx)] + future_y = env_info["{:}-y".format(idx)] + future_container = hypernet(task_embeds[idx]) + w_container_per_epoch[idx] = future_container.no_grad_clone() + with torch.no_grad(): + future_y_hat = model.forward_with_container( + future_x, w_container_per_epoch[idx] + ) + future_loss = 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(".") + parser.add_argument( + "--save_dir", + type=str, + default="./outputs/lfna-synthetic/lfna-battle", + 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( + "--layer_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.", + ) + parser.add_argument( + "--device", + type=str, + default="cpu", + help="", + ) + # 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.layer_dim + args.save_dir = "{:}-{:}-d{:}".format( + args.save_dir, args.env_version, args.hidden_dim + ) + main(args) diff --git a/exps/LFNA/lfna_models_v2.py b/exps/LFNA/lfna_models_v2.py new file mode 100644 index 0000000..8cdbe97 --- /dev/null +++ b/exps/LFNA/lfna_models_v2.py @@ -0,0 +1,72 @@ +##################################################### +# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.04 # +##################################################### +import copy +import torch + +import torch.nn.functional as F + +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, + meta_timestamps, + return_container: bool = 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( + config=dict(model_type="dual_norm_mlp"), + input_dim=layer_embeding + task_embedding, + output_dim=max(self._numel_per_layer), + hidden_dims=[layer_embeding * 4] * 3, + act_cls="gelu", + norm_cls="layer_norm_1d", + dropout=0.1, + ) + import pdb + + pdb.set_trace() + self._generator = get_model(**model_kwargs) + self._return_container = return_container + print("generator: {:}".format(self._generator)) + + def forward_raw(self, task_embed): + # task_embed = F.normalize(task_embed, dim=-1, p=2) + # layer_embed = F.normalize(self._super_layer_embed, dim=-1, p=2) + layer_embed = self._super_layer_embed + task_embed = task_embed.view(1, -1).expand(self._num_layers, -1) + + joint_embed = torch.cat((task_embed, 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)) diff --git a/lib/datasets/synthetic_env.py b/lib/datasets/synthetic_env.py index 6c2c6ed..506e1f2 100644 --- a/lib/datasets/synthetic_env.py +++ b/lib/datasets/synthetic_env.py @@ -55,6 +55,10 @@ class SyntheticDEnv(data.Dataset): def timestamp_interval(self): return self._timestamp_generator.interval + def get_timestamp(self, index): + index, timestamp = self._timestamp_generator[index] + return timestamp + def set_oracle_map(self, functor): self._oracle_map = functor diff --git a/lib/datasets/synthetic_utils.py b/lib/datasets/synthetic_utils.py index 7c95d4b..93e7b2b 100644 --- a/lib/datasets/synthetic_utils.py +++ b/lib/datasets/synthetic_utils.py @@ -60,7 +60,7 @@ class TimeStamp(UnifiedSplit, data.Dataset): @property def max_timestamp(self): return self._max_timestamp - + @property def interval(self): return self._interval