import os,time import torch,wandb from tqdm import trange from einops import rearrange import vis,geometry from copy import deepcopy import numpy as np import piqa,kornia from torchvision.utils import make_grid from einops import rearrange, repeat from models import ch_sec from torch.cuda.amp import autocast, GradScaler import torch.nn.functional as F import getpass from glob import glob import data,models def to_gpu(ob): return {k: to_gpu(torch.tensor(v)) for k, v in ob.items()} if isinstance(ob, dict) else ob.cuda() def train(run,until_img=25,until_vid=100,until_save=500,optim=None,single_data=None): def loss_fn(model_out, gt, data_sample,model,step,args): rays = lambda x,y: torch.stack([x[i,:,y[i].long()] for i in range(len(x))]) #losses = { "metrics/action_loss" : (data_sample["joint_state" if not run.args.endeffector_action else "end_link"] - model_out["action_state"]).square().mean()*1e1 } losses = { "metrics/action_loss" : (data_sample["joint_state"] - model_out["action_state"]).square().mean()*1e1 } print(losses) return losses losses_agg=[] optim = torch.optim.Adam(lr=run.args.lr, params=run.model.parameters()) dataset=data.TmpTraj(run.args.dirpath) dataloader = iter(torch.utils.data.DataLoader(dataset, batch_size=run.args.batch_size, num_workers=min(run.args.n_workers,run.args.batch_size),pin_memory=True)) # Train loop step=0 for step_ in trange(run.args.n_train_steps, desc="Fitting"): # train until user interruption overfit=0 # Get data if overfit:print("overfitting on sample") if step_==0 or not overfit: data_sample=to_gpu(next(dataloader)) # Run model and calculate losses total_loss = 0. out=run.model(data_sample) losses = loss_fn(out, data_sample, data_sample,run.model,step,run.args) for loss_name, loss in losses.items(): wandb.log({loss_name: loss.item()}, step=step) total_loss += loss wandb.log({"loss": total_loss.item()}, step=step) # Optim step total_loss.backward();optim.step();optim.zero_grad(); # Vis/save with torch.no_grad(): wandb_imgs=None if step%until_img==0: wandb_imgs=vis.wandb_summary( 0, out, data_sample, data_sample, None,step=step) if step%until_save == 0 and step and run.args.save_model: # save model print(f"Saving to {run.save_dir}"); torch.save({ 'step': step, 'model_state_dict': run.model.state_dict(), }, os.path.join(run.save_dir, f"checkpoint.pt")) step+=1 # Data/args setup and run import argparse parser = argparse.ArgumentParser(description='simple training job') # logging parameters parser.add_argument('-n','--name', type=str,default="",required=False,help="wandb training name") parser.add_argument('-c','--init_ckpt', type=str,default=None,required=False,help="File for checkpoint loading. If folder specific, will use latest .pt file") parser.add_argument('-o','--online', default=False, action='store_true') parser.add_argument('-s','--save_model', default=True, action='store_true') parser.add_argument('--viser', default=False, action='store_true') parser.add_argument('--save_opt_vis', default=False, action='store_true') # data/training parameters parser.add_argument('-d','--dataset', type=str,default="hydrant") parser.add_argument('--imgpath', type=str,default="") parser.add_argument('--dirpath', type=str,default="/data/cameron/rotation_testing/duck_dataset/")#rotation_gripper_dataset/") parser.add_argument('-b','--batch_size', type=int,default=1,help="number of videos/sequences per training step") parser.add_argument('-v','--vid_len', type=int,default=6,help="video length or number of images per batch") parser.add_argument('--n_workers',type=int,default=4,help="number of workers per dataloader") parser.add_argument('--until_save',type=int,default=500,help="number of steps until model save") parser.add_argument('--lr',type=float,default=1e-4,help="learning rate") parser.add_argument('--n_train_steps',type=int,default=10000,help="n train steps ") parser.add_argument('--overfit', default=True, action='store_true',help="Whether to overfit on a single scene") parser.add_argument('--until_img', type=int,default=50,help="Number of steps until image summary. ") parser.add_argument('--sf', type=float,default=1,help="Image resolution scale factor (fractional is cheaper)") parser.add_argument('--load_save', default=False, action='store_true',help="Whether to load the previously saved data if overfitting (to avoid running flow again)") # model parameters parser.add_argument('--cam_pointmap_inp', default=False, action='store_true',help="use camera space pointmap img") parser.add_argument('--rob_pointmap_inp', default=False, action='store_true',help="use robot space pointmap img") parser.add_argument('--cam_pointcloud_inp', default=False, action='store_true',help="use camera space pointcloud") parser.add_argument('--rob_pointcloud_inp', default=False, action='store_true',help="use robot space pointcloud") parser.add_argument('--rob_rerender_inp', default=False, action='store_true',help="use robot space re-render image") parser.add_argument('--endeffector_action', default=False, action='store_true',help="use endeffector prediction instead of joint angle prediction") def make_run(args=None,val=False): args = parser.parse_args(args) self = argparse.Namespace() user = getpass.getuser() print(f"user={user}") # Wandb init run = wandb.init(entity="cameronsmithbusiness",project="biasing",mode="online" if args.online else "disabled",name=args.name,dir=f"/tmp/wandb") wandb.run.log_code(".") self.save_dir = "/tmp/"+args.name#os.path.join(os.environ.get('LOGDIR', "") , run.name) os.makedirs(self.save_dir,exist_ok=True) wandb.save(os.path.join(self.save_dir, "checkpoint*")) wandb.save(os.path.join(self.save_dir, "video*")) self.args=args self.wandb=run # Make model and load checkpoint self.model = (models.PolicyModel)(args).cuda() if args.init_ckpt is not None: ckpt_file = args.init_ckpt if os.path.isfile(os.path.expanduser(args.init_ckpt)) else max(glob(os.path.join(args.init_ckpt,"*.pt")), key=os.path.getctime) self.model.load_state_dict(torch.load(ckpt_file)["model_state_dict"],strict=False) return self run = make_run() torch.autograd.set_detect_anomaly(False) train(run,until_save=run.args.until_save, until_vid=100 if not run.args.overfit else 300, until_img=run.args.until_img)