train.py

import random
from collections import defaultdict

import numpy as np
import torch
import torch_geometric
from sacred import Experiment
from sacred.observers import FileStorageObserver
from tqdm import trange

from datasets import load_temporal_data
from model.tgne import TGNE

ex = Experiment("tgne")
ex.observers.append(FileStorageObserver.create("my_runs"))


def set_seed(seed):
    print("Setting seed to", seed)
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)

    torch.cuda.manual_seed_all(seed)
    torch_geometric.seed_everything(seed)


@ex.config
def config():
    dataset = "toy"
    seed = 42
    dim = 2
    prior_scale_init = None
    prior_scale = 1.0
    n_ticks = 15
    n_epochs = 200
    lr_z = 0.1
    lr_bias = 0.001
    cuda = False
    batch_size = None


@ex.automain
def train(
    dataset,
    dim,
    prior_scale_init,
    prior_scale,
    n_ticks,
    n_epochs,
    lr_z,
    lr_bias,
    cuda,
    batch_size,
):
    temporal_data = load_temporal_data(dataset)
    set_seed(42)
    save = True
    import sys

    if "-d" in sys.argv or "--debug" in sys.argv:
        print("Debug mode")
        n_epochs = 1
        save = False
    tgne = TGNE(
        n_nodes=temporal_data.num_nodes,
        n_ticks=n_ticks,
        dim=dim,
        prior_scale=prior_scale,
        prior_scale_init=prior_scale if prior_scale_init is None else prior_scale_init,
        cuda=cuda,
        lr_bias=lr_bias,
        lr_z=lr_z,
    )
    train_data = temporal_data
    train_loader = torch.utils.data.DataLoader(
        train_data,
        batch_size=len(train_data) if not batch_size else batch_size,
        collate_fn=train_data.collate,
    )

    steps = trange(n_epochs)

    losses = []
    grads = defaultdict(list)
    for step in steps:
        for batch in train_loader:
            batch_logs = tgne.training_step(batch)
        loss = batch_logs["loss"]
        for param_name, grad in batch_logs["grads"].items():
            grads[param_name].append(grad)

        losses.append(batch_logs["loss"])
        steps.set_description("loss: {:.4f}".format(loss))
        ex.log_scalar("loss", loss, step)
        for param_name, grad in batch_logs["grads"].items():
            ex.log_scalar(f"grad_{param_name}", grad, step)

    out = {
        "state_dict": tgne.state_dict(),
        "losses": losses,
        "grads": grads,
    }

    if save:
        fname = f"tgne_{dataset}_{n_epochs}.ckpt"
        outpath = f"output/{fname}"
        torch.save(out, outpath)
        # This will copy and paste the trained model to the run-specific folder
        ex.add_artifact(outpath, name=fname)