File size: 9,810 Bytes
b300acf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
"""Fit stable target clusters, NumPy RF, and neural classifiers."""

import argparse
import json
import os
import random
import sys
from pathlib import Path

import numpy as np
import torch
import yaml
from torch import nn
from torch.nn.parallel import DistributedDataParallel
from torch.utils.data import DataLoader, TensorDataset
from torch.utils.data.distributed import DistributedSampler


ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))
from model.cesm_seasonal_ml import (CHECKPOINT_FORMAT_VERSION, DATA_FORMAT_VERSION, FeedForwardNN,
                                    NumpyGradientBoostedTrees, NumpyRandomForest, SeasonalLSTM,
                                    StablePrecipKMeans, classification_metrics)


def setup(config):
    distributed = int(os.environ.get("WORLD_SIZE", "1")) > 1
    local_world = int(os.environ.get("LOCAL_WORLD_SIZE", "1"))
    use_cuda = (torch.cuda.is_available() and config["runtime"]["device"] != "cpu"
                and torch.cuda.device_count() >= local_world)
    if distributed:
        torch.distributed.init_process_group(config["runtime"]["ddp_backend_gpu"] if use_cuda else config["runtime"]["ddp_backend_cpu"])
    rank = torch.distributed.get_rank() if distributed else 0
    device = torch.device(f"cuda:{int(os.environ.get('LOCAL_RANK', 0))}" if use_cuda else "cpu")
    if use_cuda:
        torch.cuda.set_device(device)
    return distributed, rank, device


def train_neural(name, model, x_train, y_train, x_val, y_val, settings, distributed, device):
    dataset = TensorDataset(torch.from_numpy(x_train).float(), torch.from_numpy(y_train).long())
    sampler = DistributedSampler(dataset, shuffle=True) if distributed else None
    local_samples = len(sampler) if sampler is not None else len(dataset)
    loader = DataLoader(dataset, batch_size=settings["batch_size"], shuffle=sampler is None,
                        sampler=sampler, drop_last=name == "nn" and local_samples % settings["batch_size"] == 1)
    model = model.to(device)
    wrapped = DistributedDataParallel(model, device_ids=[device.index] if device.type == "cuda" else None) if distributed else model
    counts = np.bincount(y_train, minlength=4)
    weights = len(y_train) / np.maximum(4 * counts, 1)
    criterion = nn.CrossEntropyLoss(weight=torch.tensor(weights, dtype=torch.float32, device=device))
    optimizer = torch.optim.Adam(wrapped.parameters(), lr=settings["learning_rate"])
    history, best_accuracy, best_state = [], -1.0, None
    for epoch in range(settings["epochs"]):
        if sampler:
            sampler.set_epoch(epoch)
        wrapped.train()
        total_loss, total_samples = 0.0, 0
        for features, target in loader:
            optimizer.zero_grad(set_to_none=True)
            loss = criterion(wrapped(features.to(device)), target.to(device))
            loss.backward()
            optimizer.step()
            total_loss += float(loss.detach()) * len(features)
            total_samples += len(features)
        loss_summary = torch.tensor([total_loss, total_samples], dtype=torch.float64, device=device)
        if distributed:
            torch.distributed.all_reduce(loss_summary)
        wrapped.eval()
        with torch.no_grad():
            logits = wrapped(torch.from_numpy(x_val).float().to(device))
            correct = (logits.argmax(1).cpu().numpy() == y_val).sum()
            count = len(y_val)
            if distributed:
                summary = torch.tensor([correct, count], dtype=torch.float64, device=device)
                torch.distributed.all_reduce(summary)
                correct, count = summary.tolist()
            accuracy = correct / count
        history.append({"epoch": epoch + 1, "loss": float(loss_summary[0] / loss_summary[1]), "validation_accuracy": accuracy})
        if accuracy > best_accuracy:
            best_accuracy = accuracy
            module = wrapped.module if distributed else wrapped
            best_state = {key: value.detach().cpu().clone() for key, value in module.state_dict().items()}
    return best_state, history


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--season", choices=["NDJ", "JFM"])
    parser.add_argument("--models", nargs="+", choices=["rf", "nn", "lstm", "xgboost"])
    args = parser.parse_args()
    config = yaml.safe_load((ROOT / "conf/config.yaml").read_text())
    season = args.season or config["season"]
    enabled = args.models or config["model"]["enabled"]
    seed = int(config["seed"])
    random.seed(seed); np.random.seed(seed); torch.manual_seed(seed)
    distributed, rank, device = setup(config)
    data = np.load(ROOT / config["data"]["path"])
    if str(data["format_version"]) != DATA_FORMAT_VERSION:
        raise ValueError("data format version mismatch")
    season_index = [str(x) for x in data["seasons"]].index(season)
    train, validation, test = data["split"] == 0, data["split"] == 1, data["split"] == 2
    cluster = StablePrecipKMeans(seed=seed).fit(data["precipitation"][season_index, train])
    labels = cluster.transform(data["precipitation"][season_index])
    models, histories, validation_probabilities = {}, {}, {}
    if "rf" in enabled:
        rf_config = config["model"]["rf"]
        forest = NumpyRandomForest(rf_config["trees"], rf_config["mtry"], rf_config["max_depth"], rf_config["min_samples_split"], seed)
        if rank == 0:
            forest.fit(data["rf_features"][season_index, train], labels[train])
            validation_probabilities["rf"] = forest.predict_proba(data["rf_features"][season_index, validation])
            importance = forest.permutation_importance(data["rf_features"][season_index, validation], labels[validation], rf_config["permutation_repeats"])
            mean_minimum_depth, root_frequency = forest.tree_structure_importance()
            models["rf"] = forest
            histories["rf"] = {"permutation_importance": importance.tolist(), "mean_minimum_depth": mean_minimum_depth.tolist(),
                               "root_frequency": root_frequency.tolist(), "feature_names": data["rf_manifest"].tolist()}
    if "xgboost" in enabled and rank == 0:
        xgb_config = config["model"]["xgboost"]
        boosted = NumpyGradientBoostedTrees(xgb_config["n_rounds"], xgb_config["max_depth"],
                                            xgb_config["eta"], xgb_config["gamma"], seed)
        boosted.fit(data["rf_features"][season_index, train], labels[train])
        models["xgboost"] = boosted
        validation_probabilities["xgboost"] = boosted.predict_proba(data["rf_features"][season_index, validation])
        histories["xgboost"] = {"rounds": xgb_config["n_rounds"]}
    history_length = 4 if season == "NDJ" else 12
    specifications = {
        "nn": (FeedForwardNN(hidden=tuple(config["model"]["nn"]["hidden"]), dropout=config["model"]["nn"]["dropout"]), data["nn_features"][season_index], config["model"]["nn"]),
        "lstm": (SeasonalLSTM(hidden_size=config["model"]["lstm"]["hidden_size"], dense_size=config["model"]["lstm"]["dense_size"], dropout=config["model"]["lstm"]["dropout"]), data["eof_sequence"][season_index, :, -history_length:], config["model"]["lstm"])
    }
    for name in ("nn", "lstm"):
        if name in enabled:
            model, features, settings = specifications[name]
            state, history = train_neural(name, model, features[train], labels[train], features[validation], labels[validation], settings, distributed, device)
            model.load_state_dict(state); model.to(device).eval()
            with torch.no_grad():
                probability = torch.softmax(model(torch.from_numpy(features[validation]).float().to(device)), dim=1).cpu().numpy()
            models[name] = state
            histories[name] = history
            validation_probabilities[name] = probability
    if rank == 0:
        metrics = {name: classification_metrics(labels[validation], probability, config["evaluation"]["random_trials"], seed) for name, probability in validation_probabilities.items()}
        model_config = {"enabled": list(enabled), "primary": config["model"]["primary"], "settings": config["model"],
                        "history_length": history_length, "classes": 4, "rf_features": 103,
                        "nn_features": 416, "eof_channels": 28}
        data_spec = {"format_version": str(data["format_version"]), "season": season,
                     "seasons": data["seasons"].tolist(), "rf_manifest": data["rf_manifest"].tolist(),
                     "nn_manifest": data["nn_manifest"].tolist(), "rf_shape": list(data["rf_features"].shape),
                     "nn_shape": list(data["nn_features"].shape), "eof_shape": list(data["eof_sequence"].shape),
                     "precipitation_shape": list(data["precipitation"].shape), "class_names": data["class_names"].tolist()}
        checkpoint = {"format_version": CHECKPOINT_FORMAT_VERSION, "model": models, "model_config": model_config,
                      "data_spec": data_spec, "season": season, "cluster": cluster.state_dict(),
                      "training_history": histories, "validation_metrics": metrics, "paper_model": config["paper_model"]}
        path = ROOT / config["paths"]["checkpoint"]
        path.parent.mkdir(parents=True, exist_ok=True)
        torch.save(checkpoint, path)
        metrics_path = ROOT / config["paths"]["training_metrics"]
        metrics_path.parent.mkdir(parents=True, exist_ok=True)
        metrics_path.write_text(json.dumps({"season": season, "models": enabled, "validation": metrics}, indent=2) + "\n")
        print(f"checkpoint={path.relative_to(ROOT)} season={season} models={','.join(enabled)}")
    if distributed:
        torch.distributed.barrier(); torch.distributed.destroy_process_group()


if __name__ == "__main__":
    main()