Route2Vec/linear_evaluation.py at master · philipph77/Route2Vec · GitHub

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
import os
from typing import Tuple
from time import time
from datetime import datetime
import json
import torch.nn as nn
import logging
import pytorch_lightning as pl
from pytorch_lightning.loggers import CSVLogger
from pytorch_lightning.loggers import TensorBoardLogger
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
from architecture import DenseClassifier
from architecture import TransformerEncoder
from datamodule import RoutesDataModule
from modules import Route2Vec

class linearEvaluation():
    def __init__(
        self,
        run_name: str,
        version: str,
        logdir: str,
        num_workers: int,
        lr: float,
        batch_size: int,
        max_epochs: int,
        tasks: dict,
    ) -> None:

        logging.getLogger("lightning").setLevel(logging.WARNING)
        pl.seed_everything(33)
        self.hparams = self._find_hyperparams(run_name, version)
        self.run_name = run_name + "_linear_evaluation"
        self.logdir = logdir
        self.num_workers = num_workers
        self.lr = lr
        self.batch_size = batch_size
        self.max_epochs = max_epochs
        self.tasks = tasks

        self.datapath = self.hparams['datapath']
        self.encoder = TransformerEncoder(
            self.hparams['enc_use_tokenizer'],
            self.hparams['enc_use_cls_token'],
            self.hparams['enc_pooling_type'],
            self.hparams['enc_pos_embeddings_alpha'],
            self.hparams['enc_input_features'],
            self.hparams['embedding_dim'],
            self.hparams['enc_n_head'],
            self.hparams['enc_depth'],
        )

        self.clas = nn.ModuleList()
        self.criteria = []
        self.loss_weights = []
        self.label_names = []
        for task_id, task_data in self.tasks.items():
            assert task_data["label_name"] in ["summed_travel_time", "summed_length", "mean_curvature", "abs_mean_grade", "roadType_share", "surface_share"]
            self.label_names.append(task_data["label_name"])
            self.clas.append(DenseClassifier(self.hparams['embedding_dim'], task_data["n_classes"]))
            self.criteria.append(task_data["criterion"])
            self.loss_weights.append(task_data["loss_weight"])

    @staticmethod
    def _find_hyperparams(runname, version):
        with open("runs_hparams.json", 'r') as f:
            buffer = json.load(f)
        buffer = [b for b in buffer if runname == b['run_name']]
        return buffer[version*4]

    def linear_evaluation(self) -> None:
        module = Route2Vec(
            self.encoder, self.clas, self.criteria, self.loss_weights, self.label_names, self.lr, self.batch_size
        )
        for param in module.encoder.parameters():
            param.requires_grad = False
        datamodule = RoutesDataModule(self.datapath, self.label_names, self.batch_size, self.num_workers)

        checkpoint_callback = ModelCheckpoint(
            monitor="val_loss", dirpath=f"{self.logdir}/checkpoints/{self.run_name}", save_last=True
        )
        csv_logger = CSVLogger(save_dir=f"{self.logdir}/csv/", name=self.run_name)
        print(csv_logger.log_dir) # necessary in ddp - else log_dir will point to a wrong version
        trainer = pl.Trainer(
            accelerator="auto",
            strategy="ddp_find_unused_parameters_false",
            default_root_dir=self.logdir,
            max_epochs=self.max_epochs,
            log_every_n_steps=1,
            callbacks=[
                checkpoint_callback,
            ],
            logger=[
                csv_logger,
                TensorBoardLogger(
                    save_dir=f"{self.logdir}/tb/",
                    name=self.run_name,
                    log_graph=False,
                    default_hp_metric=False,
                ),
            ],
            limit_train_batches=None,
            limit_val_batches=None,
        )

        trainer.fit(module, datamodule)

        with open(os.path.join(csv_logger.log_dir, "best_model_path.txt"), "w") as f:
            f.write(checkpoint_callback.best_model_path)


if __name__ == "__main__":
    for runname in ["smallest", "small", "medium", "large", "x-large"]:
        le = linearEvaluation(
            run_name = runname,
            version = 0,
            logdir = "logs_linear_evaluation",
            num_workers = 4,
            lr = 1e-4,
            batch_size = 32,
            max_epochs = 100,
            tasks = {
                0: {
                    "label_name": "summed_travel_time",
                    "n_classes": 1,
                    "criterion": nn.MSELoss(reduction="mean"),
                    "loss_weight": 0.2
                    },
                1: {
                    "label_name": "summed_length",
                    "n_classes": 1,
                    "criterion": nn.MSELoss(reduction="mean"),
                    "loss_weight": 0.2
                    },
                2: {
                    "label_name": "mean_curvature",
                    "n_classes": 1,
                    "criterion": nn.MSELoss(reduction="mean"),
                    "loss_weight": 0.2
                    },

                3: {
                    "label_name": "abs_mean_grade",
                    "n_classes": 1,
                    "criterion": nn.MSELoss(reduction="mean"),
                    "loss_weight": 0.2
                    },
                4: {
                    "label_name": "roadType_share",
                    "n_classes": 15,
                    "criterion": nn.CrossEntropyLoss(reduction="mean"),
                    "loss_weight": 0.2
                    },
                }
        )
        le.linear_evaluation()