DSensDp/train_multi.py at main · fmenat/DSensDp · 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
import yaml
import argparse
import os
import sys
import time
import gc
import copy
from pathlib import Path
import numpy as np
import pandas as pd

from code.training.learn_pipeline import MultiFusion_train
from code.training.utils import assign_multifusion_name, output_name, assign_labels_weights
from code.datasets.views_structure import Dataset_MultiView
from code.datasets.utils import create_dataloader, load_structure

def main_run(config_file, just_return_first_model=False):
    start_time = time.time()
    input_dir_folder = config_file["input_dir_folder"]
    output_dir_folder = config_file["output_dir_folder"]
    data_name = config_file["data_name"]
    runs_seed = config_file["experiment"].get("runs_seed", [])
    if len(runs_seed) == 0:
        runs = config_file["experiment"].get("runs", 1)
        runs_seed = [np.random.randint(50000) for _ in range(runs)]

    BS = config_file["training"]["batch_size"]
    if "loss_args" not in config_file["training"]:
        config_file["training"]["loss_args"] = {}
    if config_file.get("task_type", "").lower() == "classification":
        config_file["training"]["loss_args"]["name"] = "ce" if "name" not in config_file["training"]["loss_args"] else config_file["training"]["loss_args"]["name"]
    elif config_file.get("task_type", "").lower() == "multilabel":
        config_file["training"]["loss_args"]["name"] = "bce" if "name" not in config_file["training"]["loss_args"] else config_file["training"]["loss_args"]["name"]
    method_name = assign_multifusion_name(config_file["training"],config_file["method"], more_info_str=config_file.get("additional_method_name", ""))

    if "train" in data_name:
        data_views_tr = load_structure(input_dir_folder, data_name, load_memory=config_file.get("load_memory", False))
        data_views_tr.load_stats(input_dir_folder, data_name)
        data_views_te = load_structure(input_dir_folder, data_name.replace("train", "test"), load_memory=config_file.get("load_memory", False))
        data_views_te.load_stats(input_dir_folder, data_name)
        try:
            data_views_va = load_structure(input_dir_folder, data_name.replace("train", "val"), load_memory=config_file.get("load_memory", False))
            data_views_va.load_stats(input_dir_folder, data_name)
        except:
            data_views_va = data_views_te
            print("No validation set found")
        kfolds = 1
    else:
        data_views_tr = load_structure(input_dir_folder, data_name, load_memory=config_file.get("load_memory", False))
        data_views_tr.load_stats(input_dir_folder, data_name)
        kfolds = config_file["experiment"].get("kfolds", 2)

    metadata_r = {"epoch_runs":[], "full_prediction_time":[], "training_time":[], "best_score":[] }
    for r,r_seed in enumerate(runs_seed):
        np.random.seed(r_seed)
        if kfolds != 1:
            indexs_ = data_views_tr.get_all_identifiers()
            np.random.shuffle(indexs_)
            indexs_runs = np.array_split(indexs_, kfolds)
        for k in range(kfolds):
            print(f"******************************** Executing model on run {r+1} and kfold {k+1}")

            if kfolds != 1:
                data_views_tr.set_val_mask(indexs_runs[k])
                data_views_te = copy.deepcopy(data_views_tr)
                data_views_te.set_data_mode(train=False)
                data_views_va = data_views_te
            data_views_tr.set_additional_info(**config_file["experiment"].get("preprocess"))
            data_views_va.set_additional_info(**config_file["experiment"].get("preprocess"))
            data_views_te.set_additional_info(**config_file["experiment"].get("preprocess"))
            print(f"Training with {len(data_views_tr)} samples and validating on {len(data_views_te)}")

            if config_file.get("task_type", "").lower() in ["classification", "multilabel"]:
                assign_labels_weights(config_file, data_views_tr)

            start_aux = time.time()
            method, trainer = MultiFusion_train(data_views_tr, val_data=data_views_va,run_id=r,fold_id=k,method_name=method_name, **config_file)
            if just_return_first_model:
                return method
            metadata_r["training_time"].append(time.time()-start_aux)
            metadata_r["epoch_runs"].append(trainer.callbacks[0].stopped_epoch)
            metadata_r["best_score"].append(trainer.callbacks[0].best_score.cpu())
            print("Training done")

            #store original predictions to calculate error-difference
            pred_time_Start = time.time()
            outputs_te = method.transform(create_dataloader(data_views_te, batch_size=BS, train=False), out_norm=output_name(config_file["task_type"]), not_return_repre=True)
            metadata_r["full_prediction_time"].append(time.time()-pred_time_Start)
            data_save_te = Dataset_MultiView([outputs_te["prediction"]], identifiers=data_views_te.get_all_identifiers(), view_names=[f"out_run-{r:02d}_fold-{k:02d}"])
            data_save_te.save(f"{output_dir_folder}/pred/{data_name}/{method_name}", ind_views=True, xarray=False)

            for v in outputs_te:
                if ":prediction" in v and "aggregated" in outputs_te[v]:
                    print(f"Warning, overwritting based on auxiliar predictions from {v}")
                    data_save_te = Dataset_MultiView([outputs_te[v]["aggregated"]], identifiers=data_views_te.get_all_identifiers(), view_names=[f"out_run-{r:02d}_fold-{k:02d}"])
                    data_save_te.save(f"{output_dir_folder}/pred_aux/{data_name}/{method_name}", ind_views=True, xarray=False) #overwrite previous pred

            if config_file.get("args_forward") and config_file["args_forward"].get("list_testing_views"):
                for (test_views, percentages) in config_file["args_forward"].get("list_testing_views"):
                    for perc_missing in percentages:
                        print("Inference with the following views ",test_views, " and percentage missing ",perc_missing)
                        if "missing_method" in config_file["args_forward"]:
                            args_forward = {"inference_views":test_views, **{k:v for k,v in config_file["args_forward"].items() if k!= "list_testing_views"}}
                        else:
                            method.set_missing_info(None, **config_file["training"].get("missing_method", {}))
                            args_forward = {"inference_views":test_views, "missing_method": method.missing_method}

                        pred_time_Start = time.time()
                        outputs_te = method.transform(create_dataloader(data_views_te, batch_size=config_file['args_forward'].get("batch_size", BS), train=False), out_norm=output_name(config_file["task_type"]), args_forward=args_forward, perc_forward=perc_missing, not_return_repre=True)
                        if f"{'_'.join(test_views)}_{perc_missing*100:.0f}_prediction_time" not in metadata_r:
                            metadata_r[f"{'_'.join(test_views)}_{perc_missing*100:.0f}_prediction_time"] = []
                        metadata_r[f"{'_'.join(test_views)}_{perc_missing*100:.0f}_prediction_time"].append(time.time()-pred_time_Start)

                        aux_name = assign_multifusion_name(config_file["training"],config_file["method"], forward_views=test_views, perc=perc_missing,
                                                        more_info_str=config_file.get("additional_method_name", ""))
                        ## EXTRA -- PREDICTIONS ##
                        data_save_te = Dataset_MultiView([outputs_te["prediction"]], identifiers=data_views_te.get_all_identifiers(), view_names=[f"out_run-{r:02d}_fold-{k:02d}"])
                        data_save_te.save(f"{output_dir_folder}/pred/{data_name}/{aux_name}", ind_views=True, xarray=False)

                        for v in outputs_te:
                            if ":prediction" in v and "aggregated" in outputs_te[v]: #in case of multi-loss they are used as auxiliar for prediction
                                print(f"Warning, overwritting based on auxiliar predictions from {v}")
                                data_save_te = Dataset_MultiView([outputs_te[v]["aggregated"]], identifiers=data_views_te.get_all_identifiers(), view_names=[f"out_run-{r:02d}_fold-{k:02d}"])
                                data_save_te.save(f"{output_dir_folder}/pred/{data_name}/{aux_name}", ind_views=True, xarray=False) #overwrite previous pred
                        print(f"Fold {k+1}/{kfolds} of Run {r+1}/{len(runs_seed)} in {aux_name} finished...")
            print(f"Fold {k+1}/{kfolds} of Run {r+1}/{len(runs_seed)} in {method_name} finished...")
    Path(f"{output_dir_folder}/metadata/{data_name}/{method_name}").mkdir(parents=True, exist_ok=True)
    pd.DataFrame(metadata_r).to_csv(f"{output_dir_folder}/metadata/{data_name}/{method_name}/metadata_runs.csv")
    print("Epochs for %s runs on average for %.2f epochs +- %.3f"%(method_name,np.mean(metadata_r["epoch_runs"]),np.std(metadata_r["epoch_runs"])))
    print(f"Finished whole execution of {len(runs_seed)} runs in {time.time()-start_time:.2f} secs")
    return metadata_r


if __name__ == "__main__":
    arg_parser = argparse.ArgumentParser()
    arg_parser.add_argument(
        "--settings_file",
        "-s",
        action="store",
        dest="settings_file",
        required=True,
        type=str,
        help="path of the settings file",
    )
    args = arg_parser.parse_args()
    with open(args.settings_file) as fd:
        config_file = yaml.load(fd, Loader=yaml.SafeLoader)

    main_run(config_file)