inatGeoModelTraining/make_thresholds.py at main · inaturalist/inatGeoModelTraining · 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
158
159
160
161
162
163
164
"""
make_thresholds.py

Generates per-taxon threshold values for a trained geospatial model using one of several thresholding strategies.

Usage:
    python make_thresholds.py \
        --train_config_file configs/config_train_inat_2_21.yaml \
        --thresholding_strategy lpt_r \
        --lpt_r_percentile 5 \
        --output_path results/thresholds_inat_2_21_lpt_r.csv

Supported strategies:
    - lpt_r (requires --lpt_r_percentile)
    - (TODO: lpt, scott_pa, fixed)

Outputs:
    CSV with columns: taxon_id, threshold_type, threshold_value
"""
import click
import h3
import h3pandas
import numpy as np
import pandas as pd
from tqdm.auto import tqdm
import yaml

from lib.models.tf_gp_model import TFGeoPriorModel
from lib.models.encoders import CoordEncoder

@click.command()
@click.option("--train_config_file", type=click.Path(exists=True), required=True, help="Path to training config YAML file.")
@click.option("--thresholding_strategy", type=str, required=True, help="Thresholding method to use (lpt_r, lpt, scott_pa, fixed).")
@click.option("--fixed_threshold", type=float, required=False, help="Threshold value if using 'fixed' strategy.")
@click.option("--lpt_r_percentile", type=int, required=False, help="Percentile value for LPT-R (e.g., 5 for 5th percentile).")
@click.option("--output_path", type=click.Path(), required=True, help="Path to output CSV file.")
@click.option("--stop_after", type=int, required=False, help="Optional: stop after evaluating N taxa (for debugging).")
def make_thresholds(train_config_file, thresholding_strategy, fixed_threshold, lpt_r_percentile, output_path, stop_after):
    valid_thresholding_strats = ["lpt_r", "lpt", "scott_pa", "fixed"]

    assert thresholding_strategy in valid_thresholding_strats, f"unsupported thresholding strategy {thresholding_strategy}, supported techniques are {valid_thresholding_strats}"

    if thresholding_strategy == "lpt_r":
        assert lpt_r_percentile is not None, "if using lpt-r, a percentile is required"

    with open(train_config_file, "r") as f:
        train_config = yaml.safe_load(f)

    # load model & basics
    print("loading model, encoder, taxonomy...")
    model = TFGeoPriorModel(train_config.get("model_save_name"))

    if train_config.get("inputs").get("covariates") == "env":
        assert train_config.get("bioclim_data") is not None, "if using env, need bioclim data"
        raster = np.load(train_config.get("bioclim_data")).astype(np.float32)
    elif train_config.get("inputs").get("covariates") == "elev":
        assert train_config.get("elev_data") is not None, "if using elev, need elevation data"
        raster = np.load(train_config.get("elev_data")).astype(np.float32)
    else:
        raster = None
    encoder = CoordEncoder(train_config.get("inputs").get("loc_feat_encoding"), raster)

    if train_config.get("dataset_type") == "inat":
        train_dataset = train_config.get("inat_dataset")
    elif train_config.get("dataset_type") == "sinr":
        train_dataset = train_config.get("sinr_dataset")
    else:
        raise NotImplementedError(f"{train_config.get('dataset_type')} dataset type not implemented yet.")

    tax = pd.read_csv(train_dataset.get("taxonomy_data"))

    # load training data
    if thresholding_strategy in ["lpt", "lpt_r", "scott_pa"]:
        print("loading spatial training data...")
        spatial_train = pd.read_parquet(train_dataset.get("train_data"))
        if train_config.get("dataset_type") == "sinr":
            # we need to manually add the leaf_class_id column in spatial_train
            print("adding leaf_class_id to sinr style spatial data dataset")
            spatial_train = pd.merge(
                spatial_train,
                tax[["taxon_id", "leaf_class_id"]],
                left_on="taxon_id",
                right_on="taxon_id",
                how="left",
            )

        print("discretizing spatial training data at resolution 4...")
        spatial_train_h3 = spatial_train.h3.geo_to_h3(
            resolution=4,
            lat_col="latitude",
            lng_col="longitude",
        )
        spatial_train_h3.reset_index(inplace=True)
        spatial_train_h3.set_index("leaf_class_id", inplace=True)

    # generate geo features for all cells at resolution 4 for evaluation
    print("make geo features for all h3 cells at r4")
    res0_cells = h3.get_res0_indexes()
    all_cells = set()
    for res0 in tqdm(res0_cells, dynamic_ncols=True):
        children = h3.h3_to_children(res0, 4)
        all_cells.update(children)
    cells_df = pd.DataFrame({"h3_04": list(all_cells)})
    cells_df.set_index("h3_04", inplace=True)
    dfh3 = cells_df.h3.h3_to_geo()
    dfh3["lng"] = dfh3.geometry.x
    dfh3["lat"] = dfh3.geometry.y
    _ = dfh3.pop("geometry")
    dfh3.reset_index(inplace=True)

    locs = dfh3[["lng", "lat"]].values
    encoded_locs = encoder.encode(locs)
    geo_model_feats = model.get_loc_emb(encoded_locs)

    # make thresholds for all taxa in the model
    print("making thresholds...")
    leaf_tax = tax[~tax.leaf_class_id.isna()]
    thresholds = []
    for i in tqdm(range(len(leaf_tax)), dynamic_ncols=True):
        leaf_taxon = leaf_tax.iloc[i]
        target_taxon_id = int(leaf_taxon.taxon_id)
        target_class_id = int(leaf_taxon.leaf_class_id)

        preds = model.eval_one_class_from_feats(
            geo_model_feats,
            class_of_interest=target_class_id
        )

        if thresholding_strategy == "lpt_r":
            assert lpt_r_percentile is not None, "lpt_r_percentile must be provided for 'lpt_r' strategy"

            dfh3["preds"] = preds[0]
            spatial_train_h3_target_taxon = spatial_train_h3.loc[target_class_id]
            presence_preds = dfh3[dfh3["h3_04"].isin(spatial_train_h3_target_taxon["h3_04"])]
            lpt_r_threshold = np.percentile(presence_preds.preds, lpt_r_percentile)
            thresholds.append({
                "taxon_id": target_taxon_id,
                "threshold_type": f"lpt_r_at_{lpt_r_percentile}_pct",
                "threshold_value": lpt_r_threshold,
            })
        elif thresholding_strategy == "fixed":
            assert fixed_threshold is not None, "fixed_threshold must be provided for 'fixed' strategy"
            thresholds.append({
                "taxon_id": target_taxon_id,
                "threshold_type": f"fixed_{fixed_threshold}",
                "threshold_value": fixed_threshold,
            })
        elif thresholding_strategy in ["lpt", "scott_pa"]:
            raise NotImplementedError(f"{thresholding_strategy} strategy not implemented yet.")

        if stop_after is not None and i >= stop_after:
            break

    print("saving thresholds...")
    thresholds_df = pd.DataFrame(thresholds)
    thresholds_df.to_csv(
        output_path, index=False
    )


if __name__ == "__main__":
    make_thresholds()