-
Notifications
You must be signed in to change notification settings - Fork 25
/
analyze_latent.py
271 lines (238 loc) · 10.3 KB
/
analyze_latent.py
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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
__all__ = ["analyze_latent"]
import re
from pathlib import Path
from typing import Sized, cast
import hydra
import numpy as np
import pandas as pd
import torch
from sklearn.base import TransformerMixin
import move.visualization as viz
from move.analysis.metrics import (
calculate_accuracy,
calculate_cosine_similarity,
)
from move.conf.schema import AnalyzeLatentConfig, MOVEConfig
from move.core.logging import get_logger
from move.core.typing import FloatArray
from move.data import io
from move.data.dataloaders import MOVEDataset, make_dataloader
from move.data.perturbations import (
perturb_categorical_data,
perturb_continuous_data,
)
from move.data.preprocessing import one_hot_encode_single
from move.models.vae import VAE
from move.training.training_loop import TrainingLoopOutput
def find_feature_values(
feature_name: str,
feature_names_lists: list[list[str]],
feature_values: list[FloatArray],
) -> tuple[int, FloatArray]:
"""Look for the feature in the list of datasets and returns its values.
Args:
feature_name: Look-up key
feature_names_lists: List of lists with feature names for each dataset
feature_values: List of data arrays, each representing a dataset
Raises:
KeyError: If feature does not exist in any dataset
Returns:
Tuple containing (1) index of dataset containing feature and (2)
values corresponding to the feature
"""
dataset_index, feature_index = [None] * 2
for dataset_index, feature_names in enumerate(feature_names_lists):
try:
feature_index = feature_names.index(feature_name)
except ValueError:
continue
break
if dataset_index is not None and feature_index is not None:
return (
dataset_index,
np.take(feature_values[dataset_index], feature_index, axis=1),
)
raise KeyError(f"Feature '{feature_name}' not in any dataset.")
def _validate_task_config(task_config: AnalyzeLatentConfig) -> None:
if "_target_" not in task_config.reducer:
raise ValueError("Reducer class not specified properly.")
def analyze_latent(config: MOVEConfig) -> None:
"""Train one model to inspect its latent space projections."""
logger = get_logger(__name__)
logger.info("Beginning task: analyze latent space")
task_config = cast(AnalyzeLatentConfig, config.task)
_validate_task_config(task_config)
raw_data_path = Path(config.data.raw_data_path)
interim_path = Path(config.data.interim_data_path)
output_path = Path(config.data.results_path) / "latent_space"
output_path.mkdir(exist_ok=True, parents=True)
logger.debug("Reading data")
sample_names = io.read_names(raw_data_path / f"{config.data.sample_names}.txt")
cat_list, cat_names, con_list, con_names = io.load_preprocessed_data(
interim_path,
config.data.categorical_names,
config.data.continuous_names,
)
test_dataloader = make_dataloader(
cat_list,
con_list,
shuffle=False,
batch_size=task_config.batch_size,
)
test_dataset = cast(MOVEDataset, test_dataloader.dataset)
df_index = pd.Index(sample_names, name="sample")
assert task_config.model is not None
device = torch.device("cuda" if task_config.model.cuda == True else "cpu")
model: VAE = hydra.utils.instantiate(
task_config.model,
continuous_shapes=test_dataset.con_shapes,
categorical_shapes=test_dataset.cat_shapes,
)
logger.debug(f"Model: {model}")
model_path = output_path / "model.pt"
if model_path.exists():
logger.debug("Re-loading model")
model.load_state_dict(torch.load(model_path))
model.to(device)
else:
logger.debug("Training model")
model.to(device)
train_dataloader = make_dataloader(
cat_list,
con_list,
shuffle=True,
batch_size=task_config.batch_size,
drop_last=True,
)
output: TrainingLoopOutput = hydra.utils.call(
task_config.training_loop,
model=model,
train_dataloader=train_dataloader,
)
losses = output[:-1]
torch.save(model.state_dict(), model_path)
logger.info("Generating visualizations")
logger.debug("Generating plot: loss curves")
fig = viz.plot_loss_curves(losses)
fig_path = str(output_path / "loss_curve.png")
fig.savefig(fig_path, bbox_inches="tight")
fig_df = pd.DataFrame(dict(zip(viz.LOSS_LABELS, losses)))
fig_df.index.name = "epoch"
fig_df.to_csv(output_path / "loss_curve.tsv", sep="\t")
model.eval()
logger.info("Projecting into latent space")
latent_space = model.project(test_dataloader)
reducer: TransformerMixin = hydra.utils.instantiate(task_config.reducer)
embedding = reducer.fit_transform(latent_space)
mappings_path = interim_path / "mappings.json"
if mappings_path.exists():
mappings = io.load_mappings(mappings_path)
else:
mappings = {}
fig_df = pd.DataFrame(
np.take(embedding, [0, 1], axis=1),
columns=["dim0", "dim1"],
index=df_index,
)
for feature_name in task_config.feature_names:
logger.debug(f"Generating plot: latent space + '{feature_name}'")
is_categorical = False
try:
dataset_index, feature_values = find_feature_values(
feature_name, cat_names, cat_list
)
is_categorical = True
except KeyError:
try:
dataset_index, feature_values = find_feature_values(
feature_name, con_names, con_list
)
except KeyError:
logger.warning(f"Feature '{feature_name}' not found in any dataset.")
continue
if is_categorical:
# Convert one-hot encoding to category codes
is_nan = feature_values.sum(axis=1) == 0
feature_values = np.argmax(feature_values, axis=1)
dataset_name = config.data.categorical_names[dataset_index]
feature_mapping = {
str(code): category for category, code in mappings[dataset_name].items()
}
fig = viz.plot_latent_space_with_cat(
embedding,
feature_name,
feature_values,
feature_mapping,
is_nan,
)
fig_df[feature_name] = np.where(is_nan, np.nan, feature_values)
else:
feature_values = feature_values
fig = viz.plot_latent_space_with_con(
embedding, feature_name, feature_values
)
fig_df[feature_name] = np.where(feature_values == 0, np.nan, feature_values)
# Remove non-alpha characters
safe_feature_name = re.sub(r"[^\w\s]", "", feature_name)
fig_path = str(output_path / f"latent_space_{safe_feature_name}.png")
fig.savefig(fig_path, bbox_inches="tight")
fig_df.to_csv(output_path / "latent_space.tsv", sep="\t")
logger.info("Reconstructing")
cat_recons, con_recons = model.reconstruct(test_dataloader)
con_recons = np.split(con_recons, np.cumsum(model.continuous_shapes[:-1]), axis=1)
logger.info("Computing reconstruction metrics")
scores = []
labels = config.data.categorical_names + config.data.continuous_names
for cat, cat_recon in zip(cat_list, cat_recons):
accuracy = calculate_accuracy(cat, cat_recon)
scores.append(accuracy)
for con, con_recon in zip(con_list, con_recons):
cosine_sim = calculate_cosine_similarity(con, con_recon)
scores.append(cosine_sim)
logger.debug("Generating plot: reconstruction metrics")
plot_scores = [np.ma.compressed(np.ma.masked_equal(each, 0)) for each in scores]
fig = viz.plot_metrics_boxplot(plot_scores, labels)
fig_path = str(output_path / "reconstruction_metrics.png")
fig.savefig(fig_path, bbox_inches="tight")
fig_df = pd.DataFrame(dict(zip(labels, scores)), index=df_index)
fig_df.to_csv(output_path / "reconstruction_metrics.tsv", sep="\t")
logger.info("Computing feature importance")
num_samples = len(cast(Sized, test_dataloader.sampler))
for i, dataset_name in enumerate(config.data.categorical_names):
logger.debug(f"Generating plot: feature importance '{dataset_name}'")
na_value = one_hot_encode_single(mappings[dataset_name], None)
dataloaders = perturb_categorical_data(
test_dataloader, config.data.categorical_names, dataset_name, na_value
)
num_features = len(dataloaders)
z = model.project(test_dataloader)
diffs = np.empty((num_samples, num_features))
for j, dataloader in enumerate(dataloaders):
z_perturb = model.project(dataloader)
diffs[:, j] = np.sum(z_perturb - z, axis=1)
feature_mapping = {
str(code): category for category, code in mappings[dataset_name].items()
}
fig = viz.plot_categorical_feature_importance(
diffs, cat_list[i], cat_names[i], feature_mapping
)
fig_path = str(output_path / f"feat_importance_{dataset_name}.png")
fig.savefig(fig_path, bbox_inches="tight")
fig_df = pd.DataFrame(diffs, columns=cat_names[i], index=df_index)
fig_df.to_csv(output_path / f"feat_importance_{dataset_name}.tsv", sep="\t")
for i, dataset_name in enumerate(config.data.continuous_names):
logger.debug(f"Generating plot: feature importance '{dataset_name}'")
dataloaders = perturb_continuous_data(
test_dataloader, config.data.continuous_names, dataset_name, 0.0
)
num_features = len(dataloaders)
z = model.project(test_dataloader)
diffs = np.empty((num_samples, num_features))
for j, dataloader in enumerate(dataloaders):
z_perturb = model.project(dataloader)
diffs[:, j] = np.sum(z_perturb - z, axis=1)
fig = viz.plot_continuous_feature_importance(diffs, con_list[i], con_names[i])
fig_path = str(output_path / f"feat_importance_{dataset_name}.png")
fig.savefig(fig_path, bbox_inches="tight")
fig_df = pd.DataFrame(diffs, columns=con_names[i], index=df_index)
fig_df.to_csv(output_path / f"feat_importance_{dataset_name}.tsv", sep="\t")