File size: 6,402 Bytes
0ec2e87 | 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 | """Utilities for saving training/testing visualizations."""
from __future__ import annotations
from typing import Iterable, Optional
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt # noqa: E402
def _to_numpy(values: Iterable) -> np.ndarray:
if values is None:
return np.array([])
if isinstance(values, np.ndarray):
return values
try:
import torch
if torch.is_tensor(values):
return values.detach().cpu().numpy()
except ImportError:
pass
return np.asarray(values)
def plot_flow_time_loss(
time_loss_pairs: Iterable[tuple[float, float]],
save_path: str,
*,
num_bins: int = 10,
) -> None:
pairs = list(time_loss_pairs)
if not pairs:
return
times = _to_numpy([pair[0] for pair in pairs]).astype(float)
losses = _to_numpy([pair[1] for pair in pairs]).astype(float)
plt.figure(figsize=(6, 4))
plt.scatter(times, losses, alpha=0.3, s=12, label="samples")
bins = np.linspace(0.0, 1.0, num_bins + 1)
bin_indices = np.digitize(times, bins) - 1
bin_centers = 0.5 * (bins[:-1] + bins[1:])
bin_means = []
for idx in range(num_bins):
mask = bin_indices == idx
if np.any(mask):
bin_means.append(np.mean(losses[mask]))
else:
bin_means.append(np.nan)
plt.plot(bin_centers, bin_means, color="red", linewidth=2, label="binned mean")
plt.xlim(0.0, 1.0)
plt.xlabel("t")
plt.ylabel("loss")
plt.title("Flow time loss vs time")
plt.legend()
plt.tight_layout()
plt.savefig(save_path)
plt.close()
def plot_uncertainty_calibration(
predicted: Iterable[float],
empirical: Iterable[float],
save_path: str,
*,
num_bins: int = 10,
) -> None:
predicted_arr = _to_numpy(predicted).astype(float)
empirical_arr = _to_numpy(empirical).astype(float)
if predicted_arr.size == 0 or empirical_arr.size == 0:
return
min_val = np.min(predicted_arr)
max_val = np.max(predicted_arr)
if np.isclose(min_val, max_val):
max_val = min_val + 1e-6
bins = np.linspace(min_val, max_val, num_bins + 1)
bin_indices = np.digitize(predicted_arr, bins) - 1
bin_means_pred = []
bin_means_emp = []
for idx in range(num_bins):
mask = bin_indices == idx
if np.any(mask):
bin_means_pred.append(np.mean(predicted_arr[mask]))
bin_means_emp.append(np.mean(empirical_arr[mask]))
plt.figure(figsize=(5, 5))
plt.plot(bin_means_pred, bin_means_emp, marker="o", label="calibration")
line_min = min(np.min(predicted_arr), np.min(empirical_arr))
line_max = max(np.max(predicted_arr), np.max(empirical_arr))
plt.plot([line_min, line_max], [line_min, line_max], linestyle="--", color="gray", label="y=x")
plt.xlabel("Mean predicted uncertainty")
plt.ylabel("Mean empirical error")
plt.title("Uncertainty calibration")
plt.legend()
plt.tight_layout()
plt.savefig(save_path)
plt.close()
def plot_training_trends(
steps: Iterable[int],
loss_fm: Iterable[float],
loss_u: Optional[Iterable[float]],
mean_pred_uncertainty: Iterable[float],
mean_residual_magnitude: Iterable[float],
save_path: str,
) -> None:
steps_arr = _to_numpy(steps)
plt.figure(figsize=(7, 4))
plt.plot(steps_arr, _to_numpy(loss_fm), label="loss_fm")
if loss_u is not None:
loss_u_arr = _to_numpy(loss_u)
if loss_u_arr.size:
plt.plot(steps_arr, loss_u_arr, label="loss_u")
plt.plot(steps_arr, _to_numpy(mean_pred_uncertainty), label="mean predicted uncertainty")
plt.plot(steps_arr, _to_numpy(mean_residual_magnitude), label="mean residual magnitude")
plt.xlabel("global step")
plt.ylabel("value")
plt.title("Training trends")
plt.legend()
plt.tight_layout()
plt.savefig(save_path)
plt.close()
def plot_uncertainty_heatmap(
uncertainty: Iterable,
save_path: str,
*,
cmap: str = "viridis",
) -> None:
uncertainty_arr = _to_numpy(uncertainty).astype(float)
if uncertainty_arr.size == 0:
return
plt.figure(figsize=(6, 4))
plt.imshow(uncertainty_arr, aspect="auto", cmap=cmap)
plt.colorbar(label="s_hat")
plt.xlabel("Action dimension")
plt.ylabel("Timestep")
plt.title("Uncertainty heatmap")
plt.tight_layout()
plt.savefig(save_path)
plt.close()
def plot_refinement_effect(
mean_s_hat: Iterable[float],
mean_residual: Optional[Iterable[float]],
save_path: str,
) -> None:
mean_s_hat_arr = _to_numpy(mean_s_hat)
if mean_s_hat_arr.size == 0:
return
iterations = np.arange(1, len(mean_s_hat_arr) + 1)
plt.figure(figsize=(6, 4))
plt.plot(iterations, mean_s_hat_arr, marker="o", label="mean s_hat")
if mean_residual is not None:
mean_residual_arr = _to_numpy(mean_residual)
if mean_residual_arr.size:
plt.plot(iterations, mean_residual_arr, marker="x", label="mean residual magnitude")
plt.xlabel("Refinement iteration")
plt.ylabel("value")
plt.title("Refinement effect")
plt.legend()
plt.tight_layout()
plt.savefig(save_path)
plt.close()
def plot_task_summary(
task_labels: Iterable[str],
success_rate: Iterable[float],
mean_uncertainty: Iterable[float],
mean_failure_uncertainty: Iterable[float],
save_path: str,
) -> None:
labels = list(task_labels)
success_arr = _to_numpy(success_rate).astype(float)
mean_uncertainty_arr = _to_numpy(mean_uncertainty).astype(float)
mean_failure_arr = _to_numpy(mean_failure_uncertainty).astype(float)
x = np.arange(len(labels))
width = 0.25
plt.figure(figsize=(10, 4))
plt.bar(x - width, success_arr, width=width, label="success rate")
plt.bar(x, mean_uncertainty_arr, width=width, label="mean uncertainty")
plt.bar(x + width, mean_failure_arr, width=width, label="mean uncertainty (failure)")
plt.xticks(x, labels, rotation=45, ha="right")
plt.ylabel("value")
plt.title("Per-task summary")
plt.legend()
plt.tight_layout()
plt.savefig(save_path)
plt.close()
|