Pikaliov/caption-test
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add Average Precision (AP) metric for train and val

Browse Source 
- Compute AP (Mean Reciprocal Rank) in _evaluate() for both q→g and g→q
- AP saved in val.csv and train_recall.csv alongside R@K
- New AP plot panel in val_metrics.png (train vs val, both directions)
- Log AP in console output for train-recall and val epochs

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
pikaliov
2026-04-22 08:08:48 +03:00
parent df60e83ead
commit 4a05d05ccd
2 changed files with 35 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
src/training/plot_metrics.py
View File

@@ -70,9 +70,9 @@ def plot_train_metrics(train_df: pd.DataFrame, out_dir: Path) -> None:
def plot_val_metrics(val_df: pd.DataFrame, out_dir: Path, train_recall_df: pd.DataFrame | None = None) -> None:
"""Plot recall metrics: train vs val R@K."""
fig, axes = plt.subplots(1, 2, figsize=(14, 5))
fig.suptitle("Recall Metrics (train vs val)", fontsize=16, fontweight="bold")
"""Plot recall + AP metrics: train vs val."""
fig, axes = plt.subplots(1, 3, figsize=(20, 5))
fig.suptitle("Retrieval Metrics (train vs val)", fontsize=16, fontweight="bold")
# 1. Recall@K (q→g): train + val.
ax = axes[0]
@@ -102,6 +102,23 @@ def plot_val_metrics(val_df: pd.DataFrame, out_dir: Path, train_recall_df: pd.Da
ax.set_ylim(0, 1)
ax.legend(fontsize=8)
# 3. AP (train vs val, both directions).
ax = axes[2]
if "ap_q2g" in val_df.columns:
sns.lineplot(data=val_df, x="epoch", y="ap_q2g", ax=ax, marker="o", linewidth=2, color="royalblue", label="val AP q→g")
if "ap_g2q" in val_df.columns:
sns.lineplot(data=val_df, x="epoch", y="ap_g2q", ax=ax, marker="s", linewidth=2, color="coral", label="val AP g→q")
if train_recall_df is not None:
if "ap_q2g" in train_recall_df.columns:
sns.lineplot(data=train_recall_df, x="epoch", y="ap_q2g", ax=ax, marker="x", linewidth=1.5, linestyle="--", color="royalblue", label="train AP q→g")
if "ap_g2q" in train_recall_df.columns:
sns.lineplot(data=train_recall_df, x="epoch", y="ap_g2q", ax=ax, marker="x", linewidth=1.5, linestyle="--", color="coral", label="train AP g→q")
ax.set_title("Average Precision")
ax.set_xlabel("Epoch")
ax.set_ylabel("AP")
ax.set_ylim(0, 1)
ax.legend(fontsize=8)
plt.tight_layout()
path = out_dir / "val_metrics.png"
fig.savefig(path, dpi=150, bbox_inches="tight")

18
src/training/train_gtauav.py
View File

@@ -228,18 +228,27 @@ def _evaluate(
if batch_losses:
metrics["loss"] = sum(batch_losses) / len(batch_losses)
# R@K and AP (q→g).
sorted_idx = sim.argsort(dim=1, descending=True)
for k in k_values:
top_k = sorted_idx[:, :k]
hit = (top_k == targets.unsqueeze(1)).any(dim=1).float()
metrics[f"r@{k}_q2g"] = float(hit.mean().item())
# AP q→g: for each query, rank of the correct gallery = 1/(rank+1).
ranks_q2g = (sorted_idx == targets.unsqueeze(1)).nonzero(as_tuple=True)[1].float()
metrics["ap_q2g"] = float((1.0 / (ranks_q2g + 1)).mean().item())
# R@K and AP (g→q).
sorted_idx_g2q = sim.t().argsort(dim=1, descending=True)
for k in k_values:
top_k = sorted_idx_g2q[:, :k]
hit = (top_k == targets.unsqueeze(1)).any(dim=1).float()
metrics[f"r@{k}_g2q"] = float(hit.mean().item())
ranks_g2q = (sorted_idx_g2q == targets.unsqueeze(1)).nonzero(as_tuple=True)[1].float()
metrics["ap_g2q"] = float((1.0 / (ranks_g2q + 1)).mean().item())
metrics["gate_q"] = model.fusion_query.gate_value
metrics["gate_g"] = model.fusion_gallery.gate_value
return metrics
@@ -696,11 +705,13 @@ def train(cfg: TrainConfigGTAUAV) -> None:
csv_logger.log_train_recall(epoch, train_recall)
tracker.log_train(epoch, {f"recall/{k}": v for k, v in train_recall.items() if k.startswith("r@")}, step=global_step)
LOGGER.info(
"train-recall epoch=%d R@1=%.4f R@5=%.4f R@10=%.4f",
"train-recall epoch=%d R@1=%.4f R@5=%.4f R@10=%.4f AP=%.4f loss=%.4f",
epoch,
train_recall.get("r@1_q2g", 0.0),
train_recall.get("r@5_q2g", 0.0),
train_recall.get("r@10_q2g", 0.0),
train_recall.get("ap_q2g", 0.0),
train_recall.get("loss", 0.0),
)
# Val R@K (full test set).
@@ -721,13 +732,14 @@ def train(cfg: TrainConfigGTAUAV) -> None:
tracker.log_scalar("val/best_r@1_q2g", best_r1, step=global_step)
LOGGER.info(
"val epoch=%d R@1=%.4f R@5=%.4f R@10=%.4f gate_q=%.4f gate_g=%.4f",
"val epoch=%d R@1=%.4f R@5=%.4f R@10=%.4f AP=%.4f loss=%.4f gate_q=%.4f",
epoch,
val_metrics.get("r@1_q2g", 0.0),
val_metrics.get("r@5_q2g", 0.0),
val_metrics.get("r@10_q2g", 0.0),
val_metrics.get("ap_q2g", 0.0),
val_metrics.get("loss", 0.0),
val_metrics.get("gate_q", 1.0),
val_metrics.get("gate_g", 1.0),
)
# --- Grad-CAM visualization ---