+
+
+
+
+
+
+In [13]:
+
+
+
+
+
+import json
+import pandas as pd
+import matplotlib.pyplot as plt
+
+# Wczytaj log
+with open("../models/binary/training_log.json") as f:
+ data = json.load(f)
+
+# Zamień na DataFrame
+df = pd.DataFrame(data)
+
+# Rozdziel na train/eval
+train_df = df[df.get("loss").notna()].copy()
+eval_df = df[df.get("eval_loss").notna()].copy()
+
+print(f"🟩 Wczytano {len(train_df)} kroków treningu, 🔵 {len(eval_df)} ewaluacji")
+df.head()
+
+
+
+
+
+
+
+
+
+
+
+🟩 Wczytano 396 kroków treningu, 🔵 7 ewaluacji ++
+
+Out[13]:
+
+
+
+
+
+
+
+
+| + | loss | +grad_norm | +learning_rate | +epoch | +step | +eval_loss | +eval_accuracy | +eval_precision | +eval_recall | +eval_f1 | +eval_runtime | +eval_samples_per_second | +eval_steps_per_second | +train_runtime | +train_samples_per_second | +train_steps_per_second | +total_flos | +train_loss | +
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | +0.3831 | +0.071747 | +0.00002 | +0.02 | +50 | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +
| 1 | +0.2821 | +0.131042 | +0.00002 | +0.03 | +100 | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +
| 2 | +0.3882 | +0.649664 | +0.00002 | +0.05 | +150 | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +
| 3 | +0.1869 | +0.065614 | +0.00002 | +0.06 | +200 | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +
| 4 | +0.3480 | +0.131579 | +0.00002 | +0.08 | +250 | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +NaN | +
+
+
+
+
+
+
+In [14]:
+
+
+
+
+
+plt.figure(figsize=(12, 6))
+plt.plot(train_df["epoch"], train_df["loss"], label="Train Loss", alpha=0.5, marker=".", markersize=3)
+plt.plot(eval_df["epoch"], eval_df["eval_loss"], label="Eval Loss", color="orange", marker="s")
+plt.title("Strata (Loss) w czasie treningu")
+plt.xlabel("Epoka")
+plt.ylabel("Loss")
+plt.legend()
+plt.grid(True)
+plt.tight_layout()
+plt.show()
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+In [15]:
+
+
+
+
+
+plt.figure(figsize=(12, 6))
+plt.plot(eval_df["epoch"], eval_df["eval_accuracy"], label="Accuracy", marker="o")
+plt.plot(eval_df["epoch"], eval_df["eval_f1"], label="F1 Score", marker="s")
+plt.plot(eval_df["epoch"], eval_df["eval_precision"], label="Precision", marker="^")
+plt.plot(eval_df["epoch"], eval_df["eval_recall"], label="Recall", marker="v")
+plt.title("Metryki ewaluacji w czasie treningu")
+plt.xlabel("Epoka")
+plt.ylabel("Wartość")
+plt.ylim(0.90, 1.01)
+plt.legend()
+plt.grid(True)
+plt.tight_layout()
+plt.show()
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+In [16]:
+
+
+
+
+
+fig, ax1 = plt.subplots(figsize=(12, 6))
+
+color = "tab:red"
+ax1.set_xlabel("Epoka")
+ax1.set_ylabel("Grad Norm", color=color)
+ax1.plot(train_df["epoch"], train_df["grad_norm"], color=color, alpha=0.6, label="Grad Norm")
+ax1.tick_params(axis='y', labelcolor=color)
+ax1.legend(loc="upper left")
+
+ax2 = ax1.twinx()
+color = "tab:blue"
+ax2.set_ylabel("Learning Rate", color=color)
+ax2.plot(train_df["epoch"], train_df["learning_rate"], color=color, alpha=0.6, linestyle="--", label="LR")
+ax2.tick_params(axis='y', labelcolor=color)
+ax2.legend(loc="upper right")
+
+plt.title("Gradient i learning rate w czasie treningu")
+fig.tight_layout()
+plt.grid(True)
+plt.show()
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+In [17]:
+
+
+
+
+
+best_idx = eval_df["eval_f1"].idxmax()
+best_row = eval_df.loc[best_idx]
+print("🏆 Najlepszy punkt ewaluacji:")
+print(best_row[["epoch", "eval_loss", "eval_accuracy", "eval_precision", "eval_recall", "eval_f1"]])
+
+
+
+
+
+
+
+
+
+
+
+🏆 Najlepszy punkt ewaluacji: +epoch 5.000000 +eval_loss 0.054792 +eval_accuracy 0.992121 +eval_precision 0.993364 +eval_recall 0.998000 +eval_f1 0.995677 +Name: 334, dtype: float64 ++
+
+
+
+
+
+
+In [18]:
+
+
+
+
+
+train_df.to_csv("../models/binary/train_log.csv", index=False)
+eval_df.to_csv("../models/binary/eval_log.csv", index=False)
+print("💾 Zapisano logi CSV")
+
+
+
+
+
+
+
+
+
+
+
+💾 Zapisano logi CSV ++
+
+
+
+
+
+
+
+In [ ]:
+
+
+
+
+
+
+