File size: 19,174 Bytes
67ba414 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 | import argparse
import csv
import json
import os
import matplotlib.pyplot as plt
import wandb
DEFAULT_BUCKETS = [
"bucket_1",
"bucket_2",
"bucket_3",
"bucket_4",
"bucket_5",
"bucket_6",
]
COMPONENTS = ["kl", "entropy", "task"]
LOSS_COMPONENTS = ["policy", "entropy", "kl", "total"]
def _sanitize_dir_name(name: str) -> str:
return "".join(c if c.isalnum() or c in ("-", "_") else "_" for c in name)
def _bucket_sort_key(bucket_name: str):
if bucket_name.startswith("bucket_"):
suffix = bucket_name.split("_", 1)[1]
if suffix.isdigit():
return (0, int(suffix))
return (1, bucket_name)
def _extract_buckets(metric_source: dict) -> list[str]:
buckets = set()
for key in metric_source.keys():
if not key.startswith("grad_norm/bucket_"):
continue
parts = key.split("/")
if len(parts) >= 2:
buckets.add(parts[1])
if not buckets:
return DEFAULT_BUCKETS
return sorted(buckets, key=_bucket_sort_key)
def get_bucket_label(bucket_name):
"""Formats bucket names for the plot axis."""
if bucket_name.startswith("bucket_"):
return bucket_name.replace("_", " ")
return bucket_name
def main():
parser = argparse.ArgumentParser(
description="Plot gradient-analysis metrics from a W&B run.",
epilog=(
"Examples:\n"
" python gradient_analysis/plot_gradient_analysis.py --wandb-path entity/project/run_id\n"
" python gradient_analysis/plot_gradient_analysis.py --wandb-path entity/project/run_id --step 1\n"
" python gradient_analysis/plot_gradient_analysis.py --wandb-path entity/project/run_id "
"--output-dir gradient_analysis_outputs/my_run\n"
" python gradient_analysis/plot_gradient_analysis.py --wandb-path entity/project/run_id --list-steps"
),
formatter_class=argparse.RawTextHelpFormatter,
)
parser.add_argument(
"--wandb-path",
required=True,
help="W&B run path like entity/project/run_id",
)
parser.add_argument(
"--output-dir",
default=None,
help=(
"Directory for generated plots and exported metrics. "
"Defaults to gradient_analysis_outputs/<run_name>_<run_id>."
),
)
parser.add_argument(
"--step",
dest="steps",
type=int,
nargs="+",
default=None,
help="One or more training steps to plot. Default: all available gradient-analysis steps.",
)
parser.add_argument(
"--list-steps",
action="store_true",
help="List available gradient-analysis steps in the run and exit.",
)
args = parser.parse_args()
print(f"Connecting to WandB run: {args.wandb_path}...")
api = wandb.Api()
try:
run = api.run(args.wandb_path)
except Exception as e:
print(f"Error accessing run: {e}")
return
summary = run.summary
step_metrics = {}
available_steps = set()
for row in run.scan_history():
bucket_items = [(k, v) for k, v in row.items() if k.startswith("grad_norm/bucket_")]
has_bucket_metrics = any(v is not None for _, v in bucket_items)
has_nonzero_bucket_metrics = any((v is not None and v != 0) for _, v in bucket_items)
if not has_bucket_metrics or not has_nonzero_bucket_metrics:
continue
step = row.get("_step")
if step is None:
continue
available_steps.add(step)
if step not in step_metrics:
step_metrics[step] = {}
for k, v in row.items():
if v is None:
continue
step_metrics[step][k] = v
if available_steps:
print(f"Found grad_norm bucket metrics at steps: {sorted(available_steps)}")
else:
print("Warning: no grad_norm metrics found in history; falling back to run summary.")
step_metrics = {"summary": summary}
if args.list_steps:
if available_steps:
print("Available gradient-analysis steps:")
for step in sorted(available_steps):
print(step)
else:
print("No gradient-analysis steps found.")
return
default_dir = os.path.join(
"gradient_analysis_outputs",
f"{_sanitize_dir_name(run.name)}_{run.id}",
)
output_dir = args.output_dir or default_dir
os.makedirs(output_dir, exist_ok=True)
titles = {
"kl": "KL Gradient Norm",
"entropy": "Entropy Gradient Norm",
"task": "Task (Policy) Gradient Norm"
}
colors = ["#3498db", "#2ecc71", "#e74c3c"] # Blue, Green, Red
loss_titles = {
"policy": "Policy (Task) Loss",
"entropy": "Entropy Loss",
"kl": "KL Loss",
"total": "Total Loss",
}
loss_colors = ["#8e44ad", "#27ae60", "#2980b9", "#c0392b"] # Purple, Green, Blue, Red
norm_titles = {
"kl": "KL Grad Norm (Per Sample vs Per Token)",
"entropy": "Entropy Grad Norm (Per Sample vs Per Token)",
"task": "Task Grad Norm (Per Sample vs Per Token)",
}
steps_to_plot = sorted(step_metrics.keys(), key=lambda x: (isinstance(x, str), x))
if args.steps is not None:
requested_steps = set(args.steps)
steps_to_plot = [s for s in steps_to_plot if s in requested_steps]
if not steps_to_plot:
print(f"Error: none of the requested steps {sorted(requested_steps)} were found.")
return
for step_key in steps_to_plot:
metric_source = step_metrics[step_key]
buckets = _extract_buckets(metric_source)
x_labels = [get_bucket_label(b) for b in buckets]
step_tag = f"step_{step_key}"
output_file = os.path.join(output_dir, f"gradient_analysis_plots_{step_tag}.png")
output_file_loss = os.path.join(output_dir, f"gradient_analysis_loss_plots_{step_tag}.png")
output_file_rv = os.path.join(output_dir, f"gradient_analysis_reward_std_{step_tag}.png")
output_file_normed = os.path.join(output_dir, f"gradient_analysis_normed_grads_{step_tag}.png")
output_file_summary = os.path.join(output_dir, f"gradient_analysis_summary_{step_tag}.png")
output_metrics_json = os.path.join(output_dir, f"gradient_analysis_metrics_{step_tag}.json")
output_rv_table = os.path.join(output_dir, f"gradient_analysis_bucket_rv_table_{step_tag}.csv")
# Create subplots for gradient norms
fig, axes = plt.subplots(1, 3, figsize=(20, 6))
plt.subplots_adjust(wspace=0.3, top=0.62, bottom=0.12)
bucket_rv = {
b: {
"mean": metric_source.get(f"grad_norm/{b}/reward_std_mean", 0),
"min": metric_source.get(f"grad_norm/{b}/reward_std_min", 0),
"max": metric_source.get(f"grad_norm/{b}/reward_std_max", 0),
}
for b in buckets
}
bucket_rv_values = {b: [] for b in buckets}
table_keys = [f"grad_norm/{b}/group_rv_table" for b in buckets]
for row in run.scan_history(keys=["_step", *table_keys]):
if row.get("_step") != step_key:
continue
for b in buckets:
key = f"grad_norm/{b}/group_rv_table"
table_meta = row.get(key)
if not isinstance(table_meta, dict) or "path" not in table_meta:
continue
table_path = table_meta["path"]
try:
file_ref = run.file(table_path)
local_path = file_ref.download(replace=True).name
with open(local_path, "r") as f:
table_json = json.load(f)
# table_json has keys: columns, data
data_rows = table_json.get("data", [])
# columns: bucket, group_id, reward_std
for row_vals in data_rows:
if len(row_vals) >= 3:
bucket_rv_values[b].append(float(row_vals[2]))
except Exception:
continue
# Save raw metric snapshot and RV table values for this step
try:
with open(output_metrics_json, "w") as f:
json.dump(metric_source, f, indent=2, sort_keys=True)
except Exception as e:
print(f"Warning: failed to write metrics json: {e}")
try:
with open(output_rv_table, "w", newline="") as f:
writer = csv.writer(f)
writer.writerow(["bucket", "reward_std"])
for b in buckets:
for rv in bucket_rv_values[b]:
writer.writerow([b, rv])
except Exception as e:
print(f"Warning: failed to write rv table csv: {e}")
legend_lines = []
for label, bucket in zip(x_labels, buckets):
rv = bucket_rv.get(bucket, {})
legend_lines.append(
f"{label}: mean={rv.get('mean', 0):.3f} min={rv.get('min', 0):.3f} max={rv.get('max', 0):.3f}"
)
for ax, comp, color in zip(axes, COMPONENTS, colors):
y_values = []
for bucket in buckets:
key = f"grad_norm/{bucket}/{comp}"
val = metric_source.get(key, 0)
y_values.append(val)
bars = ax.bar(x_labels, y_values, color=color, alpha=0.8, edgecolor='black', linewidth=1)
ax.set_title(titles[comp], fontsize=16, fontweight='bold', pad=15)
ax.set_ylabel("Grad Norm Magnitude", fontsize=12)
ax.set_xlabel("Reward Variance Bucket", fontsize=12)
ax.grid(axis='y', linestyle='--', alpha=0.6)
for bar in bars:
height = bar.get_height()
ax.text(bar.get_x() + bar.get_width()/2., height + (max(y_values)*0.01 if y_values else 0.01),
f'{height:.4f}', ha='center', va='bottom', fontsize=10, fontweight='bold')
fig.suptitle(f"Gradient Norms - Run: {run.name} (Step {step_key})", fontsize=20, y=0.98)
fig.text(0.5, 0.88, "\n".join(legend_lines), ha="center", va="top", fontsize=8)
plt.savefig(output_file, bbox_inches='tight', dpi=300)
print(f"\nSuccess! Results visualization saved to: {os.path.abspath(output_file)}")
plt.close(fig)
# Create subplots for per-component losses
fig2, axes2 = plt.subplots(2, 2, figsize=(16, 10))
plt.subplots_adjust(hspace=0.35, wspace=0.25)
for ax, comp, color in zip(axes2.flatten(), LOSS_COMPONENTS, loss_colors):
y_values = []
for bucket in buckets:
key = f"grad_norm/{bucket}/loss/{comp}"
val = metric_source.get(key, 0)
y_values.append(val)
bars = ax.bar(x_labels, y_values, color=color, alpha=0.8, edgecolor="black", linewidth=1)
ax.set_title(loss_titles[comp], fontsize=14, fontweight="bold", pad=10)
ax.set_ylabel("Loss", fontsize=11)
ax.set_xlabel("Reward Variance Bucket", fontsize=11)
ax.grid(axis="y", linestyle="--", alpha=0.6)
for bar in bars:
height = bar.get_height()
ax.text(
bar.get_x() + bar.get_width() / 2.0,
height + (max(y_values) * 0.01 if y_values else 0.01),
f"{height:.4f}",
ha="center",
va="bottom",
fontsize=9,
fontweight="bold",
)
fig2.suptitle(f"Per-Component Losses - Run: {run.name} (Step {step_key})", fontsize=18, y=1.02)
plt.tight_layout()
plt.savefig(output_file_loss, bbox_inches="tight", dpi=300)
print(f"Success! Loss visualization saved to: {os.path.abspath(output_file_loss)}")
plt.close(fig2)
# Create plot for per-bucket mean reward variance (std)
rv_values = []
for bucket in buckets:
rv_values.append(metric_source.get(f"grad_norm/{bucket}/reward_std_mean", 0))
if any(v != 0 for v in rv_values):
fig3, ax3 = plt.subplots(1, 1, figsize=(10, 5))
bars = ax3.bar(x_labels, rv_values, color="#f39c12", alpha=0.85, edgecolor="black", linewidth=1)
ax3.set_title(f"Reward Std Mean by Bucket - Run: {run.name} (Step {step_key})", fontsize=14, fontweight="bold", pad=10)
ax3.set_ylabel("Reward Std (Mean)", fontsize=11)
ax3.set_xlabel("Reward Variance Bucket", fontsize=11)
ax3.grid(axis="y", linestyle="--", alpha=0.6)
for bar in bars:
height = bar.get_height()
ax3.text(
bar.get_x() + bar.get_width() / 2.0,
height + (max(rv_values) * 0.01 if rv_values else 0.01),
f"{height:.4f}",
ha="center",
va="bottom",
fontsize=9,
fontweight="bold",
)
plt.tight_layout()
plt.savefig(output_file_rv, bbox_inches="tight", dpi=300)
print(f"Success! Reward std visualization saved to: {os.path.abspath(output_file_rv)}")
plt.close(fig3)
else:
print(f"Warning: No reward std mean metrics found at step {step_key}; skipping reward std plot.")
# Create plots for per-sample and per-token grad norms (combined per component)
fig4, axes4 = plt.subplots(1, 3, figsize=(20, 6))
plt.subplots_adjust(wspace=0.3)
for ax, comp in zip(axes4, COMPONENTS):
per_sample = []
per_token = []
for bucket in buckets:
per_sample.append(metric_source.get(f"grad_norm/{bucket}/per_sample/{comp}", 0))
per_token.append(metric_source.get(f"grad_norm/{bucket}/per_token/{comp}", 0))
x = range(len(x_labels))
width = 0.38
bars1 = ax.bar([i - width / 2 for i in x], per_sample, width=width, label="per_sample", color="#16a085", alpha=0.85)
bars2 = ax.bar([i + width / 2 for i in x], per_token, width=width, label="per_token", color="#f39c12", alpha=0.85)
ax.set_xticks(list(x))
ax.set_xticklabels(x_labels)
ax.set_title(norm_titles[comp], fontsize=14, fontweight="bold", pad=10)
ax.set_ylabel("Grad Norm", fontsize=11)
ax.set_xlabel("Reward Variance Bucket", fontsize=11)
ax.grid(axis="y", linestyle="--", alpha=0.6)
ax.legend(frameon=False, fontsize=9)
for bar in list(bars1) + list(bars2):
height = bar.get_height()
ax.text(
bar.get_x() + bar.get_width() / 2.0,
height + (max(per_sample + per_token) * 0.01 if (per_sample + per_token) else 0.01),
f"{height:.4f}",
ha="center",
va="bottom",
fontsize=8,
)
fig4.suptitle(f"Normalized Grad Norms - Run: {run.name} (Step {step_key})", fontsize=18, y=1.03)
plt.tight_layout()
plt.savefig(output_file_normed, bbox_inches="tight", dpi=300)
print(f"Success! Normalized grad visualization saved to: {os.path.abspath(output_file_normed)}")
plt.close(fig4)
# Summary 3-panel plot with available aggregates
rv_means = [bucket_rv[b]["mean"] for b in buckets]
rv_mins = [bucket_rv[b]["min"] for b in buckets]
rv_maxs = [bucket_rv[b]["max"] for b in buckets]
task_grads = [metric_source.get(f"grad_norm/{b}/task", 0) for b in buckets]
kl_grads = [metric_source.get(f"grad_norm/{b}/kl", 0) for b in buckets]
ent_grads = [metric_source.get(f"grad_norm/{b}/entropy", 0) for b in buckets]
reg_grads = [k + e for k, e in zip(kl_grads, ent_grads)]
fig5, axes5 = plt.subplots(1, 3, figsize=(20, 6))
plt.subplots_adjust(wspace=0.35, top=0.80, bottom=0.15)
# Left: RV mean with min/max error bars per bucket
ax = axes5[0]
use_boxplot = any(bucket_rv_values[b] for b in buckets)
if use_boxplot:
data = [bucket_rv_values[b] for b in buckets]
data_mins = [min(v) if v else 0 for v in data]
data_maxs = [max(v) if v else 0 for v in data]
# sanity check: compare against logged min/max (per-sample)
mismatch = any(
abs(dm - rm) > 1e-3 or abs(dx - rx) > 1e-3
for dm, rm, dx, rx in zip(data_mins, rv_mins, data_maxs, rv_maxs)
)
if mismatch:
print(f"Warning: bucket RV table min/max mismatch at step {step_key}; falling back to error bars.")
use_boxplot = False
if use_boxplot:
ax.boxplot(data, tick_labels=x_labels, showfliers=False)
ax.set_title("RV by Bucket (Boxplot)", fontsize=13, fontweight="bold")
else:
# Guard against negative error bars from inconsistent min/max logging.
yerr = [
[max(0.0, m - lo) for m, lo in zip(rv_means, rv_mins)],
[max(0.0, hi - m) for m, hi in zip(rv_means, rv_maxs)],
]
ax.errorbar(x_labels, rv_means, yerr=yerr, fmt="o-", color="#6c5ce7", ecolor="#2d3436", capsize=4)
ax.set_title("RV by Bucket (Mean ± Min/Max)", fontsize=13, fontweight="bold")
ax.set_xlabel("Bucket")
ax.set_ylabel("Reward Variance (Std)")
ax.grid(axis="y", linestyle="--", alpha=0.5)
# Middle: task grad norm vs mean RV
ax = axes5[1]
ax.plot(rv_means, task_grads, "o-", color="#e67e22")
for i, (xv, yv) in enumerate(zip(rv_means, task_grads), start=1):
ax.text(xv, yv, f"Q{i}", fontsize=8, ha="left", va="bottom")
ax.set_title("Task Grad Norm vs RV Mean", fontsize=13, fontweight="bold")
ax.set_xlabel("RV Mean")
ax.set_ylabel("Task Grad Norm")
ax.grid(axis="y", linestyle="--", alpha=0.5)
mid_ylim = ax.get_ylim()
# Right: regularizer grad norm (KL+Entropy) vs mean RV
ax = axes5[2]
ax.plot(rv_means, reg_grads, "o-", color="#16a085")
for i, (xv, yv) in enumerate(zip(rv_means, reg_grads), start=1):
ax.text(xv, yv, f"Q{i}", fontsize=8, ha="left", va="bottom")
ax.set_title("Reg Grad Norm vs RV Mean (KL+Ent)", fontsize=13, fontweight="bold")
ax.set_xlabel("RV Mean")
ax.set_ylabel("KL+Entropy Grad Norm")
ax.set_ylim(0, 1)
ax.grid(axis="y", linestyle="--", alpha=0.5)
fig5.suptitle(f"Gradient Summary - Run: {run.name} (Step {step_key})", fontsize=16, y=0.98)
plt.tight_layout()
plt.savefig(output_file_summary, bbox_inches="tight", dpi=300)
print(f"Success! Summary visualization saved to: {os.path.abspath(output_file_summary)}")
plt.close(fig5)
print(f"All outputs written to: {os.path.abspath(output_dir)}")
if __name__ == "__main__":
main()
|