diff options
Diffstat (limited to 'experiments/analyze_resnet_crossover_r2.py')
| -rw-r--r-- | experiments/analyze_resnet_crossover_r2.py | 430 |
1 files changed, 430 insertions, 0 deletions
diff --git a/experiments/analyze_resnet_crossover_r2.py b/experiments/analyze_resnet_crossover_r2.py new file mode 100644 index 0000000..3c2bf46 --- /dev/null +++ b/experiments/analyze_resnet_crossover_r2.py @@ -0,0 +1,430 @@ +#!/usr/bin/env python3 +"""Audit the complete failure-retaining 27-cell ResNet R2 panel.""" +import argparse +import hashlib +import json +import math +import os +import sys + + +ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) +sys.path.insert(0, ROOT) +from experiments.resnet_crossover_r2 import ( + DEPTHS, + METHODS, + DEFAULT_SELECTOR, + r2_jobs, + registry_sha256, + selector_report, +) + + +def sha256(path): + digest = hashlib.sha256() + with open(path, "rb") as handle: + for chunk in iter(lambda: handle.read(1024 * 1024), b""): + digest.update(chunk) + return digest.hexdigest() + + +def read_json(path): + with open(path, encoding="utf-8") as handle: + return json.load(handle) + + +def finite(value): + return ( + isinstance(value, (int, float)) + and not isinstance(value, bool) + and math.isfinite(float(value)) + ) + + +def expected_mode(method): + return { + "bp": "bp", + "fa": "hfa", + "dfa": "dfa", + "clean_kp": "kp", + "sdil": "kp_traffic", + }[method] + + +def history_metrics(record, job): + method = job["method"] + expected_epochs = job["epochs"] + losses = [] + validation_accuracies = [] + history_examples = 0 + if method == "ff": + layers = record.get("layers", []) + layer_indices = [row.get("layer") for row in layers] + assert layer_indices == list(range(len(layers))) + for layer in layers: + epochs = layer.get("epochs", []) + assert [row.get("epoch") for row in epochs] == list( + range(1, len(epochs) + 1) + ) + assert len(epochs) <= expected_epochs + losses.extend(row.get("loss") for row in epochs) + history_examples += sum( + int(row.get("examples", 0)) for row in epochs + ) + complete = ( + len(layers) == job["depth"] + and all( + len(layer.get("epochs", [])) == expected_epochs + for layer in layers + ) + ) + completed_epochs = sum( + len(layer.get("epochs", [])) for layer in layers + ) + else: + epochs = record.get("epochs", []) + assert [row.get("epoch") for row in epochs] == list( + range(1, len(epochs) + 1) + ) + assert len(epochs) <= expected_epochs + complete = len(epochs) == expected_epochs + completed_epochs = len(epochs) + if method in ("pepita", "ep", "dualprop"): + losses = [row.get("train_loss") for row in epochs] + history_examples = sum( + int(row.get("train_examples", 0)) for row in epochs + ) + validation_accuracies = [ + row["validation"]["accuracy"] + for row in epochs + if row.get("validation") is not None + ] + else: + losses = [row.get("train_loss") for row in epochs] + history_examples = sum( + int(row.get("train_examples", 0)) for row in epochs + ) + validation_accuracies = [ + row["eval_accuracy"] + for row in epochs + if row.get("eval_accuracy") is not None + ] + final = record.get("final", {}) + final_accuracy = final.get("accuracy") + final_loss = final.get("loss") + finite_accuracies = [ + float(value) + for value in validation_accuracies + [final_accuracy] + if finite(value) + ] + best_accuracy = max(finite_accuracies) if finite_accuracies else None + losses_finite = bool(losses) and all(finite(value) for value in losses) + final_finite = ( + final.get("finite") is True + and finite(final_accuracy) + and (final_loss is None or finite(final_loss)) + ) + first_nonfinite = record.get("first_nonfinite_step") + return { + "complete_trajectory": complete, + "epochs_completed": completed_epochs, + "history_training_examples": history_examples, + "all_training_losses_finite": losses_finite, + "final_validation_accuracy": + float(final_accuracy) if finite(final_accuracy) else None, + "best_validation_accuracy": + float(best_accuracy) if best_accuracy is not None else None, + "final_validation_loss": + float(final_loss) if finite(final_loss) else None, + "first_nonfinite_step": first_nonfinite, + "finite": ( + complete + and losses_finite + and final_finite + and first_nonfinite is None + ), + } + + +def work_metrics(record, job, history): + method = job["method"] + if method in ("pepita", "ff", "ep", "dualprop"): + work = record["work"] + ordinary = int(work["ordinary_training_examples"]) + validation = int(work["ordinary_validation_examples"]) + presentations = int(work["training_example_presentations"]) + relaxation = int(work["relaxation_example_passes"]) + candidate = int(work["candidate_label_evaluation_presentations"]) + queries = int(work["logical_task_loss_queries"]) + local_vjp = int(work["local_vjp_example_evaluations"]) + local_backward = int( + work["local_target_backward_example_evaluations"] + ) + assert ordinary == history["history_training_examples"] + if method == "ff": + assert presentations == 2 * ordinary + assert candidate == 10 * validation + assert relaxation == 0 + assert local_backward == 2 * ordinary + assert local_vjp == 0 + elif method == "pepita": + assert presentations == 2 * ordinary + assert relaxation == candidate == local_vjp == local_backward == 0 + elif method == "ep": + assert presentations == ordinary + assert relaxation == 24 * ordinary + 20 * validation + assert local_vjp == relaxation + assert candidate == local_backward == 0 + else: + assert presentations == ordinary + assert relaxation == 16 * ordinary + assert local_vjp == relaxation + assert candidate == local_backward == 0 + total_work = { + "forward_macs_per_example": work["forward_macs_per_example"], + "feedforward_example_passes": work["feedforward_example_passes"], + "relaxation_example_passes": relaxation, + "local_vjp_example_evaluations": local_vjp, + "candidate_label_evaluation_presentations": candidate, + } + else: + work = record["work"] + counters = record["counters"] + ordinary = int(counters["ordinary_examples"]) + validation = ( + int(record["split"]["validation_examples"]) + * int(record["evaluation_protocol"]["validation_evaluations"]) + ) + presentations = ordinary + queries = int(work["logical_batch_loss_queries"]) + total_work = { + "forward_macs_per_example": work["forward_macs_per_example"], + "total_macs_estimate": work["total_macs_estimate"], + "elementwise_operations_estimate": + work["elementwise_operations_estimate"], + "total_forward_equivalent_examples": + work["total_forward_equivalent_examples"], + } + assert ordinary == history["history_training_examples"] + assert queries == 0 + return { + "ordinary_training_examples": ordinary, + "ordinary_validation_examples": validation, + "training_example_presentations": presentations, + "logical_task_loss_queries": queries, + "work": total_work, + } + + +def audit_job(job, expected_source, expected_selector): + manifest_path = job["output"] + ".manifest.json" + if not os.path.isfile(manifest_path): + raise AssertionError(f"missing R2 manifest: {job['experiment_name']}") + manifest = read_json(manifest_path) + for key in ( + "stage", + "method", + "architecture", + "depth", + "rate", + "epochs", + "lr_schedule", + "experiment_name", + "output", + "timeout_seconds", + "command", + ): + assert manifest[key] == job[key], ( + f"{job['experiment_name']}: manifest drift in {key}" + ) + assert manifest["source"] == expected_source + assert manifest["selector"] == expected_selector + hardware = manifest["hardware_lock"] + assert hardware["physical_gpu_index"] in (5, 7) + assert hardware["physical_gpu_uuid"] + common = { + "cell_id": f"resnet{job['depth']}::{job['method']}", + "method": job["method"], + "depth": job["depth"], + "rate": job["rate"], + "status": manifest["status"], + "manifest": os.path.relpath(manifest_path, ROOT), + "driver_wall_seconds": float(manifest["driver_wall_seconds"]), + "physical_gpu_index": hardware["physical_gpu_index"], + "physical_gpu_uuid": hardware["physical_gpu_uuid"], + "output_sha256": manifest["output_sha256"], + } + if manifest["status"] != "completed": + assert manifest["status"] in { + "timeout", "nonzero_exit", "missing_output" + } + assert ( + manifest["output_sha256"] is None + if not manifest["output_exists"] + else manifest["output_sha256"] == sha256(job["output"]) + ) + return { + **common, + "finite": False, + "complete_trajectory": False, + "epochs_completed": 0, + "best_validation_accuracy": None, + "final_validation_accuracy": None, + "final_validation_loss": None, + "first_nonfinite_step": None, + "ordinary_training_examples": None, + "ordinary_validation_examples": None, + "training_example_presentations": None, + "logical_task_loss_queries": None, + "forward_parameter_count": None, + "peak_memory_allocated_bytes": None, + "total_wall_seconds": None, + "work": None, + } + assert manifest["output_exists"] is True + assert manifest["output_sha256"] == sha256(job["output"]) + record = read_json(job["output"]) + provenance = record["provenance"] + assert provenance["git_commit"] == expected_source["git_commit"] + assert provenance["git_tracked_dirty"] is False + args = record["args"] + for key, expected in { + "depth": job["depth"], + "width": 16, + "seed": 0, + "loader_seed": 0, + "split_seed": 2027, + "batch_size": 128, + "epochs": job["epochs"], + "train_limit": 0, + "val_examples": 5000, + "eval_split": "validation", + "eval_every": 1, + "augment_train": 1, + "lr": job["rate"], + "lr_schedule": job["lr_schedule"], + "momentum": 0.9, + "weight_decay": 1e-4, + }.items(): + assert args[key] == expected, ( + f"{job['experiment_name']}: argument drift in {key}" + ) + if job["method"] in ("pepita", "ff", "ep", "dualprop"): + assert args["method"] == job["method"] + else: + assert args["mode"] == expected_mode(job["method"]) + split = record["split"] + assert split["train_examples"] == 45_000 + assert split["validation_examples"] == 5_000 + assert split["split_seed"] == 2027 + evaluation = record["evaluation_protocol"] + assert evaluation["test_evaluations"] == 0 + assert evaluation["test_used_for_selection"] is False + architecture = record["architecture"] + assert architecture["depth"] == job["depth"] + assert architecture["base_width"] == 16 + history = history_metrics(record, job) + work = work_metrics(record, job, history) + if history["complete_trajectory"]: + expected_ordinary = 45_000 * job["epochs"] + if job["method"] == "ff": + expected_ordinary *= job["depth"] + assert work["ordinary_training_examples"] == expected_ordinary + hardware_record = record["hardware"] + assert hardware_record["cuda_visible_devices"] in ("5", "7") + assert hardware_record["cuda_device_name"] == "NVIDIA GeForce GTX 1080" + parameter_count = architecture.get("forward_parameter_count") + if parameter_count is None: + parameter_count = architecture["forward_parameters"] + return { + **common, + **history, + **work, + "forward_parameter_count": int(parameter_count), + "peak_memory_allocated_bytes": int( + hardware_record["peak_memory_allocated_bytes"] + ), + "total_wall_seconds": float( + record.get("total_wall_seconds") + if "total_wall_seconds" in record + else record["timing"]["total_timed_wall_s"] + ), + } + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--selector", default=DEFAULT_SELECTOR) + parser.add_argument( + "--out", default="results/resnet_crossover/r2_audit.json" + ) + args = parser.parse_args() + selector = selector_report(args.selector) + jobs = r2_jobs(selector["selected_rates"]) + launch_path = os.path.join( + ROOT, "results", "resnet_crossover", "r2_launch.json" + ) + assert os.path.isfile(launch_path), "missing ResNet R2 launch lock" + launch = read_json(launch_path) + assert launch["stage"] == "r2" + assert launch["selector"] == selector + assert launch["registry_sha256"] == registry_sha256(jobs) + assert launch["num_jobs"] == 27 + assert launch["allowed_physical_gpus"] == [5, 7] + source = launch["source"] + records = [audit_job(job, source, selector) for job in jobs] + assert len(records) == 27 + assert len({record["cell_id"] for record in records}) == 27 + assert { + (record["method"], record["depth"]) for record in records + } == {(method, depth) for method in METHODS for depth in DEPTHS} + for depth in DEPTHS: + counts = { + record["forward_parameter_count"] + for record in records + if record["depth"] == depth + and record["forward_parameter_count"] is not None + } + assert len(counts) <= 1, f"forward parameter mismatch at depth {depth}" + failures = [ + record["cell_id"] for record in records if not record["finite"] + ] + report = { + "audit_status": "passed", + "stage": "resnet_crossover_r2", + "complete_grid": True, + "failure_retaining": True, + "num_expected_cells": 27, + "num_audited_cells": len(records), + "num_finite_cells": len(records) - len(failures), + "failed_or_incomplete_cells": failures, + "test_policy": "none", + "source": source, + "selector": selector, + "launch_lock": { + "path": os.path.relpath(launch_path, ROOT), + "sha256": sha256(launch_path), + "registry_sha256": launch["registry_sha256"], + }, + "records": records, + } + os.makedirs(os.path.dirname(os.path.abspath(args.out)), exist_ok=True) + with open(args.out, "w", encoding="utf-8") as handle: + json.dump(report, handle, indent=2, sort_keys=True) + handle.write("\n") + print( + json.dumps( + { + "audit_status": report["audit_status"], + "num_audited_cells": len(records), + "num_finite_cells": report["num_finite_cells"], + "failed_or_incomplete_cells": failures, + }, + indent=2, + sort_keys=True, + ) + ) + + +if __name__ == "__main__": + main() |