Visual Studio Code + Claude Sonnet 4<\/p>\n\n\n\n
bgilbert1984: Proceed with next LaTeX paper for ‘\/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue\/code\/ensemble_ml_classifier.py’: ‘Fallback Paths: Hierarchical \u2192 Frequency-Based Rescue \u2014 When parent super().classify_signal() fails, you drop to SignalProcessor frequency classification; figs: failure modes and rescue rate. Hooks: exception branch in classify_signal().’ > Let\u2019s spin up the \u201cfallback\u201d paper + bench in the same house style as your suite: grayscale figs, crop-safe setup badges, JSON\u2192captions, and a clean IEEEtran note.<\/p>\n\n\n\n
Below are drop-in files\/patches<\/strong> you can paste into your repo. They do three things:<\/p>\n\n\n\n Wrap the parent call; on exception, run a fast frequency classifier and annotate the path\/timing.<\/p>\n<\/blockquote>\n\n\n\n If your parent classifier already returns Create Create Create This gives you a clean \u201chierarchical\u2192frequency rescue\u201d<\/strong> paper with two honest figures, CI bands, and the exact hook reviewers will ask about: the GitHub Copilot: I’ll help you create the complete “Fallback Paths” paper with the hierarchical frequency-based rescue system. Let me implement all the components you’ve outlined.<\/p>\n\n\n\n Created 5 todos<\/p>\n\n\n\n Starting (1\/5) Patch ensemble_ml_classifier.py with rescue logic<\/em><\/p>\n\n\n\n Let me first examine the current Read <\/a>, lines 1 to 100<\/p>\n\n\n\n Read <\/a>, lines 200 to 300<\/p>\n\n\n\n Now let me search for the classify_signal method to see its current structure:<\/p>\n\n\n\n Searched text for Read <\/a>, lines 340 to 450<\/p>\n\n\n\n Read <\/a>, lines 450 to 550<\/p>\n\n\n\n Now I understand the structure. Let me check the imports and see if we need time module:<\/p>\n\n\n\n Read <\/a>, lines 1 to 30<\/p>\n\n\n\n Now I’ll apply the rescue patch to add the time import and modify the Using “Replace String in File”<\/p>\n\n\n\n Good, now let me find the Searched text for Read <\/a>, lines 60 to 120<\/p>\n\n\n\n Now let me add the rescue_mode parameter to the init<\/strong> method:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Now I need to completely replace the Read <\/a>, lines 620 to 680<\/p>\n\n\n\n Let me find where the classify_signal method ends by looking for the next method:<\/p>\n\n\n\n Searched text for So the Read <\/a>, lines 345 to 586<\/p>\n\n\n\n Now I’ll replace the entire classify_signal method with the new rescue-enabled version:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Starting (2\/5) Create fallback rescue benchmark script<\/em><\/p>\n\n\n\n Great! Now let me create the benchmark script to simulate failure modes and measure rescue performance:<\/p>\n\n\n\n Created <\/a><\/p>\n\n\n\n Created <\/a><\/p>\n\n\n\n Starting (3\/5) Create captions builder script<\/em><\/p>\n\n\n\n Now let me create the captions builder script:<\/p>\n\n\n\n Created <\/a><\/p>\n\n\n\n Starting (4\/5) Create main LaTeX paper<\/em><\/p>\n\n\n\n Now let me create the main LaTeX paper:<\/p>\n\n\n\n Created <\/a><\/p>\n\n\n\n Starting (5\/5) Run complete benchmark and build pipeline<\/em><\/p>\n\n\n\n Now let me run the complete benchmark and build pipeline. I’ll create the necessary directories first:<\/p>\n\n\n\n Created <\/a><\/p>\n\n\n\n Continue to iterate? Made changes.<\/p>\n\n\n\n bgilbert1984: @agent Continue: “Continue to iterate?”<\/p>\n\n\n\n GitHub Copilot: Created <\/a><\/p>\n\n\n\n Let me configure the Python environment for this workspace and then run the benchmark:<\/p>\n\n\n\n Configuring a Python Environment Now let me run the benchmark script:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/run_fallback_rescue_bench.py –runs 5 –n 24000 –snr-db 8.0 –p-fail 0.0 0.1 0.2 0.4 0.6 0.8 –modes load_error shape_mismatch nan_input timeout class_mismatch –output data\/fallback_metrics.json<\/p>\n\n\n\n bgilbert1984: Ensure GPU is being use with ‘cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/run_fallback_rescue_bench.py –runs 5 –n 24000 –snr-db 8.0 –p-fail 0.0 0.1 0.2 0.4 0.6 0.8 –modes load_error shape_mismatch nan_input timeout class_mismatch –output data\/fallback_metrics.json’ > Absolutely\u2014here\u2019s a tight patch that makes the fallback benchmark run on your RTX 3060 (12 GB)<\/strong>. It adds a What changes with If you prefer a strict failure on non-GPU hosts, keep If you used my earlier If you want the classifier\u2019s<\/strong> in-library GitHub Copilot: Perfect! Let me apply the GPU-enabled patch to the benchmark script to utilize your RTX 3060. I’ll update the script to support CUDA for both signal synthesis and the frequency rescue FFT operations.<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Now I need to add the device argument and update the main function:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Now let me update the signal generation call and the rescue FFT logic:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Now let me update the configuration output and figure annotations to include the GPU information:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n I need to fix a small issue with the frequency grid generation:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Now let me also update the captions builder to include the GPU information:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Perfect! Now let’s run the GPU-enabled benchmark on your RTX 3060:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/run_fallback_rescue_bench.py –runs 5 –n 24000 –snr-db 8.0 –p-fail 0.0 0.1 0.2 0.4 0.6 0.8 –modes load_error shape_mismatch nan_input timeout class_mismatch –device cuda –output data\/fallback_metrics.json<\/p>\n\n\n\n It seems PyTorch doesn’t have CUDA support in this environment. Let me run it on CPU instead and then build the rest of the pipeline:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/run_fallback_rescue_bench.py –runs 5 –n 24000 –snr-db 8.0 –p-fail 0.0 0.1 0.2 0.4 0.6 0.8 –modes load_error shape_mismatch nan_input timeout class_mismatch –device cpu –output data\/fallback_metrics.json<\/p>\n\n\n\n Excellent! The benchmark completed successfully. Now let me generate the captions and build the LaTeX paper:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/build_captions_fallback.py –metrics data\/fallback_metrics.json –out data\/captions_fallback.tex<\/p>\n\n\n\n Let me check what’s in the metrics file and fix the captions script:<\/p>\n\n\n\n Read <\/a>, lines 1 to 20<\/p>\n\n\n\n Let me check if the gpu field is included:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && head -30 data\/fallback_metrics.json | grep -A 5 -B 5 device<\/p>\n\n\n\n Good! The device and gpu fields are there. Let me debug the caption script by checking what’s in the TEX template:<\/p>\n\n\n\n Read <\/a>, lines 1 to 10<\/p>\n\n\n\n The issue is that I need to convert the modes list to a string. Let me fix the captions script:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/build_captions_fallback.py –metrics data\/fallback_metrics.json –out data\/captions_fallback.tex<\/p>\n\n\n\n Let me debug this by printing what’s expected vs what’s provided. Let me create a simple test script:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python -c “ TEX = r”’ d = json.loads(Path(‘data\/fallback_metrics.json’).read_text()) print(‘Config keys:’, list(cfg.keys())) data = dict(runs=cfg[‘runs’], n=cfg[‘n’], snr_db=cfg[‘snr_db’], try: I see the issue – the LaTeX has Using “Replace String in File”<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/build_captions_fallback.py –metrics data\/fallback_metrics.json –out data\/captions_fallback.tex<\/p>\n\n\n\n Great! Now let’s verify the generated captions:<\/p>\n\n\n\n Read <\/a><\/p>\n\n\n\n Perfect! Now let’s build the LaTeX paper. First, let me create the data directory if needed:<\/p>\n\n\n\n Created <\/a><\/p>\n\n\n\n Now let’s compile the LaTeX paper:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && pdflatex -interaction=nonstopmode -halt-on-error main_fallback.tex<\/p>\n\n\n\n There’s an issue with the LaTeX captions having dollar signs. Let me fix the captions by properly escaping the math mode:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Continue to iterate? Made changes.<\/p>\n\n\n\n bgilbert1984: @agent Continue: “Continue to iterate?”<\/p>\n\n\n\n GitHub Copilot: Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/build_captions_fallback.py –metrics data\/fallback_metrics.json –out data\/captions_fallback.tex<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && pdflatex -interaction=nonstopmode -halt-on-error main_fallback.tex<\/p>\n\n\n\n Let me check what’s in the generated captions file:<\/p>\n\n\n\n Read <\/a><\/p>\n\n\n\n I think the issue is with the underscore in the modes. Let me escape the underscores:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/build_captions_fallback.py –metrics data\/fallback_metrics.json –out data\/captions_fallback.tex<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && pdflatex -interaction=nonstopmode -halt-on-error main_fallback.tex<\/p>\n\n\n\n Excellent! The PDF was successfully generated! There are some overfull hbox warnings due to long mode names, but the paper compiled successfully. Let me complete the final todo and then check the generated files:<\/p>\n\n\n\n Completed (5\/5) Run complete benchmark and build pipeline<\/em><\/p>\n\n\n\n Perfect! Let me verify what files were generated and show you what we’ve accomplished:<\/p>\n\n\n\n Read <\/a><\/p>\n\n\n\n Made changes.<\/p>\n\n\n\n bgilbert1984: Good news: the PDF built cleanly. The two things you\u2019re seeing:<\/p>\n\n\n\n Here\u2019s a tiny, surgical fix for both.<\/p>\n\n\n\n Then regenerate the captions + rebuild the paper:<\/p>\n\n\n\n That kills the overfull hbox by letting TeX break after commas, and it stamps the actual GPU name into From the same If CUDA is False<\/strong> but Run the bench on GPU and regenerate everything:<\/p>\n\n\n\n You should now see:<\/p>\n\n\n\n If you still want even tighter captions, drop the enumerated modes entirely and show a count:<\/p>\n\n\n\n Want me to wire a Table I (median rescue latency per mode)<\/strong> here too, like we did for the mismatch paper? I can drop a 15-line helper that reads your JSON and emits a tiny table macro.<\/p>\n\n\n\n GitHub Copilot: Perfect! Let me implement the surgical fixes for GPU-awareness and line-break friendly captions. I’ll start with updating the captions builder script:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Now let’s check what CUDA support is available in the current environment:<\/p>\n\n\n\n Ran terminal command: nvidia-smi<\/p>\n\n\n\n Great! The RTX 3060 is available. Now let’s check PyTorch CUDA support:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python -c “ The PyTorch installation doesn’t have CUDA support. Let’s install a CUDA-enabled version:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && source \/home\/bgilbert\/rf_quantum_env\/bin\/activate && pip uninstall -y torch torchvision torchaudio<\/p>\n\n\n\n There’s a permission issue. Let me try a different approach – let’s install the CUDA version directly over the CPU version:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && source \/home\/bgilbert\/rf_quantum_env\/bin\/activate && pip install –index-url https:\/\/download.pytorch.org\/whl\/cu121 torch torchvision torchaudio –force-reinstall<\/p>\n\n\n\n Let me try with –user flag:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && source \/home\/bgilbert\/rf_quantum_env\/bin\/activate && pip install –index-url https:\/\/download.pytorch.org\/whl\/cu121 torch torchvision torchaudio –upgrade –user<\/p>\n\n\n\n The virtual environment has permission issues. Let me proceed with the current CPU version but use the improved captions. Let me regenerate the captions with the breakable modes and then show you the complete solution:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/build_captions_fallback.py –metrics data\/fallback_metrics.json –out data\/captions_fallback.tex<\/p>\n\n\n\n Read <\/a><\/p>\n\n\n\n Perfect! Now let’s rebuild the PDF with the improved captions:<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && pdflatex -interaction=nonstopmode -halt-on-error main_fallback.tex<\/p>\n\n\n\n There’s still an issue with the LaTeX syntax. Let me fix the backslash escaping in the TeXtemplate:<\/p>\n\n\n\n Using “Replace String in File”<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && \/home\/bgilbert\/rf_quantum_env\/bin\/python scripts\/build_captions_fallback.py –metrics data\/fallback_metrics.json –out data\/captions_fallback.tex<\/p>\n\n\n\n Ran terminal command: cd \/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue && pdflatex -interaction=nonstopmode -halt-on-error main_fallback.tex<\/p>\n\n\n\n Let me check what’s in the actual generated captions file:<\/p>\n\n\n\n\n
classify_signal()<\/code> with a rescue branch that falls back to a frequency classifier when the parent path throws.<\/li>\n\n\n\n
\n\n\n\n1) Patch: exception\u2192rescue in
classify_signal()<\/code><\/h1>\n\n\n\n\n
--- a\/ensemble_ml_classifier.py\n+++ b\/ensemble_ml_classifier.py\n@@\n import time, numpy as np\n@@\n class EnsembleMLClassifier(...):\n def __init__(self, *args, rescue_mode='frequency', **kw):\n super().__init__(*args, **kw)\n self.rescue_mode = rescue_mode\n@@\n- def classify_signal(self, iq):\n- return super().classify_signal(iq)\n+ def classify_signal(self, iq):\n+ \"\"\"Primary path -> on failure, frequency-based rescue.\"\"\"\n+ t0 = time.perf_counter()\n+ try:\n+ out = super().classify_signal(iq)\n+ out = dict(out) if isinstance(out, dict) else {\"label\": out}\n+ out.setdefault(\"path\", \"primary\")\n+ out.setdefault(\"latency_ms\", (time.perf_counter()-t0)*1e3)\n+ return out\n+ except Exception as e:\n+ if self.rescue_mode != 'frequency':\n+ raise\n+ rt0 = time.perf_counter()\n+ label, conf = self._frequency_rescue(iq)\n+ return {\n+ \"label\": label,\n+ \"confidence\": float(conf),\n+ \"path\": \"rescue\",\n+ \"error\": str(e)[:200],\n+ \"latency_ms\": (time.perf_counter()-t0)*1e3,\n+ \"rescue_ms\": (time.perf_counter()-rt0)*1e3,\n+ }\n+\n+ # --- Minimal frequency rescue: spectral centroid + band energy ratios\n+ def _frequency_rescue(self, iq, nfft=256):\n+ x = np.asarray(iq).astype(np.complex64)\n+ spec = np.abs(np.fft.fftshift(np.fft.fft(x, n=nfft)))\n+ spec = spec \/ (spec.max() + 1e-9)\n+ freqs = np.linspace(-0.5, 0.5, nfft, endpoint=False)\n+ centroid = float((spec * np.abs(freqs)).sum() \/ (spec.sum() + 1e-9))\n+ low = spec[(np.abs(freqs) < 0.1)].mean()\n+ mid = spec[(np.abs(freqs) >= 0.1) & (np.abs(freqs) < 0.2)].mean()\n+ high = spec[(np.abs(freqs) >= 0.2)].mean()\n+ # Heuristic mapping\u2014tweak per your class set\n+ scores = {\n+ \"AM\": 0.6*low + 0.2*mid + 0.2*high,\n+ \"FM\": 0.2*low + 0.6*mid + 0.2*high,\n+ \"PSK\": 0.2*low + 0.2*mid + 0.6*high,\n+ \"CW\": 1.0 - centroid,\n+ }\n+ # restrict to known runtime mapping if provided\n+ classes = getattr(self, \"class_mapping\", None) or list(scores.keys())\n+ scores = {k: v for k, v in scores.items() if k in classes}\n+ label, conf = max(scores.items(), key=lambda kv: kv[1])\n+ return label, conf<\/code><\/pre>\n\n\n\n\n
{\"label\":..., \"confidence\":...}<\/code>, this wrapper preserves it and just tags path<\/code>.<\/p>\n<\/blockquote>\n\n\n\n
\n\n\n\n2) Bench script: simulate failures<\/strong> & measure rescue<\/h1>\n\n\n\n
scripts\/run_fallback_rescue_bench.py<\/code>:<\/p>\n\n\n\n# scripts\/run_fallback_rescue_bench.py\nimport argparse, json, time\nimport numpy as np\nfrom pathlib import Path\nimport matplotlib.pyplot as plt\n\n# ---- tiny signal sim (reuse your generator if available)\ndef sim_iq(N=128, snr_db=8.0, kind=\"AM\", seed=0):\n rng = np.random.RandomState(seed)\n t = np.arange(N)\n if kind == \"AM\":\n m = 0.5 + 0.5*np.cos(2*np.pi*0.07*t)\n s = m * np.exp(1j*2*np.pi*0.18*t)\n elif kind == \"FM\":\n m = np.cos(2*np.pi*0.05*t)\n s = np.exp(1j*2*np.pi*(0.18*t + 0.3*np.cumsum(m)))\n elif kind == \"PSK\":\n bits = rng.randint(0, 4, size=N)\n s = np.exp(1j*np.pi\/2 * bits)\n else: # CW\n s = np.exp(1j*2*np.pi*0.08*t)\n # noise\n sigp = np.mean(np.abs(s)**2)\n noip = sigp \/ (10**(snr_db\/10))\n n = np.sqrt(noip\/2)*(rng.randn(N)+1j*rng.randn(N))\n return s + n\n\n# ---- failure injector for primary path\ndef primary_predict(iq, label, p_fail, mode, rng):\n \"\"\"Return (ok, y_pred, latency_ms) or raise simulated failure to trigger rescue.\"\"\"\n t0 = time.perf_counter()\n if rng.rand() < p_fail:\n if mode == \"load_error\":\n raise RuntimeError(\"checkpoint load error\")\n if mode == \"shape_mismatch\":\n raise ValueError(\"tensor shape mismatch\")\n if mode == \"nan_input\":\n raise FloatingPointError(\"NaN in input\")\n if mode == \"timeout\":\n time.sleep(0.02) # emulate hang\n raise TimeoutError(\"inference timeout\")\n if mode == \"class_mismatch\":\n raise KeyError(\"class_mapping mismatch\")\n raise RuntimeError(\"generic failure\")\n # trivial primary predictor: gets 70% right baseline\n ok = True\n y_pred = label if rng.rand() < 0.70 else rng.choice([\"AM\",\"FM\",\"PSK\",\"CW\"])\n return ok, y_pred, (time.perf_counter()-t0)*1e3\n\n# ---- plotting utils (grayscale & setup badge)\ndef annotate(ax, text, fontsize=7):\n ax.text(0.99, 0.02, text, ha='right', va='bottom',\n transform=ax.transAxes, fontsize=fontsize,\n bbox=dict(boxstyle=\"round,pad=0.2\", fc=\"white\", ec=\"0.5\", lw=0.5))\n\ndef main():\n ap = argparse.ArgumentParser()\n ap.add_argument(\"--runs\", type=int, default=5)\n ap.add_argument(\"--n\", type=int, default=24000)\n ap.add_argument(\"--snr-db\", type=float, default=8.0)\n ap.add_argument(\"--p-fail\", type=float, nargs=\"+\", default=[0.0,0.1,0.2,0.4,0.6,0.8])\n ap.add_argument(\"--modes\", nargs=\"+\",\n default=[\"load_error\",\"shape_mismatch\",\"nan_input\",\"timeout\",\"class_mismatch\"])\n ap.add_argument(\"--seed\", type=int, default=7)\n ap.add_argument(\"--output\", type=str, default=\"data\/fallback_metrics.json\")\n Path(\"figs\").mkdir(exist_ok=True, parents=True)\n Path(\"data\").mkdir(exist_ok=True, parents=True)\n args = ap.parse_args()\n\n classes = [\"AM\",\"FM\",\"PSK\",\"CW\"]\n rng0 = np.random.RandomState(args.seed)\n series = {} # mode -> list of (p_fail, rescue_rate_mean, ci, acc_final_mean, ci, extra)\n # raw to compute medians if needed\n raw = {m: {str(p): {\"rescued\":[], \"failures\":[], \"lat_rescue_ms\":[]} for p in args.p_fail} for m in args.modes}\n\n for mode in args.modes:\n xs, resc_means, resc_cis, acc_means, acc_cis = [], [], [], [], []\n for p in args.p_fail:\n resc, fails, accs, lat_rs = [], [], [], []\n for r in range(args.runs):\n rng = np.random.RandomState(args.seed + 1000*r + hash(mode)%997)\n correct = total = rescued = failed = 0\n rescue_lat = []\n for i in range(args.n\/\/args.runs):\n y = rng.choice(classes)\n iq = sim_iq(N=128, snr_db=args.snr_db, kind=y, seed=rng.randint(1<<30))\n # primary path with failure injection\n try:\n ok, y_pred, lat_ms = primary_predict(iq, y, p, mode, rng)\n pred = y_pred\n path = \"primary\"\n if pred == y: correct += 1\n except Exception:\n failed += 1\n # frequency rescue (same heuristic as in classifier)\n t1 = time.perf_counter()\n # minimal rescue impl (same as patch)\n nfft=256\n spec = np.abs(np.fft.fftshift(np.fft.fft(iq, n=nfft)))\n spec = spec\/(spec.max()+1e-9)\n freqs = np.linspace(-0.5,0.5,nfft,endpoint=False)\n centroid = float((spec*np.abs(freqs)).sum()\/(spec.sum()+1e-9))\n low = spec[(np.abs(freqs)<0.1)].mean()\n mid = spec[(np.abs(freqs)>=0.1)&(np.abs(freqs)<0.2)].mean()\n high = spec[(np.abs(freqs)>=0.2)].mean()\n scores = {\"AM\":0.6*low+0.2*mid+0.2*high,\n \"FM\":0.2*low+0.6*mid+0.2*high,\n \"PSK\":0.2*low+0.2*mid+0.6*high,\n \"CW\":1.0-centroid}\n pred = max(scores.items(), key=lambda kv: kv[1])[0]\n rlat = (time.perf_counter()-t1)*1e3\n rescue_lat.append(rlat)\n if pred == y:\n rescued += 1\n correct += 1\n total += 1\n resc.append(rescued \/ max(failed,1))\n accs.append(correct \/ total)\n fails.append(failed)\n lat_rs.extend(rescue_lat)\n # aggregate\n m_resc = float(np.mean(resc)); se_resc = float(np.std(resc, ddof=1)\/np.sqrt(len(resc)))\n m_acc = float(np.mean(accs)); se_acc = float(np.std(accs, ddof=1)\/np.sqrt(len(accs)))\n xs.append(p); resc_means.append(m_resc); resc_cis.append(1.96*se_resc)\n acc_means.append(m_acc); acc_cis.append(1.96*se_acc)\n raw[mode][str(p)][\"rescued\"] = resc\n raw[mode][str(p)][\"failures\"] = fails\n raw[mode][str(p)][\"lat_rescue_ms\"] = lat_rs\n series[mode] = dict(p_fail=xs, rescue_mean=resc_means, rescue_ci=resc_cis,\n acc_mean=acc_means, acc_ci=acc_cis)\n\n # save metrics\n out = {\n \"config\": dict(runs=args.runs, n=args.n, snr_db=args.snr_db,\n modes=args.modes, p_fail=args.p_fail, seed=args.seed),\n \"series\": series,\n \"raw\": raw\n }\n Path(args.output).write_text(json.dumps(out, indent=2))\n print(\"[OK] wrote\", args.output)\n\n # ==== Fig. 1: failure modes & rescue rate (bars per mode @ p_fail=0.4) ====\n pf = 0.4\n idx_pf = args.p_fail.index(pf) if pf in args.p_fail else np.argmin(np.abs(np.array(args.p_fail)-0.4))\n modes = args.modes\n vals = [series[m][\"rescue_mean\"][idx_pf] for m in modes]\n cis = [series[m][\"rescue_ci\"][idx_pf] for m in modes]\n plt.figure(figsize=(3.4,2.6))\n ax = plt.gca()\n ax.bar(np.arange(len(modes)), vals, yerr=cis, capsize=3, color=\"gray\")\n ax.set_xticks(np.arange(len(modes))); ax.set_xticklabels(modes, rotation=20, ha='right', fontsize=8)\n ax.set_ylabel(\"Rescue rate @ p_fail\u22480.4\"); ax.set_ylim(0,1)\n annotate(ax, f\"runs {args.runs}; N {args.n}; SNR {args.snr_db} dB; p_fail~{args.p_fail[idx_pf]:.2f}\")\n plt.tight_layout(); plt.savefig(\"figs\/failure_modes.pdf\", bbox_inches='tight')\n print(\"[OK] wrote figs\/failure_modes.pdf\")\n\n # ==== Fig. 2: rescue rate vs p_fail (lines per mode) ====\n plt.figure(figsize=(3.4,2.6))\n ax = plt.gca()\n for m in modes:\n xs = series[m][\"p_fail\"]; ys = series[m][\"rescue_mean\"]; ci = series[m][\"rescue_ci\"]\n ax.plot(xs, ys, '-o', label=m, linewidth=1.1)\n ax.fill_between(xs, np.array(ys)-np.array(ci), np.array(ys)+np.array(ci), alpha=0.15)\n ax.set_xlabel(\"Injected failure probability p_fail\")\n ax.set_ylabel(\"Rescue rate (fraction of failures corrected)\")\n ax.set_ylim(0,1)\n ax.legend(fontsize=7, ncol=2, frameon=False)\n annotate(ax, f\"runs {args.runs}; N {args.n}; SNR {args.snr_db} dB\")\n plt.tight_layout(); plt.savefig(\"figs\/rescue_rate.pdf\", bbox_inches='tight')\n print(\"[OK] wrote figs\/rescue_rate.pdf\")\nif __name__ == \"__main__\":\n main()<\/code><\/pre>\n\n\n\n
\n\n\n\n3) Captions builder \u2192 LaTeX macros<\/h1>\n\n\n\n
scripts\/build_captions_fallback.py<\/code>:<\/p>\n\n\n\n# scripts\/build_captions_fallback.py\nimport json\nfrom pathlib import Path\nimport argparse\n\nTEX = r\"\"\"\n\\newcommand{\\fbSetup}{runs \\texttt{%(runs)s}; $N$ \\texttt{%(n)s}; SNR \\texttt{%(snr_db).1f} dB; modes \\texttt{%(modes)s}}\n\\newcommand{\\fbFigOne}{Rescue rate by failure mode at $p_\\mathrm{fail}\\approx0.4$. Bars show mean with 95\\% CIs over seeds. (Setup: \\fbSetup)}\n\\newcommand{\\fbFigTwo}{Rescue rate vs injected failure probability $p_\\mathrm{fail}$. Lines are means; bands are 95\\% CIs; one curve per failure mode. (Setup: \\fbSetup)}\n\"\"\"\n\ndef main():\n ap = argparse.ArgumentParser()\n ap.add_argument(\"--metrics\", default=\"data\/fallback_metrics.json\")\n ap.add_argument(\"--out\", default=\"data\/captions_fallback.tex\")\n a = ap.parse_args()\n d = json.loads(Path(a.metrics).read_text())\n cfg = d[\"config\"]\n tex = TEX % dict(runs=cfg[\"runs\"], n=cfg[\"n\"], snr_db=cfg[\"snr_db\"],\n modes=\",\".join(cfg[\"modes\"]))\n Path(a.out).write_text(tex)\n print(\"[OK] wrote\", a.out)\n\nif __name__ == \"__main__\":\n main()<\/code><\/pre>\n\n\n\n
\n\n\n\n4) LaTeX paper (IEEEtran),
main_fallback.tex<\/code><\/h1>\n\n\n\nmain_fallback.tex<\/code>:<\/p>\n\n\n\n\\documentclass[journal]{IEEEtran}\n\\usepackage{graphicx,booktabs,amsmath,amssymb,siunitx,listings}\n\\usepackage{hyperref}\n\\begin{document}\n\\title{Fallback Paths: Hierarchical $\\rightarrow$ Frequency-Based Rescue for RF Modulation Inference}\n\\author{Ben Gilbert}\n\\maketitle\n\n\\input{data\/captions_fallback.tex}\n\n\\begin{abstract}\nWe implement a hierarchical fallback in \\texttt{classify\\_signal()} that attempts the parent deep path and, on exception, drops to a fast frequency-based classifier. On synthetic RF IQ, we inject realistic failure modes (load error, shape mismatch, NaN input, timeout, class mismatch) at probability $p_\\mathrm{fail}$ and report rescue rate and accuracy impacts. The rescue adds $<1$\\,ms median latency and corrects a majority of failed inferences across modes.\n\\end{abstract}\n\n\\section{Method}\nWe wrap the parent call in a \\texttt{try\/except}; failures trigger a frequency rescue that classifies using magnitude FFT features (centroid and band energies). The wrapper annotates \\texttt{path} $\\in\\{\\text{primary},\\text{rescue}\\}$ and timings. Failure injection covers: \\emph{load\\_error, shape\\_mismatch, nan\\_input, timeout, class\\_mismatch}. Each run uses $N$ signals, 5 seeds, and SNR as stamped in the figure badges.\n\n\\section{Results}\n\\begin{figure}[t]\\centering\n\\includegraphics[width=\\linewidth]{figs\/failure_modes.pdf}\n\\caption{\\fbFigOne}\\label{fig:modes}\n\\end{figure}\n\n\\begin{figure}[t]\\centering\n\\includegraphics[width=\\linewidth]{figs\/rescue_rate.pdf}\n\\caption{\\fbFigTwo}\\label{fig:rescue}\n\\end{figure}\n\n\\section{Discussion}\nRescue rates are highest for \\emph{load\\_error} and \\emph{timeout} (the deep model never ran), and lower for \\emph{nan\\_input} (input corruption harms spectral cues). Median rescue latency $<1$\\,ms keeps end-to-end TTFB within budget. Future work: learned frequency heads, per-class rescue policies, and integrating mismatch calibration before rescue.\n\n\\bibliographystyle{IEEEtran}\n\\small\n\\begin{thebibliography}{1}\\setlength{\\itemsep}{0pt}\n\\bibitem{your2025suite} B. Gilbert, ``Open RF ML Papers Suite,'' 2025.\n\\end{thebibliography}\n\\end{document}<\/code><\/pre>\n\n\n\n
\n\n\n\n5) Run it end-to-end<\/h1>\n\n\n\n
# 1) Bench & figs\npython scripts\/run_fallback_rescue_bench.py \\\n --runs 5 --n 24000 --snr-db 8.0 \\\n --p-fail 0.0 0.1 0.2 0.4 0.6 0.8 \\\n --modes load_error shape_mismatch nan_input timeout class_mismatch \\\n --output data\/fallback_metrics.json\n\n# 2) Captions\npython scripts\/build_captions_fallback.py --metrics data\/fallback_metrics.json\n\n# 3) Build paper\npdflatex -interaction=nonstopmode -halt-on-error main_fallback.tex<\/code><\/pre>\n\n\n\n
\n\n\n\nNotes \/ switches<\/h2>\n\n\n\n
\n
--snr-grid 0 4 8 12<\/code> and bucket in the JSON; easy to extend the script.<\/li>\n\n\n\nbuild_table_rescue.py<\/code> mirroring your mismatch table generator.<\/li>\n<\/ul>\n\n\n\nclassify_signal()<\/code> exception branch.<\/p>\n\n\n\nensemble_ml_classifier.py<\/code> to understand its structure before applying the rescue patch.<\/p>\n\n\n\ndef classify_signal<\/code> (**\/ensemble_ml_classifier.py<\/code>), 13 results<\/p>\n\n\n\n__init__<\/code> method and classify_signal<\/code> method:<\/p>\n\n\n\n__init__<\/code> method to add the rescue_mode parameter:<\/p>\n\n\n\ndef __init__<\/code> (**\/ensemble_ml_classifier.py<\/code>), 13 results<\/p>\n\n\n\nclassify_signal<\/code> method with the new rescue-enabled version. Let me first find its full extent:<\/p>\n\n\n\ndef<\/code> (**\/ensemble_ml_classifier.py<\/code>), 20 results<\/p>\n\n\n\nclassify_signal<\/code> method starts at line 345 and the next method _create_temporal_input<\/code> is at line 587. Let me replace the entire classify_signal method with the rescue-enabled version:<\/p>\n\n\n\n
[object Object]<\/p>\n\n\n\n
Completed with input: {
“resourcePath”: “\/home\/bgilbert\/paper_Fallback_Paths_Hierarchical_Frequency-Based_Rescue”
}<\/p>\n\n\n\n--device<\/code> flag (cpu|cuda<\/code>), moves the rescue path\u2019s FFT\/features to PyTorch CUDA<\/strong>, syncs the GPU for honest timings, and stamps the device into the figure badges. It keeps full CPU fallback if you pass --device cpu<\/code> (or if CUDA isn\u2019t available).<\/p>\n\n\n\n
\n\n\n\nPatch: GPU-enable
run_fallback_rescue_bench.py<\/code><\/h1>\n\n\n\n--- a\/scripts\/run_fallback_rescue_bench.py\n+++ b\/scripts\/run_fallback_rescue_bench.py\n@@\n-import argparse, json, time\n+import argparse, json, time, sys\n import numpy as np\n from pathlib import Path\n import matplotlib.pyplot as plt\n+try:\n+ import torch\n+except Exception:\n+ torch = None\n\n # ---- tiny signal sim (reuse your generator if available)\n-def sim_iq(N=128, snr_db=8.0, kind=\"AM\", seed=0):\n+def sim_iq_cpu(N=128, snr_db=8.0, kind=\"AM\", seed=0):\n rng = np.random.RandomState(seed)\n t = np.arange(N)\n if kind == \"AM\":\n m = 0.5 + 0.5*np.cos(2*np.pi*0.07*t)\n s = m * np.exp(1j*2*np.pi*0.18*t)\n@@\n n = np.sqrt(noip\/2)*(rng.randn(N)+1j*rng.randn(N))\n return s + n\n+\n+def sim_iq_cuda(N=128, snr_db=8.0, kind=\"AM\", seed=0, device=\"cuda\"):\n+ # torch-based version to avoid CPU<->GPU shuttling; keeps iq on device as complex64\n+ g = torch.Generator(device=device)\n+ g.manual_seed(int(seed) & 0x7FFFFFFF)\n+ t = torch.arange(N, device=device, dtype=torch.float32)\n+ if kind == \"AM\":\n+ m = 0.5 + 0.5*torch.cos(2*torch.pi*0.07*t)\n+ s = m * torch.exp(1j*2*torch.pi*0.18*t)\n+ elif kind == \"FM\":\n+ m = torch.cos(2*torch.pi*0.05*t)\n+ s = torch.exp(1j*2*torch.pi*(0.18*t + 0.3*torch.cumsum(m, dim=0)))\n+ elif kind == \"PSK\":\n+ bits = torch.randint(0, 4, (N,), generator=g, device=device)\n+ s = torch.exp(1j*torch.pi\/2 * bits.to(torch.float32))\n+ else: # CW\n+ s = torch.exp(1j*2*torch.pi*0.08*t)\n+ sigp = (s.abs()**2).mean()\n+ noip = sigp \/ (10**(snr_db\/10))\n+ n = torch.randn(N, device=device, dtype=torch.float32, generator=g)\n+ n = (n + 1j*torch.randn(N, device=device, dtype=torch.float32, generator=g)) * torch.sqrt(noip\/2)\n+ return (s + n).to(torch.complex64)\n@@\n-def annotate(ax, text, fontsize=7):\n+def annotate(ax, text, fontsize=7):\n ax.text(0.99, 0.02, text, ha='right', va='bottom',\n transform=ax.transAxes, fontsize=fontsize,\n bbox=dict(boxstyle=\"round,pad=0.2\", fc=\"white\", ec=\"0.5\", lw=0.5))\n\n def main():\n ap = argparse.ArgumentParser()\n@@\n- ap.add_argument(\"--seed\", type=int, default=7)\n+ ap.add_argument(\"--seed\", type=int, default=7)\n+ ap.add_argument(\"--device\", type=str, default=\"cpu\", choices=[\"cpu\",\"cuda\"],\n+ help=\"Compute device for rescue path and signal sim (FFT, features)\")\n ap.add_argument(\"--output\", type=str, default=\"data\/fallback_metrics.json\")\n Path(\"figs\").mkdir(exist_ok=True, parents=True)\n Path(\"data\").mkdir(exist_ok=True, parents=True)\n args = ap.parse_args()\n\n+ # Device gating \/ checks\n+ dev = args.device\n+ if dev == \"cuda\":\n+ if torch is None:\n+ print(\"[ERR] PyTorch not available but --device cuda was requested.\", file=sys.stderr); sys.exit(1)\n+ if not torch.cuda.is_available():\n+ print(\"[ERR] CUDA not available but --device cuda was requested.\", file=sys.stderr); sys.exit(1)\n+ cuda_name = torch.cuda.get_device_name(0)\n+ torch.set_default_dtype(torch.float32)\n+ torch.backends.cudnn.benchmark = True\n+ else:\n+ cuda_name = \"CPU\"\n+\n classes = [\"AM\",\"FM\",\"PSK\",\"CW\"]\n rng0 = np.random.RandomState(args.seed)\n series = {} # mode -> list of (p_fail, rescue_rate_mean, ci, acc_final_mean, ci, extra)\n # raw to compute medians if needed\n raw = {m: {str(p): {\"rescued\":[], \"failures\":[], \"lat_rescue_ms\":[]} for p in args.p_fail} for m in args.modes}\n@@\n- for i in range(args.n\/\/args.runs):\n+ for i in range(args.n\/\/args.runs):\n y = rng.choice(classes)\n- iq = sim_iq(N=128, snr_db=args.snr_db, kind=y, seed=rng.randint(1<<30))\n+ if dev == \"cuda\":\n+ iq = sim_iq_cuda(N=128, snr_db=args.snr_db, kind=y,\n+ seed=int(rng.randint(1<<30)), device=\"cuda\")\n+ else:\n+ iq = sim_iq_cpu(N=128, snr_db=args.snr_db, kind=y, seed=rng.randint(1<<30))\n # primary path with failure injection\n try:\n ok, y_pred, lat_ms = primary_predict(iq, y, p, mode, rng)\n pred = y_pred\n path = \"primary\"\n if pred == y: correct += 1\n except Exception:\n failed += 1\n- # frequency rescue (same heuristic as in classifier)\n- t1 = time.perf_counter()\n- # minimal rescue impl (same as patch)\n- nfft=256\n- spec = np.abs(np.fft.fftshift(np.fft.fft(iq, n=nfft)))\n- spec = spec\/(spec.max()+1e-9)\n- freqs = np.linspace(-0.5,0.5,nfft,endpoint=False)\n- centroid = float((spec*np.abs(freqs)).sum()\/(spec.sum()+1e-9))\n- low = spec[(np.abs(freqs)<0.1)].mean()\n- mid = spec[(np.abs(freqs)>=0.1)&(np.abs(freqs)<0.2)].mean()\n- high = spec[(np.abs(freqs)>=0.2)].mean()\n- scores = {\"AM\":0.6*low+0.2*mid+0.2*high,\n- \"FM\":0.2*low+0.6*mid+0.2*high,\n- \"PSK\":0.2*low+0.2*mid+0.6*high,\n- \"CW\":1.0-centroid}\n- pred = max(scores.items(), key=lambda kv: kv[1])[0]\n- rlat = (time.perf_counter()-t1)*1e3\n+ # frequency rescue (CUDA if requested)\n+ t1 = time.perf_counter()\n+ nfft = 256\n+ if dev == \"cuda\":\n+ if iq.dtype != torch.complex64:\n+ iq = iq.to(torch.complex64, non_blocking=True)\n+ spec = torch.fft.fft(iq, n=nfft)\n+ spec = torch.fft.fftshift(spec).abs()\n+ spec = spec \/ (spec.max() + 1e-9)\n+ # frequency grid on device\n+ freqs = torch.linspace(-0.5, 0.5, nfft, device=\"cuda\", dtype=torch.float32, requires_grad=False)[:-0] # length nfft\n+ centroid = ((spec * freqs.abs()).sum() \/ (spec.sum() + 1e-9)).item()\n+ mask_low = (freqs.abs() < 0.1)\n+ mask_mid = (freqs.abs() >= 0.1) & (freqs.abs() < 0.2)\n+ mask_high = (freqs.abs() >= 0.2)\n+ low = spec[mask_low].mean().item()\n+ mid = spec[mask_mid].mean().item()\n+ high = spec[mask_high].mean().item()\n+ if torch.cuda.is_available():\n+ torch.cuda.synchronize()\n+ else:\n+ iq_np = np.asarray(iq)\n+ spec = np.abs(np.fft.fftshift(np.fft.fft(iq_np, n=nfft)))\n+ spec = spec\/(spec.max()+1e-9)\n+ freqs = np.linspace(-0.5,0.5,nfft,endpoint=False)\n+ centroid = float((spec*np.abs(freqs)).sum()\/(spec.sum()+1e-9))\n+ low = float(spec[(np.abs(freqs)<0.1)].mean())\n+ mid = float(spec[(np.abs(freqs)>=0.1)&(np.abs(freqs)<0.2)].mean())\n+ high = float(spec[(np.abs(freqs)>=0.2)].mean())\n+ scores = {\"AM\":0.6*low+0.2*mid+0.2*high,\n+ \"FM\":0.2*low+0.6*mid+0.2*high,\n+ \"PSK\":0.2*low+0.2*mid+0.6*high,\n+ \"CW\":1.0-centroid}\n+ pred = max(scores.items(), key=lambda kv: kv[1])[0]\n+ rlat = (time.perf_counter()-t1)*1e3\n rescue_lat.append(rlat)\n if pred == y:\n rescued += 1\n correct += 1\n total += 1\n@@\n- out = {\n- \"config\": dict(runs=args.runs, n=args.n, snr_db=args.snr_db,\n- modes=args.modes, p_fail=args.p_fail, seed=args.seed),\n+ out = {\n+ \"config\": dict(runs=args.runs, n=args.n, snr_db=args.snr_db,\n+ modes=args.modes, p_fail=args.p_fail, seed=args.seed,\n+ device=dev, gpu=cuda_name),\n \"series\": series,\n \"raw\": raw\n }\n@@\n- annotate(ax, f\"runs {args.runs}; N {args.n}; SNR {args.snr_db} dB; p_fail~{args.p_fail[idx_pf]:.2f}\")\n+ annotate(ax, f\"runs {args.runs}; N {args.n}; SNR {args.snr_db} dB; device {cuda_name}; p_fail~{args.p_fail[idx_pf]:.2f}\")\n@@\n- annotate(ax, f\"runs {args.runs}; N {args.n}; SNR {args.snr_db} dB\")\n+ annotate(ax, f\"runs {args.runs}; N {args.n}; SNR {args.snr_db} dB; device {cuda_name}\")<\/code><\/pre>\n\n\n\n
\n\n\n\nRun it on your RTX 3060 (12 GB)<\/h1>\n\n\n\n
# (Optional) ensure only the 3060 is visible\nexport CUDA_VISIBLE_DEVICES=0\n\n# GPU run (CUDA)\npython scripts\/run_fallback_rescue_bench.py \\\n --runs 5 --n 24000 --snr-db 8.0 \\\n --p-fail 0.0 0.1 0.2 0.4 0.6 0.8 \\\n --modes load_error shape_mismatch nan_input timeout class_mismatch \\\n --device cuda \\\n --output data\/fallback_metrics.json<\/code><\/pre>\n\n\n\n--device cuda<\/code>:<\/p>\n\n\n\n\n
rescue_ms<\/code> is honest.<\/li>\n\n\n\ndevice NVIDIA GeForce RTX 3060<\/code> (or your exact CUDA string).<\/li>\n<\/ul>\n\n\n\n\n
--device cuda<\/code>; the script now exits with a clear message if CUDA isn\u2019t present.<\/p>\n<\/blockquote>\n\n\n\n
\n\n\n\n(Optional) Ensure the paper<\/em> captions include the device<\/h1>\n\n\n\n
build_captions_fallback.py<\/code>, it\u2019ll now inject device<\/code> automatically (it reads config.gpu<\/code>). If you want to hard-stamp it in LaTeX body text too, just add a one-liner like:<\/p>\n\n\n\n\\emph{Device:} NVIDIA RTX 3060 (12\\,GB).<\/code><\/pre>\n\n\n\n
\n\n\n\n_frequency_rescue()<\/code> to run on CUDA as well (outside the benchmark), I can drop a tiny torch<\/code> branch there too so production path matches the bench.<\/p>\n\n\n\n
import json
from pathlib import Path<\/p>\n\n\n\n
\\newcommand{\\fbSetup}{runs \\texttt{%(runs)s}; \\$N\\$ \\texttt{%(n)s}; SNR \\texttt{%(snr_db).1f} dB; device \\texttt{%(gpu)s}; modes \\texttt{%(modes)s}}
\\newcommand{\\fbFigOne}{Rescue rate by failure mode at \\$p_\\mathrm{fail}\\approx0.4\\$. Bars show mean with 95% CIs over seeds. (Setup: \\fbSetup)}
\\newcommand{\\fbFigTwo}{Rescue rate vs injected failure probability \\$p_\\mathrm{fail}\\$. Lines are means; bands are 95% CIs; one curve per failure mode. (Setup: \\fbSetup)}
”’<\/p>\n\n\n\n
cfg = d[‘config’]
modes_str = ‘,’.join(cfg[‘modes’])<\/p>\n\n\n\n
print(‘runs:’, cfg[‘runs’])
print(‘n:’, cfg[‘n’])
print(‘snr_db:’, cfg[‘snr_db’])
print(‘gpu:’, cfg.get(‘gpu’, ‘CPU’))
print(‘modes_str:’, modes_str)<\/p>\n\n\n\n
gpu=cfg.get(‘gpu’, ‘CPU’), modes=modes_str)
print(‘Data dict:’, data)<\/p>\n\n\n\nTry the format<\/h1>\n\n\n\n
result = TEX % data
print(‘SUCCESS’)
print(result)
except Exception as e:
print(‘ERROR:’, e)
“<\/p>\n\n\n\n$<\/code> symbols which are being interpreted as format placeholders. Let me fix the TEX template by using different string formatting:<\/p>\n\n\n\n
[object Object]<\/p>\n\n\n\n\n
--device cuda<\/code>, the figures + captions will stamp the RTX 3060<\/strong> string.<\/li>\n\n\n\n\\texttt{load_error,shape_mismatch,...}<\/code> \u2192 that\u2019s just TeX refusing to break a long monospaced, comma-separated token. Make the commas breakable (or shorten the string) and it\u2019ll go away.<\/li>\n<\/ol>\n\n\n\n
\n\n\n\nPatch: make captions GPU-aware and line-break friendly<\/h2>\n\n\n\n
scripts\/build_captions_fallback.py<\/code><\/strong> \u2014 update the template and add breakable commas:<\/p>\n\n\n\n# scripts\/build_captions_fallback.py\nimport json\nfrom pathlib import Path\nimport argparse\n\nTEX = r\"\"\"\n\\newcommand{\\fbGPU}{\\texttt{%(gpu)s}}\n\\newcommand{\\fbSetup}{device \\fbGPU; runs \\texttt{%(runs)s}; $N$ \\texttt{%(n)s}; SNR \\texttt{%(snr_db).1f} dB; modes \\texttt{%(modes)s}}\n\\newcommand{\\fbFigOne}{Rescue rate by failure mode at $p_\\mathrm{fail}\\approx0.4$. Bars: mean with 95\\%% CIs over seeds. (Setup: \\fbSetup)}\n\\newcommand{\\fbFigTwo}{Rescue rate vs injected failure probability $p_\\mathrm{fail}$. Lines: mean; bands: 95\\%% CIs; one curve per failure mode. (Setup: \\fbSetup)}\n\"\"\"\n\ndef make_breakable_modes(modes_list, max_show=5):\n # 1) keep it short; 2) allow breaks after commas inside \\texttt\n modes = modes_list[:max_show]\n s = \",\\\\allowbreak \".join(modes)\n if len(modes_list) > max_show:\n s += \",\\\\allowbreak \u2026\"\n return s\n\ndef main():\n ap = argparse.ArgumentParser()\n ap.add_argument(\"--metrics\", default=\"data\/fallback_metrics.json\")\n ap.add_argument(\"--out\", default=\"data\/captions_fallback.tex\")\n a = ap.parse_args()\n d = json.loads(Path(a.metrics).read_text())\n cfg = d[\"config\"]\n modes_tex = make_breakable_modes(cfg[\"modes\"])\n tex = TEX % dict(\n runs=cfg[\"runs\"], n=cfg[\"n\"], snr_db=cfg[\"snr_db\"],\n modes=modes_tex,\n gpu=cfg.get(\"gpu\", cfg.get(\"device\", \"CPU\"))\n )\n Path(a.out).write_text(tex)\n print(\"[OK] wrote\", a.out)\n\nif __name__ == \"__main__\":\n main()<\/code><\/pre>\n\n\n\n# after you re-run the bench (see next section)\npython scripts\/build_captions_fallback.py --metrics data\/fallback_metrics.json\npdflatex -interaction=nonstopmode -halt-on-error main_fallback.tex<\/code><\/pre>\n\n\n\n\\fbGPU<\/code>.<\/p>\n\n\n\n
\n\n\n\nGet the GPU recognized (so the JSON says 3060, not CPU)<\/h2>\n\n\n\n
rf_quantum_env<\/code>:<\/p>\n\n\n\n# Sanity checks\nnvidia-smi\npython - <<'PY'\nimport torch\nprint(\"cuda available:\", torch.cuda.is_available(), \"build:\", getattr(torch.version,'cuda',None))\nprint(\"device:\", torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"\u2014\")\nPY<\/code><\/pre>\n\n\n\nnvidia-smi<\/code> works, install a CUDA wheel:<\/p>\n\n\n\n# Prefer cu121 if your driver is recent (R535+)\npip uninstall -y torch torchvision torchaudio\npip install --index-url https:\/\/download.pytorch.org\/whl\/cu121 torch torchvision torchaudio\n\n# Or cu118 for older drivers (R520+)\n# pip install --index-url https:\/\/download.pytorch.org\/whl\/cu118 torch torchvision torchaudio<\/code><\/pre>\n\n\n\nexport CUDA_VISIBLE_DEVICES=0 # the 3060\npython scripts\/run_fallback_rescue_bench.py \\\n --runs 5 --n 24000 --snr-db 8.0 \\\n --p-fail 0.0 0.1 0.2 0.4 0.6 0.8 \\\n --modes load_error shape_mismatch nan_input timeout class_mismatch \\\n --device cuda \\\n --output data\/fallback_metrics.json\n\npython scripts\/build_captions_fallback.py --metrics data\/fallback_metrics.json\npdflatex -interaction=nonstopmode -halt-on-error main_fallback.tex<\/code><\/pre>\n\n\n\n\n
device NVIDIA GeForce RTX 3060<\/code>.<\/li>\n\n\n\n% swap modes ... with:\nmodes \\texttt{\\#%d} % and set that from JSON in the builder<\/code><\/pre>\n\n\n\n
import torch
print(‘cuda available:’, torch.cuda.is_available(), ‘build:’, getattr(torch.version,’cuda’,None))
if torch.cuda.is_available():
print(‘device:’, torch.cuda.get_device_name(0))
else:
print(‘device: \u2014’)
“<\/p>\n\n\n\n