{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "Inference @128: 100%|██████████| 782/782 [4:10:15<00:00, 19.20s/it] \n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "Finished inference with batch=128 in 250.25 min\n", "Saved probabilities to 100k_probs.npy\n", "\n", "bottom-100000 metrics: ROC-AUC=0.9123 | PR-AUC=0.6204 (batch=128)\n", "Confusion (thr=0.5):\n", " [[83418 9395]\n", " [ 1097 6090]]\n", "Best-F1 thr=0.5832 -> Confusion:\n", " [[89911 2902]\n", " [ 2323 4864]]\n", "Brier score: 0.1676\n", "Recall@Top1%: 0.11172951161819951\n", "Recall@Top5%: 0.4886600807012661\n", "Recall@Top10%: 0.7528871573674689\n", "Recall @ FPR=1%: 0.3568 (thr=0.6822442412376404)\n" ] } ], "source": [ "# =============================\n", "# RUN: Bottom-100k Evaluation (single cell)\n", "# - loads model + feat scalers\n", "# - takes the last 100k rows from 900k_test.parquet\n", "# - normalizes features exactly like training\n", "# - batched tokenization & inference with GPU-safe fallback\n", "# - metrics, leaderboard, ROC/PR, calibration, ranking metrics\n", "# =============================\n", "\n", "import os, time, gc, json\n", "from tqdm.auto import tqdm\n", "import numpy as np\n", "import pandas as pd\n", "import torch\n", "import torch.nn as nn\n", "import matplotlib.pyplot as plt\n", "import seaborn as sns\n", "from safetensors.torch import load_file\n", "from transformers import AutoTokenizer, AutoModel\n", "from sklearn.metrics import (\n", " roc_auc_score, average_precision_score, roc_curve,\n", " precision_recall_curve, confusion_matrix, brier_score_loss\n", ")\n", "from sklearn.calibration import calibration_curve\n", "\n", "# ---------- CONFIG ----------\n", "MODEL_DIR = \"PathoPreter_Ready_V1\"\n", "PARQUET_900K = \"900k_test.parquet\"\n", "BOTTOM_K = 100_000 # <--- using bottom 100k as requested\n", "OUT_PROBS = \"100k_probs.npy\"\n", "\n", "# feature columns exactly as training\n", "feature_cols = [\n", " \"gnomad_af\",\n", " \"GERP++_RS_rankscore\",\n", " \"GERP_91_mammals_rankscore\",\n", " \"phyloP100way_vertebrate_rankscore\",\n", " \"phyloP470way_mammalian_rankscore\",\n", " \"phyloP17way_primate_rankscore\",\n", " \"phastCons100way_vertebrate_rankscore\",\n", " \"phastCons470way_mammalian_rankscore\",\n", " \"phastCons17way_primate_rankscore\",\n", "]\n", "\n", "bench_cols = [\n", " \"CADD_raw_rankscore\",\"REVEL_rankscore\",\"AlphaMissense_rankscore\",\n", " \"ClinPred_rankscore\",\"PrimateAI_rankscore\",\"MPC_rankscore\",\n", " \"BayesDel_addAF_rankscore\",\"EVE_rankscore\",\"ESM1b_rankscore\"\n", "]\n", "\n", "DEVICE = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n", "print(\"Device:\", DEVICE)\n", "if DEVICE.type == \"cuda\":\n", " print(torch.cuda.get_device_properties(0))\n", "\n", "# ---------- helpers ----------\n", "def map_label_str(x):\n", " if pd.isna(x): return np.nan\n", " s = str(x).lower()\n", " if s.startswith(\"path\"): return 1\n", " if s.startswith(\"ben\"): return 0\n", " return np.nan\n", "\n", "class SequenceClassificationWithFeatures(nn.Module):\n", " def __init__(self, encoder, hidden_size, feature_dim, num_labels):\n", " super().__init__()\n", " self.encoder = encoder\n", " self.classifier = nn.Sequential(\n", " nn.Linear(hidden_size + feature_dim, hidden_size),\n", " nn.GELU(),\n", " nn.Dropout(0.1),\n", " nn.Linear(hidden_size, num_labels)\n", " )\n", " def forward(self, input_ids=None, attention_mask=None, features=None):\n", " out = self.encoder(input_ids=input_ids, attention_mask=attention_mask, return_dict=True)\n", " last = out.last_hidden_state\n", " mask = attention_mask.unsqueeze(-1)\n", " pooled = (last * mask).sum(1) / mask.sum(1).clamp(min=1e-9)\n", " features = features.to(pooled.device).to(pooled.dtype)\n", " logits = self.classifier(torch.cat([pooled, features], dim=1))\n", " return logits\n", "\n", "# ---------- load metadata + scalers + tokenizer ----------\n", "print(\"Loading model metadata and scalers...\")\n", "meta_path = os.path.join(MODEL_DIR, \"model_metadata.json\")\n", "if not os.path.exists(meta_path):\n", " raise FileNotFoundError(\"Missing model metadata: \" + meta_path)\n", "with open(meta_path, \"r\") as f:\n", " meta = json.load(f)\n", "\n", "feat_means_path = os.path.join(MODEL_DIR, \"feat_means.npy\")\n", "feat_stds_path = os.path.join(MODEL_DIR, \"feat_stds.npy\")\n", "if not (os.path.exists(feat_means_path) and os.path.exists(feat_stds_path)):\n", " raise FileNotFoundError(\"Missing feat_means.npy or feat_stds.npy in model directory.\")\n", "feat_means = np.load(feat_means_path)\n", "feat_stds = np.load(feat_stds_path)\n", "\n", "# sanity check feature lengths\n", "if len(feat_means) != len(feature_cols):\n", " print(\"Warning: feat_means length != feature_cols. Using feature_cols from meta if present.\")\n", " if \"feature_cols\" in meta:\n", " feature_cols = meta[\"feature_cols\"]\n", " print(\"Using meta feature_cols length:\", len(feature_cols))\n", " else:\n", " raise RuntimeError(\"feature_cols mismatch; check model metadata.\")\n", "\n", "print(\"Loading tokenizer...\")\n", "tokenizer = AutoTokenizer.from_pretrained(\"InstaDeepAI/nucleotide-transformer-500m-human-ref\", trust_remote_code=True)\n", "\n", "# ---------- load model weights safely ----------\n", "print(\"Loading encoder and model weights (safetensors)...\")\n", "base_encoder = AutoModel.from_pretrained(MODEL_DIR, trust_remote_code=True)\n", "hidden_size = base_encoder.config.hidden_size\n", "FEATURE_DIM = len(feature_cols)\n", "NUM_LABELS = meta.get(\"num_labels\", 2)\n", "\n", "model = SequenceClassificationWithFeatures(base_encoder, hidden_size, FEATURE_DIM, NUM_LABELS)\n", "state_path = os.path.join(MODEL_DIR, \"model.safetensors\")\n", "if not os.path.exists(state_path):\n", " raise FileNotFoundError(\"Missing model.safetensors in \" + MODEL_DIR)\n", "state = load_file(state_path)\n", "state_t = {k: torch.as_tensor(v) for k, v in state.items()}\n", "missing, unexpected = model.load_state_dict(state_t, strict=False)\n", "print(\"Loaded weights (strict=False). Missing keys:\", len(missing), \"Unexpected keys:\", len(unexpected))\n", "model.to(DEVICE).eval()\n", "torch.cuda.empty_cache()\n", "\n", "# ---------- prepare bottom-100k dataframe ----------\n", "print(f\"\\nLoading {PARQUET_900K} and slicing last {BOTTOM_K} rows (bottom)...\")\n", "if not os.path.exists(PARQUET_900K):\n", " raise FileNotFoundError(\"Missing parquet file: \" + PARQUET_900K)\n", "df_all = pd.read_parquet(PARQUET_900K)\n", "print(\"Total rows in file:\", len(df_all))\n", "df = df_all.tail(BOTTOM_K).reset_index(drop=True)\n", "print(\"Sliced rows:\", len(df))\n", "del df_all\n", "gc.collect()\n", "\n", "# ensure labels exist and map\n", "df[\"labels\"] = df[\"clean_label\"].map(map_label_str)\n", "valid_mask = ~df[\"labels\"].isna()\n", "print(\"Rows with labels present:\", valid_mask.sum())\n", "df = df.loc[valid_mask].reset_index(drop=True)\n", "df[\"labels\"] = df[\"labels\"].astype(int)\n", "\n", "# ---------- inference routine ----------\n", "def run_inference_on_df(df, name=\"bottom100k\", batch_candidates=(2048,1536,1024,768,512,256,128)):\n", " print(f\"\\nStarting inference on {name} ({len(df)} rows)\")\n", " # prepare sequences and features\n", " seqs = df[\"raw_sequence\"].astype(str).tolist()\n", " feats = df[feature_cols].copy()\n", " # impute missing with training means\n", " feats = feats.fillna(pd.Series(feat_means, index=feature_cols))\n", " feats_arr = feats.values.astype(np.float32)\n", " feats_norm = (feats_arr - feat_means.reshape(1,-1)) / (feat_stds.reshape(1,-1) + 1e-9)\n", " feats_tensor = torch.from_numpy(feats_norm)\n", "\n", " y_true = df[\"labels\"].values\n", "\n", " probs = None\n", " chosen_bs = None\n", " for bs in batch_candidates:\n", " try:\n", " print(f\"\\nTrying batch size {bs} ...\")\n", " probs_chunks = []\n", " t0 = time.time()\n", " for i in tqdm(range(0, len(seqs), bs), desc=f\"Inference @{bs}\"):\n", " j = min(len(seqs), i+bs)\n", " batch_seqs = seqs[i:j]\n", " tok = tokenizer(batch_seqs, padding=\"max_length\", truncation=True,\n", " max_length=tokenizer.model_max_length, return_tensors=\"pt\")\n", " ids = tok[\"input_ids\"].to(DEVICE)\n", " mask = tok[\"attention_mask\"].to(DEVICE)\n", " feats_batch = feats_tensor[i:j].to(DEVICE, non_blocking=True)\n", "\n", " with torch.no_grad():\n", " logits = model(ids, mask, feats_batch)\n", " probs_batch = torch.softmax(logits, dim=1)[:, 1].cpu()\n", " probs_chunks.append(probs_batch)\n", "\n", " # cleanup\n", " del ids, mask, feats_batch, tok, logits, probs_batch\n", " torch.cuda.empty_cache()\n", "\n", " probs = torch.cat(probs_chunks).numpy()\n", " chosen_bs = bs\n", " elapsed = (time.time() - t0) / 60.0\n", " print(f\"Finished inference with batch={bs} in {elapsed:.2f} min\")\n", " break\n", "\n", " except RuntimeError as e:\n", " msg = str(e).lower()\n", " print(f\"RuntimeError at bs={bs}: {msg}\")\n", " if \"out of memory\" in msg or \"cuda\" in msg:\n", " print(\" -> OOM, clearing cache and trying next smaller batch\")\n", " torch.cuda.empty_cache(); gc.collect(); time.sleep(2)\n", " continue\n", " else:\n", " raise\n", "\n", " if probs is None:\n", " raise RuntimeError(\"All batch sizes failed.\")\n", "\n", " # save probs\n", " np.save(OUT_PROBS, probs)\n", " print(\"Saved probabilities to\", OUT_PROBS)\n", "\n", " # metrics\n", " roc = roc_auc_score(y_true, probs)\n", " ap = average_precision_score(y_true, probs)\n", " print(f\"\\n{name} metrics: ROC-AUC={roc:.4f} | PR-AUC={ap:.4f} (batch={chosen_bs})\")\n", "\n", " # confusion at 0.5\n", " y_pred05 = (probs >= 0.5).astype(int)\n", " cm05 = confusion_matrix(y_true, y_pred05)\n", " print(\"Confusion (thr=0.5):\\n\", cm05)\n", "\n", " # best F1 threshold via precision-recall curve\n", " prec, rec, thr = precision_recall_curve(y_true, probs)\n", " f1 = 2 * prec * rec / (prec + rec + 1e-12)\n", " if len(thr) > 0:\n", " best_idx = np.nanargmax(f1[:-1])\n", " best_thr = thr[best_idx]\n", " else:\n", " best_thr = 0.5\n", " y_pred_best = (probs >= best_thr).astype(int)\n", " cm_best = confusion_matrix(y_true, y_pred_best)\n", " print(f\"Best-F1 thr={best_thr:.4f} -> Confusion:\\n\", cm_best)\n", "\n", " # calibration\n", " brier = brier_score_loss(y_true, probs)\n", " prob_true, prob_pred = calibration_curve(y_true, probs, n_bins=10, strategy=\"uniform\")\n", " print(f\"Brier score: {brier:.4f}\")\n", "\n", " # ranking metrics\n", " def recall_at_top_frac(scores, labels, frac):\n", " k = max(1, int(len(scores) * frac))\n", " idx = np.argsort(scores)[-k:]\n", " return labels[idx].sum() / (labels.sum() + 1e-12)\n", "\n", " print(\"Recall@Top1%:\", recall_at_top_frac(probs, y_true, 0.01))\n", " print(\"Recall@Top5%:\", recall_at_top_frac(probs, y_true, 0.05))\n", " print(\"Recall@Top10%:\", recall_at_top_frac(probs, y_true, 0.10))\n", "\n", " # recall at fixed FPR=1%\n", " fpr_vals, tpr_vals, t_thresh = roc_curve(y_true, probs)\n", " valid = np.where(fpr_vals <= 0.01)[0]\n", " if len(valid) > 0:\n", " idx = valid[np.argmax(tpr_vals[valid])]\n", " rec_at_1pct, thr_at_1pct = tpr_vals[idx], t_thresh[idx]\n", " else:\n", " rec_at_1pct, thr_at_1pct = 0.0, None\n", " print(f\"Recall @ FPR=1%: {rec_at_1pct:.4f} (thr={thr_at_1pct})\")\n" ] } ], "metadata": { "kernelspec": { "display_name": ".venv", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.10.11" } }, "nbformat": 4, "nbformat_minor": 2 }