6.3 KiB
闂涓€ B0-B4 妯″瀷瀵规瘮缁撴灉
鏈姤鍛婃寜 final/Q1/闂涓€.pdf 绗?4.8.3 鑺傚疄鏂戒簲缁勫崟鍥犵礌瀵圭収銆傚叏閮?100 鏉¢檮浠朵竴瑙嗛鍧囦娇鐢ㄥ悓涓€濂楀喕缁撶壒寰佹娊鍙栧櫒銆?.1 绉掍富鏃堕棿缃戞牸銆佽娴嬫帺鐮佷笌鎸?video_id 鍒嗙粍鐨勪簲鎶樺垝鍒嗐€傛儏鎰熸爣绛句粎鐢ㄤ簬鎶樺唴鐨勮交閲忛娴嬫帰閽堬紝涓嶈繘鍏ユ椂闂村榻愩€丆TC 鍚庨獙鎴栧姩鎬佺鍚嶈绠椼€?
瀹為檯鐗瑰緛鍜屾瘮杈冪増鏈?
| 鐗堟湰 | 瀹炴柦鍐呭 |
|---|---|
| B0 | CTC Viterbi 璇嶅尯闂翠笌闊抽/瑙嗚鐗╃悊鏃堕棿鍖洪棿鎶曞奖锛涜缁冩姌鍧囧€?鏍囧噯宸缉鏀撅紱瑙嗚鏃嬭浆鍚戦噺鍜岃绾垮悜閲忎綔鏅€氬垎閲忓潎鍊笺€? |
| B1 | 鍦?B0 涓婃敼鐢ㄨ缁冩姌涓綅鏁颁笌 MAD 鐨勭ǔ鍋ュ昂搴︼紙MAD 涓洪浂鏃跺洖閫€鍒拌缁冩姌鏍囧噯宸級锛屽苟瀵硅瑙夋棆杞敤 SO(3) 鍧囧€笺€佽绾跨敤褰掍竴鍖栧悜閲忓潎鍊笺€? |
| B2 | 鍦?B1 涓婂彧灏嗘枃鏈瘝鍚戦噺鐨勭‖杈圭晫鎶曞奖鏇挎崲涓哄浐瀹氳浆鍐?CTC 鐘舵€佸浘鐨勫墠鍚戔€撳悗鍚戝崰鎹鐜囨姇褰憋紱濯掍綋鏃堕棿鎴充粛鎸夌墿鐞嗘椂闂淬€? |
| B3 | 鍦?B1 涓婁粎闄勫姞闊抽 log-F0/鑳介噺涓庤瑙?jaw-open/鍢撮儴寮€鍚堢巼鐨勬椂闂村骞夸竴銆佷簩闃惰矾寰勭鍚嶏紱缂烘祴鐐逛箣闂翠笉杩炵嚎銆? |
| B4 | 鍦?B1 涓婁粎闄勫姞鍐荤粨 Wav2Vec2-base-960h 鏈€鍚庡洓灞傜殑鏃堕棿鎶曞奖琛ㄧず銆? |
闄勪欢涓€鐜版湁鏍囩琛ㄥ拰鍘熷瑙嗛琚洿鎺ヤ娇鐢ㄣ€傛枃鏈负 768 缁?BERT-base-uncased 鏈洓灞傚潎鍊硷紱闊抽涓?74 缁达紙log-Mel 40銆丮FCC 13銆佄擬FCC 13銆侀煹寰?璋辩粺璁?8锛夛紱瑙嗚涓?35 缁达紙17 涓?MediaPipe blendshape 浠g悊銆? 缁村ご濮裤€? 缁磋繎浼艰绾裤€? 缁撮潰閮ㄦ瘮渚嬪嚑浣曪級銆傝瑙変唬鐞嗗拰鍑犱綍绱㈠紩瀹氫箟瑙?compare_models.py锛屽畠浠笌 OpenFace AU 瀹氫箟骞朵笉绛夊悓锛涜闄愬埗椤诲湪璁烘枃涓槑绀恒€?
缁撴灉
鍒嗙被鎸夎繛缁爣绛剧殑涓ユ牸绗﹀彿鏋勯€?Negative/Neutral/Positive锛屽己搴﹂娴嬮檺骞呭埌 [-3, 3]銆傛瘡鎶樿缁冩姌鍗曠嫭鎷熷悎鏍囧噯鍖栧櫒銆侀€昏緫鍥炲綊锛圕=0.05锛変笌 Ridge锛坅lpha=25锛夛紱鍥哄畾浜旀姌鐢?GroupKFold 鎸夊師濮?video_id 鍒嗙粍銆傛€讳綋 OOF 鎸囨爣鎸?100 鏉$暀缁勯娴嬭绠椼€?
| 鏂规硶 | OOF Accuracy | OOF Macro-F1 | OOF MAE | OOF Pearson | Macro-F1 鎶樺潎鍊悸盨D |
|---|---|---|---|---|---|
| B0 | 0.600 | 0.361 | 0.564 | 0.346 | 0.337 卤 0.098 |
| B1 | 0.610 | 0.387 | 0.599 | 0.238 | 0.349 卤 0.119 |
| B2 | 0.580 | 0.331 | 0.588 | 0.279 | 0.313 卤 0.087 |
| B3 | 0.590 | 0.359 | 0.600 | 0.230 | 0.333 卤 0.096 |
| B4 | 0.600 | 0.422 | 0.593 | 0.199 | 0.398 卤 0.089 |
鎸夊綋鍓?OOF 鎺㈤拡锛孧acro-F1 鏈€楂樼殑鏄?B4锛?.422锛夛紝MAE 鏈€浣庣殑鏄?B0锛?.564锛夈€傜浉瀵?B1锛孊4 鐨?Macro-F1 宸负 +0.035锛岃棰戠粍 Bootstrap 95% 鍖洪棿 [-0.051, +0.137]锛屽尯闂磋法杩囬浂銆侭0 鐨?MAE 宸负 -0.035锛屽尯闂?[-0.063, -0.012]锛孭earson 宸负 +0.108銆侭2 鐨?Macro-F1 宸负 -0.057锛屽尯闂?[-0.109, -0.001]锛孧AE 宸负 -0.011锛汢3 鐨?Macro-F1 宸负 -0.029锛屽尯闂?[-0.072, +0.000]銆傜粍 Bootstrap 鍙弽鏄犲綋鍓?37 涓潵婧愯棰戜笂鐨勬娊鏍蜂笉纭畾鎬э紝鏈綔澶氶噸姣旇緝鏍℃銆?
瑕嗙洊涓庡榻愭牳楠?
鏍锋湰鏁帮細100锛涙爣绛捐〃杩炵画鏍囩涓庢瀬鎬т竴鑷?100/100锛汣TC 纭榻愬畬鏁寸殑鏍锋湰 91/100銆侭0 纭尯闂村钩鍧囨枃鏈鐩栫巼涓?0.673锛孊2 骞冲潎璇嶅崰鎹悗楠岃川閲忎负 0.229锛堣繖鏄湡鏈涘崰鎹巼锛屼笉鏄鐩栫巼锛夛紱瑙嗚妫€娴嬭鐩栫巼锛堟寜 0.1 绉掔綉鏍煎钩鍧囷級涓?0.859銆傚悗楠屽唴閮?90% 杈圭晫鍖洪棿骞冲潎瀹藉害涓鸿捣鐐?0.131 绉掋€佺粓鐐?0.111 绉掋€傛病鏈変汉宸ヨ竟鐣屽瓙闆嗭紝鍥犳杩欎簺瀹藉害鍙槸妯″瀷鍐呴儴涓嶇‘瀹氭€ф憳瑕侊紝涓嶆槸缁忛獙鏍″噯鐜囨垨杈圭晫璇樊銆? B0鈥揃4 鐨勭壒寰佹潵婧愩€佹湁鏁堟帺鐮併€佽繛缁鐩栫巼銆佽瘝杈圭晫鍚庨獙瀹藉害銆佹姌鍐呴娴嬪強瑙嗛鍝堝笇鍧囧垎鍒繚瀛樺湪閰嶅 CSV/JSON 涓€傛病鏈変汉宸ヨ瘝杈圭晫鏃朵笉鎶ュ憡 IoU/MATE锛屼篃涓嶆妸涓嶅悓鏂规硶鐨勯娴嬫帰閽堝垎鏁拌В閲婁负瀵归綈鐪熷€笺€?
鏂囦欢
comparison_summary.csv锛氫簲绉嶆柟娉曟€讳綋 OOF 涓庢姌鍧囧€兼寚鏍囥€?-fold_metrics.csv锛氭瘡鎶樺垎绫?鍥炲綊鎸囨爣涓庣壒寰佺淮鏁般€?-oof_predictions.csv锛氭瘡鏉℃牱鏈殑鐣欑粍棰勬祴銆?-group_bootstrap_deltas.csv锛氱浉瀵?B1 鐨勮棰戠粍閰嶅 Bootstrap 95% 鍖洪棿銆?-sample_alignment_summary.csv锛?00 鏉℃牱鏈€佺‖/姒傜巼鏂囨湰瑕嗙洊鍜屽悗楠岃竟鐣屽搴︺€?-modality_summary.csv锛?00 琛屾牱鏈?妯℃€佹槑缁嗐€?-word_alignment_posterior.csv锛氶€愯瘝纭竟鐣屻€佸疄闄呯浉瀵硅川閲忔潈閲嶅強鍚庨獙杈圭晫鍖洪棿銆?-split_assignments.csv锛氶€愭牱鏈姌鍙枫€?-comparison.png锛氭牳蹇?OOF 鎸囨爣鍥俱€?-typical_alignment_example.png锛氫腑浣嶆椂闀挎牱鏈殑璇嶈竟鐣屻€佽鐩栫巼銆佸0瀛?瑙嗚杞ㄨ抗涓庤棰戝抚鏍搁獙鍥俱€?-run_manifest.json锛氱幆澧冦€佹ā鍨?revision銆佸弬鏁般€佸搱甯屼笌杩愯鏃堕棿銆?
澶嶇幇
鍦ㄤ粨搴撴牴鐩綍鎵ц锛?
cd final/Q1
uv sync
uv run python compare_models.py
鐗瑰緛缂撳瓨鍐欏湪 final/Q1/cache/native/锛屾寮忕粨鏋滃彧鍐欏湪 final/Q1/results/model_comparison/銆傚垹闄ょ紦瀛樺悗浼氫粠闄勪欢涓€閲嶆柊鎻愬彇銆傞娆¤繍琛岄渶瑕佷笅杞?BERT銆乄av2Vec2 鍜?MediaPipe Face Landmarker 鏉冮噸銆?