{"id":1809,"date":"2026-09-18T20:24:05","date_gmt":"2026-09-18T12:24:05","guid":{"rendered":"https:\/\/www.izhuhn.cn\/?p=1809"},"modified":"2026-09-18T20:24:05","modified_gmt":"2026-09-18T12:24:05","slug":"pi0-5-%e5%be%ae%e8%b0%83%e8%ae%ad%e7%bb%83%e4%b8%8e%e6%95%b0%e6%8d%ae%e6%b5%81%e6%b0%b4%e7%ba%bf%e8%b8%a9%e5%9d%91%e7%ae%80%e8%ae%b0","status":"publish","type":"post","link":"https:\/\/www.izhuhn.cn\/index.php\/2026\/09\/18\/pi0-5-%e5%be%ae%e8%b0%83%e8%ae%ad%e7%bb%83%e4%b8%8e%e6%95%b0%e6%8d%ae%e6%b5%81%e6%b0%b4%e7%ba%bf%e8%b8%a9%e5%9d%91%e7%ae%80%e8%ae%b0\/","title":{"rendered":"PI0.5 \u5fae\u8c03\u8bad\u7ec3\u4e0e\u6570\u636e\u6d41\u6c34\u7ebf\u8e29\u5751\u7b80\u8bb0"},"content":{"rendered":"<figure class=\"wp-block-table\"><table><thead><tr><th>------<\/th><\/tr><\/thead><tbody><tr><td>\u6570\u636e<\/td><td>task_5867\uff08\u6392\u9664\u7b2c 355 \u96c6\uff09+ task_5867-1\uff0c\u5408\u5e76\u540e **681 \u96c6 \/ 305,425 \u5e27**\uff0830fps\uff0c3 \u8def\u76f8\u673a\uff09<\/td><\/tr><tr><td>\u914d\u7f6e<\/td><td>global batch 32\uff084\/\u5361 \u00d7 8 \u5361\uff09\u300130000 \u6b65\u3001action horizon 32\u3001bf16<\/td><\/tr><tr><td>\u72b6\u6001\/\u52a8\u4f5c<\/td><td>\u5404 **16 \u7ef4**\uff0814 \u5173\u8282 + 2 \u5939\u722a\uff09<\/td><\/tr><tr><td>loss<\/td><td>1.47 \u2192 **0.03 \u91cf\u7ea7**\uff08step 7000 \u65f6 0.029\u20130.043\uff09<\/td><\/tr><tr><td>\u541e\u5410<\/td><td>0.838 s\/step\uff08\u626b\u63cf\uff09\u300138.1 samples\/s<\/td><\/tr><tr><td>\u663e\u5b58<\/td><td>\u6bcf\u5361 22\u201348 GB\uff0864 GB \u5361\uff09\uff0cbatch \u4e0a\u9650 = **\u6bcf\u5361 4**<\/td><\/tr><\/tbody><\/table><\/figure>\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u73af\u5883\u4e0e\u6743\u91cd<\/h2>\n<ul class=\"wp-block-list\"><li>\u5bb9\u5668\uff1a`drivingsdk:26.1.0-9.1.0-torch_npu2.7.1.post6-910b-openeuler24.03-py3.10`<\/li><li>\u4ee3\u7801\uff1a`\/DrivingSDK\/model_examples\/Pi-0.5`\uff08lerobot 0.3.4 + pi05 \u9002\u914d\uff09<\/li><li>\u73af\u5883\uff1a`conda activate torch2.7.1`\uff08README \u91cc\u5199\u7684 `pi05` \u73af\u5883\u5e76\u4e0d\u5b58\u5728\uff09<\/li><li>\u6743\u91cd\uff1a`\/data\/huawei\/models\/pi05_base`<\/li><li>\u79bb\u7ebf\u673a\u5668\uff08\u8bbf\u95ee\u4e0d\u4e86 huggingface.co\uff09\uff0c\u5fc5\u987b\u5148\u628a\u6240\u6709\u4f9d\u8d56\u672c\u5730\u5316<\/li><\/ul>\n<p class=\"wp-block-paragraph\">**\u4e00\u4e2a\u5bb9\u6613\u5ffd\u7565\u7684\u70b9**\uff1api05 \u7684\u5206\u8bcd\u5668\u662f\u786c\u7f16\u7801\u7684 `google\/paligemma-3b-pt-224`\uff0c\u79bb\u7ebf\u73af\u5883\u4f1a\u76f4\u63a5\u62a5<\/p>\n<p class=\"wp-block-paragraph\">`OSError: We couldn't connect to 'https:\/\/huggingface.co'`\u3002\u800c\u4e14 PaliGemma \u662f HF \u7684 gated \u4ed3\u5e93\uff0c\u955c\u50cf\u7ad9\u4e5f\u62ff\u4e0d\u5230\u3002<\/p>\n<p class=\"wp-block-paragraph\">\u6700\u540e\u4ece **ModelScope** \u4e0b\u8f7d\u4e86\u5168\u5957\uff08`tokenizer.model` 4.26MB + `tokenizer.json` 17.5MB \u7b49 4 \u4e2a\u6587\u4ef6\uff09\uff0c<\/p>\n<p class=\"wp-block-paragraph\">\u628a `processor_pi05.py` \u91cc\u7684 `tokenizer_name` \u6307\u5411\u672c\u5730\u8def\u5f84\u89e3\u51b3\u3002<\/p>\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u6570\u636e\u51c6\u5907\uff1a11 \u4e2a\u5751<\/h2>\n<h3 class=\"wp-block-heading\">3.1 \u4e3a\u4ec0\u4e48\u4e0d\u80fd\u76f4\u63a5\u7528<\/h3>\n<p class=\"wp-block-paragraph\">\u4e24\u4ef6\u4e8b\u5fc5\u987b\u5148\u505a\uff1a<\/p>\n<ol class=\"wp-block-list\"><li>**\u4e24\u4efd\u6570\u636e\u96c6\u8981\u4e00\u8d77\u8bad** \u2192 \u4f46\u8fd9\u4efd lerobot \u7684 `make_dataset` \u5bf9\u591a\u6570\u636e\u96c6\u76f4\u63a5<\/li><\/ol>\n<p class=\"wp-block-paragraph\">   `raise NotImplementedError(\"The MultiLeRobotDataset isn't supported for now.\")`\uff0c\u53ea\u80fd\u5408\u5e76\u6210\u4e00\u4efd\u3002<\/p>\n<ol class=\"wp-block-list\"><li>**\u6570\u636e\u662f v2.1\uff0c\u8bad\u7ec3\u53ea\u5403 v3.0** \u2192 \u5b98\u65b9\u63d0\u793a `Please update your dataset to the new format`\u3002<\/li><\/ol>\n<h3 class=\"wp-block-heading\">3.2 \u5b98\u65b9\u5408\u5e76\u8def\u5f84\u8e29\u6b7b\uff08\u91cd\u8981\uff09<\/h3>\n<p class=\"wp-block-paragraph\">\u6309\u5b98\u65b9 `merge_datasets` \u8d70\uff0c\u8fde\u7eed\u5d29\u4e86\u591a\u6b21\uff1a<\/p>\n<pre class=\"wp-block-code\"><code>ValueError: cannot convert float NaN to integer          # \u5220\u96c6\u65f6\nTypeError: loop of ufunc does not support argument 0 of type numpy.ndarray   # np.sqrt\uff08numpy2 \u517c\u5bb9\u6027\uff09\nTypeError: Object of type ndarray is not JSON serializable\nValueError: Same features is expected, but got ...\nValueError: all input arrays must have the same shape\nValueError: Shape of 'count' must be (1), but is (1, 1) instead.<\/code><\/pre>\n<p class=\"wp-block-paragraph\">\u9010\u4e2a\u4fee\u5b8c\u4e4b\u540e\uff0c**\u5408\u5e76\u7ec8\u4e8e\u6210\u529f\uff08681 \u96c6 \/ 305425 \u5e27\uff09\uff0c\u4f46\u89c6\u9891\u5143\u6570\u636e\u662f\u574f\u7684**\u2014\u2014\u6821\u9a8c\u53d1\u73b0\u6709<\/p>\n<p class=\"wp-block-paragraph\">**414\/681 \u96c6\u7684\u89c6\u9891\u65f6\u95f4\u6233\u8d85\u51fa\u4e86\u5b9e\u9645\u6587\u4ef6\u65f6\u957f**\uff0c\u8bad\u7ec3\u76f4\u63a5\u62a5\uff1a<\/p>\n<pre class=\"wp-block-code\"><code>AssertionError: One or several query timestamps unexpectedly violate the tolerance\nqueried timestamps: tensor([1422.6666])  \u2192  \u5b9e\u9645\u6587\u4ef6\u53ea\u5230 766.5s<\/code><\/pre>\n<p class=\"wp-block-paragraph\">\u4e5f\u5c31\u662f\u8bf4\u5b98\u65b9 `aggregate_videos` \u7684\u62fc\u63a5\u903b\u8f91\u6709\u95ee\u9898\uff08\u5199\u8fdb\u53bb\u7684\u65f6\u95f4\u6233\u4e0e\u4ea7\u51fa\u7684\u89c6\u9891\u5bf9\u4e0d\u4e0a\uff09\u3002<\/p>\n<p class=\"wp-block-paragraph\">**\u6700\u7ec8\u65b9\u6848\uff1a\u7ed5\u5f00\u5b98\u65b9\u805a\u5408\uff0c\u6539\u7528\u624b\u5de5\u6587\u4ef6\u7ea7\u5408\u5e76**\u2014\u2014\u5728 v2.1 \u5c42\u9762\u628a\u4e24\u4efd\u6570\u636e\u6309 episode \u9010\u6587\u4ef6\u590d\u5236\u5e76\u91cd\u7f16\u53f7<\/p>\n<p class=\"wp-block-paragraph\">\uff08`data\/chunk-000\/episode_XXXXXX.parquet` + \u4e09\u8def `videos\/...\/episode_XXXXXX.mp4`\uff09\uff0c\u6539\u5199<\/p>\n<p class=\"wp-block-paragraph\">`episode_index` \/ `index` \u5217\u4e0e `meta\/{episodes,episodes_stats,tasks,info}.json`\uff0c\u518d\u7528\u5b98\u65b9\u8f6c\u6362\u5668\u8f6c v3.0\u3002<\/p>\n<p class=\"wp-block-paragraph\">\u8fd9\u6837**\u89c6\u9891\u62fc\u63a5\u4ea4\u7ed9\u8f6c\u6362\u5668\u505a**\uff08\u5df2\u9a8c\u8bc1\u6b63\u786e\uff09\uff0c\u7ed3\u679c\uff1a**681 \u96c6 \/ 305,425 \u5e27\uff0c\u89c6\u9891\u5143\u6570\u636e 0 \u5f02\u5e38**\uff0c\u800c\u4e14\u5408\u5e76\u672c\u8eab\u53ea\u82b1\u4e86\u7ea6 1 \u5206\u949f\u3002<\/p>\n<h3 class=\"wp-block-heading\">3.3 \u5b98\u65b9\u5206\u4f4d\u6570\u811a\u672c\u4e0d\u53ef\u7528<\/h3>\n<p class=\"wp-block-paragraph\">pi05 \u7684 STATE\/ACTION \u5f3a\u5236\u4f7f\u7528 **QUANTILES \u5f52\u4e00\u5316**\uff08\u7f3a q01\/q99 \u76f4\u63a5\u62a5\u9519\uff09\uff0c\u5b98\u65b9\u7ed9\u4e86<\/p>\n<p class=\"wp-block-paragraph\">`augment_dataset_quantile_stats.py`\u3002\u4f46\u5b83\u662f**\u9010\u96c6\u89e3\u7801 3 \u8def\u89c6\u9891**\uff1a<\/p>\n<pre class=\"wp-block-code\"><code>Processing episodes: 6%|\u258c | 44\/681 [4:20:13&lt;62:47:22, 354.85s\/it]   \u2190 330+ \u79d2\/\u96c6\uff0c\u5916\u63a8 62 \u5c0f\u65f6<\/code><\/pre>\n<p class=\"wp-block-paragraph\">\u8dd1\u5230\u7b2c 44 \u96c6\u8fd8\u5d29\u4e86\u3002**\u6539\u7528\u76f4\u63a5\u8bfb parquet \u8ba1\u7b97**\uff08215K \u5e27 \u00d7 199 \u7ef4\uff0c\u79d2\u7ea7\u5b8c\u6210\uff09\uff1a<\/p>\n<pre class=\"wp-block-code\"><code>for q in (0.01, 0.10, 0.50, 0.90, 0.99):\n    st[f'q{int(q*100):02d}'] = np.percentile(X, q * 100, axis=0)<\/code><\/pre>\n<p class=\"wp-block-paragraph\">\u987a\u4fbf\u8865\u4e00\u53e5\uff1a\u56fe\u50cf\u7279\u5f81\u4e0d\u9700\u8981\u6570\u636e\u96c6\u7edf\u8ba1\u2014\u2014`DatasetConfig.use_imagenet_stats=True` \u4f1a\u5728\u5efa\u6570\u636e\u96c6\u65f6\u81ea\u52a8\u5957 ImageNet \u5747\u503c\u65b9\u5dee\u3002<\/p>\n<p class=\"wp-block-paragraph\">**\u6240\u4ee5\u89c6\u9891\u7279\u5f81\u90a3\u4efd `count=0` \u7684\u5047\u7edf\u8ba1\u6839\u672c\u6ca1\u5fc5\u8981\u7b97**\uff0c\u800c\u5b83\u6b63\u662f\u628a\u5b98\u65b9\u5408\u5e76\u8def\u5f84\u641e\u5d29\u7684\u5143\u51f6\u4e4b\u4e00\u3002<\/p>\n<h3 class=\"wp-block-heading\">3.4 \u5751\u6e05\u5355\uff0811 \u9879\uff09<\/h3>\n<figure class=\"wp-block-table\"><table><thead><tr><th>#<\/th><th>\u73b0\u8c61<\/th><th>\u6839\u56e0<\/th><th>\u5904\u7f6e<\/th><th>\u4e0a\u6e38 bug<\/th><\/tr><\/thead><tbody><tr><td>1<\/td><td>\u8981\u6c42 v3.0 \u683c\u5f0f<\/td><td>\u6570\u636e\u662f v2.1<\/td><td>\u5b98\u65b9\u8f6c\u6362\u5668\u8f6c\u6362<\/td><td>\u2014<\/td><\/tr><tr><td>2<\/td><td>`Couldn't cast list<uint64> to uint64`<\/td><td>`ts`\/`action_src_status` \u5b58\u6210 `list`[1]<\/td><td>\u538b\u5e73\u4e3a\u6807\u91cf<\/td><td>\u2014<\/td><\/tr><tr><td>3<\/td><td>\u5220\u96c6\u5d29 `int(NaN)`<\/td><td>**\u8f6c\u6362\u5668\u628a ep312 \u7684 `data\/file_index` \u5199\u9519**<\/td><td>\u6309\u5b9e\u9645\u6587\u4ef6\u4fee\u6b63<\/td><td>\u2705<\/td><\/tr><tr><td>4<\/td><td>\u805a\u5408\u9664\u96f6 \u2192 NaN \u2192 `np.sqrt` \u5d29<\/td><td>\u89c6\u9891\u7279\u5f81 stats \u662f `count=0` \u5360\u4f4d\u7b26<\/td><td>\u6e05\u9664\u5360\u4f4d<\/td><td>\u2014<\/td><\/tr><tr><td>5<\/td><td>`ndarray is not JSON serializable`<\/td><td>stats \u542b\u5d4c\u5957 ndarray<\/td><td>`write_json` \u9012\u5f52 sanitize<\/td><td>\u2705<\/td><\/tr><tr><td>6<\/td><td>`np.sqrt(object array)` \u5d29<\/td><td>numpy 2 \u517c\u5bb9\u6027<\/td><td>\u663e\u5f0f `float64`<\/td><td>\u2705<\/td><\/tr><tr><td>7<\/td><td>`Same features is expected`<\/td><td>\u4e24\u4efd\u6570\u636e features \u7684 `fps` \u5b57\u6bb5\u4e0d\u4e00\u81f4<\/td><td>\u7edf\u4e00\u5b57\u6bb5<\/td><td>\u2014<\/td><\/tr><tr><td>8<\/td><td>`all input arrays must have the same shape`<\/td><td>\u540c\u4e00\u4e2a\u7279\u5f81\u7684 std\/count \u5f62\u72b6 (1,) vs (1,1)<\/td><td>\u6309 `min` \u5f62\u72b6\u5bf9\u9f50<\/td><td>\u2014<\/td><\/tr><tr><td>9<\/td><td>**414\/681 \u96c6\u89c6\u9891\u65f6\u95f4\u6233\u9519\u8bef**<\/td><td>\u5b98\u65b9 `aggregate_videos` \u62fc\u63a5 bug<\/td><td>**\u6539\u624b\u5de5\u6587\u4ef6\u7ea7\u5408\u5e76**<\/td><td>\u2705<\/td><\/tr><tr><td>10<\/td><td>\u5206\u4f4d\u6570\u811a\u672c 330 \u79d2\/\u96c6<\/td><td>\u9010\u96c6\u89e3\u7801\u89c6\u9891<\/td><td>\u76f4\u63a5\u8bfb parquet \u8ba1\u7b97<\/td><td>\u5de5\u5177\u6027\u80fd<\/td><\/tr><tr><td>11<\/td><td>**\u8bad\u7ec3 loss \u7206\u70b8\u5230 8\u00d710\u2075**<\/td><td>\u89c1\u4e0b\u8282<\/td><td>\u88c1\u526a\u7ef4\u5ea6<\/td><td>\u2014<\/td><\/tr><\/tbody><\/table><\/figure>\n<h2 class=\"wp-block-heading\">\u56db\u3001\u6700\u9690\u853d\u7684\u4e00\u4e2a\u5751\uff1aloss \u7206\u70b8<\/h2>\n<p class=\"wp-block-paragraph\">\u6570\u636e\u90fd\u5bf9\u9f50\u4e4b\u540e\u5f00\u8bad\uff0c\u7ed3\u679c\uff1a<\/p>\n<pre class=\"wp-block-code\"><code>step:20 loss:1.733\nstep:40 loss:813270.345        \u2190 \u76f4\u63a5\u70b8\u4e86<\/code><\/pre>\n<p class=\"wp-block-paragraph\">\u6839\u56e0\u4e0d\u5728\u8d85\u53c2\uff0c\u800c\u5728**\u6570\u636e\u91cf\u7eb2**\u3002AgiBot \u5bfc\u51fa\u7684 `observation.state` \u662f **163 \u7ef4**\uff0c\u91cc\u9762\u9664\u4e86\u5173\u8282\u89d2\u8fd8\u6df7\u7740\uff1a<\/p>\n<ul class=\"wp-block-list\"><li>\u529b\/\u529b\u77e9\uff08wrench\uff0c\u00b150\uff09<\/li><li>\u5173\u8282\u7535\u6d41\uff08\u00b13000\uff09<\/li><li>\u76f8\u673a\u5916\u53c2\uff08\u00b11000 \u91cf\u7ea7\uff09<\/li><li>\u7b1b\u5361\u5c14\u9065\u64cd\u4f5c\u91cf<\/li><\/ul>\n<p class=\"wp-block-paragraph\">pi05 \u7528 QUANTILES \u5f52\u4e00\u5316\uff1a\u628a [q01, q99] \u6620\u5230 [-1, 1]\uff0c**\u8d85\u51fa\u8303\u56f4\u7684\u503c\u505a\u7ebf\u6027\u5916\u63a8**\u3002<\/p>\n<p class=\"wp-block-paragraph\">\u4e8e\u662f\u8fd9\u4e9b\"\u6781\u7aef\u91cf\u7eb2\"\u901a\u9053\u88ab\u653e\u5927\u6210\u51e0\u767e\u4e0a\u5343\u7684\u8f93\u5165\uff0c\u8bad\u7ec3\u7acb\u523b\u53d1\u6563\u3002<\/p>\n<p class=\"wp-block-paragraph\">**\u5904\u7f6e**\uff1a\u628a state\/action \u88c1\u5230\u771f\u6b63\u6709\u610f\u4e49\u7684\u63a7\u5236\u91cf\u2014\u2014<\/p>\n<pre class=\"wp-block-code\"><code>state  = joint\/position[28:42) + effector[0:2)   \u2192 16 \u7ef4\uff0814 \u5173\u8282 + \u5de6\u53f3\u5939\u722a\uff09\naction = joint\/position[16:30) + effector[0:2)   \u2192 16 \u7ef4<\/code><\/pre>\n<p class=\"wp-block-paragraph\">\u987a\u5e26\u4e00\u4e2a\u597d\u5904\uff1a16 \u7ef4\u843d\u56de pi05 \u9884\u8bad\u7ec3\u5934\u539f\u751f\u7684 `max_state_dim = max_action_dim = 32`\uff0c<\/p>\n<p class=\"wp-block-paragraph\">\u4e0d\u518d\u9700\u8981 `--policy.max_state_dim=256 --policy.max_action_dim=64` \u8fd9\u79cd\u8865\u4e01<\/p>\n<p class=\"wp-block-paragraph\">\uff08\u90a3\u4e2a\u8865\u4e01\u867d\u7136\u80fd\u8dd1\u901a\u2014\u2014\u5b83\u89e3\u51b3\u7684\u662f\"36 \u7ef4 action > 32 \u4e0a\u9650\"\u5bfc\u81f4\u7684<\/p>\n<p class=\"wp-block-paragraph\">`aclnnAddmm: k-axis different`\u2014\u2014\u4f46\u8d85\u51fa\u9884\u8bad\u7ec3\u7ef4\u5ea6\u7684\u90e8\u5206\u662f\u968f\u673a\u521d\u59cb\u5316\u7684\uff0c\u8d28\u91cf\u4e0d\u4fdd\u8bc1\uff09\u3002<\/p>\n<p class=\"wp-block-paragraph\">\u88c1\u5b8c\u91cd\u8bad\uff0closs \u7acb\u523b\u6b63\u5e38\uff1a<\/p>\n<pre class=\"wp-block-code\"><code>step:160  loss:1.473\nstep:200  loss:1.403\nstep:220  loss:1.346   updt_s:0.846  data_s:0.018\n...\nstep:7K   loss:0.029~0.043<\/code><\/pre>\n<h2 class=\"wp-block-heading\">\u4e94\u3001batch size \u4e0e\u663e\u5b58\u5b9e\u6d4b<\/h2>\n<p class=\"wp-block-paragraph\">\u6362\u786c\u4ef6\u6216\u6362\u914d\u7f6e\u65f6\uff0c\u5efa\u8bae\u5148\u8dd1\u4e00\u904d\u626b\u63cf\uff08\u811a\u672c\uff1a\u5355\u5361 `pi05_sweep.sh`\u30018 \u5361 `pi05_sweep8.sh`\uff09\u3002<\/p>\n<h3 class=\"wp-block-heading\">5.1 \u5355\u5361\uff08910B3 64GB\uff09<\/h3>\n<figure class=\"wp-block-table\"><table><thead><tr><th>batch<\/th><th>s\/step<\/th><th>\u541e\u5410<\/th><th>\u5cf0\u503c HBM<\/th><th>\u7ed3\u679c<\/th><\/tr><\/thead><tbody><tr><td>2<\/td><td>0.529<\/td><td>3.7 samples\/s<\/td><td>38.4 GB<\/td><td>\u2705<\/td><\/tr><tr><td>**4**<\/td><td>**0.705**<\/td><td>**5.6 samples\/s**<\/td><td>**47.5 GB**<\/td><td>\u2705 \u4e0a\u9650<\/td><\/tr><tr><td>8<\/td><td>\u2014<\/td><td>\u2014<\/td><td>62.8 GB<\/td><td>\u274c \u5d29\uff08\u8fd1 OOM\uff09<\/td><\/tr><tr><td>16<\/td><td>\u2014<\/td><td>\u2014<\/td><td>65.3 GB<\/td><td>\u274c OOM<\/td><\/tr><\/tbody><\/table><\/figure>\n<h3 class=\"wp-block-heading\">5.2 8 \u5361<\/h3>\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u6bcf\u5361 batch<\/th><th>global batch<\/th><th>s\/step<\/th><th>\u541e\u5410<\/th><th>\u5cf0\u503c HBM<\/th><th>\u7ed3\u679c<\/th><\/tr><\/thead><tbody><tr><td>2<\/td><td>16<\/td><td>0.748<\/td><td>21.3 samples\/s<\/td><td>27.9 GB<\/td><td>\u2705<\/td><\/tr><tr><td>**4**<\/td><td>**32**<\/td><td>**0.838**<\/td><td>**38.1 samples\/s**<\/td><td>22.2 GB<\/td><td>\u2705 \u4e0a\u9650<\/td><\/tr><tr><td>8<\/td><td>64<\/td><td>\u2014<\/td><td>\u2014<\/td><td>\u2014<\/td><td>\u274c OOM<\/td><\/tr><\/tbody><\/table><\/figure>\n<p class=\"wp-block-paragraph\">**\u7ed3\u8bba**\uff1a\u672c\u5e73\u53f0\uff0864 GB\/\u5361\uff09\u7684\u4e0a\u9650\u662f**\u6bcf\u5361 4 \u2192 global batch 32**\u3002\u8fd9\u6070\u597d\u7b49\u4e8e\u5ba2\u6237\u7ed9\u5b9a\u914d\u7f6e\u3002<\/p>\n<h3 class=\"wp-block-heading\">5.3 \u4e00\u4e2a\u6709\u610f\u601d\u7684\u5bf9\u7167<\/h3>\n<p class=\"wp-block-paragraph\">\u5ba2\u6237\u7528 **RTX PRO 6000\uff0896 GB\uff09** \u8dd1\u540c\u6837 batch 32\uff0c\u4f46**\u663e\u5b58\u5360\u7528\u660e\u663e\u504f\u4f4e**\uff08\u6709\u5927\u534a\u5bcc\u4f59\uff09\u3002<\/p>\n<p class=\"wp-block-paragraph\">\u6211\u4eec\u5728 64 GB \u5361\u4e0a\u8dd1\u540c\u6837\u7684 global batch 32 \u5df2\u7ecf\u8d34\u7740\u4e0a\u9650\u3002<\/p>\n<p class=\"wp-block-paragraph\">\u5dee\u5f02\u6765\u81ea**\u5355\u6837\u672c\u5f00\u9500**\uff1a\u6211\u4eec\u8fd9\u4efd\u6570\u636e\u7684 3 \u8def\u76f8\u673a\u662f 1280\u00d7800 + 848\u00d7480\uff08\u5934\u90e8\u9c7c\u773c + \u53cc\u624b D405\uff09\uff0c<\/p>\n<p class=\"wp-block-paragraph\">\u800c\u5bf9\u65b9\u90a3\u5957\u5206\u8fa8\u7387\/\u76f8\u673a\u6570\u66f4\u7701\u3002**\u6362\u53e5\u8bdd\u8bf4\uff0c\u4ed6\u4eec\u90a3\u5f20\u5361\u7684\u663e\u5b58\u8fd8\u6709\u5f88\u5927\u7a7a\u95f4\uff0c\u52a0\u5927 batch \u662f\u63d0\u5347\u541e\u5410\u6700\u76f4\u63a5\u7684\u624b\u6bb5\u3002**<\/p>\n<h2 class=\"wp-block-heading\">\u516d\u3001\u5c0f\u7ed3<\/h2>\n<ol class=\"wp-block-list\"><li>**\u5177\u8eab\u6a21\u578b\u7684\u9002\u914d\u6210\u672c\uff0c\u5927\u5934\u5728\u6570\u636e\u4e0d\u662f\u6a21\u578b**\u3002\u8fd9\u6b21 11 \u4e2a\u5751\u91cc\u6709 9 \u4e2a\u5728\u6570\u636e\u6d41\u6c34\u7ebf\u3002<\/li><li>**\u5bf9\u5ba2\u4ea4\u4ed8\u7684\u6570\u636e\u8981\u6309\"\u80fd\u76f4\u63a5\u5582\u8bad\u7ec3\"\u7684\u6807\u51c6\u9a8c\u6536**\uff1a\u683c\u5f0f\u7248\u672c\u3001\u7c7b\u578b\u3001\u7edf\u8ba1\u91cf\u3001\u89c6\u9891\u65f6\u95f4\u6233\uff0c\u7f3a\u4e00\u9879\u90fd\u4f1a\u5728\u8bad\u7ec3\u65f6\u624d\u70b8\u3002<\/li><li>**\u5b98\u65b9\u5de5\u5177\u94fe\u4e0d\u80fd\u76f2\u4fe1**\uff1a`aggregate_videos`\uff08\u89c6\u9891\u65f6\u95f4\u6233\u9519\uff09\u3001`augment_dataset_quantile_stats`\uff08330 \u79d2\/\u96c6\uff09\u3001<\/li><\/ol>\n<p class=\"wp-block-paragraph\">   \u8f6c\u6362\u5668\u7684 `file_index`\u3001`np.sqrt` \u7684 numpy2 \u517c\u5bb9\u6027\u3001`write_json` \u7684 ndarray\u2014\u2014\u5efa\u8bae\u53cd\u9988\u4e0a\u6e38\u3002<\/p>\n<ol class=\"wp-block-list\"><li>**loss \u7206\u70b8\u5148\u770b\u6570\u636e\u91cf\u7eb2**\uff0c\u522b\u6025\u7740\u8c03\u5b66\u4e60\u7387\uff1a\u6781\u7aef\u91cf\u7eb2\u901a\u9053\u5728\u91cf\u5316\u5f52\u4e00\u5316\u4e0b\u4f1a\u88ab\u7ebf\u6027\u5916\u63a8\u653e\u5927\u3002<\/li><li>16 \u7ef4\uff08\u5173\u8282 + \u5939\u722a\uff09\u7684\u63a7\u5236\u91cf\u65e2\u591f\u7528\u3001\u53c8\u6070\u597d\u843d\u5728\u9884\u8bad\u7ec3\u5934\u7ef4\u5ea6\u5185\uff0c\u662f\u8fd9\u6b21\u6700\"\u987a\"\u7684\u9009\u62e9\u3002<\/li><\/ol>\n<p class=\"wp-block-paragraph\">**\u9644**\uff1a\u5b8c\u6574\u590d\u73b0\u6b65\u9aa4\u3001\u811a\u672c\u6e05\u5355\u4e0e\u6821\u9a8c\u6807\u51c6\u89c1\u300aPI0.5 \u5fae\u8c03\u8bad\u7ec3\u590d\u73b0\u624b\u518c\u300b\u3002<\/p>","protected":false},"excerpt":{"rendered":"<p>&#8212;&#8212; \u6570\u636e task_5867\uff08\u6392\u9664\u7b2c 355 \u96c6\uff09+ task_5867-1\uff0c\u5408\u5e76\u540e **681 \u96c6 \/ 305,425 &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[1],"tags":[25,51,45,17,67,16,68],"class_list":["post-1809","post","type-post","status-publish","format-standard","hentry","category-uncategorized","tag-910b","tag-lerobot","tag-pi0-5","tag-vla","tag-67","tag-16","tag-68"],"jetpack_featured_media_url":"","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1809","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1809"}],"version-history":[{"count":1,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1809\/revisions"}],"predecessor-version":[{"id":1810,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1809\/revisions\/1810"}],"wp:attachment":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1809"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1809"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1809"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}