![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250815120953-689f23914c2a0.png\")
<\/p>\nNumPy\u4f5c\u4e3aPython\u6570\u503c\u8ba1\u7b97\u9886\u57df\u7684\u57fa\u7840\u6846\u67b6，\u51ed\u501f\u5176\u5f3a\u5927\u7684N\u7ef4\u6570\u7ec4\u7ed3\u6784\u548c\u4e30\u5bcc\u7684\u51fd\u6570\u751f\u6001\u7cfb\u7edf，\u6210\u4e3a\u79d1\u5b66\u5bb6\u3001\u5de5\u7a0b\u5e08\u548c\u6570\u636e\u5206\u6790\u5e08\u7684\u6838\u5fc3\u5de5\u5177\u3002\u7136\u800c，\u968f\u7740\u8ba1\u7b97\u9700\u6c42\u7684\u5feb\u901f\u589e\u957f，\u7279\u522b\u662f\u5728\u673a\u5668\u5b66\u4e60\u548c\u5927\u89c4\u6a21\u79d1\u5b66\u6a21\u62df\u9886\u57df，NumPy\u57fa\u4e8eCPU\u7684\u6267\u884c\u6a21\u5f0f\u4ee5\u53ca\u7f3a\u4e4f\u5185\u7f6e\u81ea\u52a8\u5fae\u5206\u529f\u80fd\u7684\u9650\u5236\u6108\u53d1\u660e\u663e\u3002<\/p>\n
JAX\u6b63\u662f\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\u800c\u8bbe\u8ba1\u7684\u3002\u4f5c\u4e3aGoogle Research\u5f00\u53d1\u7684\u6570\u503c\u8ba1\u7b97\u5e93，JAX\u81f4\u529b\u4e8e\u5c06NumPy\u5f15\u5165\u73b0\u4ee3\u786c\u4ef6\u52a0\u901f\u5668\u548c\u57fa\u4e8e\u68af\u5ea6\u4f18\u5316\u7684\u8ba1\u7b97\u8303\u5f0f\u4e2d\u3002\u9700\u8981\u8bf4\u660e\u7684\u662f，JAX\u76ee\u524d\u4ecd\u662f\u4e00\u4e2a\u7814\u7a76\u9879\u76ee，\u800c\u975eGoogle\u7684\u5b98\u65b9\u4ea7\u54c1，\u56e0\u6b64\u5728\u4f7f\u7528\u8fc7\u7a0b\u4e2d\u53ef\u80fd\u9047\u5230\u4e00\u4e9b\u4e0d\u7a33\u5b9a\u56e0\u7d20\u548c\u6f5c\u5728\u95ee\u9898\u3002 <\/p>\n
\u5bf9\u4e8e\u4ece\u4e8b\u79d1\u5b66\u8ba1\u7b97\u7684Python\u5f00\u53d1\u8005\u800c\u8a00，NumPy\u51e0\u4e4e\u662f\u5fc5\u5907\u5de5\u5177\u3002\u5176\u6838\u5fc3ndarray\u5bf9\u8c61\u4e3a\u9ad8\u6548\u5b58\u50a8\u548c\u64cd\u4f5c\u5bc6\u96c6\u6570\u503c\u6570\u7ec4\u63d0\u4f9b\u4e86\u57fa\u7840，\u914d\u5408\u5927\u91cf\u7ecf\u8fc7\u4f18\u5316\u7684\u6570\u5b66\u51fd\u6570\u5e93（\u901a\u5e38\u901a\u8fc7C\u3001C++\u6216Fortran\u4ee3\u7801\u5b9e\u73b0），\u5728\u5904\u7406\u6570\u7ec4\u8fd0\u7b97\u65f6\u8fdc\u8d85\u7eafPython\u5faa\u73af\u7684\u6027\u80fd\u3002<\/p>\n
NumPy\u6784\u6210\u4e86\u6574\u4e2a\u79d1\u5b66Python\u751f\u6001\u7cfb\u7edf\u7684\u57fa\u77f3，\u5305\u62ecSciPy\u3001Pandas\u3001Scikit-learn\u3001Matplotlib\u7b49\u91cd\u8981\u5e93\u90fd\u5efa\u7acb\u5728\u5176\u4e4b\u4e0a\u3002\u7136\u800c，\u5176\u8bbe\u8ba1\u4e3b\u8981\u9762\u5411CPU\u6267\u884c\u73af\u5883，\u4e14\u672c\u8eab\u4e0d\u5177\u5907\u81ea\u52a8\u68af\u5ea6\u8ba1\u7b97\u80fd\u529b\u3002<\/p>\n
\u68af\u5ea6\u8ba1\u7b97\u7684\u91cd\u8981\u6027\u4e0d\u5bb9\u5ffd\u89c6\u3002\u68af\u5ea6\u8861\u91cf\u51fd\u6570\u8f93\u51fa\u76f8\u5bf9\u4e8e\u8f93\u5165\u7684\u53d8\u5316\u7387，\u6307\u793a\u51fd\u6570\u503c\u589e\u51cf\u6700\u5feb\u7684\u65b9\u5411\u3002\u8fd9\u4e00\u4fe1\u606f\u5bf9\u4f18\u5316\u7b97\u6cd5\u81f3\u5173\u91cd\u8981，\u7279\u522b\u662f\u5728\u673a\u5668\u5b66\u4e60\u9886\u57df，\u68af\u5ea6\u6307\u5bfc\u6a21\u578b\u53c2\u6570\u7684\u8c03\u6574\u8fc7\u7a0b，\u4ee5\u5728\u8bad\u7ec3\u671f\u95f4\u6700\u5c0f\u5316\u635f\u5931\u51fd\u6570\u3002\u9ad8\u6548\u7684\u68af\u5ea6\u8ba1\u7b97\u80fd\u529b\u4f7f\u6a21\u578b\u80fd\u591f\u4ece\u6570\u636e\u4e2d\u5b66\u4e60\u590d\u6742\u7684\u6a21\u5f0f\u3002<\/p>\n
JAX\u662fGoogle\u5f00\u53d1\u7684\u9ad8\u6027\u80fdPython\u6570\u503c\u8ba1\u7b97\u5e93，\u5c06\u7c7b\u4f3cNumPy\u7684API\u4e0e\u81ea\u52a8\u5fae\u5206（autodiff）\u548c\u52a0\u901f\u786c\u4ef6\u6267\u884c（GPU\/TPU）\u76f8\u7ed3\u5408\u3002<\/p>\n
JAX\u7684\u4e3b\u8981\u6280\u672f\u7279\u6027\u4f53\u73b0\u5728\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762：<\/p>\n
jax.numpy<\/p>\n
\u6a21\u5757\u63d0\u4f9b\u4e86NumPy\u7684\u76f4\u63a5\u66ff\u4ee3\u65b9\u6848，\u4fdd\u6301\u76f8\u540c\u7684API\u63a5\u53e3\u4f46\u652f\u6301GPU\/TPU\u52a0\u901f；<\/p>\n
jax.grad<\/p>\n
\u6a21\u5757\u5b9e\u73b0\u51fd\u6570\u7684\u81ea\u52a8\u5fae\u5206\u529f\u80fd，\u7c7b\u4f3c\u4e8eTensorFlow\u6216PyTorch\u7684\u68af\u5ea6\u8ba1\u7b97\u673a\u5236；<\/p>\n
jax.jit<\/p>\n
\u6a21\u5757\u901a\u8fc7\u52a0\u901f\u7ebf\u6027\u4ee3\u6570（XLA）\u5e93\u63d0\u4f9b\u5373\u65f6\u7f16\u8bd1\u529f\u80fd，\u5b9e\u73b0\u6781\u9ad8\u7684\u6267\u884c\u6548\u7387；<\/p>\n
jax.vmap<\/p>\n
\u548c<\/p>\n
jax.pmap<\/p>\n
\u6a21\u5757\u652f\u6301\u81ea\u52a8\u5411\u91cf\u5316\u548c\u5e76\u884c\u5316\u5904\u7406；\u6b64\u5916JAX\u5728GPU\/TPU\u4e0a\u65e0\u7f1d\u8fd0\u884c，\u65e0\u9700\u4fee\u6539\u73b0\u6709\u4ee3\u7801\u3002<\/p>\n
JAX\u5728\u4ee5\u4e0b\u573a\u666f\u4e2d\u8868\u73b0\u51fa\u663e\u8457\u4f18\u52bf：\u5f53\u9700\u8981\u901a\u8fc7GPU\u6216TPU\u663e\u8457\u52a0\u901f\u7c7b\u4f3cNumPy\u7684\u8ba1\u7b97\u4efb\u52a1\u65f6；\u5f53\u9700\u8981\u81ea\u52a8\u8ba1\u7b97\u6570\u503cPython\u51fd\u6570\u7684\u68af\u5ea6\u4ee5\u8fdb\u884c\u4f18\u5316\u65f6（\u5982\u673a\u5668\u5b66\u4e60\u3001\u7269\u7406\u6a21\u62df\u7b49）；\u5f53\u9700\u8981\u901a\u8fc7JIT\u7f16\u8bd1\u5173\u952ePython\u4ee3\u7801\u6bb5\u4ee5\u83b7\u5f97\u8fdb\u4e00\u6b65\u52a0\u901f\u65f6；\u5f53\u9700\u8981\u8f7b\u677e\u5b9e\u73b0\u51fd\u6570\u5411\u91cf\u5316\u4ee5\u5904\u7406\u6570\u636e\u6279\u6b21\u6216\u5728\u591a\u4e2a\u52a0\u901f\u5668\u8bbe\u5907\u4e0a\u5e76\u884c\u5316\u8ba1\u7b97\u65f6\u3002<\/p>\n
\u5728\u9009\u62e9JAX\u66ff\u4ee3NumPy\u4e4b\u524d，\u9700\u8981\u4e86\u89e3\u4e24\u4e2a\u5e93\u4e4b\u95f4\u7684\u5173\u952e\u5dee\u5f02\u3002\u867d\u7136<\/p>\n
jax.numpy<\/p>\n
\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u6a21\u4effNumPy API，\u4f46\u5728\u4ee5\u4e0b\u65b9\u9762\u5b58\u5728\u663e\u8457\u533a\u522b：<\/p>\n
\u5728\u6267\u884c\u540e\u7aef\u548c\u7f16\u8bd1\u65b9\u9762，NumPy\u5728CPU\u4e0a\u91c7\u7528\u5373\u65f6\u6267\u884c\u6a21\u5f0f，\u901a\u5e38\u4f7f\u7528\u9884\u7f16\u8bd1\u7684C\u3001C++\u6216Fortran\u6269\u5c55\u4ee5\u53ca\u4f18\u5316\u7684\u7ebf\u6027\u4ee3\u6570\u5e93（\u5982OpenBLAS）\u3002\u76f8\u6bd4\u4e4b\u4e0bJAX\u4f7f\u7528XLA\u7f16\u8bd1\u5668\u5c06\u4ee3\u7801\u8f6c\u6362\u4e3a\u9488\u5bf9CPU\u3001GPU\u6216TPU\u4f18\u5316\u7684\u673a\u5668\u4ee3\u7801，\u652f\u6301\u901a\u8fc7<\/p>\n
jax.jit<\/p>\n
\u8fdb\u884c\u5373\u65f6\u7f16\u8bd1，\u5e76\u901a\u5e38\u91c7\u7528\u5f02\u6b65\u5206\u6d3e\u65b9\u5f0f\u3002<\/p>\n
\u5728\u6267\u884c\u6a21\u578b\u65b9\u9762，NumPy\u64cd\u4f5c\u901a\u5e38\u540c\u6b65\u6267\u884c，Python\u89e3\u91ca\u5668\u7b49\u5f85\u64cd\u4f5c\u5b8c\u6210\u540e\u7ee7\u7eed\u6267\u884c\u3002\u800cJAX\u64cd\u4f5c\u5f02\u6b65\u5206\u6d3e\u5230\u52a0\u901f\u5668，Python\u4ee3\u7801\u53ef\u80fd\u5728\u8ba1\u7b97\u8fdb\u884c\u65f6\u7ee7\u7eed\u8fd0\u884c\u3002\u56e0\u6b64，\u901a\u5e38\u9700\u8981\u4f7f\u7528<\/p>\n
result.block_until_ready()<\/p>\n
\u6765\u786e\u4fdd\u51c6\u786e\u8ba1\u65f6\u6216\u5728\u5176\u4ed6\u5730\u65b9\u4f7f\u7528\u7ed3\u679c\u4e4b\u524d\u786e\u4fdd\u7ed3\u679c\u53ef\u7528\u3002<\/p>\n
\u5728\u6570\u636e\u53ef\u53d8\u6027\u65b9\u9762，NumPy\u6570\u7ec4（ndarray）\u662f\u53ef\u53d8\u7684，\u5141\u8bb8\u5c31\u5730\u4fee\u6539\u5143\u7d20\u3002JAX\u6570\u7ec4\u5219\u662f\u4e0d\u53ef\u53d8\u7684，\u4e0d\u5141\u8bb8\u5c31\u5730\u66f4\u65b0\u3002\u8fd9\u79cd\u51fd\u6570\u5f0f\u7f16\u7a0b\u65b9\u6cd5\u786e\u4fddJAX\u7684\u8f6c\u6362\u529f\u80fd\u80fd\u591f\u53ef\u9760\u5de5\u4f5c\u800c\u4e0d\u4ea7\u751f\u526f\u4f5c\u7528，\u66f4\u65b0\u64cd\u4f5c\u9700\u8981\u4f7f\u7528\u7d22\u5f15\u66f4\u65b0\u8bed\u6cd5\u521b\u5efa\u65b0\u6570\u7ec4\u3002<\/p>\n
\u5728\u968f\u673a\u6570\u751f\u6210\u65b9\u9762，NumPy\u4f7f\u7528\u5168\u5c40\u968f\u673a\u6570\u751f\u6210\u5668\u72b6\u6001，\u8fd9\u5728\u5e76\u884c\u6216\u8f6c\u6362\u4ee3\u7801\u4e2d\u53ef\u80fd\u5f71\u54cd\u53ef\u91cd\u73b0\u6027\u3002JAX\u5219\u9700\u8981\u663e\u5f0f\u5904\u7406\u968f\u673a\u5bc6\u94a5，\u5fc5\u987b\u624b\u52a8\u7ba1\u7406\u548c\u5206\u5272\u5bc6\u94a5\u4ee5\u786e\u4fdd\u968f\u673a\u6027\u7684\u53ef\u91cd\u73b0\u6027\u3002<\/p>\n
\u5728API\u8986\u76d6\u8303\u56f4\u65b9\u9762，NumPy\u63d0\u4f9b\u6db5\u76d6\u6570\u503c\u8ba1\u7b97\u5404\u4e2a\u65b9\u9762\u7684\u5168\u9762API，\u800cJAX\u8986\u76d6\u6700\u5e38\u89c1NumPy API\u7684\u5927\u90e8\u5206\u5b50\u96c6\u4e14\u5728\u4e0d\u65ad\u6269\u5c55，\u4f46\u5e76\u975e100%\u7684\u76f4\u63a5\u66ff\u4ee3\u54c1\u3002\u4e00\u4e9b\u4e0d\u5e38\u89c1\u7684\u51fd\u6570\u3001\u7279\u5b9a\u6570\u636e\u7c7b\u578b（\u5982\u5bf9\u8c61\u6570\u7ec4）\u6216\u7279\u5b9a\u884c\u4e3a\u53ef\u80fd\u5b58\u5728\u5dee\u5f02\u6216\u7f3a\u5931\u3002<\/p>\n
\u672c\u6587\u7684\u4ee3\u7801\u793a\u4f8b\u57fa\u4e8eWSL2 Ubuntu for Windows\u5f00\u53d1\u73af\u5883\u3002\u5bf9\u4e8e\u62e5\u6709Nvidia GPU\u7684\u7cfb\u7edf，\u53ef\u4ee5\u5145\u5206\u5229\u7528GPU\u52a0\u901f\u529f\u80fd\u3002\u5373\u4f7f\u6ca1\u6709GPU，JAX\u4ecd\u80fd\u5728CPU\u4e0a\u63d0\u4f9b\u76f8\u6bd4NumPy\u66f4\u597d\u7684\u6027\u80fd\u3002\u5bf9\u4e8e\u4e0d\u540c\u7684GPU\u54c1\u724c\u6216\u914d\u7f6e，\u5efa\u8bae\u53c2\u8003\u5b98\u65b9\u6587\u6863\u83b7\u53d6\u8be6\u7ec6\u7684\u5b89\u88c5\u8bf4\u660e\u3002<\/p>\n
\u9996\u5148\u521b\u5efa\u4e13\u7528\u7684\u5f00\u53d1\u73af\u5883，\u8fd9\u91cc\u4f7f\u7528conda\u8fdb\u884c\u73af\u5883\u7ba1\u7406：<\/p>\n
conda create -n jax_test python=3.13 -y <\/p>\n
\u6fc0\u6d3b\u73af\u5883\u5e76\u5b89\u88c5\u5fc5\u8981\u7684\u5e93\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f，\u5bf9\u4e8eNVIDIA GPU，\u9700\u8981\u786e\u4fdd\u5b89\u88c5\u4e86\u9002\u5f53\u7684NVIDIA\u9a71\u52a8\u7a0b\u5e8f\u548cCUDA\u73af\u5883（\u5982CUDA 11\u6216CUDA 12）：<\/p>\n
conda activate jax_test
\n pip install jupyter numpy "jax[cuda12]" matplotlib pillow<\/p>\n
JAX\u7684\u5b89\u88c5\u8fc7\u7a0b\u53ef\u80fd\u8f83\u4e3a\u8017\u65f6，\u5b8c\u6210\u540e\u53ef\u4ee5\u542f\u52a8Jupyter notebook\u3002\u5982\u679c\u6d4f\u89c8\u5668\u672a\u81ea\u52a8\u6253\u5f00，\u53ef\u4ee5\u4ece\u547d\u4ee4\u884c\u8f93\u51fa\u4e2d\u627e\u5230\u7c7b\u4f3c\u4ee5\u4e0b\u683c\u5f0f\u7684URL：<\/p>\n
http:\/\/127.0.0.1:8888\/tree?token=3b9f7bd07b6966b41b68e2350721b2d0b6f388d248cc69d<\/p>\n
\u7b2c\u4e00\u4e2a\u6848\u4f8b\u5c55\u793a\u4e86JAX\u5982\u4f55\u901a\u8fc7\u5373\u65f6\u7f16\u8bd1\u6280\u672f\u63d0\u5347NumPy\u7684\u6027\u80fd\u3002\u8fd9\u4e2a\u6848\u4f8b\u5b9e\u73b0\u4e86\u5e94\u7528\u4e8e10,000 x 10,000\u6570\u7ec4\u7684SELU（Scaled Exponential Linear Unit）\u51fd\u6570\u3002SELU\u662f\u81ea\u5f52\u4e00\u5316\u795e\u7ecf\u7f51\u7edc\u4e2d\u5e7f\u6cdb\u4f7f\u7528\u7684\u6fc0\u6d3b\u51fd\u6570，\u5176\u6570\u5b66\u5b9a\u4e49\u5982\u4e0b\u56fe\u6240\u793a： ![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250815120954-689f2392ee1eb.jpg\")
<\/p>\n
\u8be5\u51fd\u6570\u901a\u8fc7<\/p>\n
np.where<\/p>\n
（NumPy）\u6216<\/p>\n
jnp.where<\/p>\n
（JAX）\u5b9e\u73b0，\u6839\u636e\u8f93\u5165\u503c<\/p>\n
x<\/p>\n
\u7684\u6b63\u8d1f\u6027\u9009\u62e9\u4e0d\u540c\u7684\u8ba1\u7b97\u516c\u5f0f\u3002<\/p>\n
\u4ee3\u7801\u5b9e\u73b0\u5305\u542b\u4e09\u79cd\u7248\u672c：<\/p>\n
selu_numpy(x)<\/p>\n
\u4e3a\u6807\u51c6NumPy\u5b9e\u73b0；<\/p>\n
selu_jax(x)<\/p>\n
\u4e3aJAX\u7248\u672c，\u4ee3\u7801\u7ed3\u6784\u76f8\u540c\u4f46\u4f7f\u7528JAX\u6570\u7ec4；<\/p>\n
selu_jax_jit(x)<\/p>\n
\u5728JAX\u7248\u672c\u57fa\u7840\u4e0a\u6dfb\u52a0<\/p>\n
@jax.jit<\/p>\n
\u88c5\u9970\u5668\u4ee5\u542f\u7528\u7f16\u8bd1\u4f18\u5316\u3002<\/p>\n
\u6d4b\u8bd5\u4f7f\u7528\u4e00\u4e2a10,000 x 10,000\u7684\u968f\u673a\u6570\u6570\u7ec4\u4f5c\u4e3a\u8f93\u5165\u6570\u636e，\u5206\u522b\u6d4b\u91cf\u5404\u5b9e\u73b0\u7684\u6267\u884c\u65f6\u95f4\u3002\u5176\u4e2d，<\/p>\n
selu_numpy<\/p>\n
\u5728CPU\u4e0a\u76f4\u63a5\u8fd0\u884c；<\/p>\n
selu_jax<\/p>\n
\u5728\u6ca1\u6709JIT\u4f18\u5316\u7684\u60c5\u51b5\u4e0b\u8fd0\u884c，\u7531\u4e8e\u9700\u8981\u89e3\u91ca\u6267\u884c\u800c\u76f8\u5bf9\u8f83\u6162；<\/p>\n
selu_jax_jit<\/p>\n
\u9996\u6b21\u8fd0\u884c\u65f6\u5305\u542b\u7f16\u8bd1\u65f6\u95f4，\u4f46\u5728\u540e\u7eed\u8fd0\u884c\u4e2d\u91cd\u7528\u7f16\u8bd1\u540e\u7684\u51fd\u6570，\u6267\u884c\u901f\u5ea6\u663e\u8457\u63d0\u5347\u3002<\/p>\n
\u9700\u8981\u7279\u522b\u6ce8\u610f\u7684\u662f，\u7531\u4e8eJAX\u91c7\u7528\u5f02\u6b65\u6267\u884c\u6a21\u5f0f，\u9700\u8981\u4f7f\u7528<\/p>\n
block_until_ready()<\/p>\n
\u7b49\u5f85\u64cd\u4f5c\u5b8c\u6210\u4ee5\u83b7\u5f97\u51c6\u786e\u7684\u6027\u80fd\u6d4b\u91cf\u7ed3\u679c\u3002\u9996\u6b21JIT\u8fd0\u884c\u5305\u542b\u7f16\u8bd1\u5f00\u9500，\u800c\u7b2c\u4e8c\u6b21\u8fd0\u884c\u5219\u76f4\u63a5\u4f7f\u7528\u7f13\u5b58\u7684\u7f16\u8bd1\u7ed3\u679c，\u5b9e\u73b0\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n
import numpy as np
\nimport jax
\nimport jax.numpy as jnp
\nfrom timeit import default_timer as timer <\/p>\n
# \u4e3aSELU\u5b9a\u4e49\u5e38\u6570
\nalpha = 1.6732632423543772848170429916717
\nscale = 1.0507009873554804934193349852946 <\/p>\n
# — NumPy\u7248\u672c —
\ndef selu_numpy(x):
\n return scale * np.where(x > 0, x, alpha * np.exp(x) – alpha) <\/p>\n
# — JAX\u7248\u672c —
\ndef selu_jax(x):
\n return scale * jnp.where(x > 0, x, alpha * jnp.exp(x) – alpha) <\/p>\n
# — JIT\u7f16\u8bd1\u7684JAX\u7248\u672c —
\n# \u5e94\u7528@jax.jit\u88c5\u9970\u5668
\n@jax.jit
\ndef selu_jax_jit(x):
\n return scale * jnp.where(x > 0, x, alpha * jnp.exp(x) – alpha) <\/p>\n
# \u751f\u6210\u6d4b\u8bd5\u6570\u636e
\nx_np = np.random.rand(10000, 10000).astype(np.float32)
\n# \u4f7f\u7528JAX\u7684\u968f\u673a\u6570\u751f\u6210（\u9700\u8981\u663e\u5f0f\u5bc6\u94a5\u7ba1\u7406）
\nkey = jax.random.PRNGKey(0)
\nx_jax = jax.random.normal(key, (10000, 10000), dtype=jnp.float32) <\/p>\n
print("\u6267\u884c\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5…") <\/p>\n
# — \u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5 — <\/p>\n
# NumPy\u57fa\u51c6\u6d4b\u8bd5
\nstart = timer()
\nresult_np = selu_numpy(x_np)
\n# NumPy\u5728CPU\u4e0a\u540c\u6b65\u6267\u884c，\u65e0\u9700\u7b49\u5f85
\nprint(f"NumPy\u6267\u884c\u65f6\u95f4: {timer()-start:.6f} \u79d2") <\/p>\n
# JAX（\u65e0JIT）\u57fa\u51c6\u6d4b\u8bd5
\nstart = timer()
\nresult_jax = selu_jax(x_jax)
\nresult_jax.block_until_ready() # \u5173\u952e\u6b65\u9aa4：\u7b49\u5f85JAX\u8ba1\u7b97\u5b8c\u6210
\nprint(f"JAX (\u65e0JIT)\u6267\u884c\u65f6\u95f4: {timer()-start:.6f} \u79d2") <\/p>\n
# JAX（JIT）\u9996\u6b21\u8fd0\u884c（\u5305\u542b\u7f16\u8bd1\u65f6\u95f4）
\nstart = timer()
\nresult_jax_jit = selu_jax_jit(x_jax)
\nresult_jax_jit.block_until_ready()
\nprint(f"JAX (JIT)\u9996\u6b21\u8fd0\u884c\u65f6\u95f4（\u542b\u7f16\u8bd1）: {timer()-start:.6f} \u79d2") <\/p>\n
# JAX（JIT）\u7b2c\u4e8c\u6b21\u8fd0\u884c（\u4f7f\u7528\u7f13\u5b58\u7f16\u8bd1\u7ed3\u679c）
\nstart = timer()
\nresult_jax_jit_2 = selu_jax_jit(x_jax)
\nresult_jax_jit_2.block_until_ready()
\nprint(f"JAX (JIT)\u7b2c\u4e8c\u6b21\u8fd0\u884c\u65f6\u95f4: {timer()-start:.6f} \u79d2") <\/p>\n
# \u9a8c\u8bc1\u8ba1\u7b97\u7ed3\u679c\u7684\u4e00\u81f4\u6027
\n print(np.allclose(selu_numpy(np.array(x_jax)), result_jax_jit_2, atol=1e-6))<\/p>\n
\u6d4b\u8bd5\u7ed3\u679c\u663e\u793a：<\/p>\n
\u6267\u884c\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5…
\n NumPy\u6267\u884c\u65f6\u95f4: 0.357104 \u79d2
\n JAX (\u65e0JIT)\u6267\u884c\u65f6\u95f4: 0.108734 \u79d2
\n JAX (JIT)\u9996\u6b21\u8fd0\u884c\u65f6\u95f4（\u542b\u7f16\u8bd1）: 0.026956 \u79d2
\n JAX (JIT)\u7b2c\u4e8c\u6b21\u8fd0\u884c\u65f6\u95f4: 0.002400 \u79d2
\n True<\/p>\n
\u7ed3\u679c\u8868\u660e，JAX\u7684JIT\u7f16\u8bd1\u7248\u672c\u5728\u7b2c\u4e8c\u6b21\u8fd0\u884c\u65f6\u76f8\u6bd4NumPy\u5b9e\u73b0\u4e86\u8d85\u8fc7100\u500d\u7684\u6027\u80fd\u63d0\u5347，\u5373\u4f7f\u4e0d\u4f7f\u7528JIT\u4f18\u5316，JAX\u7684\u6027\u80fd\u4e5f\u6bd4NumPy\u63d0\u5347\u7ea63\u500d\u3002<\/p>\n
\u4e0b\u9762\u8fd9\u4e2a\u6848\u4f8b\u5c06\u5c55\u793a\u4e86JAX\u5f3a\u5927\u7684\u81ea\u52a8\u5fae\u5206\u80fd\u529b，\u8fd9\u662f\u5176\u533a\u522b\u4e8eNumPy\u7684\u6838\u5fc3\u7279\u6027\u4e4b\u4e00\u3002<\/p>\n
import jax
\nimport jax.numpy as jnp <\/p>\n
# \u4f7f\u7528jax.numpy\u5b9a\u4e49\u76ee\u6807\u51fd\u6570
\ndef cubic_sum(x):
\n return jnp.sum(x**3) <\/p>\n
# \u901a\u8fc7jax.grad\u83b7\u53d6\u68af\u5ea6\u51fd\u6570
\ngrad_cubic_sum = jax.grad(cubic_sum) <\/p>\n
# \u521b\u5efa\u8f93\u5165\u6570\u636e
\nx_input = jnp.arange(1.0, 5.0) <\/p>\n
# \u8ba1\u7b97\u68af\u5ea6
\ngradient = grad_cubic_sum(x_input)
\nprint(f"\\\\n— \u81ea\u52a8\u5fae\u5206\u793a\u4f8b —")
\nprint(f"\u539f\u59cb\u51fd\u6570\u8f93\u5165: {x_input}")
\nprint(f"\u51fd\u6570\u8f93\u51fa f(x): {cubic_sum(x_input)}")
\n print(f"\u68af\u5ea6 df\/dx: {gradient}")<\/p>\n
\u6267\u884c\u7ed3\u679c：<\/p>\n
— \u81ea\u52a8\u5fae\u5206\u793a\u4f8b —
\n \u539f\u59cb\u51fd\u6570\u8f93\u5165: [1. 2. 3. 4.]
\n \u51fd\u6570\u8f93\u51fa f(x): 100.0
\n \u68af\u5ea6 df\/dx: [ 3. 12. 27. 48.]<\/p>\n
\u5728\u8fd9\u4e2a\u793a\u4f8b\u4e2d，\u76ee\u6807\u51fd\u6570\u4e3ax\u00b3，\u5176\u5bfc\u6570\u4e3a3x\u00b2\u3002\u5bf9\u4e8e\u8f93\u5165\u5e8f\u5217[1, 2, 3, 4]，\u76f8\u5e94\u7684\u68af\u5ea6\u8ba1\u7b97\u4e3a：<\/p>\n
\u8fd9\u4e2a\u7b80\u5355\u7684\u4f8b\u5b50\u5c55\u793a\u4e86JAX\u81ea\u52a8\u5fae\u5206\u7684\u5f3a\u5927\u529f\u80fd，\u65e0\u9700\u624b\u52a8\u63a8\u5bfc\u548c\u5b9e\u73b0\u68af\u5ea6\u8ba1\u7b97，JAX\u80fd\u591f\u81ea\u52a8\u5904\u7406\u590d\u6742\u51fd\u6570\u7684\u5fae\u5206\u8fd0\u7b97\u3002<\/p>\n
\u6211\u4eec\u8fd8\u901a\u8fc7\u5411\u91cf\u5316\u64cd\u4f5c\u5c55\u793a\u4e86JAX\u7684\u6027\u80fd\u4f18\u52bf，\u5177\u4f53\u5b9e\u73b0\u4e86\u5c06\u4e00\u4e2a10,000\u5143\u7d20\u77e9\u9635\u4e0e128\u4e2a10,000\u5143\u7d20\u5411\u91cf\u6279\u6b21\u76f8\u4e58\u7684\u8ba1\u7b97\u4efb\u52a1\u3002<\/p>\n
import numpy as np
\nimport jax
\nimport jax.numpy as jnp
\nfrom timeit import default_timer as timer <\/p>\n
# — \u5355\u4e2a\u6570\u636e\u70b9\u7684\u57fa\u7840\u51fd\u6570 —
\ndef mat_vec_product(matrix, vector):
\n """\u8ba1\u7b97\u77e9\u9635-\u5411\u91cf\u4e58\u79ef"""
\n return jnp.dot(matrix, vector) <\/p>\n
# — \u4f7f\u7528vmap\u521b\u5efa\u6279\u5904\u7406\u7248\u672c —
\n# \u76ee\u6807\u662f\u5c06mat_vec_product\u5e94\u7528\u5230\u6279\u6b21\u4e2d\u7684\u6bcf\u4e2a\u5411\u91cf
\n# \u77e9\u9635\u5bf9\u6279\u6b21\u4e2d\u7684\u6240\u6709\u5411\u91cf\u4fdd\u6301\u4e0d\u53d8
\n# in_axes=(None, 0)\u7684\u542b\u4e49：
\n# None: \u4e0d\u5bf9\u7b2c\u4e00\u4e2a\u53c2\u6570(matrix)\u8fdb\u884c\u6620\u5c04，\u91c7\u7528\u5e7f\u64ad\u65b9\u5f0f
\n# 0: \u5bf9\u7b2c\u4e8c\u4e2a\u53c2\u6570\u7684\u7b2c0\u8f74(\u5411\u91cf\u6279\u6b21)\u8fdb\u884c\u6620\u5c04
\nbatched_mat_vec = jax.vmap(mat_vec_product, in_axes=(None, 0)) <\/p>\n
# — JIT\u7f16\u8bd1\u7684\u5411\u91cf\u5316\u51fd\u6570 —
\n@jax.jit
\ndef batched_mat_vec_jit(matrix, vectors):
\n """JIT\u7f16\u8bd1\u7684\u6279\u5904\u7406\u77e9\u9635-\u5411\u91cf\u4e58\u6cd5"""
\n return jax.vmap(mat_vec_product, in_axes=(None, 0))(matrix, vectors) <\/p>\n
# — \u6570\u636e\u914d\u7f6e —
\nmatrix_size = 10000
\nvector_size = 10000
\nbatch_size = 128
\ndtype = jnp.float32 <\/p>\n
# JAX\u968f\u673a\u6570\u751f\u6210\u9700\u8981\u663e\u5f0f\u5bc6\u94a5\u7ba1\u7406
\nkey = jax.random.PRNGKey(0)
\nkey, subkey1, subkey2 = jax.random.split(key, 3) <\/p>\n
# \u751f\u6210\u6d4b\u8bd5\u6570\u636e
\nmatrix_jax = jax.random.normal(subkey1, (matrix_size, vector_size), dtype=dtype)
\nvectors_jax = jax.random.normal(subkey2, (batch_size, vector_size), dtype=dtype) <\/p>\n
# \u8f6c\u6362\u4e3aNumPy\u683c\u5f0f\u4ee5\u8fdb\u884c\u5bf9\u6bd4\u6d4b\u8bd5
\nmatrix_np = np.array(matrix_jax)
\nvectors_np = np.array(vectors_jax) <\/p>\n
print(f"\\\\n— vmap\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5 (\u77e9\u9635: {matrix_size}x{vector_size}, \u6279\u6b21\u5927\u5c0f: {batch_size}) —")
\nprint(f"\u53ef\u7528JAX\u8bbe\u5907: {jax.devices()}") <\/p>\n
# — \u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5 — <\/p>\n
# NumPy\u65b9\u6cd51：Python\u5faa\u73af\u5b9e\u73b0（\u4ec5\u4f5c\u6f14\u793a，\u901a\u5e38\u6027\u80fd\u8f83\u5dee）
\nstart_np_loop = timer()
\noutput_np_loop = np.array([np.dot(matrix_np, v) for v in vectors_np])
\nend_np_loop = timer()
\nprint(f"NumPy (Python\u5faa\u73af)\u6267\u884c\u65f6\u95f4: {end_np_loop – start_np_loop:.6f} \u79d2") <\/p>\n
# NumPy\u65b9\u6cd52：\u77e9\u9635\u4e58\u6cd5\u4e0e\u8f6c\u7f6e\u64cd\u4f5c（\u9ad8\u6548\u5b9e\u73b0\u65b9\u5f0f）
\nstart_np_matmul = timer()
\n# \u77e9\u9635\u4e58\u6cd5\u8981\u6c42vectors_np\u7684\u5f62\u72b6\u4e3a(vector_size, batch_size)
\noutput_np_matmul = (matrix_np @ vectors_np.T).T
\nend_np_matmul = timer()
\nprint(f"NumPy (\u77e9\u9635\u4e58\u6cd5@)\u6267\u884c\u65f6\u95f4: {end_np_matmul – start_np_matmul:.6f} \u79d2") <\/p>\n
# JAX vmap（\u65e0JIT\u4f18\u5316）
\nstart_jax_vmap = timer()
\noutput_jax_vmap = batched_mat_vec(matrix_jax, vectors_jax)
\noutput_jax_vmap.block_until_ready()
\nend_jax_vmap = timer()
\nprint(f"JAX (vmap, \u65e0JIT)\u6267\u884c\u65f6\u95f4: {end_jax_vmap – start_jax_vmap:.6f} \u79d2") <\/p>\n
# JAX vmap（JIT\u4f18\u5316）\u9996\u6b21\u8fd0\u884c（\u5305\u542b\u7f16\u8bd1\u5f00\u9500）
\nstart_jax_vmap_jit_compile = timer()
\noutput_jax_vmap_jit_compile = batched_mat_vec_jit(matrix_jax, vectors_jax)
\noutput_jax_vmap_jit_compile.block_until_ready()
\nend_jax_vmap_jit_compile = timer()
\nprint(f"JAX (vmap+JIT)\u9996\u6b21\u8fd0\u884c\u65f6\u95f4（\u542b\u7f16\u8bd1）: {end_jax_vmap_jit_compile – start_jax_vmap_jit_compile:.6f} \u79d2") <\/p>\n
# JAX vmap（JIT\u4f18\u5316）\u7b2c\u4e8c\u6b21\u8fd0\u884c
\nstart_jax_vmap_jit = timer()
\noutput_jax_vmap_jit = batched_mat_vec_jit(matrix_jax, vectors_jax)
\noutput_jax_vmap_jit.block_until_ready()
\nend_jax_vmap_jit = timer()
\n print(f"JAX (vmap+JIT)\u7b2c\u4e8c\u6b21\u8fd0\u884c\u65f6\u95f4: {end_jax_vmap_jit – start_jax_vmap_jit:.6f} \u79d2")<\/p>\n
\u6027\u80fd\u6d4b\u8bd5\u7ed3\u679c：<\/p>\n
— vmap\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5 (\u77e9\u9635: 10000×10000, \u6279\u6b21\u5927\u5c0f: 128) —
\n \u53ef\u7528JAX\u8bbe\u5907: [CudaDevice(id=0)]
\n NumPy (Python\u5faa\u73af)\u6267\u884c\u65f6\u95f4: 1.129315 \u79d2
\n NumPy (\u77e9\u9635\u4e58\u6cd5@)\u6267\u884c\u65f6\u95f4: 0.029319 \u79d2
\n JAX (vmap, \u65e0JIT)\u6267\u884c\u65f6\u95f4: 0.901569 \u79d2
\n JAX (vmap+JIT)\u9996\u6b21\u8fd0\u884c\u65f6\u95f4（\u542b\u7f16\u8bd1）: 0.539354 \u79d2
\n JAX (vmap+JIT)\u7b2c\u4e8c\u6b21\u8fd0\u884c\u65f6\u95f4: 0.001776 \u79d2<\/p>\n
\u867d\u7136\u9996\u6b21JIT\u7f16\u8bd1\u9700\u8981\u76f8\u5bf9\u8f83\u957f\u7684\u65f6\u95f4，\u4f46\u540e\u7eed\u8fd0\u884c\u7684\u6027\u80fd\u63d0\u5347\u6781\u4e3a\u663e\u8457，\u4f53\u73b0\u4e86JAX\u7f16\u8bd1\u4f18\u5316\u7684\u5f3a\u5927\u6548\u679c\u3002<\/p>\n
\u6700\u540e\u4e00\u4e2a\u6848\u4f8b\u4ee5\u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u65af\u6a21\u7cca\u4e3a\u4f8b，\u5c55\u793a\u4e86JAX\u5728\u5b9e\u9645\u5e94\u7528\u573a\u666f\u4e2d\u7684\u6027\u80fd\u8868\u73b0\u3002\u5377\u79ef\u662f\u56fe\u50cf\u5904\u7406\u7684\u57fa\u7840\u64cd\u4f5c，\u5e7f\u6cdb\u5e94\u7528\u4e8e\u6a21\u7cca\u3001\u9510\u5316\u548c\u8fb9\u7f18\u68c0\u6d4b\u7b49\u4efb\u52a1\u3002\u8be5\u64cd\u4f5c\u6d89\u53ca\u5728\u56fe\u50cf\u4e0a\u6ed1\u52a8\u5c0f\u77e9\u9635（\u5377\u79ef\u6838），\u5e76\u5728\u6bcf\u4e2a\u4f4d\u7f6e\u8ba1\u7b97\u6838\u4e0b\u50cf\u7d20\u7684\u52a0\u6743\u548c\u3002\u672c\u6848\u4f8b\u901a\u8fc7\u6570\u7ec4\u5207\u7247\u548c\u9010\u5143\u7d20\u64cd\u4f5c\u5b9e\u73b0\u9ad8\u65af\u6a21\u7cca\u7684\u57fa\u672c\u7248\u672c，\u4ee5\u8bc4\u4f30jax.jit\u5bf9\u6b64\u7c7b\u64cd\u4f5c\u5e8f\u5217\u7684\u4f18\u5316\u6548\u679c\u3002<\/p>\n
\u8f93\u5165\u56fe\u50cf\u793a\u4f8b： [\u5916\u94fe\u56fe\u7247\u8f6c\u5b58\u4e2d\u2026(img-b59nsNpi-1755002028476)]<\/p>\n
\u539f\u59cb\u56fe\u50cf\u7531Yury Taranik\u63d0\u4f9b<\/p>\n
import numpy as np
\nimport jax
\nimport jax.numpy as jnp
\nfrom timeit import default_timer as timer
\nfrom PIL import Image
\nimport matplotlib.pyplot as plt
\nimport os <\/p>\n
# — \u914d\u7f6e\u53c2\u6570 —
\nimage_path = "\/mnt\/d\/images\/taj_mahal.png"
\nkernel_size = 9 # \u589e\u5927\u5377\u79ef\u6838\u4ee5\u83b7\u5f97\u66f4\u660e\u663e\u7684\u6a21\u7cca\u6548\u679c
\nsigma = 2.5
\ndtype = jnp.float32 <\/p>\n
# — \u56fe\u50cf\u6587\u4ef6\u68c0\u67e5 —
\nif not os.path.exists(image_path):
\n print(f"\u9519\u8bef：\u5728\u8def\u5f84 '{image_path}' \u672a\u627e\u5230\u56fe\u50cf\u6587\u4ef6")
\n print("\u8bf7\u66f4\u65b0\u811a\u672c\u4e2d\u7684 'image_path' \u53d8\u91cf")
\n exit() <\/p>\n
# — \u56fe\u50cf\u52a0\u8f7d\u548c\u9884\u5904\u7406 —
\nprint(f"\u4ece\u4ee5\u4e0b\u8def\u5f84\u52a0\u8f7d\u56fe\u50cf: {image_path}")
\ntry:
\n # \u6253\u5f00\u56fe\u50cf，\u8f6c\u6362\u4e3a\u7070\u5ea6\u6a21\u5f0f，\u7136\u540e\u8f6c\u6362\u4e3aNumPy\u6570\u7ec4
\n with Image.open(image_path) as img:
\n image_np_uint8 = np.array(img.convert('L')) <\/p>\n
# \u5f52\u4e00\u5316\u4e3a0.0\u52301.0\u8303\u56f4\u7684float32\u7c7b\u578b
\n image_np = image_np_uint8.astype(np.float32) \/ 255.0
\n image_jax = jnp.array(image_np) <\/p>\n
image_size_h, image_size_w = image_np.shape
\n print(f"\u56fe\u50cf\u52a0\u8f7d\u6210\u529f ({image_size_h}x{image_size_w})") <\/p>\n
except Exception as e:
\n print(f"\u9519\u8bef：\u65e0\u6cd5\u52a0\u8f7d\u6216\u5904\u7406\u56fe\u50cf '{image_path}'\u3002\u9519\u8bef\u4fe1\u606f: {e}")
\n exit() <\/p>\n
# — \u9ad8\u65af\u5377\u79ef\u6838\u751f\u6210 —
\ndef gaussian_kernel(size, sigma=1.0):
\n """\u751f\u62102D\u9ad8\u65af\u5377\u79ef\u6838"""
\n ax = jnp.arange(-size \/\/ 2 + 1., size \/\/ 2 + 1.)
\n xx, yy = jnp.meshgrid(ax, ax)
\n kernel = jnp.exp(-(xx**2 + yy**2) \/ (2. * sigma**2))
\n return (kernel \/ jnp.sum(kernel)).astype(dtype) <\/p>\n
# — \u624b\u52a8\u5377\u79ef\u5b9e\u73b0 —
\ndef convolve_2d_manual(image, kernel):
\n """\u4f7f\u7528\u57fa\u672c\u6570\u7ec4\u64cd\u4f5c\u5b9e\u73b02D\u5377\u79ef"""
\n im_h, im_w = image.shape
\n ker_h, ker_w = kernel.shape
\n pad_h, pad_w = ker_h \/\/ 2, ker_w \/\/ 2
\n padded_image = jnp.pad(image, ((pad_h, pad_h), (pad_w, pad_w)), mode='edge')
\n output = jnp.zeros_like(image)
\n for i in range(ker_h):
\n for j in range(ker_w):
\n # \u4f7f\u7528dynamic_slice\u786e\u4fdd\u4e0eJIT\u517c\u5bb9
\n image_slice = jax.lax.dynamic_slice(padded_image, (i, j), (im_h, im_w))
\n output += kernel[i, j] * image_slice
\n return output <\/p>\n
# — JIT\u7f16\u8bd1\u7248\u672c —
\n@jax.jit
\ndef convolve_2d_manual_jit(image, kernel):
\n """JIT\u7f16\u8bd1\u76842D\u5377\u79ef\u5b9e\u73b0"""
\n im_h, im_w = image.shape
\n ker_h, ker_w = kernel.shape
\n pad_h, pad_w = ker_h \/\/ 2, ker_w \/\/ 2
\n padded_image = jnp.pad(image, ((pad_h, pad_h), (pad_w, pad_w)), mode='edge')
\n output = jnp.zeros_like(image)
\n for i in range(ker_h):
\n for j in range(ker_w):
\n # \u5f53\u5207\u7247\u5927\u5c0f\u52a8\u6001\u53d8\u5316\u65f6，\u4f7f\u7528jax.lax.dynamic_slice\u786e\u4fddJIT\u517c\u5bb9\u6027
\n image_slice = jax.lax.dynamic_slice(padded_image, (i, j), (im_h, im_w))
\n output += kernel[i, j] * image_slice
\n return output <\/p>\n
# — NumPy\u5bf9\u7167\u5b9e\u73b0 —
\ndef convolve_2d_manual_np(image, kernel):
\n """NumPy\u7248\u672c\u76842D\u5377\u79ef\u5b9e\u73b0"""
\n im_h, im_w = image.shape
\n ker_h, ker_w = kernel.shape
\n pad_h, pad_w = ker_h \/\/ 2, ker_w \/\/ 2
\n padded_image = np.pad(image, ((pad_h, pad_h), (pad_w, pad_w)), mode='edge')
\n output = np.zeros_like(image)
\n for i in range(ker_h):
\n for j in range(ker_w):
\n image_slice = padded_image[i:i + im_h, j:j + im_w]
\n output += kernel[i, j] * image_slice
\n return output <\/p>\n
# — \u5377\u79ef\u6838\u51c6\u5907 —
\nkernel_jax = gaussian_kernel(kernel_size, sigma=sigma)
\nkernel_np = np.array(kernel_jax) <\/p>\n
print(f"\\\\n— \u5377\u79ef\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5 (\u56fe\u50cf: {image_size_h}x{image_size_w}, \u5377\u79ef\u6838: {kernel_size}x{kernel_size}) —")
\nprint(f"\u53ef\u7528JAX\u8bbe\u5907: {jax.devices()}") <\/p>\n
# — \u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5 — <\/p>\n
# NumPy（CPU）\u5b9e\u73b0
\nstart_np = timer()
\noutput_np = convolve_2d_manual_np(image_np, kernel_np)
\nend_np = timer()
\nprint(f"NumPy (\u624b\u52a8\u5377\u79ef)\u6267\u884c\u65f6\u95f4: {end_np – start_np:.6f} \u79d2") <\/p>\n
# JAX（\u65e0JIT）\u5b9e\u73b0
\nstart_jax = timer()
\noutput_jax = convolve_2d_manual(image_jax, kernel_jax)
\noutput_jax.block_until_ready()
\nend_jax = timer()
\nprint(f"JAX (\u65e0JIT, \u624b\u52a8\u5377\u79ef)\u6267\u884c\u65f6\u95f4: {end_jax – start_jax:.6f} \u79d2") <\/p>\n
# JAX（JIT）\u9996\u6b21\u8fd0\u884c（\u5305\u542b\u7f16\u8bd1\u65f6\u95f4）
\nstart_jax_compile = timer()
\noutput_jax_jit_compile = convolve_2d_manual_jit(image_jax, kernel_jax)
\noutput_jax_jit_compile.block_until_ready()
\nend_jax_compile = timer()
\nprint(f"JAX (JIT, \u624b\u52a8\u5377\u79ef)\u9996\u6b21\u8fd0\u884c\u65f6\u95f4（\u542b\u7f16\u8bd1）: {end_jax_compile – start_jax_compile:.6f} \u79d2") <\/p>\n
# JAX（JIT）\u7b2c\u4e8c\u6b21\u8fd0\u884c
\nstart_jax_jit = timer()
\noutput_jax_jit = convolve_2d_manual_jit(image_jax, kernel_jax)
\noutput_jax_jit.block_until_ready()
\nend_jax_jit = timer()
\nprint(f"JAX (JIT, \u624b\u52a8\u5377\u79ef)\u7b2c\u4e8c\u6b21\u8fd0\u884c\u65f6\u95f4: {end_jax_jit – start_jax_jit:.6f} \u79d2") <\/p>\n
# \u7ed3\u679c\u9a8c\u8bc1（\u8003\u8651float32\u7d2f\u79ef\u8bef\u5dee）
\nmax_diff_conv = np.max(np.abs(output_np – output_jax_jit))
\nprint(f"\u5377\u79ef\u7ed3\u679c\u6700\u5927\u7edd\u5bf9\u5dee\u503c: {max_diff_conv:.6f}")
\nprint(f"\u5377\u79ef\u7ed3\u679c\u8fd1\u4f3c\u7a0b\u5ea6 (atol=1e-3, rtol=1e-3): {np.allclose(output_np, output_jax_jit, atol=1e-3, rtol=1e-3)}") <\/p>\n
# — \u7ed3\u679c\u53ef\u89c6\u5316 —
\nprint("\\\\n— \u8f93\u5165\u8f93\u51fa\u56fe\u50cf\u53ef\u89c6\u5316 —") <\/p>\n
fig, axes = plt.subplots(1, 2, figsize=(12, 6)) <\/p>\n
# \u663e\u793a\u539f\u59cb\u7070\u5ea6\u56fe\u50cf
\naxes[0].imshow(image_np, cmap='gray', vmin=0, vmax=1)
\naxes[0].set_title('\u539f\u59cb\u7070\u5ea6\u56fe\u50cf (\u8f93\u5165)')
\naxes[0].axis('off') <\/p>\n
# \u663e\u793a\u6a21\u7cca\u5904\u7406\u540e\u7684\u56fe\u50cf
\naxes[1].imshow(output_np, cmap='gray', vmin=0, vmax=1)
\naxes[1].set_title(f'\u6a21\u7cca\u5904\u7406\u56fe\u50cf (\u8f93\u51fa, \u5377\u79ef\u6838\u5927\u5c0f={kernel_size})')
\naxes[1].axis('off') <\/p>\n
plt.tight_layout()
\n plt.show()<\/p>\n
\u6027\u80fd\u6d4b\u8bd5\u7ed3\u679c：<\/p>\n
\u4ece\u4ee5\u4e0b\u8def\u5f84\u52a0\u8f7d\u56fe\u50cf: \/mnt\/d\/images\/taj_mahal.png
\n\u56fe\u50cf\u52a0\u8f7d\u6210\u529f (473×716) <\/p>\n
— \u5377\u79ef\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5 (\u56fe\u50cf: 473×716, \u5377\u79ef\u6838: 9×9) —
\n\u53ef\u7528JAX\u8bbe\u5907: [CudaDevice(id=0)]
\nNumPy (\u624b\u52a8\u5377\u79ef)\u6267\u884c\u65f6\u95f4: 0.025815 \u79d2
\nJAX (\u65e0JIT, \u624b\u52a8\u5377\u79ef)\u6267\u884c\u65f6\u95f4: 0.234791 \u79d2
\nJAX (JIT, \u624b\u52a8\u5377\u79ef)\u9996\u6b21\u8fd0\u884c\u65f6\u95f4（\u542b\u7f16\u8bd1）: 0.366345 \u79d2
\nJAX (JIT, \u624b\u52a8\u5377\u79ef)\u7b2c\u4e8c\u6b21\u8fd0\u884c\u65f6\u95f4: 0.000238 \u79d2
\n\u5377\u79ef\u7ed3\u679c\u6700\u5927\u7edd\u5bf9\u5dee\u503c: 0.000000
\n \u5377\u79ef\u7ed3\u679c\u8fd1\u4f3c\u7a0b\u5ea6 (atol=1e-3, rtol=1e-3): True<\/p>\n
\u518d\u6b21\u9a8c\u8bc1\u4e86JAX JIT\u7f16\u8bd1\u5728\u7b2c\u4e8c\u6b21\u8fd0\u884c\u65f6\u5b9e\u73b0\u7ea6100\u500d\u7684\u6027\u80fd\u63d0\u5347，\u5c55\u73b0\u4e86\u5176\u5728\u56fe\u50cf\u5904\u7406\u7b49\u8ba1\u7b97\u5bc6\u96c6\u578b\u4efb\u52a1\u4e2d\u7684\u5de8\u5927\u6f5c\u529b\u3002<\/p>\n
\u5904\u7406\u7ed3\u679c\u56fe\u50cf： ![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250815120955-689f2393926b0.jpg\")
<\/p>\n
JAX\u4ee3\u8868\u4e86Python\u9ad8\u6027\u80fd\u6570\u503c\u8ba1\u7b97\u9886\u57df\u7684\u91cd\u8981\u8fdb\u5c55\u3002\u901a\u8fc7\u63d0\u4f9b\u4e0eNumPy\u517c\u5bb9\u7684\u63a5\u53e3，\u7ed3\u5408\u5f3a\u5927\u7684\u51fd\u6570\u8f6c\u6362\u80fd\u529b（\u5305\u62ecgrad\u3001jit\u3001vmap\u7b49）\u4ee5\u53ca\u57fa\u4e8e\u52a0\u901f\u7ebf\u6027\u4ee3\u6570（XLA）\u7684\u9ad8\u6548\u786c\u4ef6\u6267\u884c，JAX\u4e3a\u73b0\u4ee3\u673a\u5668\u5b66\u4e60\u548c\u5927\u89c4\u6a21\u79d1\u5b66\u8ba1\u7b97\u63d0\u4f9b\u4e86\u5fc5\u8981\u7684\u6280\u672f\u652f\u6491\u3002<\/p>\n
JAX\u7684\u663e\u8457\u6027\u80fd\u63d0\u5347\u548c\u5185\u7f6e\u81ea\u52a8\u5fae\u5206\u529f\u80fd\u4f7f\u5176\u6210\u4e3a\u8ba1\u7b97\u79d1\u5b66\u7814\u7a76\u4eba\u5458\u548c\u5de5\u7a0b\u5e08\u7684\u91cd\u8981\u5de5\u5177，\u7279\u522b\u9002\u7528\u4e8e\u5b58\u5728\u6027\u80fd\u74f6\u9888\u6216\u9700\u8981\u68af\u5ea6\u8ba1\u7b97\u7684NumPy\u5e94\u7528\u573a\u666f\u3002<\/p>\n
\u4f5c\u4e3aGoogle\u7684\u5b9e\u9a8c\u6027\u9879\u76ee，JAX\u7684\u672a\u6765\u53d1\u5c55\u4ecd\u5b58\u5728\u4e0d\u786e\u5b9a\u6027\u3002\u867d\u7136\u5176\u6280\u672f\u4f18\u52bf\u660e\u663e，\u4f46\u8003\u8651\u5230\u73b0\u6709NumPy\u4ee3\u7801\u7684\u5e9e\u5927\u57fa\u6570，\u5373\u4f7fJAX\u83b7\u5f97\u5e7f\u6cdb\u91c7\u7528，\u5b8c\u5168\u66ff\u4ee3NumPy\u4ecd\u9700\u8981\u76f8\u5f53\u957f\u7684\u65f6\u95f4\u3002\u4e0e\u6b64\u540c\u65f6，NumPy\u5f00\u53d1\u56e2\u961f\u4e5f\u5728\u6301\u7eed\u6539\u8fdb\u5176\u529f\u80fd\u548c\u6027\u80fd，\u8fd9\u79cd\u826f\u6027\u7ade\u4e89\u5c06\u63a8\u52a8\u6574\u4e2a\u6570\u503c\u8ba1\u7b97\u751f\u6001\u7cfb\u7edf\u7684\u53d1\u5c55\u3002<\/p>\n
\u4ece\u6280\u672f\u53d1\u5c55\u8d8b\u52bf\u6765\u770b，JAX\u4ee3\u8868\u4e86\u6570\u503c\u8ba1\u7b97\u5e93\u5411\u786c\u4ef6\u52a0\u901f\u3001\u81ea\u52a8\u5fae\u5206\u548c\u7f16\u8bd1\u4f18\u5316\u65b9\u5411\u6f14\u8fdb\u7684\u91cd\u8981\u91cc\u7a0b\u7891，\u4e3aPython\u5728\u9ad8\u6027\u80fd\u8ba1\u7b97\u9886\u57df\u7684\u5e94\u7528\u5f00\u8f9f\u4e86\u65b0\u7684\u53ef\u80fd\u6027\u3002<\/p>\n
https:\/\/avoid.overfit.cn\/post\/c4a32ba84033446cace5fed8a7d32e62<\/p>\n
\u4f5c\u8005：Thomas Reid<\/p>\n","protected":false},"excerpt":{"rendered":"
\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb1.1k\u6b21\uff0c\u70b9\u8d5e32\u6b21\uff0c\u6536\u85cf8\u6b21\u3002JAX\u901a\u8fc7GPU\u52a0\u901f\u548cJIT\u7f16\u8bd1\u663e\u8457\u63d0\u5347\u6570\u503c\u8ba1\u7b97\u6027\u80fd\u3002\u76f8\u6bd4NumPy\u7684\u540c\u6b65CPU\u6267\u884c\uff0cJAX\u5728\u65e0JIT\u4f18\u5316\u65f6\u5df2\u83b7\u5f973\u500d\u52a0\u901f\uff0c\u7ecfJIT\u7f16\u8bd1\u540e\u6027\u80fd\u63d0\u5347\u66f4\u4e3a\u663e\u8457\u3002<\/p>\n","protected":false},"author":2,"featured_media":57804,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[6073,1818,81,50,86],"topic":[],"class_list":["post-57807","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-jax","tag-numpy","tag-python","tag-50","tag-86"],"yoast_head":"\n