{"id":3327,"date":"2022-03-31T15:52:36","date_gmt":"2022-03-31T07:52:36","guid":{"rendered":"http:\/\/139.9.1.231\/?p=3327"},"modified":"2022-03-31T15:52:38","modified_gmt":"2022-03-31T07:52:38","slug":"torchsnooper","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/03\/31\/torchsnooper\/","title":{"rendered":"PyTorch\u4ee3\u7801\u8c03\u8bd5\u5229\u5668&#8211;TorchSnooper &#038; Captum-pytorch\u6a21\u578b\u53ef\u89e3\u91ca\u6027\u5e93"},"content":{"rendered":"\n\n\n<h2>TorchSnooper<\/h2>\n\n\n\n<p><a href=\"https:\/\/github.com\/zasdfgbnm\/TorchSnooper\">https:\/\/github.com\/zasdfgbnm\/TorchSnooper<\/a><\/p>\n\n\n\n<p>\u5927\u5bb6\u53ef\u80fd\u9047\u5230\u8fd9\u6837\u5b50\u7684\u56f0\u6270\uff1a\u6bd4\u5982\u8bf4\u8fd0\u884c\u81ea\u5df1\u7f16\u5199\u7684 PyTorch \u4ee3\u7801\u7684\u65f6\u5019\uff0cPyTorch \u63d0\u793a\u4f60\u8bf4\u6570\u636e\u7c7b\u578b\u4e0d\u5339\u914d\uff0c\u9700\u8981\u4e00\u4e2a double \u7684 tensor \u4f46\u662f\u4f60\u7ed9\u7684\u5374\u662f float\uff1b\u518d\u6216\u8005\u5c31\u662f\u9700\u8981\u4e00\u4e2a CUDA tensor, \u4f60\u7ed9\u7684\u5374\u662f\u4e2a CPU tensor\u3002\u6bd4\u5982\u4e0b\u9762\u8fd9\u79cd\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>RuntimeError:&nbsp;Expected&nbsp;object&nbsp;of&nbsp;scalar&nbsp;type&nbsp;Double&nbsp;but&nbsp;got&nbsp;scalar&nbsp;type&nbsp;Float\n<\/code><\/pre>\n\n\n\n<p>\u8fd9\u79cd\u95ee\u9898\u8c03\u8bd5\u8d77\u6765\u5f88\u9ebb\u70e6\uff0c\u56e0\u4e3a\u4f60\u4e0d\u77e5\u9053\u4ece\u54ea\u91cc\u5f00\u59cb\u51fa\u95ee\u9898\u7684\u3002\u6bd4\u5982\u4f60\u53ef\u80fd\u5728\u4ee3\u7801\u7684\u7b2c\u4e09\u884c\u7528 torch.zeros \u65b0\u5efa\u4e86\u4e00\u4e2a CPU tensor, \u7136\u540e\u8fd9\u4e2a tensor \u8fdb\u884c\u4e86\u82e5\u5e72\u8fd0\u7b97\uff0c\u5168\u662f\u5728 CPU \u4e0a\u8fdb\u884c\u7684\uff0c\u4e00\u76f4\u6ca1\u6709\u62a5\u9519\uff0c\u76f4\u5230\u7b2c\u5341\u884c\u9700\u8981\u8ddf\u4f60\u4f5c\u4e3a\u8f93\u5165\u4f20\u8fdb\u6765\u7684 CUDA tensor \u8fdb\u884c\u8fd0\u7b97\u7684\u65f6\u5019\uff0c\u624d\u62a5\u9519\u3002\u8981\u8c03\u8bd5\u8fd9\u79cd\u9519\u8bef\uff0c\u6709\u65f6\u5019\u5c31\u4e0d\u5f97\u4e0d\u4e00\u884c\u884c\u5730\u624b\u5199 print \u8bed\u53e5\uff0c\u975e\u5e38\u9ebb\u70e6\u3002<\/p>\n\n\n\n<p>\u518d\u6216\u8005\uff0c\u4f60\u53ef\u80fd\u8111\u5b50\u91cc\u60f3\u8c61\u7740\u5c06\u4e00\u4e2a tensor \u8fdb\u884c\u4ec0\u4e48\u6837\u5b50\u7684\u64cd\u4f5c\uff0c\u5c31\u4f1a\u5f97\u5230\u4ec0\u4e48\u6837\u5b50\u7684\u7ed3\u679c\uff0c\u4f46\u662f PyTorch \u4e2d\u9014\u62a5\u9519\u8bf4 tensor \u7684\u5f62\u72b6\u4e0d\u5339\u914d\uff0c\u6216\u8005\u538b\u6839\u6ca1\u62a5\u9519\u4f46\u662f\u6700\u7ec8\u51fa\u6765\u7684\u5f62\u72b6\u4e0d\u662f\u6211\u4eec\u60f3\u8981\u7684\u3002\u8fd9\u4e2a\u65f6\u5019\uff0c\u6211\u4eec\u5f80\u5f80\u4e5f\u4e0d\u77e5\u9053\u662f\u4ec0\u4e48\u5730\u65b9\u5f00\u59cb\u8ddf\u6211\u4eec\u300c\u9884\u671f\u7684\u53d1\u751f\u504f\u79bb\u7684\u300d\u3002\u6211\u4eec\u6709\u65f6\u5019\u4e5f\u5f97\u9700\u8981\u63d2\u5165\u4e00\u5927\u5806 print \u8bed\u53e5\u624d\u80fd\u627e\u5230\u539f\u56e0\u3002<\/p>\n\n\n\n<p>TorchSnooper \u5c31\u662f\u4e00\u4e2a\u8bbe\u8ba1\u4e86\u7528\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u7684\u5de5\u5177\u3002TorchSnooper \u7684\u5b89\u88c5\u975e\u5e38\u7b80\u5355\uff0c\u53ea\u9700\u8981\u6267\u884c\u6807\u51c6\u7684 Python \u5305\u5b89\u88c5\u6307\u4ee4\u5c31\u597d\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>pip install snoop\npip install torchsnooper<\/code><\/pre>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">1\u3001\u76d1\u6d4b\u51fd\u6570\u4e2d\u7684\u53d8\u91cf\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">import torch\nimport torchsnooper\n\n@torchsnooper.snoop()\ndef myfunc(mask, x):\n    y = torch.zeros(6)\n    y.masked_scatter_(mask, x)\n    return y\n\nmask = torch.tensor([0, 1, 0, 1, 1, 0], device='cuda')\nsource = torch.tensor([1.0, 2.0, 3.0], device='cuda')\ny = myfunc(mask, source)<\/pre>\n\n\n\n<p>Run our script, and we will see:<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>Starting var:.. mask = tensor&lt;(6,), int64, cuda:0&gt;\nStarting var:.. x = tensor&lt;(3,), float32, cuda:0&gt;\n21:41:42.941668 call         5 def myfunc(mask, x):\n21:41:42.941834 line         6     y = torch.zeros(6)\nNew var:....... y = tensor&lt;(6,), float32, cpu&gt;\n21:41:42.943443 line         7     y.masked_scatter_(mask, x)\n21:41:42.944404 exception    7     y.masked_scatter_(mask, x)<\/code><\/pre>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">2\u3001\u76d1\u6d4bfor\u5faa\u73af\u4e2d\u7684\u53d8\u91cf\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">with torchsnooper.snoop():\n    for _ in range(100):\n        optimizer.zero_grad()\n        pred = model(x)\n        squared_diff = (y - pred) ** 2\n        loss = squared_diff.mean()\n        print(loss.item())\n        loss.backward()\n        optimizer.step()<\/pre>\n\n\n\n<p>Part of the trace looks like:<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>New var:....... x = tensor&lt;(4, 2), float32, cpu&gt;\nNew var:....... y = tensor&lt;(4,), float32, cpu&gt;\nNew var:....... model = Model(  (layer): Linear(in_features=2, out_features=1, bias=True))\nNew var:....... optimizer = SGD (Parameter Group 0    dampening: 0    lr: 0....omentum: 0    nesterov: False    weight_decay: 0)\n22:27:01.024233 line        21     for _ in range(100):\nNew var:....... _ = 0\n22:27:01.024439 line        22         optimizer.zero_grad()\n22:27:01.024574 line        23         pred = model(x)\nNew var:....... pred = tensor&lt;(4, 1), float32, cpu, grad&gt;\n22:27:01.026442 line        24         squared_diff = (y - pred) ** 2\nNew var:....... squared_diff = tensor&lt;(4, 4), float32, cpu, grad&gt;\n22:27:01.027369 line        25         loss = squared_diff.mean()\nNew var:....... loss = tensor&lt;(), float32, cpu, grad&gt;\n22:27:01.027616 line        26         print(loss.item())\n22:27:01.027793 line        27         loss.backward()\n22:27:01.050189 line        28         optimizer.step()<\/code><\/pre>\n\n\n\n<h2><strong>Captum\uff1aPyTorch \u7684\u7edf\u4e00\u901a\u7528\u6a21\u578b\u53ef\u89e3\u91ca\u6027\u5e93<\/strong><\/h2>\n\n\n\n<p><a href=\"https:\/\/captum.ai\/tutorials\/CIFAR_TorchVision_Captum_Insights\">https:\/\/captum.ai\/tutorials\/CIFAR_TorchVision_Captum_Insights<\/a><\/p>\n\n\n\n<p>\u53ef\u89e3\u91ca\u6027\uff0c\u5373\u7406\u89e3\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u4e3a\u4ec0\u4e48\u505a\u51fa\u51b3\u5b9a\u7684\u80fd\u529b\uff0c\u5bf9\u4e8e\u5f00\u53d1\u4eba\u5458\u89e3\u91ca\u6a21\u578b\u4e3a\u4ec0\u4e48\u505a\u51fa\u67d0\u4e2a\u51b3\u5b9a\u662f\u5f88\u91cd\u8981\u7684\u3002\u5b83\u53ef\u4ee5\u4f7f\u4eba\u5de5\u667a\u80fd\u7b26\u5408\u76d1\u7ba1\u6cd5\u5f8b\uff0c\u4ee5\u5e94\u7528\u4e8e\u9700\u8981\u89e3\u91ca\u6027\u7684\u4e1a\u52a1\u3002<\/p>\n\n\n\n<p>Captum\u65e8\u5728\u5b9e\u73b0AI\u6a21\u578b\u7684\u6700\u65b0\u7248\u672c\uff0c\u5982\u96c6\u6210\u68af\u5ea6\u3001\u6df1\u5ea6\u5f2f\u66f2\u548c\u4f20\u5bfc\u7b49\u7b49\uff0c\u5b83\u53ef\u4ee5\u5e2e\u52a9\u7814\u7a76\u4eba\u5458\u548c\u5f00\u53d1\u4eba\u5458\u89e3\u91ca\u4eba\u5de5\u667a\u80fd\u5728\u591a\u6a21\u6001\u73af\u5883\u4e2d\u505a\u51fa\u7684\u51b3\u7b56\uff0c\u5e76\u80fd\u5e2e\u52a9\u7814\u7a76\u4eba\u5458\u628a\u7ed3\u679c\u4e0e\u6570\u636e\u5e93\u4e2d\u73b0\u6709\u7684\u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img src=\"http:\/\/imgcdn.atyun.com\/2019\/10\/Captum_1.png\" alt=\"Captum\u2014\u2014\u65b0\u7684\u4eba\u5de5\u667a\u80fd\u53ef\u89e3\u91ca\u6027\u5de5\u5177\"\/><\/figure><\/div>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"704\" height=\"389\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/03\/image-93.png\" alt=\"\" class=\"wp-image-3333\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/03\/image-93.png 704w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/03\/image-93-300x166.png 300w\" sizes=\"(max-width: 704px) 100vw, 704px\" \/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>TorchSnooper https:\/\/github.com\/zasdfgbnm\/TorchSnooper  &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2022\/03\/31\/torchsnooper\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">PyTorch\u4ee3\u7801\u8c03\u8bd5\u5229\u5668&#8211;TorchSnooper &#038; Captum-pytorch\u6a21\u578b\u53ef\u89e3\u91ca\u6027\u5e93<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[11],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/3327"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=3327"}],"version-history":[{"count":6,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/3327\/revisions"}],"predecessor-version":[{"id":3335,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/3327\/revisions\/3335"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=3327"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=3327"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=3327"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}