mlvbench 1.0.0__tar.gz

mlvbench-1.0.0/.cache/.gitignore

@@ -0,0 +1 @@
+*

mlvbench-1.0.0/.cache/10063063919107421344

@@ -0,0 +1,247 @@
+{
+  "data": {
+    "meta": {},
+    "content": "<h1 id=\"mlvbenchprobes\"><code>mlvbench.probes</code><a class=\"headerlink\" href=\"#mlvbenchprobes\" title=\"Permanent link\">&para;</a></h1>\n\n\n<div class=\"doc doc-object doc-module\">\n\n\n\n<a id=\"mlvbench.probes\"></a>\n    <div class=\"doc doc-contents first\">\n\n        <p>Probes.</p>\n<p>Probes are small models that readout information from intermediate features of a\npretrained vision model.</p>\n\n\n\n\n\n\n\n\n\n\n<div class=\"doc doc-children\">\n\n\n\n\n\n\n\n\n\n<div class=\"doc doc-object doc-class\">\n\n\n\n<h2 id=\"mlvbench.probes.LinearProbe\" class=\"doc doc-heading\">\n            <code>LinearProbe</code>\n\n\n<a href=\"#mlvbench.probes.LinearProbe\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h2>\n\n\n    <div class=\"doc doc-contents \">\n            <p class=\"doc doc-class-bases\">\n              Bases: <code><autoref identifier=\"mlvbench.probes.Probe\" backlink-type=\"subclassed-by\" backlink-anchor=\"mlvbench.probes.LinearProbe\" optional hover>Probe</autoref></code></p>\n\n\n\n        <p>Linear probe.</p>\n<p>A linear projection of the input features to a scalar, optionally with a learnable\nbias:</p>\n<div class=\"arithmatex\">\\[\n    y = W x + b\n\\]</div>\n\n\n\n\n\n\n\n\n\n\n\n<div class=\"doc doc-children\">\n\n\n\n\n\n\n\n<div class=\"doc doc-object doc-attribute\">\n\n\n\n<h3 id=\"mlvbench.probes.LinearProbe.name\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">name</span></code>\n\n  <span class=\"doc doc-labels\">\n      <small class=\"doc doc-label doc-label-property\"><code>property</code></small>\n  </span>\n\n<a href=\"#mlvbench.probes.LinearProbe.name\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Name of the probe type (<code>\"linear\"</code>).</p>\n\n    </div>\n\n</div>\n\n\n\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h3 id=\"mlvbench.probes.LinearProbe.__init__\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"fm\">__init__</span><span class=\"p\">(</span><span class=\"n\">layer</span><span class=\"p\">,</span> <span class=\"n\">input_dim</span><span class=\"p\">,</span> <span class=\"n\">bias</span><span class=\"o\">=</span><span class=\"s1\">&#39;scalar&#39;</span><span class=\"p\">,</span> <span class=\"n\">num_patches</span><span class=\"o\">=</span><span class=\"kc\">None</span><span class=\"p\">,</span> <span class=\"n\">metadata</span><span class=\"o\">=</span><span class=\"kc\">None</span><span class=\"p\">)</span></code>\n\n<a href=\"#mlvbench.probes.LinearProbe.__init__\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Initialize the probe.</p>\n\n\n<p><span class=\"doc-section-title\">Parameters:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Name</th>\n          <th>Type</th>\n          <th>Description</th>\n          <th>Default</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>layer</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.__init__\" optional>str</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Name of the model layer the probe reads from.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>input_dim</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"int\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.__init__\" optional>int</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Number of dimensions of the input features.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>bias</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"typing.Literal\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.__init__\" optional hover>Literal</autoref>[&#39;none&#39;, &#39;scalar&#39;, &#39;spatial&#39;] | <autoref identifier=\"torch.Tensor\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.__init__\" optional hover>Tensor</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The kind of learnable bias added to the output:</p>\n<ul>\n<li><code>\"none\"</code>: no bias.</li>\n<li><code>\"scalar\"</code>: a single learnable scalar shared across all patches.</li>\n<li><code>\"spatial\"</code>: a learnable per-patch bias with shape <code>(num_patches, 1)</code>,\n    initialized to zero. Requires <code>num_patches</code>.</li>\n<li>a <code>torch.Tensor</code> (a prior with shape <code>(N, 1)</code>): a per-patch bias\n    initialized from the tensor; equivalent to <code>\"spatial\"</code> with\n    <code>num_patches = N</code>.</li>\n</ul>\n              </div>\n            </td>\n            <td>\n                  <code>&#39;scalar&#39;</code>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>num_patches</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"int\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.__init__\" optional>int</autoref> | None</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Number of patch tokens of the input features. Required when\nusing a <code>\"spatial\"</code> bias and ignored otherwise.</p>\n              </div>\n            </td>\n            <td>\n                  <code>None</code>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>metadata</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"dict\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.__init__\" optional>dict</autoref>[<autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.__init__\" optional>str</autoref>, <autoref identifier=\"typing.Any\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.__init__\" optional hover>Any</autoref>] | None</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Free-form metadata to attach to the probe.</p>\n              </div>\n            </td>\n            <td>\n                  <code>None</code>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    </div>\n\n</div>\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h3 id=\"mlvbench.probes.LinearProbe.forward\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">forward</span><span class=\"p\">(</span><span class=\"n\">features</span><span class=\"p\">)</span></code>\n\n<a href=\"#mlvbench.probes.LinearProbe.forward\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Forward pass.</p>\n\n\n<p><span class=\"doc-section-title\">Parameters:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Name</th>\n          <th>Type</th>\n          <th>Description</th>\n          <th>Default</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>features</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"mlvbench.models.Features\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.forward\" optional hover>Features</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The features extracted from the model. The probe uses the patch\ntokens of <code>self.layer</code>, expecting shape <code>(B, N, C)</code> and dtype <code>float32</code>.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    <p><span class=\"doc-section-title\">Returns:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Type</th>\n          <th>Description</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                  <code><autoref identifier=\"torch.Tensor\" backlink-type=\"returned-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.forward\" optional hover>Tensor</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The output with shape <code>(B, N, 1)</code> and dtype <code>float32</code>.</p>\n              </div>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    </div>\n\n</div>\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h3 id=\"mlvbench.probes.LinearProbe.select_features\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">select_features</span><span class=\"p\">(</span><span class=\"n\">features</span><span class=\"p\">)</span></code>\n\n<a href=\"#mlvbench.probes.LinearProbe.select_features\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Select the part of the feature representation used by the probe.</p>\n<p>Defaults to the patch tokens of <code>self.layer</code> with shape <code>(B, N, C)</code>. Override to\nread from a different part of the representation (e.g. the cls token or a\ndifferent token format).</p>\n\n\n<p><span class=\"doc-section-title\">Parameters:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Name</th>\n          <th>Type</th>\n          <th>Description</th>\n          <th>Default</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>features</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"mlvbench.models.Features\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.select_features\" optional hover>Features</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The features extracted from the backbone.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    <p><span class=\"doc-section-title\">Returns:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Type</th>\n          <th>Description</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                  <code><autoref identifier=\"torch.Tensor\" backlink-type=\"returned-by\" backlink-anchor=\"mlvbench.probes.LinearProbe.select_features\" optional hover>Tensor</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The selected feature tensor.</p>\n              </div>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    </div>\n\n</div>\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h3 id=\"mlvbench.probes.LinearProbe.serialize\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">serialize</span><span class=\"p\">()</span></code>\n\n<a href=\"#mlvbench.probes.LinearProbe.serialize\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Serialize the probe.</p>\n<p>See <autoref identifier=\"mlvbench.probes.Probe.serialize\"><code>Probe.serialize</code></autoref>.</p>\n\n\n    </div>\n\n</div>\n\n\n\n  </div>\n\n    </div>\n\n</div>\n\n<div class=\"doc doc-object doc-class\">\n\n\n\n<h2 id=\"mlvbench.probes.Probe\" class=\"doc doc-heading\">\n            <code>Probe</code>\n\n\n<a href=\"#mlvbench.probes.Probe\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h2>\n\n\n    <div class=\"doc doc-contents \">\n            <p class=\"doc doc-class-bases\">\n              Bases: <code><autoref identifier=\"torch.nn.Module\" backlink-type=\"subclassed-by\" backlink-anchor=\"mlvbench.probes.Probe\" optional hover>Module</autoref></code>, <code><autoref identifier=\"abc.ABC\" backlink-type=\"subclassed-by\" backlink-anchor=\"mlvbench.probes.Probe\" optional hover>ABC</autoref></code></p>\n\n\n\n        <p>Abstract base class for probes.</p>\n\n\n<p><span class=\"doc-section-title\">Attributes:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Name</th>\n          <th>Type</th>\n          <th>Description</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td><code><autoref identifier=\"mlvbench.probes.Probe.name\" optional hover>name</autoref></code></td>\n            <td>\n                  <code><autoref identifier=\"str\" optional>str</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Name of the probe type.</p>\n              </div>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td><code><autoref identifier=\"mlvbench.probes.Probe.layer\" optional hover>layer</autoref></code></td>\n            <td>\n                  <code><autoref identifier=\"str\" optional>str</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Name of the backbone layer the probe reads from.</p>\n              </div>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td><code><autoref identifier=\"mlvbench.probes.Probe.metadata\" optional hover>metadata</autoref></code></td>\n            <td>\n                  <code><autoref identifier=\"dict\" optional>dict</autoref>[<autoref identifier=\"str\" optional>str</autoref>, <autoref identifier=\"typing.Any\" optional hover>Any</autoref>]</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Free-form dict of training/experiment information attached to the\nprobe. Persisted in full when the probe is saved.</p>\n              </div>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n\n\n\n\n\n\n\n\n\n<div class=\"doc doc-children\">\n\n\n\n\n\n\n\n<div class=\"doc doc-object doc-attribute\">\n\n\n\n<h3 id=\"mlvbench.probes.Probe.name\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">name</span></code>\n\n  <span class=\"doc doc-labels\">\n      <small class=\"doc doc-label doc-label-abstractmethod\"><code>abstractmethod</code></small>\n      <small class=\"doc doc-label doc-label-property\"><code>property</code></small>\n  </span>\n\n<a href=\"#mlvbench.probes.Probe.name\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Name of the probe type.</p>\n\n    </div>\n\n</div>\n\n\n\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h3 id=\"mlvbench.probes.Probe.__init__\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"fm\">__init__</span><span class=\"p\">(</span><span class=\"n\">layer</span><span class=\"p\">,</span> <span class=\"n\">input_dim</span><span class=\"p\">,</span> <span class=\"o\">*</span><span class=\"p\">,</span> <span class=\"n\">metadata</span><span class=\"o\">=</span><span class=\"kc\">None</span><span class=\"p\">)</span></code>\n\n<a href=\"#mlvbench.probes.Probe.__init__\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Initialize the probe.</p>\n\n\n<p><span class=\"doc-section-title\">Parameters:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Name</th>\n          <th>Type</th>\n          <th>Description</th>\n          <th>Default</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>layer</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.Probe.__init__\" optional>str</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Name of the model layer the probe reads from.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>input_dim</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"int\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.Probe.__init__\" optional>int</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Number of dimensions of the input features.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>metadata</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"dict\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.Probe.__init__\" optional>dict</autoref>[<autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.Probe.__init__\" optional>str</autoref>, <autoref identifier=\"typing.Any\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.Probe.__init__\" optional hover>Any</autoref>] | None</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Free-form dict of training/experiment information.</p>\n              </div>\n            </td>\n            <td>\n                  <code>None</code>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    </div>\n\n</div>\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h3 id=\"mlvbench.probes.Probe.forward\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">forward</span><span class=\"p\">(</span><span class=\"n\">features</span><span class=\"p\">)</span></code>\n\n  <span class=\"doc doc-labels\">\n      <small class=\"doc doc-label doc-label-abstractmethod\"><code>abstractmethod</code></small>\n  </span>\n\n<a href=\"#mlvbench.probes.Probe.forward\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Apply the probe to the given features.</p>\n<p>The probe selects the part of the representation it reads from via\n<autoref identifier=\"mlvbench.probes.Probe.select_features\"><code>select_features</code></autoref>.</p>\n\n\n<p><span class=\"doc-section-title\">Parameters:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Name</th>\n          <th>Type</th>\n          <th>Description</th>\n          <th>Default</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>features</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"mlvbench.models.Features\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.Probe.forward\" optional hover>Features</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The features extracted from the backbone.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    <p><span class=\"doc-section-title\">Returns:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Type</th>\n          <th>Description</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                  <code><autoref identifier=\"torch.Tensor\" backlink-type=\"returned-by\" backlink-anchor=\"mlvbench.probes.Probe.forward\" optional hover>Tensor</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The prediction from the probe as torch.Tensor. Shape and dtype very per\nprobe.</p>\n              </div>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    </div>\n\n</div>\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h3 id=\"mlvbench.probes.Probe.select_features\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">select_features</span><span class=\"p\">(</span><span class=\"n\">features</span><span class=\"p\">)</span></code>\n\n<a href=\"#mlvbench.probes.Probe.select_features\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Select the part of the feature representation used by the probe.</p>\n<p>Defaults to the patch tokens of <code>self.layer</code> with shape <code>(B, N, C)</code>. Override to\nread from a different part of the representation (e.g. the cls token or a\ndifferent token format).</p>\n\n\n<p><span class=\"doc-section-title\">Parameters:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Name</th>\n          <th>Type</th>\n          <th>Description</th>\n          <th>Default</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>features</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"mlvbench.models.Features\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.Probe.select_features\" optional hover>Features</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The features extracted from the backbone.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    <p><span class=\"doc-section-title\">Returns:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Type</th>\n          <th>Description</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                  <code><autoref identifier=\"torch.Tensor\" backlink-type=\"returned-by\" backlink-anchor=\"mlvbench.probes.Probe.select_features\" optional hover>Tensor</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The selected feature tensor.</p>\n              </div>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    </div>\n\n</div>\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h3 id=\"mlvbench.probes.Probe.serialize\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">serialize</span><span class=\"p\">()</span></code>\n\n  <span class=\"doc doc-labels\">\n      <small class=\"doc doc-label doc-label-abstractmethod\"><code>abstractmethod</code></small>\n  </span>\n\n<a href=\"#mlvbench.probes.Probe.serialize\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h3>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Return all information needed to reconstruct the probe.</p>\n<p>This returns a self-describing dictionary that includes both the torch\n<code>state_dict</code> and all config options. The probe can be rebuilt from this\nrepresentation via <autoref identifier=\"mlvbench.probes.build_probe\"><code>build_probe</code></autoref>.</p>\n\n\n    <p><span class=\"doc-section-title\">Returns:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Type</th>\n          <th>Description</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                  <code><autoref identifier=\"dict\" backlink-type=\"returned-by\" backlink-anchor=\"mlvbench.probes.Probe.serialize\" optional>dict</autoref>[<autoref identifier=\"str\" backlink-type=\"returned-by\" backlink-anchor=\"mlvbench.probes.Probe.serialize\" optional>str</autoref>, <autoref identifier=\"typing.Any\" backlink-type=\"returned-by\" backlink-anchor=\"mlvbench.probes.Probe.serialize\" optional hover>Any</autoref>]</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>A dictionary describing the probe, with detached CPU tensors for <code>weights</code>.</p>\n              </div>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    </div>\n\n</div>\n\n\n\n  </div>\n\n    </div>\n\n</div>\n\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h2 id=\"mlvbench.probes.build_probe\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">build_probe</span><span class=\"p\">(</span><span class=\"n\">name</span><span class=\"p\">,</span> <span class=\"n\">layer</span><span class=\"p\">,</span> <span class=\"n\">input_dim</span><span class=\"p\">,</span> <span class=\"n\">metadata</span><span class=\"o\">=</span><span class=\"kc\">None</span><span class=\"p\">,</span> <span class=\"n\">state_dict</span><span class=\"o\">=</span><span class=\"kc\">None</span><span class=\"p\">,</span> <span class=\"n\">probe_types</span><span class=\"o\">=</span><span class=\"kc\">None</span><span class=\"p\">,</span> <span class=\"o\">**</span><span class=\"n\">kwargs</span><span class=\"p\">)</span></code>\n\n<a href=\"#mlvbench.probes.build_probe\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h2>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Build the probe with the given name and arguments.</p>\n\n\n<p><span class=\"doc-section-title\">Parameters:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Name</th>\n          <th>Type</th>\n          <th>Description</th>\n          <th>Default</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>name</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>str</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Name of the probe type (e.g. <code>\"linear\"</code>).</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>layer</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>str</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Name of the model layer the probe reads from.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>input_dim</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"int\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>int</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Number of dimensions of the input features.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>metadata</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"dict\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>dict</autoref>[<autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>str</autoref>, <autoref identifier=\"typing.Any\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional hover>Any</autoref>] | None</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Free-form metadata to attach to the probe.</p>\n              </div>\n            </td>\n            <td>\n                  <code>None</code>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>state_dict</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"dict\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>dict</autoref>[<autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>str</autoref>, <autoref identifier=\"typing.Any\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional hover>Any</autoref>] | None</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Optional state dict for restoring pretrained probes.</p>\n              </div>\n            </td>\n            <td>\n                  <code>None</code>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>probe_types</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"dict\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>dict</autoref>[<autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>str</autoref>, <autoref identifier=\"type\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>type</autoref>[<autoref identifier=\"mlvbench.probes.Probe\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional hover>Probe</autoref>]] | None</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Optional mapping of custom probe type names to <code>Probe</code> subclasses,\nmerged with the built-in types. Each class must be constructible as\n<code>Cls(layer, input_dim, metadata=metadata, **kwargs)</code>.</p>\n              </div>\n            </td>\n            <td>\n                  <code>None</code>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>**kwargs</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"typing.Any\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional hover>Any</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>Additional keyword arguments forwarded to the probe constructor (e.g.\n<code>bias</code> and <code>num_patches</code>).</p>\n              </div>\n            </td>\n            <td>\n                  <code>{}</code>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n<p><span class=\"doc-section-title\">Raises:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Type</th>\n          <th>Description</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                  <code><autoref identifier=\"ValueError\" backlink-type=\"raised-by\" backlink-anchor=\"mlvbench.probes.build_probe\" optional>ValueError</autoref></code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>If <code>name</code> is not a known built-in or custom probe type.</p>\n              </div>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    </div>\n\n</div>\n\n\n\n  </div>\n\n    </div>\n\n</div>",
+    "search": [
+      {
+        "location": null,
+        "level": 1,
+        "title": "<code>mlvbench.probes</code>",
+        "text": "<p>Probes.</p> <p>Probes are small models that readout information from intermediate features of a pretrained vision model.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.LinearProbe",
+        "level": 2,
+        "title": "<code>LinearProbe</code>",
+        "text": "<p>               Bases: <code>Probe</code></p> <p>Linear probe.</p> <p>A linear projection of the input features to a scalar, optionally with a learnable bias:</p> \\[     y = W x + b \\]",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.LinearProbe.name",
+        "level": 3,
+        "title": "<code>name</code>  <code>property</code>",
+        "text": "<p>Name of the probe type (<code>\"linear\"</code>).</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.LinearProbe.__init__",
+        "level": 3,
+        "title": "<code>__init__(layer, input_dim, bias='scalar', num_patches=None, metadata=None)</code>",
+        "text": "<p>Initialize the probe.</p> <p>Parameters:</p> Name Type Description Default <code>layer</code> <code>str</code> <p>Name of the model layer the probe reads from.</p> required <code>input_dim</code> <code>int</code> <p>Number of dimensions of the input features.</p> required <code>bias</code> <code>Literal['none', 'scalar', 'spatial'] | Tensor</code> <p>The kind of learnable bias added to the output:</p> <ul> <li><code>\"none\"</code>: no bias.</li> <li><code>\"scalar\"</code>: a single learnable scalar shared across all patches.</li> <li><code>\"spatial\"</code>: a learnable per-patch bias with shape <code>(num_patches, 1)</code>,     initialized to zero. Requires <code>num_patches</code>.</li> <li>a <code>torch.Tensor</code> (a prior with shape <code>(N, 1)</code>): a per-patch bias     initialized from the tensor; equivalent to <code>\"spatial\"</code> with     <code>num_patches = N</code>.</li> </ul> <code>'scalar'</code> <code>num_patches</code> <code>int | None</code> <p>Number of patch tokens of the input features. Required when using a <code>\"spatial\"</code> bias and ignored otherwise.</p> <code>None</code> <code>metadata</code> <code>dict[str, Any] | None</code> <p>Free-form metadata to attach to the probe.</p> <code>None</code>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.LinearProbe.forward",
+        "level": 3,
+        "title": "<code>forward(features)</code>",
+        "text": "<p>Forward pass.</p> <p>Parameters:</p> Name Type Description Default <code>features</code> <code>Features</code> <p>The features extracted from the model. The probe uses the patch tokens of <code>self.layer</code>, expecting shape <code>(B, N, C)</code> and dtype <code>float32</code>.</p> required <p>Returns:</p> Type Description <code>Tensor</code> <p>The output with shape <code>(B, N, 1)</code> and dtype <code>float32</code>.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.LinearProbe.select_features",
+        "level": 3,
+        "title": "<code>select_features(features)</code>",
+        "text": "<p>Select the part of the feature representation used by the probe.</p> <p>Defaults to the patch tokens of <code>self.layer</code> with shape <code>(B, N, C)</code>. Override to read from a different part of the representation (e.g. the cls token or a different token format).</p> <p>Parameters:</p> Name Type Description Default <code>features</code> <code>Features</code> <p>The features extracted from the backbone.</p> required <p>Returns:</p> Type Description <code>Tensor</code> <p>The selected feature tensor.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.LinearProbe.serialize",
+        "level": 3,
+        "title": "<code>serialize()</code>",
+        "text": "<p>Serialize the probe.</p> <p>See <code>Probe.serialize</code>.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.Probe",
+        "level": 2,
+        "title": "<code>Probe</code>",
+        "text": "<p>               Bases: <code>Module</code>, <code>ABC</code></p> <p>Abstract base class for probes.</p> <p>Attributes:</p> Name Type Description <code>name</code> <code>str</code> <p>Name of the probe type.</p> <code>layer</code> <code>str</code> <p>Name of the backbone layer the probe reads from.</p> <code>metadata</code> <code>dict[str, Any]</code> <p>Free-form dict of training/experiment information attached to the probe. Persisted in full when the probe is saved.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.Probe.name",
+        "level": 3,
+        "title": "<code>name</code>  <code>abstractmethod</code> <code>property</code>",
+        "text": "<p>Name of the probe type.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.Probe.__init__",
+        "level": 3,
+        "title": "<code>__init__(layer, input_dim, *, metadata=None)</code>",
+        "text": "<p>Initialize the probe.</p> <p>Parameters:</p> Name Type Description Default <code>layer</code> <code>str</code> <p>Name of the model layer the probe reads from.</p> required <code>input_dim</code> <code>int</code> <p>Number of dimensions of the input features.</p> required <code>metadata</code> <code>dict[str, Any] | None</code> <p>Free-form dict of training/experiment information.</p> <code>None</code>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.Probe.forward",
+        "level": 3,
+        "title": "<code>forward(features)</code>  <code>abstractmethod</code>",
+        "text": "<p>Apply the probe to the given features.</p> <p>The probe selects the part of the representation it reads from via <code>select_features</code>.</p> <p>Parameters:</p> Name Type Description Default <code>features</code> <code>Features</code> <p>The features extracted from the backbone.</p> required <p>Returns:</p> Type Description <code>Tensor</code> <p>The prediction from the probe as torch.Tensor. Shape and dtype very per probe.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.Probe.select_features",
+        "level": 3,
+        "title": "<code>select_features(features)</code>",
+        "text": "<p>Select the part of the feature representation used by the probe.</p> <p>Defaults to the patch tokens of <code>self.layer</code> with shape <code>(B, N, C)</code>. Override to read from a different part of the representation (e.g. the cls token or a different token format).</p> <p>Parameters:</p> Name Type Description Default <code>features</code> <code>Features</code> <p>The features extracted from the backbone.</p> required <p>Returns:</p> Type Description <code>Tensor</code> <p>The selected feature tensor.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.Probe.serialize",
+        "level": 3,
+        "title": "<code>serialize()</code>  <code>abstractmethod</code>",
+        "text": "<p>Return all information needed to reconstruct the probe.</p> <p>This returns a self-describing dictionary that includes both the torch <code>state_dict</code> and all config options. The probe can be rebuilt from this representation via <code>build_probe</code>.</p> <p>Returns:</p> Type Description <code>dict[str, Any]</code> <p>A dictionary describing the probe, with detached CPU tensors for <code>weights</code>.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.probes.build_probe",
+        "level": 2,
+        "title": "<code>build_probe(name, layer, input_dim, metadata=None, state_dict=None, probe_types=None, **kwargs)</code>",
+        "text": "<p>Build the probe with the given name and arguments.</p> <p>Parameters:</p> Name Type Description Default <code>name</code> <code>str</code> <p>Name of the probe type (e.g. <code>\"linear\"</code>).</p> required <code>layer</code> <code>str</code> <p>Name of the model layer the probe reads from.</p> required <code>input_dim</code> <code>int</code> <p>Number of dimensions of the input features.</p> required <code>metadata</code> <code>dict[str, Any] | None</code> <p>Free-form metadata to attach to the probe.</p> <code>None</code> <code>state_dict</code> <code>dict[str, Any] | None</code> <p>Optional state dict for restoring pretrained probes.</p> <code>None</code> <code>probe_types</code> <code>dict[str, type[Probe]] | None</code> <p>Optional mapping of custom probe type names to <code>Probe</code> subclasses, merged with the built-in types. Each class must be constructible as <code>Cls(layer, input_dim, metadata=metadata, **kwargs)</code>.</p> <code>None</code> <code>**kwargs</code> <code>Any</code> <p>Additional keyword arguments forwarded to the probe constructor (e.g. <code>bias</code> and <code>num_patches</code>).</p> <code>{}</code> <p>Raises:</p> Type Description <code>ValueError</code> <p>If <code>name</code> is not a known built-in or custom probe type.</p>",
+        "path": [],
+        "tags": []
+      }
+    ],
+    "title": "mlvbench.probes",
+    "toc": [
+      {
+        "title": "mlvbench.probes",
+        "content": "<code>mlvbench.probes</code>",
+        "id": "mlvbenchprobes",
+        "url": "#mlvbenchprobes",
+        "children": [
+          {
+            "title": "probes",
+            "content": "probes",
+            "id": "mlvbench.probes",
+            "url": "#mlvbench.probes",
+            "children": [],
+            "level": 2
+          },
+          {
+            "title": "LinearProbe",
+            "content": "LinearProbe",
+            "id": "mlvbench.probes.LinearProbe",
+            "url": "#mlvbench.probes.LinearProbe",
+            "children": [
+              {
+                "title": "name",
+                "content": "name",
+                "id": "mlvbench.probes.LinearProbe.name",
+                "url": "#mlvbench.probes.LinearProbe.name",
+                "children": [],
+                "level": 3
+              },
+              {
+                "title": "__init__",
+                "content": "__init__",
+                "id": "mlvbench.probes.LinearProbe.__init__",
+                "url": "#mlvbench.probes.LinearProbe.__init__",
+                "children": [],
+                "level": 3
+              },
+              {
+                "title": "forward",
+                "content": "forward",
+                "id": "mlvbench.probes.LinearProbe.forward",
+                "url": "#mlvbench.probes.LinearProbe.forward",
+                "children": [],
+                "level": 3
+              },
+              {
+                "title": "select_features",
+                "content": "select_features",
+                "id": "mlvbench.probes.LinearProbe.select_features",
+                "url": "#mlvbench.probes.LinearProbe.select_features",
+                "children": [],
+                "level": 3
+              },
+              {
+                "title": "serialize",
+                "content": "serialize",
+                "id": "mlvbench.probes.LinearProbe.serialize",
+                "url": "#mlvbench.probes.LinearProbe.serialize",
+                "children": [],
+                "level": 3
+              }
+            ],
+            "level": 2
+          },
+          {
+            "title": "Probe",
+            "content": "Probe",
+            "id": "mlvbench.probes.Probe",
+            "url": "#mlvbench.probes.Probe",
+            "children": [
+              {
+                "title": "name",
+                "content": "name",
+                "id": "mlvbench.probes.Probe.name",
+                "url": "#mlvbench.probes.Probe.name",
+                "children": [],
+                "level": 3
+              },
+              {
+                "title": "__init__",
+                "content": "__init__",
+                "id": "mlvbench.probes.Probe.__init__",
+                "url": "#mlvbench.probes.Probe.__init__",
+                "children": [],
+                "level": 3
+              },
+              {
+                "title": "forward",
+                "content": "forward",
+                "id": "mlvbench.probes.Probe.forward",
+                "url": "#mlvbench.probes.Probe.forward",
+                "children": [],
+                "level": 3
+              },
+              {
+                "title": "select_features",
+                "content": "select_features",
+                "id": "mlvbench.probes.Probe.select_features",
+                "url": "#mlvbench.probes.Probe.select_features",
+                "children": [],
+                "level": 3
+              },
+              {
+                "title": "serialize",
+                "content": "serialize",
+                "id": "mlvbench.probes.Probe.serialize",
+                "url": "#mlvbench.probes.Probe.serialize",
+                "children": [],
+                "level": 3
+              }
+            ],
+            "level": 2
+          },
+          {
+            "title": "build_probe",
+            "content": "build_probe",
+            "id": "mlvbench.probes.build_probe",
+            "url": "#mlvbench.probes.build_probe",
+            "children": [],
+            "level": 2
+          }
+        ],
+        "level": 1
+      }
+    ]
+  },
+  "hash": 17128207827365371114
+}

mlvbench-1.0.0/.cache/11345666923742321217

@@ -0,0 +1,85 @@
+{
+  "data": {
+    "meta": {},
+    "content": "<h1 id=\"mlvbenchinfrastructure\"><code>mlvbench.infrastructure</code><a class=\"headerlink\" href=\"#mlvbenchinfrastructure\" title=\"Permanent link\">&para;</a></h1>\n\n\n<div class=\"doc doc-object doc-module\">\n\n\n\n<a id=\"mlvbench.infrastructure\"></a>\n    <div class=\"doc doc-contents first\">\n\n        <p>Infrastructure utilities.</p>\n\n\n\n\n\n\n\n\n\n\n<div class=\"doc doc-children\">\n\n\n\n\n\n\n\n\n\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h2 id=\"mlvbench.infrastructure.resolve_layers\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">resolve_layers</span><span class=\"p\">(</span><span class=\"n\">all_layers</span><span class=\"p\">,</span> <span class=\"n\">layers_spec</span><span class=\"p\">)</span></code>\n\n<a href=\"#mlvbench.infrastructure.resolve_layers\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h2>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Resolve a layer specification string to a list of layer names.</p>\n<p>Intended for experiment scripts that expose a <code>--layers</code> command line argument.\nThe resolved list can be passed directly to\n<autoref identifier=\"mlvbench.trainer.ProbeTrainer.fit\"><code>ProbeTrainer.fit</code></autoref>.</p>\n\n\n<p><span class=\"doc-section-title\">Parameters:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Name</th>\n          <th>Type</th>\n          <th>Description</th>\n          <th>Default</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>all_layers</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"list\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.infrastructure.resolve_layers\" optional>list</autoref>[<autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.infrastructure.resolve_layers\" optional>str</autoref>]</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>The full list of layer names from the model\n(<code>model.layer_names</code>).</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n          <tr class=\"doc-section-item\">\n            <td>\n                <code>layers_spec</code>\n            </td>\n            <td>\n                  <code><autoref identifier=\"str\" backlink-type=\"used-by\" backlink-anchor=\"mlvbench.infrastructure.resolve_layers\" optional>str</autoref> | None</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>A specification of a list of layers, provided as a comma-separated\nstring. Each value may be a non-negative index (e.g. 0), a negative index\n(e.g. -1), or a layer name (e.g. block.12). If None, all layers are\nreturned.</p>\n              </div>\n            </td>\n            <td>\n                <em>required</em>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    <p><span class=\"doc-section-title\">Returns:</span></p>\n    <table>\n      <thead>\n        <tr>\n          <th>Type</th>\n          <th>Description</th>\n        </tr>\n      </thead>\n      <tbody>\n          <tr class=\"doc-section-item\">\n            <td>\n                  <code><autoref identifier=\"list\" backlink-type=\"returned-by\" backlink-anchor=\"mlvbench.infrastructure.resolve_layers\" optional>list</autoref>[<autoref identifier=\"str\" backlink-type=\"returned-by\" backlink-anchor=\"mlvbench.infrastructure.resolve_layers\" optional>str</autoref>]</code>\n            </td>\n            <td>\n              <div class=\"doc-md-description\">\n                <p>A deduplicated list of resolved layer names.</p>\n              </div>\n            </td>\n          </tr>\n      </tbody>\n    </table>\n\n\n    </div>\n\n</div>\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h2 id=\"mlvbench.infrastructure.save_metadata\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">save_metadata</span><span class=\"p\">(</span><span class=\"n\">output_path</span><span class=\"p\">,</span> <span class=\"n\">parameters</span><span class=\"p\">)</span></code>\n\n<a href=\"#mlvbench.infrastructure.save_metadata\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h2>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Save the metadata for an experiment run.</p>\n\n\n    </div>\n\n</div>\n\n<div class=\"doc doc-object doc-function\">\n\n\n<h2 id=\"mlvbench.infrastructure.save_model_info\" class=\"doc doc-heading\">\n            <code class=\"highlight language-python\"><span class=\"n\">save_model_info</span><span class=\"p\">(</span><span class=\"n\">output_path</span><span class=\"p\">,</span> <span class=\"n\">model</span><span class=\"p\">)</span></code>\n\n<a href=\"#mlvbench.infrastructure.save_model_info\" class=\"headerlink\" title=\"Permanent link\">&para;</a></h2>\n\n\n    <div class=\"doc doc-contents \">\n\n        <p>Save model info for an experiment run.</p>\n\n\n    </div>\n\n</div>\n\n\n\n  </div>\n\n    </div>\n\n</div>",
+    "search": [
+      {
+        "location": null,
+        "level": 1,
+        "title": "<code>mlvbench.infrastructure</code>",
+        "text": "<p>Infrastructure utilities.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.infrastructure.resolve_layers",
+        "level": 2,
+        "title": "<code>resolve_layers(all_layers, layers_spec)</code>",
+        "text": "<p>Resolve a layer specification string to a list of layer names.</p> <p>Intended for experiment scripts that expose a <code>--layers</code> command line argument. The resolved list can be passed directly to <code>ProbeTrainer.fit</code>.</p> <p>Parameters:</p> Name Type Description Default <code>all_layers</code> <code>list[str]</code> <p>The full list of layer names from the model (<code>model.layer_names</code>).</p> required <code>layers_spec</code> <code>str | None</code> <p>A specification of a list of layers, provided as a comma-separated string. Each value may be a non-negative index (e.g. 0), a negative index (e.g. -1), or a layer name (e.g. block.12). If None, all layers are returned.</p> required <p>Returns:</p> Type Description <code>list[str]</code> <p>A deduplicated list of resolved layer names.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.infrastructure.save_metadata",
+        "level": 2,
+        "title": "<code>save_metadata(output_path, parameters)</code>",
+        "text": "<p>Save the metadata for an experiment run.</p>",
+        "path": [],
+        "tags": []
+      },
+      {
+        "location": "mlvbench.infrastructure.save_model_info",
+        "level": 2,
+        "title": "<code>save_model_info(output_path, model)</code>",
+        "text": "<p>Save model info for an experiment run.</p>",
+        "path": [],
+        "tags": []
+      }
+    ],
+    "title": "mlvbench.infrastructure",
+    "toc": [
+      {
+        "title": "mlvbench.infrastructure",
+        "content": "<code>mlvbench.infrastructure</code>",
+        "id": "mlvbenchinfrastructure",
+        "url": "#mlvbenchinfrastructure",
+        "children": [
+          {
+            "title": "infrastructure",
+            "content": "infrastructure",
+            "id": "mlvbench.infrastructure",
+            "url": "#mlvbench.infrastructure",
+            "children": [],
+            "level": 2
+          },
+          {
+            "title": "resolve_layers",
+            "content": "resolve_layers",
+            "id": "mlvbench.infrastructure.resolve_layers",
+            "url": "#mlvbench.infrastructure.resolve_layers",
+            "children": [],
+            "level": 2
+          },
+          {
+            "title": "save_metadata",
+            "content": "save_metadata",
+            "id": "mlvbench.infrastructure.save_metadata",
+            "url": "#mlvbench.infrastructure.save_metadata",
+            "children": [],
+            "level": 2
+          },
+          {
+            "title": "save_model_info",
+            "content": "save_model_info",
+            "id": "mlvbench.infrastructure.save_model_info",
+            "url": "#mlvbench.infrastructure.save_model_info",
+            "children": [],
+            "level": 2
+          }
+        ],
+        "level": 1
+      }
+    ]
+  },
+  "hash": 12273485971954117959
+}