TensorFlow v0.11.0rc0 发布,一个表达机器学习算法的接口

jopen 5年前
   <p style="text-align: center;"><img alt="" src="https://simg.open-open.com/show/326eaa9a539ec5575022ac39de585d3d.png" /></p>    <p> </p>    <p>TensorFlow 是一个表达机器学习算法的接口,并且是执行算法的实现框架。使用 TensorFlow 表示的计算可以在众多异构的系统上方便地移植,从移动设别如手机或者平板电脑到成千的GPU计算集群上都可以执行。该系统灵活,可以被用来表示很多的算法包括,深度神经网络的训练和推断算法,也已经被用作科研和应用机器学习系统在若干的计算机科学领域或者其他领域中,例如语言识别、计算机视觉、机器人、信息检索、自然语言理解、地理信息抽取和计算药物发现。</p>    <p style="text-align: center;"><a href="https://simg.open-open.com/show/4a67e12961d71d510c83c2aa35a8febb.gif"><img alt="" src="https://simg.open-open.com/show/4a67e12961d71d510c83c2aa35a8febb.gif" /></a></p>    <h2>更新日志</h2>    <p>主要功能和改进</p>    <ul>     <li>cuDNN 5 support.</li>     <li>HDFS Support.</li>     <li>Adds Fused LSTM support via cuDNN 5 in <code>tensorflow/contrib/cudnn_rnn</code>.</li>     <li>Improved support for NumPy style basic slicing including non-1 strides, ellipses, newaxis, and negative indices. For example complicated expressions like <code>foo[1, 2:4, tf.newaxis, ..., :-3:-1, :]</code> are now supported. In addition we have preliminary (non-broadcasting) support for sliced assignment to variables. In particular one can write <code>var[1:3].assign([1,11,111])</code>.</li>     <li>Introducing <code>core/util/tensor_bundle</code> module: a module to efficiently serialize/deserialize tensors to disk. Will be used in TF's new checkpoint format.</li>     <li>Added tf.svd for computing the singular value decomposition (SVD) of dense matrices or batches of matrices (CPU only).</li>     <li>Added gradients for eigenvalues and eigenvectors computed using <code>self_adjoint_eig</code> or<code>self_adjoint_eigvals</code>.</li>     <li>Eliminated <code>batch_*</code> methods for most linear algebra and FFT ops and promoted the non-batch version of the ops to handle batches of matrices.</li>     <li>Tracing/timeline support for distributed runtime (no GPU profiler yet).</li>     <li>C API gives access to inferred shapes with <code>TF_GraphGetTensorNumDims</code> and <code>TF_GraphGetTensorShape</code>.</li>     <li>Shape functions for core ops have moved to C++ via <code>REGISTER_OP(...).SetShapeFn(...)</code>. Python shape inference RegisterShape calls use the C++ shape functions with<code>common_shapes.call_cpp_shape_fn</code>. A future release will remove <code>RegisterShape</code> from python.</li>    </ul>    <h2>Bug修正等变化</h2>    <ul>     <li>Documentation now includes operator overloads on Tensor and Variable.</li>     <li><code>tensorflow.__git_version__</code> now allows users to identify the version of the code that TensorFlow was compiled with. We also have <code>tensorflow.__git_compiler__</code> which identifies the compiler used to compile TensorFlow's core.</li>     <li>Improved multi-threaded performance of <code>batch_matmul</code>.</li>     <li>LSTMCell, BasicLSTMCell, and MultiRNNCell constructors now default to <code>state_is_tuple=True</code>. For a quick fix while transitioning to the new default, simply pass the argument <code>state_is_tuple=False</code>.</li>     <li>DeviceFactory's AddDevices and CreateDevices functions now return a Status instead of void.</li>     <li>Int32 elements of list(type) arguments are no longer placed in host memory by default. If necessary, a list(type) argument to a kernel can be placed in host memory using a HostMemory annotation.</li>     <li><code>uniform_unit_scaling_initializer()</code> no longer takes a <code>full_shape</code> arg, instead relying on the partition info passed to the initializer function when it's called.</li>     <li>The NodeDef protocol message is now defined in its own file <code>node_def.proto</code> <code>instead of graph.proto</code>.</li>     <li><code>ops.NoGradient</code> was renamed <code>ops.NotDifferentiable</code>. <code>ops.NoGradient</code> will be removed soon.</li>     <li><code>dot.h</code> / DotGraph was removed (it was an early analysis tool prior to TensorBoard, no longer that useful). It remains in history should someone find the code useful.</li>    </ul>    <p> </p>    <h2>下载</h2>    <ul>     <li><a href="/misc/goto?guid=4958994140524243187" rel="nofollow"><strong>Source code</strong> (zip)</a></li>     <li><a href="/misc/goto?guid=4958994140653872664" rel="nofollow"><strong>Source code</strong> (tar.gz)</a></li>    </ul>    <p> </p>    <p>本站原创,转载时保留以下信息:<br /> 本文转自:深度开源(open-open.com)<br /> 原文地址:<a href="http://www.open-open.com/news/view/16db7e2c">http://www.open-open.com/news/view/16db7e2c</a></p>