ponder-lab/ML: 0.52.107

What's Changed

• Call a Sequential model's layers from its call, so a user layer in the list receives its input through the model (wala/ML#832) by @khatchad in https://github.com/ponder-lab/ML/pull/1019
• Type the element of a from_generator dataset that declares a dict as no tensor (wala/ML#1088) by @khatchad in https://github.com/ponder-lab/ML/pull/1021
• Call the function tf.map_fn is given with an element of elems, so the function has a call-graph node and its parameter the element's type (wala/ML#942) by @khatchad in https://github.com/ponder-lab/ML/pull/1020
• Answer ⊥ rather than UNKNOWN for an element-wise operand whose dtype read is empty while it may still grow (wala/ML#1094) by @khatchad in https://github.com/ponder-lab/ML/pull/1023
• Hand a map function a dict when the receiver's element may be one, with each key it reads typed from the receiver's element (wala/ML#1088) by @khatchad in https://github.com/ponder-lab/ML/pull/1022
• Declare Layer as the superclass of the built-in Keras layers, so a program class that subclasses one is a Keras layer (wala/ML#1098) by @khatchad in https://github.com/ponder-lab/ML/pull/1024
• Model the reversed builtin as iterating its argument, so a loop over reversed(y) reads the elements of y (wala/ML#1099) by @khatchad in https://github.com/ponder-lab/ML/pull/1025
• Record whether a property read or write came from subscript syntax or attribute syntax (wala/ML#1100) by @khatchad in https://github.com/ponder-lab/ML/pull/1026
• Record a Keras symbolic tensor in a TensorOrigin carried through operations, layer applications and parameters (wala/ML#1101) by @khatchad in https://github.com/ponder-lab/ML/pull/1027
• Return a list from tape.gradient when its sources may be a list, and a tuple when they may be a tuple, each holding the gradient (wala/ML#464). by @khatchad in https://github.com/ponder-lab/ML/pull/1028
• Model Layer.add_loss as a write of the loss to the layer's _losses (wala/ML#1095) by @khatchad in https://github.com/ponder-lab/ML/pull/1029
• Read a + over Python strings as a string concatenation with no tensor result (wala/ML#1103) by @khatchad in https://github.com/ponder-lab/ML/pull/1030
• Copy a dictionary written over another's items() field by field, so each key keeps its own value and a ** unpacking or a constant-key read of the copy no longer returns every value (wala/ML#1107). by @khatchad in https://github.com/ponder-lab/ML/pull/1031
• Give a Keras model's variable and weight properties a list, through the model shell's __init__ for a subclassed model and through the functional constructor, so tape.gradient over model.trainable_variables returns a list of gradients (wala/ML#1111). by @khatchad in https://github.com/ponder-lab/ML/pull/1032
• Call a program subclass of a built-in Keras layer through the layer's call, typed as its __call__, and read a Dense subclass's size from its constructor's call (wala/ML#1108) by @khatchad in https://github.com/ponder-lab/ML/pull/1033
• Keep the Keras symbolic origin out of a program-defined layer's call input and put it on the layer application's result (wala/ML#1109) by @khatchad in https://github.com/ponder-lab/ML/pull/1034
• Leave a directly invoked call body's input out of the Keras call input transfer, so layer.call(x) keeps its argument's dtype and Keras symbolic origin (wala/ML#1114) by @khatchad in https://github.com/ponder-lab/ML/pull/1035
• Summarize tf.executing_eagerly() and fold a call to it in a module's own body to True, so a branch there takes its eager arm (wala/ML#1110) by @khatchad in https://github.com/ponder-lab/ML/pull/1036
• Type a tensor's numpy() result as a NumPy array of the tensor's shape and dtype, and the bound method t.numpy as no tensor (wala/ML#1105) by @khatchad in https://github.com/ponder-lab/ML/pull/1037
• Name each variable temporary of a file from one counter, so a loop nested in another loop's body no longer overwrites the outer loop's iterable and iterator (wala/ML#1106) by @khatchad in https://github.com/ponder-lab/ML/pull/1038
• Model Dataset.cardinality(), tf.data.experimental.cardinality and the cardinality constants, so a dataset's cardinality is a scalar int64 tensor (wala/ML#1072) by @khatchad in https://github.com/ponder-lab/ML/pull/1040
• Read a Python number computed from a range loop index, or from a parameter every caller passes one, as a Python number, so a tensor times it keeps the tensor operand's type (wala/ML#1083) by @khatchad in https://github.com/ponder-lab/ML/pull/1039
• Summarize the time, random and cv2 modules and tf.io.gfile.GFile as reading or writing module state, so a mod-ref query sees a call's effect (wala/ML#1080) by @khatchad in https://github.com/ponder-lab/ML/pull/1041
• Read a slice of a receiver that is only a Python list or tuple as no tensor, so a list of tensors is no longer sliced as if it were their stack (wala/ML#1121) by @khatchad in https://github.com/ponder-lab/ML/pull/1042
• Read a Sequential model's weights off its layer list, and as unknown when the list holds a layer other than Dense (wala/ML#832) by @khatchad in https://github.com/ponder-lab/ML/pull/1044
• Treat a None among a Dense layer's possible units as unresolved in both weight walks, so the model's weights are of unknown shape rather than a crash by @khatchad in https://github.com/ponder-lab/ML/pull/1048
• Forget the settled recomputations of the roots a declined binding re-evaluates, so the settlement skip no longer keeps a root's pre-decline value when the declined evaluation leaves its state unchanged. by @khatchad in https://github.com/ponder-lab/ML/pull/1049
• Restrict the settlement pass's walks to the readers, transitively, of the queries whose settled values can change, so a pass no longer re-walks the whole query graph to re-confirm unchanged cyclic queries. by @khatchad in https://github.com/ponder-lab/ML/pull/1050
• Bump black from 26.10.0 to 26.10.1 in the python-tooling group by @dependabot[bot] in https://github.com/ponder-lab/ML/pull/1051
• Bump org.codehaus.mojo:versions-maven-plugin from 2.22.0 to 2.23.0 in the maven-plugins group by @dependabot[bot] in https://github.com/ponder-lab/ML/pull/1052
• Model a dictionary's copy() as a new dictionary holding the receiver's entries, so a value read off the copy is the original's (wala/ML#1082) by @khatchad in https://github.com/ponder-lab/ML/pull/1045
• Test how a script is named and an import bound when PYTHONPATH entries overlap or share a relative path by @khatchad in https://github.com/ponder-lab/ML/pull/1053

Full Changelog: https://github.com/ponder-lab/ML/compare/0.52.106...0.52.107