Is followup to: commit HEAD~1 (@`cxx/ClassCns.hxx:classCnsTests<>()`: general use)`), which instruments more forms of tests. TODO: `squash`?
@`posts/VirusAnalysis.md`: include this.

	+`class TensorFlowCns : public Cns`; implements `class Cns : public Object`.
	+`initSynapse()`; helper function for `pseudoRandomSynapses()`.
	+`initScopeRootBack()`, +`pseudoRandomSynapses()`; helper functions for `TensorFlowCns::setupSynapses()`. Is followup to HEAD~10 (+`cxx/ClassCns.hxx:pseudoRandomSynapses()`, helper), which introduced `Cns::pseudoRandomSynapses`.
	+`setupSynapsesImpl()`; helper function for `TensorFlowCns::setupSynapses()`. Is followup to HEAD~5 (+`cxx/ClassCns.hxx:setupSynapsesPostProcess()`), which introduced `Cns::setupSynapsesPostProcess()`.
	+`initScopeRootForward()`; helper function for `TensorFLowCns:process*()`.
	+`printGraphNodes()`; helper function for stacktraces.

[_Cisco-Talos_ asked for a pull request which includes our analysis code](https://github.com/SwuduSusuwu/SusuLib/issues//1) (code which is available though dual licenses; [Apache 2](https://github.com/SwuduSusuwu/SusuLib/blob/trunk/LICENSE) or [GPLv2](https://github.com/SwuduSusuwu/SusuLib/blob/trunk/LICENSE_GPLv2)), but this code relies on an [experimental backend for _TensorFlow_](https://github.com/SwuduSusuwu/SusuLib/blob/preview/cxx/ClassTensorFlowCns.hxx).

@`./compile_commands.json`; include `ClassTensorFlowCns` (required for `clang-tidy cxx/ClassTensorFlowCns.hxx`)
@`./posts/VirusAnalysis.md`; include all this.
@`./README.md#purposes`, @`./cxx/README.md#purposes`; document `./cxx/ClassTensorFlowCns.hxx`.
TODO: through [_Termux_ (smartphone) must have](https://github.com/termux/termux-packages/issues//24657) [`libtensorflow`](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/tools/android) for `SUSUWU_BUILD_OBJECTS()`, `SUSUWU_BUILD_EXECUTABLE()`, `susuwuUnitTests()` with `-DSUSUWU_USE_TENSORFLOW=true`.

+`SUSUWU_CNS_LOCAL_COEFFICIENTS`; local `coefficients`
+`SUSUWU_TENSORFLOW_HAS_DATATYPETOENUM`; use `tensorflow::DataTypeToEnum`.
@`./README.md#optionssetup`; document `-DSUSUWU_CNS_LOCAL_COEFFICIENTS`, `-DSUSUWU_TENSORFLOW_HAS_DATATYPETOENUM`.
@`./cxx/ClassTensorFlowCns.hxx`:
	`TensorFlowCns::{coefficients,biases}`; wrap with `#if SUSUWU_CNS_LOCAL_COEFFICIENTS`.
//	`TensorFlowCns::dumpTo()`; `#if !SUSUWU_CNS_LOCAL_COEFFICIENTS`, use local stack to process `root.WithOpName("coefficients")` and `root.WithOpName("biases")`

@`cxx/AssistantCns.*xx`, @`cxx/VirusAnalysis.*xx`: if `SUSUWU_USE_TENSORFLOW`, use `class TensorFlowCns`.
@`cxx/ClassTensorFlowCns.hxx`: `#ifdef SUSUWU_USE_TENSORFLOW` wraps whole file, fixes "tensorflow/cc/framework/ops.h not found".
@`posts/VirusAnalysis.md`: include this.
Is followup to: commit HEAD~1 (+`cxx/ClassTensorFlowCns.hxx`; `Cns` on TensorFlow)
TODO?: @`cxx/Macros.hxx`: +`SUSUWU_USE_TENSORFLOW` (`#include` comments list this, but for now definition is just from `./build.sh`)

	on inputs (`outputNorms()` on outputs) for `setupSynapses()` and `processTo*()`.
	Purpose: solves issues which some optimizers and activation functions have with large values.

@`posts/VirusAnalysis.md`; include this.
Is followup to: commit HEAD~3 (+`cxx/ClassTensorFlowCns.hxx`; `Cns` on TensorFlow), which introduced `class TensorFlowCns`.
	TODO: `squash` into commit HEAD~3 --- or into commit HEAD~4 (+`ClassCns::dumpTo()`, +`ClassCns::loadFrom()`), if that commit is moved after what is now HEAD~3.

Unused. Current architecture just implements SLP (single-layer-perceptron).
TODO: Those macros serve as reminders to [choose how to implement `layersOfNeurons`](https://github.com/copilot/share/427f408e-08e0-8c80-9151-f24920212857).

@`posts/VirusAnalysis.md`; include all this.

@`cxx/ClassTensorFlowCns.hxx`:
	@`setupSynapses()`; store `root.WithOpName("biases")` (membrane potentials) into `TensorFlowCns.biases`.
	@`processTo*()`; add `root.WithOpName("biases")` to output.
	+`getParameterCount()`; include count of biases into total parameter count.

Is followup to commit HEAD~1 (+`cxx/ClassTensorFlowCns.hxx`; `Cns` on TensorFlow)`; loss), which stores connectome coefficients (also known as "synaptic strengths") into `TensorFlowCns.coefficients`. TODO: squash this?
@`posts/VirusAnalysis.md`; include all this.

	allows to disable `biases` (matrix add), for formulas which just use `MatMul`.

@`cxx/ClassTensorFlowCns.hxx`: wrap all definitions (and uses) of `biases` with `SUSUWU_CNS_IF_BIAS()`.
Is followup to: commit HEAD~1 (+`cxx/ClassTensorFlowCns.hxx:biases` (potentials)). TODO: `squash` this?

@`posts/VirusAnalysis.md`: include all this.

	Is followup to: commit b10cabe (+`sh/Includes.sh`, @`sh/README.md#purposes`)
	Use `SUSUWU_DEPENDENCY_INCLUDE()` to search for `libtensorflow` include paths. If found, put into `FLAGS_USER` (which goes into `CXXFLAGS`). If not found, suggest how to install `libtensorflow`.
		Is followup to: commit b7c33e0 (+`sh/make.sh:SUSUWU_DEPENDENCY_INCLUDE()`).

	If `SUSUWU_SETUP_BUILD_FLAGS_CONDITIONAL "" "-DUSE_TENSORFLOW_CNS" "-ltensorflow" cxx/ClassTensorFlowCns.hxx` is success, store. If not, restore backup flags plus suggest how to troubleshoot.
		Is followup to: commit HEAD~5 (@`README.md#optionssetup`; +`USE_TENSORFLOW_CNS`), which introduced `USE_TENSORFLOW_CNS`. TODO: `squash` this?

	+`SUSUWU_INCLUDES_LIBTENSORFLOW_ERROR`, +`SUSUWU_INCLUDES_LIBTENSORFLOW_PASS`; prevents re-execution of `${CXX} ${CXXFLAGS} ${SUSUWU_INCLUDES_LIBTENSORFLOW_FLAGS} -c "${SUSUWU_INCLUDES_LIBTENSORFLOW_TEST_PATH}"` from `SUSUWU_TEST_BASH()`.

@`build.sh`: use `SUSUWU_INCLUDES_LIBTENSORFLOW()`.
@`README.md#purposes`, @`sh/README.md#purposes`: document `SUSUWU_INCLUDES_LIBTENSORFLOW()`.

Improves <https://github.com/SwuduSusuwu/SusuLib/issues//6>

@`build.sh`: +`SUSUWU_IS_VIRTUAL`, +`SUSUWU_INSTALL_TENSORFLOW`; installs prerequisites + TensorFlow. TODO: move those into `sh/Includes.sh`?
@`README.md#optionssetup`; document this.
@`.gitignore`, @`sh/make.sh:SUSUWU_BUILD_CTAGS()`; exclude `./tensorflow/`, `./ml_dtypes/`.
@`sh/Includes.sh:SUSUWU_INCLUDES_LIBTENSORFLOW()`: use `SUSUWU_INSTALL_TENSORFLOW`
Is followup to: commit HEAD~1 (+`sh/Includes.sh:SUSUWU_INCLUDES_LIBTENSORFLOW()`), which gives TensorFlow paths to `${CXXFLAGS}` (if found).

@`sh/Includes.sh:SUSUWU_INCLUDES_LIBTENSORFLOW()`:
	+`export TF_CPP_MIN_LOG_LEVEL=0`; so TensorFlow debug messages are shown.
	If not `${USE_GOOGLEAPIS_TENSORFLOW}`, prefer `libtensorflow_cc` C++ package (which has system headers to use).
	`libeigen3-dev` search improved.
	Prefer system `libeigen3-dev`; fixes "tensorflow/core/framework/tensor.h:26:10: fatal error: /framework/tensor.h:26:10: fatal error:  unsupported/Eigen/CXX11/Tensor: No such file or directory" on Ubuntu.
	Is followup to: commit HEAD~2 (+`sh/Includes.sh:SUSUWU_INCLUDES_LIBTENSORFLOW()`). TODO: `squash`?

@`build.sh`:
	@`SUSUWU_IS_VIRTUAL`: improved.
	@`SUSUWU_INSTALL_TENSORFLOW`: moved to `sh/Includes.sh:SUSUWU_INCLUDES_LIBTENSORFLOW()`.
	Is followup to: commit HEAD~1 (+`SUSUWU_IS_VIRTUAL`, +`SUSUWU_INSTALL_TENSORFLOW`). TODO: `squash`?

Notice: not tested (Termux nor Ubuntu's package managers have this).

	+`SUSUWU_INCLUDES_LIB_TENSORFLOW_HAS_PROTOBUF`: is set if `.pb.h` source (which `bazel` produces) found.
	If `!SUSUWU_INCLUDES_LIB_TENSORFLOW_HAS_PROTOBUF`, use `protoc` to convert `.proto` sources into those.
	TODO: `squash` this?

Fixes "absl/status/status.h: No such file or directory" on Ubuntu.

Uses `-ltensorflow_cc` instead of `-ltensorflow`.
Fixes `ld.lld: error: undefined symbol` errors.
TODO: scan for include paths to use. Improve codeflow.
Is followup to: commit HEAD~6 (+`sh/Includes.sh:SUSUWU_INCLUDES_LIBTENSORFLOW()`). TODO: `squash`?

@`build.sh`; if in GitHub workflow, `SUSUWU_ABORT_ON_FIRST_ERROR=true`.
@`sh/Includes.sh:SUSUWU_INCLUDES_LIBTENSORFLOW`; use `SUSUWU_ABORT_ON_FIRST_ERROR`.
TODO: `squash`?

+`./cxx/ClassTensorFlowCns.hxx:SigmoidCrossEntropyWithLogits()`; `tensorflow::ops::SigmoidCrossEntropyWithLogits()` was not found.
@`./cxx/ClassTensorFlowCns.hxx:setupSynapses()`;
	For binary classification loop, use `SigmoidCrossEntropyWithLogits()`.
	For linear regression loop, use `tensorflow::ops::SquaredDifference()` (`tensorflow::ops::MeanSquaredError()` was not found)

Is followup to HEAD~5 (+`cxx/ClassTensorFlowCns.hxx`; `Cns` on TensorFlow). TODO: `squash`?
TODO: include arguments, which choose regression formulas.
TODO: replace checks of `outputMode` with +`Cns::isClassification()`, +`Cns::isRegression()`.

@`posts/VirusAnalysis.md`; include all this

	@`./cxx/ClassTensorFlowCns.hxx`: +`TensorFlowCns::dumpTo()`, `TensorFlowCns::loadFrom()`

@`./posts/VirusAnalysis.md#source-code`; include all this.

	validation holdout values (TODO: use those for early stops).
	+`SUSUWU_CLASSCNS_SHUFFLE`; shuffle inputs to `setupSynapses()` so validation set is random.
	@`README.md#optionssetup`; document `-DSUSUWU_CLASSCNS_SHUFFLE`.
	Is followup to: commit HEAD~16 (+`cxx/ClassTensorFlowCns.hxx`; `Cns` on TensorFlow), which introduced `TensorFlowCns::setupSynapses()`. TODO: squash this?

@`posts/VirusAnalysis.md`; include all this.

@`cxx/ClassTensorFlowCns.hxx:setupSynapses()`; return if `loss < desiredLossThreshold`.
Is followup to commit HEAD~2 (@`class TensorFlowCns::setupSynapses()`; loss), which introduced `desiredLossThreshold`.
@`posts/VirusAnalysis.md`; include this.

@`cxx/ClassTensorFlowCns.hxx:setupSynapses()`; if `loss` improves less than `Cns.minLossDelta` for `Cns.patience` loops, abort training loop.
Is followup to commit HEAD~3 (@`class TensorFlowCns::setupSynapses()`; loss), which introduced `desiredLossThreshold`.
@`posts/VirusAnalysis.md`; include all this.

@`cxx/ClassTensorFlowCns.*xx`;
	+`classTensorFlowCnsTests()`, +`classTensorFlowCnsTestsNoexcept()`
	Is followup to: commit HEAD~2 (@`cxx/ClassCns.hxx`: +`classCnsTests<>()`), which introduces the template those use.

@`cxx/main.hxx:SusuwuUnitTestsBitmask`; +`susuwuUnitTestsTensorFlowCnsBit`
@`cxx/main.cxx`; use `classTensorFlowCnsTestsNoexcept()`
@`sh/Includes.sh:SUSUWU_INCLUDES_LIBTENSORFLOW()`; try to compile `cxx/ClassTensorFlowCns.cxx` to determine whether or not to set `-DSUSUWU_USE_TENSORFLOW`.
@`README.md#purposes`, @`cxx/README.md#purposes`; document `class TensorFlowCns`, `classTensorFlowCnsTests()`, `classTensorFlowCnsTestsNoexcept()`.
@`posts/VirusAnalysis.md#source-code`; include all this.

Purpose: diagnosis of `[2025-06-12 06:51:15] [build-stdout] [./cxx/ClassSys.hxx:93: Error: classTensorFlowCnsTests() {throw std::exception("classTensorFlowCnsTests(); label[2000] - output[2000] == 4002.000000, but float epsilon is just 0.000200.");}]`
TODO: `squash` this?
@`posts/VirusAnalysis.md`; include this.

This reduces the first of `cxx/ClassCns.hxx:classCnsTests()`'s `processToInt()` test's `worstLoss` from 2242 to 251, the second from almost 2552 to 564.
TODO: figure out what was wrong with the original codeflow.

@`classCnsTests()`; reduce `epsilon` arguments to new `worstLoss` values.
@`posts/VirusAnalysis.md`; include this.

Is followup to: commit HEAD~1 (@`posts/CnsCompress.md`; English+Markdown improved), which is set to `git push` into `trunk` branch (but this is not, until `class TensorFlowCns` is).

	This is just to ensure source code is still compatible with C++20 projects. `./build.sh`'s minimum requirement is still C++11, not C++20.
TODO: have [_GitHub_'s scan tools](https://github.com/SwuduSusuwu/SusuLib/actions/runs/13170590727/job/36760093264) build with minimum + maximum compatible versions of C++.

@`sh/Includes.sh:SUSUWU_INCLUDES_LIBTENSORFLOW_FLAGS`: ditto (`libtensorflow`'s minimum requirement is still C++17, not C++20).

This reverts commit HEAD~1

@`./posts/VirusAnalysis.md`: include this.
@`./cxx/ResultList.hxx`:
	@`struct ResultList`: is now `typedef<class Set> struct ResultListBase`.
	+`typedef struct ResultListBase<std::unordered_set<ResultListHash>> ResultList; /* `unordered_set` is the O(1) formula. */
	+`typedef struct ResultListBase<std::set<ResultListHash>> ResultListSorted; /* `set` is an O(log n) formula, but is deterministic (sorted). */`