Hello Falken demo does not launch AI evaluation

[DrunkReaperMatt]

I built the demo from source and ran it. i cannot seem to train the AI with prolonged play and when i start the evaluation process, the console gets flooded with this error:

i don't know how this came to be, but i built the project as listed in the readmes.

[hmoraldo]

Do all tests pass? None of the source files have been modified at all?

[DrunkReaperMatt]

what tests? how to run tests?

source files have not been modified

[hinasakazaki]

You can run tests using python3 service/run_tests.py

I guess we never mentioned in our docs.

But if you haven't modified source code, probably doesn't make sense to run them.

[hmoraldo]

Which demo are you running? cpp/hello_falken/?

So you move the agent, and then try changing control, and the agent doesn't move? or what does it do?

If the agent never acts when changing control, I would debug that problem before moving to the evaluation issue.

[DrunkReaperMatt]

i can move the agent and complete the objectives. once i start an evaluation session, those are the messages that pop up

[hmoraldo]

All demos have a key that allow you to switch between controlling the game yourself and letting the agent play. Before starting an evaluation session it is important to verify that works.

[DrunkReaperMatt]

what kind of key? the api key?

i can control the ship, complete the objective a few times and then press space to activate the evaluation session.

then those messages keep popping up

[bookmagnus]

I am having the same issue, any help would be appreciated. I am running tensorflow 2.10. Test gave me this error

`ERROR: test_min_train_examples (learner.learner_test.LearnerTest) learner.learner_test.LearnerTest.test_min_train_examples

Traceback (most recent call last): File "C:\Users\magnu\miniconda3\envs\py3-TF2.10\lib\unittest\mock.py", line 1369, in patched return func(*newargs, **newkeywargs) File "C:\Projects\Falken\service\learner\learner_test.py", line 612, in test_min_train_examples learner.process_assignment_until( File "C:\Projects\Falken\service\learner\learner.py", line 177, in process_assignment_until s, m = next(assignment_generator) File "C:\Projects\Falken\service\learner\learner.py", line 292, in _process_assignment_incrementally raise e File "C:\Projects\Falken\service\learner\learner.py", line 270, in _process_assignment_incrementally for y in proc.process(): File "C:\Projects\Falken\service\learner\assignment_processor.py", line 718, in process for status, metadata in self._process_step(fetcher): File "C:\Projects\Falken\service\learner\assignment_processor.py", line 669, in _process_step saved_model_id = self._save_and_evaluate_policy(exporter) File "C:\Projects\Falken\service\learner\assignment_processor.py", line 507, in _save_and_evaluate_policy exporter.export_model(tmp_checkpoint_path, evals, self._stats, model_id, File "C:\Projects\Falken\service\learner\model_exporter.py", line 136, in export_model self._export_model(model_export_task) File "C:\Projects\Falken\service\learner\model_exporter.py", line 201, in _export_model export_brain.convert_model_to_tflite(tmp_model_path, tf_lite_path) File "C:\Projects\Falken\service\learner\brains\continuous_imitation_brain.py", line 730, in convert_model_to_tflite f.write(saved_model_to_tflite_model.convert(saved_model_path, ['action'])) File "C:\Projects\Falken\service\learner\brains\saved_model_to_tflite_model.py", line 65, in convert return _patch_tflite_model(converter.convert(), main_graph) File "C:\Projects\Falken\service\learner\brains\saved_model_to_tflite_model.py", line 101, in _patch_tflite_model builder.Finish(lite_model.Pack(builder), File "C:\Projects\Falken\service\generated_flatbuffers\tflite\Model.py", line 322, in Pack operatorCodes = builder.EndVector(len(self.operatorCodes)) TypeError: Builder.EndVector() takes 1 positional argument but 2 were given`

[scotty007]

@bookmagnus The service requires older versions of the Python dependencies (and their respective dependencies), but they are not pinned in service/common/pip_installer.py. So, it installs the latest versions from PyPi.

After giving up to solve the version hell by hand, I found a working solution: https://github.com/astrofrog/pypi-timemachine

Steps on my Linux system (Ubuntu 20.04 LTS):

1. Install and run PyPI time machine in a virtual env:

$ python3 -m venv venv_tm
$ source venv_tm/bin/activate
(venv_tm)$ pip install pypi-timemachine
(venv_tm)$ pypi-timemachine --port 5000 2021-06-30
Starting pypi-timemachine server at http://localhost:5000

2. Patch service/common/pip_installer.py to use the time machine (and allow installation inside a virtual env). Here's the diff:

diff --git a/service/common/pip_installer.py b/service/common/pip_installer.py
index 4296565..8e82d44 100644
--- a/service/common/pip_installer.py
+++ b/service/common/pip_installer.py
@@ -67,7 +67,7 @@ _REQUIRED_PYTHON_MODULES = [
     ModuleInfo(pip_module_name='flufl.lock', import_module_name='flufl.lock'),
 ]
 
-_PIP_INSTALL_ARGS = [sys.executable, '-m', 'pip', 'install', '--user']
+_PIP_INSTALL_ARGS = [sys.executable, '-m', 'pip', 'install', '--index-url', 'http://localhost:5000']
 # Cache of installed modules populated by _module_installed().
 _INSTALLED_MODULE_LIST = []

3. Install and run the Falken service in another virtual env:

$ python3 -m venv venv
$ source venv/bin/activate
(venv)$ python service/launcher.py --project_ids falken_test --generate_sdk_config