Upgrade to tensorflow 2

[holub008]

Unclear if tensorflow will support v1 as python versions progress, but best not to find out.

[holub008]

worth noting that (currently, at least) tensorflow needs to be version 2.2 or higher to run on python 3.8 (via pypi distribution)

https://www.tensorflow.org/install/pip

[holub008]

After a bit of hacking around, I am not tensorflow-capable. Here are the things I tried:

• Update to tensorflow V2 via API drop-ins
  • Couldn't figure out how to move saved checkpoint weights into something v2 understands
  • Also had issues with the tf.compat functions not performing the same actions as their v1 counterparts.
• Writing the model to a SavedModel in v1, then reading into v2
  • In theory, this should be supported. But, for reasons my understanding / time to learn, it doesn't work. See below.

import tensorflow as tf
import os

from ml.ner.ner_model import NERModel
from ml.ner.ner_config import Config
import data as rpd

config = Config()
m = NERModel(config)
m.build()

m.restore_session(os.path.join(rpd.DATA_ROOT, "pico_spans/model.weights/"))

inputs = {
    "word_ids": m.word_ids,
    "sequence_lengths": m.sequence_lengths,
    "char_ids": m.char_ids,
    "word_lengths": m.word_lengths,
    "dropout": m.dropout,
}

outputs = {
    "logits": m.logits,
    "trans_params": m.trans_params,
}
tf.compat.v1.enable_resource_variables()
tf.compat.v1.saved_model.simple_save(m.sess, '/Users/kholub/robopico_savedmodel', inputs, outputs)

Then, when I m = tf.saved_model.load('/Users/kholub/robopico_savedmodel') in v2, got the following:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/saved_model/load.py", line 578, in load
    return load_internal(export_dir, tags)
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/saved_model/load.py", line 613, in load_internal
    root = load_v1_in_v2.load(export_dir, tags)
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/saved_model/load_v1_in_v2.py", line 263, in load
    return loader.load(tags=tags)
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/saved_model/load_v1_in_v2.py", line 246, in load
    signature_functions = self._extract_signatures(wrapped, meta_graph_def)
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/saved_model/load_v1_in_v2.py", line 158, in _extract_signatures
    signature_fn = wrapped.prune(feeds=feeds, fetches=fetches)
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/eager/wrap_function.py", line 358, in prune
    pruned_graph, variable_holder=self._variable_holder)
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/eager/wrap_function.py", line 230, in __init__
    fn_graph, attrs=attrs, signature=signature)
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 1565, in __init__
    func_graph, self._attrs, self._garbage_collector)
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 654, in __init__
    self._func_graph.inputs, self._func_graph.outputs, attrs)
  File "/Users/kholub/rp2/.venv/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 496, in __init__
    compat.as_str(""))
tensorflow.python.framework.errors_impl.InvalidArgumentError: Output 0 of node 'transitions' has a reference type float_ref
    Encountered while creating function '__inference_pruned_2293'

Without actually digging into tf, I can't do much more here. I will see about adding a dockerfile to at least make deploys with older version easier.

[holub008]

I've added a python 3.7 version requirement back. This is the latest version of python which tensorflow v1 is supported for. Note that continued support for both of these versions has an expiry (and may already be expired for tf).