I have model trained with tensorflow 1.15 and saved as checkpoint (with .meta, .index and .data files).
What I need is to add some additional operations to the begin and to the end of this graph. Some of this operations exists only in tensorflow 2.0 and tensorflow_text 2.0. After that I want to save this model for tensorflow-serving.
What I tried to do: using tensorflow 2.0 I saved it as .pb file like this.
trained_checkpoint_prefix = 'path/to/model'
export_dir = os.path.join('path/to/export', '0')
graph = tf.Graph()
with tf.compat.v1.Session(graph=graph) as sess:
# Restore from checkpoint
loader = tf.compat.v1.train.import_meta_graph(trained_checkpoint_prefix + '.meta')
loader.restore(sess, trained_checkpoint_prefix)
# Export checkpoint to SavedModel
builder = tf.compat.v1.saved_model.builder.SavedModelBuilder(export_dir)
classification_signature = tf.compat.v1.saved_model.signature_def_utils.build_signature_def(
inputs={
'token_indices': get_tensor_info('token_indices_ph:0'),
'token_mask': get_tensor_info('token_mask_ph:0'),
'y_mask': get_tensor_info('y_mask_ph:0'),
},
outputs={'probas': get_tensor_info('ner/Softmax:0'), 'seq_lengths': get_tensor_info('ner/Sum:0')},
method_name='predict',
)
builder.add_meta_graph_and_variables(sess,
[tf.saved_model.TRAINING, tf.saved_model.SERVING],
strip_default_attrs=True, saver=loader,
signature_def_map={'predict': classification_signature}) # , clear_devices=True)
builder.save()
After that I created a tf.keras.Model that load .pb model and do all staff that I need:
import os
from pathlib import Path
import tensorflow as tf
import tensorflow_text as tf_text
class BertPipeline(tf.keras.Model):
def __init__(self):
super().__init__()
vocab_file = Path('path/to/vocab.txt')
vocab = vocab_file.read_text().split('\n')[:-1]
self.vocab_table = self.create_table(vocab)
export_dir = 'path/to/pb/model'
self.model = tf.saved_model.load(export_dir)
self.bert_tokenizer = BertTokenizer(
self.vocab_table,
max_chars_per_token=15,
token_out_type=tf.int64
,
lower_case=True,
)
self.to_dense = tf_text.keras.layers.ToDense()
def call(self, texts):
tokens = self.bert_tokenizer.tokenize(texts)
tokens = tf.cast(tokens, dtype=tf.int32)
mask = self.make_mask(tokens)
token_ids = self.make_token_ids(tokens)
token_indices = self.to_dense(token_ids)
token_mask = self.to_dense(tf.ones_like(mask))
y_mask = self.to_dense(mask)
res = self.model.signatures['predict'](
token_indices=token_indices,
token_mask=token_mask,
y_mask=y_mask,
)
starts_range = tf.range(0, tf.shape(res['seq_lengths'])[0]) * tf.shape(res['probas'])[1]
row_splits = tf.reshape(
tf.stack(
[
starts_range,
starts_range + res['seq_lengths'],
],
axis=1,
),
[-1],
)
row_splits = tf.concat(
[
row_splits,
tf.expand_dims(tf.shape(res['probas'])[0] * tf.shape(res['probas'])[1], 0),
],
axis=0,
)
probas = tf.RaggedTensor.from_row_splits(
tf.reshape(res['probas'], [-1, 2]),
row_splits,
)[::2]
probas
return probas
def make_mask(self, tokens):
masked_suff = tf.concat(
[
tf.ones_like(tokens[:, :, :1], dtype=tf.int32),
tf.zeros_like(tokens[:, :, 1:], dtype=tf.int32),
],
axis=-1,
)
joined_mask = self.join_wordpieces(masked_suff)
return tf.concat(
[
tf.zeros_like(joined_mask[:, :1], dtype=tf.int32),
joined_mask,
tf.zeros_like(joined_mask[:, :1], dtype=tf.int32),
],
axis=-1,
)
def make_token_ids(self, tokens):
joined_tokens = self.join_wordpieces(tokens)
return tf.concat(
[
tf.fill(
[joined_tokens.nrows(), 1],
tf.dtypes.cast(
self.vocab_table.lookup(tf.constant('[CLS]')),
dtype=tf.int32,
)
),
self.join_wordpieces(tokens),
tf.fill(
[joined_tokens.nrows(), 1],
tf.dtypes.cast(
self.vocab_table.lookup(tf.constant('[SEP]')),
dtype=tf.int32,
)
),
],
axis=-1,
)
def join_wordpieces(self, wordpieces):
return tf.RaggedTensor.from_row_splits(
wordpieces.flat_values, tf.gather(wordpieces.values.row_splits,
wordpieces.row_splits))
def create_table(self, vocab, num_oov=1):
init = tf.lookup.KeyValueTensorInitializer(
vocab,
tf.range(tf.size(vocab, out_type=tf.int64), dtype=tf.int64),
key_dtype=tf.string,
value_dtype=tf.int64)
return tf.lookup.StaticVocabularyTable(init, num_oov, lookup_key_dtype=tf.string)
When I call this code it works perfectly:
bert_pipeline = BertPipeline()
print(bbert_pipeline(["Some test string", "another string"]))
---
<tf.RaggedTensor [[[0.17896245419979095, 0.8210375308990479], [0.8825045228004456, 0.11749550700187683], [0.9141901731491089, 0.0858098641037941]], [[0.2768123149871826, 0.7231876850128174], [0.9391192197799683, 0.060880810022354126]]]>
But I have no idea how to save it. If I understand right tf.keras.Model don't treat self.model and self.bert_tokenizer as part of model. If I call bert_pipeline.summary() there are no ops:
bert_pipeline.build([])
bert_pipeline.summary()
---
Model: "bert_pipeline_3"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
to_dense (ToDense) multiple 0
=================================================================
Total params: 0
Trainable params: 0
Non-trainable params: 0
_________________________________________________________________
Besides I tried to run it with tensorflow.compat.v1 using explicit Session and Graph, but in this case I just can't load model properly. Same code with import tensorflow.compat.v1 as tf and boilerplate for tensorflow 1.xx can't initialize some variables:
# tf.saved_model.load(export_dir) changed to tf.saved_model.load_v2(export_dir) above
import tensorflow.compat.v1 as tf
graph = tf.Graph()
with tf.Session(graph=graph) as sess:
bert_pipeline = BertPipeline()
texts = tf.placeholder(tf.string, shape=[None], name='texts')
res_tensor = bert_pipeline(texts)
sess.run(tf.tables_initializer())
sess.run(tf.global_variables_initializer())
sess.run(res_tensor, feed_dict={texts: ["Some test string", "another string"]})
---
FailedPreconditionError Traceback (most recent call last)
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/session.py in _do_call(self, fn, *args)
1364 try:
-> 1365 return fn(*args)
1366 except errors.OpError as e:
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/session.py in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata)
1349 return self._call_tf_sessionrun(options, feed_dict, fetch_list,
-> 1350 target_list, run_metadata)
1351
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/session.py in _call_tf_sessionrun(self, options, feed_dict, fetch_list, target_list, run_metadata)
1442 fetch_list, target_list,
-> 1443 run_metadata)
1444
FailedPreconditionError: [_Derived_]{{function_node __inference_pruned_77348}} {{function_node __inference_pruned_77348}} Attempting to use uninitialized value bert/encoder/layer_3/attention/self/query/kernel
[[{{node bert/encoder/layer_3/attention/self/query/kernel/read}}]]
[[bert_pipeline/StatefulPartitionedCall]]
During handling of the above exception, another exception occurred:
FailedPreconditionError Traceback (most recent call last)
<ipython-input-15-5a0a45327337> in <module>
21 sess.run(tf.global_variables_initializer())
22
---> 23 sess.run(res_tensor, feed_dict={texts: ["Some test string", "another string"]})
24
25 # print(res)
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)
954 try:
955 result = self._run(None, fetches, feed_dict, options_ptr,
--> 956 run_metadata_ptr)
957 if run_metadata:
958 proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)
1178 if final_fetches or final_targets or (handle and feed_dict_tensor):
1179 results = self._do_run(handle, final_targets, final_fetches,
-> 1180 feed_dict_tensor, options, run_metadata)
1181 else:
1182 results = []
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/session.py in _do_run(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)
1357 if handle is None:
1358 return self._do_call(_run_fn, feeds, fetches, targets, options,
-> 1359 run_metadata)
1360 else:
1361 return self._do_call(_prun_fn, handle, feeds, fetches)
/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/client/session.py in _do_call(self, fn, *args)
1382 '\nsession_config.graph_options.rewrite_options.'
1383 'disable_meta_optimizer = True')
-> 1384 raise type(e)(node_def, op, message)
1385
1386 def _extend_graph(self):
FailedPreconditionError: [_Derived_] Attempting to use uninitialized value bert/encoder/layer_3/attention/self/query/kernel
[[{{node bert/encoder/layer_3/attention/self/query/kernel/read}}]]
[[bert_pipeline/StatefulPartitionedCall]]
Please, if you have some thoughts how to fix my approach to save graph or maybe you know how to do it better - tell me. Thanks!
I solved it. First of all I couldn't make it with tf.keras. I used
import tensorflow.compat.v1 as tf
Besides that I used .meta, .index and bla bla checkpoint without saing to '.pb'.
The main thing that I used is described here: Tensorflow: How to replace a node in a calculation graph?
I made 2 different graphs and after that merged them like in this part of code:
def _build_model(self):
with tf.Graph().as_default() as g_1:
self.lookup_table = self._make_lookup_table()
init_table = tf.initialize_all_tables()
self.bert_tokenizer = BertTokenizer(
self.lookup_table,
max_chars_per_token=15,
token_out_type=tf.int64,
lower_case=True,
)
self.texts_ph = tf.placeholder(tf.string, shape=(None,), name="texts_ph") # input
words_without_name, tokens_int_64 = self.bert_tokenizer.tokenize(self.texts_ph)
words = words_without_name.to_tensor(default_value='', name='tokens')
tokens = tf.cast(tokens_int_64, dtype=tf.int32)
mask = self._make_mask(tokens)
token_ids = self._make_token_ids(tokens)
self.token_indices = token_ids.to_tensor(default_value=0, name='token_indices') # output 1
self.token_mask = tf.ones_like(mask).to_tensor(default_value=0, name='token_mask') # output 2
self.y_mask = mask.to_tensor(default_value=0, name='y_mask') # output 3
with tf.Graph().as_default() as g_2:
sess = tf.Session()
path_to_model = 'path/to/model'
self._load_model(sess, path_to_model)
token_indices_2 = g_2.get_tensor_by_name('token_indices_ph:0'),
token_mask_2 = g_2.get_tensor_by_name('token_mask_ph:0'),
y_mask_2 = g_2.get_tensor_by_name('y_mask_ph:0'),
probas = g_2.get_tensor_by_name('ner/Softmax:0')
seq_lengths = g_2.get_tensor_by_name('ner/Sum:0')
exclude_scopes = ('Optimizer', 'learning_rate', 'momentum', 'EMA/BackupVariables')
all_vars = variables._all_saveable_objects()
self.vars_to_save = [var for var in all_vars if all(sc not in var.name for sc in exclude_scopes)]
self.saver = tf.train.Saver(self.vars_to_save
g_1_def = g_1.as_graph_def()
g_2_def = g_2.as_graph_def()
with tf.Graph().as_default() as g_combined:
self.texts = tf.placeholder(tf.string, shape=(None,), name="texts")
y1, y2, y3, self.init_table, self.words = tf.import_graph_def(
g_1_def, input_map={"texts_ph:0": self.texts},
return_elements=["token_indices/GatherV2:0", "token_mask/GatherV2:0", "y_mask/GatherV2:0", 'init_all_tables', 'tokens/GatherV2:0'],
name='',
)
self.dense_probas, self.lengths = tf.import_graph_def(
g_2_def, input_map={"token_indices_ph:0": y1, "token_mask_ph:0": y2, "y_mask_ph:0": y3},
return_elements=["ner/Softmax:0", "ner/Sum:0"],
name='',
)
self.sess = tf.Session(graph=g_combined)
self.graph = g_combined
self.sess.run(self.init_table)
vars_dict_to_save = {v.name[:-2]: g_2.get_tensor_by_name(v.name) for v in self.vars_to_save}
self.saver.restore(self.sess, path_to_model)
You may notice that I call self._load_model(sess, path_to_model) to load model, create saver with needed variables and after that load model again with self.saver.save(sess, path_to_model). First load is needed to read presaved graph and have access to it's tensors. Second is needed to load weights in another session with g_combined merged graph. I think there is a way to do it without loading of data from disk two times, but it works and I don't want to break it :-).
One more important thing is vars_dict_to_save. This dict is needed to make mapping between loaded weights and tensors in graphs.
After that you have complete graph with all operations, so you can call it like this:
def __call__(self, texts):
lengths, words, probs = self.sess.run(
[self.lengths, self.words, self.dense_probas],
feed_dict={
self.texts: texts
},
)
return lengths, words, probs
Pay attention to implementation of __call__ method. It uses session that I created with merged graph.
Once you have full graph with loaded weights it is easy to export graph for serving:
def export(self, export_dir):
with self.graph.as_default():
builder = tf.saved_model.builder.SavedModelBuilder(export_dir)
predict_signature = tf.saved_model.signature_def_utils.predict_signature_def(
inputs={
'texts': self.texts,
},
outputs={
'lengths': self.lengths,
'tokens': self.words,
'probs': self.dense_probas,
},
)
builder.add_meta_graph_and_variables(
self.sess,
[tf.saved_model.SERVING],
strip_default_attrs=True,
signature_def_map={'predict': predict_signature},
saver=self.saver,
main_op=self.init_table,
)
builder.save()
There are a few important moments:
- Use the merged graph .as_default()
- Use the same sessions that you used with merged graph.
- Use the same saver that you used to load weights in merged graph.
- Add main main_op if you have tables that need to be initialized.
I will be happy if it will help somebody :-). It wasn't trivial for me and i spent much time to make it works.
P.S. BertTokenizer in this code slightly differs from such class from tensorflow_text, but it is not related with the problem.