DRAFT: Export TFLITE Models #92
Draft
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
parjong
commented
Apr 24, 2022
export_ALBERT.py
Outdated
| converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS] | ||
| converter._experimental_lower_tensor_list_ops = False | ||
|
|
||
| tflite_model = converter.convert() |
Owner
Author
There was a problem hiding this comment.
2022-04-24 13:55:05.748163: W tensorflow/compiler/mlir/lite/flatbuffer_export.cc:1880] Graph contains the following resource op(s), that use(s) resource type. Currently, the resource type is not natively supported in TFLite. Please consider not using the resource type if there are issues with either TFLite converter or TFLite runtime: Resource ops: SentencepieceOp, SentencepieceTokenizeOp
Details: tf.SentencepieceOp() -> (tensor<!tf_type.resource>) : {container = "", device = "", model = ...
shared_name = "SentenceTokenizerInitializer/SentencepieceOp_load_3", use_node_name_sharing = false}
tf.SentencepieceTokenizeOp(tensor<!tf_type.resource>, tensor<?x!tf_type.string>, tensor<i32>, tensor<f32>, tensor<i1>, tensor<i1>, tensor<i1>) -> (tensor<?xi32>, tensor<?xi64>) : {Tsplits = i64, device = "", out_type = i32, return_nbest = false}
tf.StaticRegexReplace(tensor<?x!tf_type.string>) -> (tensor<?x!tf_type.string>) : {device = "", pattern = "\\p{Mn}", replace_global = true, rewrite = ""}
tf.TensorListFromTensor(tensor<?x!tf_type.variant>, tensor<0xi32>) -> (tensor<!tf_type.variant<tensor<!tf_type.variant>>>) : {device = ""}
tf.TensorListFromTensor(tensor<?xi64>, tensor<0xi32>) -> (tensor<!tf_type.variant<tensor<i64>>>) : {device = ""}
tf.TensorListGetItem(tensor<!tf_type.variant<tensor<!tf_type.variant>>>, tensor<i32>, tensor<0xi32>) -> (tensor<!tf_type.variant>) : {device = ""}
tf.TensorListGetItem(tensor<!tf_type.variant<tensor<i64>>>, tensor<i32>, tensor<0xi32>) -> (tensor<i64>) : {device = ""}
tf.TensorListReserve(tensor<i32>, tensor<i32>) -> (tensor<!tf_type.variant<tensor<*x!tf_type.variant>>>) : {device = ""}
tf.TensorListSetItem(tensor<!tf_type.variant<tensor<*x!tf_type.variant>>>, tensor<i32>, tensor<!tf_type.variant>) -> (tensor<!tf_type.variant<tensor<*x!tf_type.variant>>>) : {device = ""}
tf.TensorListStack(tensor<!tf_type.variant<tensor<*x!tf_type.variant>>>, tensor<0xi32>) -> (tensor<?x!tf_type.variant>) : {device = "", num_elements = -1 : i64}
See instructions: https://www.tensorflow.org/lite/guide/ops_select
,.
parjong
force-pushed
the
draft/export_tflite_model
branch
2 times, most recently
from
d3d898e to
44ef833
Compare
parjong
commented
Apr 24, 2022
|
|
||
| converter = tf.lite.TFLiteConverter.from_keras_model(embedding_model) | ||
|
|
||
| tflite_model = converter.convert() |
Owner
Author
There was a problem hiding this comment.
I ran this script under WSL2 with 4GB memory constraint, and it fails as below.
2022-04-24 14:25:55.503549: W tensorflow/python/util/util.cc:368] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.
WARNING:absl:Found untraced functions such as restored_function_body, restored_function_body, restored_function_body, restored_function_body, restored_function_body while saving (showing 5 of 3339). These functions will not be directly callable after loading.
Killed
44ef833 to
3c466cf
Compare
Let's build a .tflite file for state-of-the-art NN models! [ci skip] Signed-off-by: Jonghyun Park <parjong@gmail.com>
parjong
force-pushed
the
draft/export_tflite_model
branch
from
3c466cf to
264c625
Compare
parjong
commented
Apr 24, 2022
| f.write(tflite_model) | ||
| # def export_to_tflite: END | ||
|
|
||
| export_to_tflite(create_encode_model(), 'models.mine/ALBERT.tflite') |
Owner
Author
There was a problem hiding this comment.
It is possible to export the encode module as a .tflite model, but the .tflite model includes flex ops (tf.Einsum):
Flex ops: FlexEinsum
Details:
tf.Einsum(tensor<?x12x?x128xf32>, tensor<?x?x12x64xf32>) -> (tensor<?x?x12x64xf32>) : {device = "", equation = "acbe,aecd->abcd"}
tf.Einsum(tensor<?x?x128xf32>, tensor<128x768xf32>) -> (tensor<?x?x768xf32>) : {device = "", equation = "...x,xy->...y"}
tf.Einsum(tensor<?x?x12x64xf32>, tensor<12x64x768xf32>) -> (tensor<?x?x768xf32>) : {device = "", equation = "abcd,cde->abe"}
tf.Einsum(tensor<?x?x12x64xf32>, tensor<?x?x12x64xf32>) -> (tensor<?x12x?x?xf32>) : {device = "", equation = "aecd,abcd->acbe"}
tf.Einsum(tensor<?x?x3072xf32>, tensor<3072x768xf32>) -> (tensor<?x?x768xf32>) : {device = "", equation = "abc,cd->abd"}
tf.Einsum(tensor<?x?x768xf32>, tensor<768x12x64xf32>) -> (tensor<?x?x12x64xf32>) : {device = "", equation = "abc,cde->abde"}
tf.Einsum(tensor<?x?x768xf32>, tensor<768x3072xf32>) -> (tensor<?x?x3072xf32>) : {device = "", equation = "abc,cd->abd"}
See instructions: https://www.tensorflow.org/lite/guide/ops_select
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Let's build a .tflite file for state-of-the-art NN models!
[ci skip]
Signed-off-by: Jonghyun Park parjong@gmail.com