Enable rewrites for quantized input models

If the input model for rewriting is quantized:

- Record de-quantized TFRecords
  - enable writing de-quantized calibration data for the training
  - re-generate augmented training data, if needed
- Use quantization-aware training (QAT) to train the replacement models
- Check if replacement model is quantized:
  If source model is quantized, we make sure rewrite's output model
  is quantized too. Right now, only int8 is supported so raising
  an error if any other datatype is present in the output.

Resolves: MLIA-907, MLIA-908, MLIA-927
Signed-off-by: Benjamin Klimczak <benjamin.klimczak@arm.com>
Change-Id: Icb4070a9e6f1fdb5ce36120d73823986e89ac955

4 files changed, 152 insertions, 37 deletions

diff --git a/src/mlia/nn/rewrite/core/extract.py b/src/mlia/nn/rewrite/core/extract.py
index f609955..4fcf735 100644
--- a/src/mlia/nn/rewrite/core/extract.py
+++ b/src/mlia/nn/rewrite/core/extract.py
@@ -2,19 +2,62 @@
 # SPDX-License-Identifier: Apache-2.0
 """Extract module."""
 # pylint: disable=too-many-arguments, too-many-locals
+from __future__ import annotations
+
 import os
+from functools import partial
+from pathlib import Path
 
 import tensorflow as tf
 from tensorflow.lite.python.schema_py_generated import SubGraphT
 
 from mlia.nn.rewrite.core.graph_edit.cut import cut_model
+from mlia.nn.rewrite.core.graph_edit.record import dequantized_path
 from mlia.nn.rewrite.core.graph_edit.record import record_model
 
-
 os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
 tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
 
 
+def _get_path(
+    ext: str, name: str, dir_path: str | Path, model_is_quantized: bool = False
+) -> Path:
+    """Create a file path for extracted files."""
+    path = Path(dir_path, f"{name}{ext}")
+    return dequantized_path(path) if model_is_quantized else path
+
+
+class TFLitePaths:  # pylint: disable=too-few-public-methods
+    """Provide safe access to TensorFlow Lite file paths."""
+
+    _get_path_tflite = partial(_get_path, ".tflite")
+
+    start = partial(_get_path_tflite, "start")
+    replace = partial(_get_path_tflite, "replace")
+    end = partial(_get_path_tflite, "end")
+
+
+class TFRecordPaths:  # pylint: disable=too-few-public-methods
+    """Provide safe access to tfrec file paths."""
+
+    _get_path_tfrec = partial(_get_path, ".tfrec")
+
+    input = partial(_get_path_tfrec, "input")
+    output = partial(_get_path_tfrec, "output")
+    end = partial(_get_path_tfrec, "end")
+
+
+class ExtractPaths:  # pylint: disable=too-few-public-methods
+    """Get paths to extract files.
+
+    This is meant to be the single source of truth regarding all file names
+    created by the extract() function in an output directory.
+    """
+
+    tflite = TFLitePaths
+    tfrec = TFRecordPaths
+
+
 def extract(
     output_path: str,
     model_file: str,
@@ -26,6 +69,7 @@ def extract(
     show_progress: bool = False,
     num_procs: int = 1,
     num_threads: int = 0,
+    dequantize_output: bool = False,
 ) -> None:
     """Extract a model after cut and record."""
     try:
@@ -33,7 +77,7 @@ def extract(
     except FileExistsError:
         pass
 
-    start_file = os.path.join(output_path, "start.tflite")
+    start_file = ExtractPaths.tflite.start(output_path)
     cut_model(
         model_file,
         input_names=None,
@@ -42,7 +86,7 @@ def extract(
         output_file=start_file,
     )
 
-    input_tfrec = os.path.join(output_path, "input.tfrec")
+    input_tfrec = ExtractPaths.tfrec.input(output_path)
     record_model(
         input_filename,
         start_file,
@@ -50,9 +94,10 @@ def extract(
         show_progress=show_progress,
         num_procs=num_procs,
         num_threads=num_threads,
+        dequantize_output=dequantize_output,
     )
 
-    replace_file = os.path.join(output_path, "replace.tflite")
+    replace_file = ExtractPaths.tflite.replace(output_path)
     cut_model(
         model_file,
         input_names=input_names,
@@ -61,7 +106,7 @@ def extract(
         output_file=replace_file,
     )
 
-    end_file = os.path.join(output_path, "end.tflite")
+    end_file = ExtractPaths.tflite.end(output_path)
     cut_model(
         model_file,
         input_names=output_names,
@@ -71,7 +116,7 @@ def extract(
     )
 
     if not skip_outputs:
-        output_tfrec = os.path.join(output_path, "output.tfrec")
+        output_tfrec = ExtractPaths.tfrec.output(output_path)
         record_model(
             input_tfrec,
             replace_file,
@@ -79,9 +124,10 @@ def extract(
             show_progress=show_progress,
             num_procs=num_procs,
             num_threads=num_threads,
+            dequantize_output=dequantize_output,
         )
 
-        end_tfrec = os.path.join(output_path, "end.tfrec")
+        end_tfrec = ExtractPaths.tfrec.end(output_path)
         record_model(
             output_tfrec,
             end_file,
@@ -89,4 +135,5 @@ def extract(
             show_progress=show_progress,
             num_procs=num_procs,
             num_threads=num_threads,
+            dequantize_output=dequantize_output,
         )
diff --git a/src/mlia/nn/rewrite/core/graph_edit/cut.py b/src/mlia/nn/rewrite/core/graph_edit/cut.py
index 13a5268..53d5389 100644
--- a/src/mlia/nn/rewrite/core/graph_edit/cut.py
+++ b/src/mlia/nn/rewrite/core/graph_edit/cut.py
@@ -1,9 +1,11 @@
 # SPDX-FileCopyrightText: Copyright 2023, Arm Limited and/or its affiliates.
 # SPDX-License-Identifier: Apache-2.0
 """Cut module."""
+from __future__ import annotations
+
 import os
 from collections import defaultdict
-from typing import Optional
+from pathlib import Path
 
 import tensorflow as tf
 from tensorflow.lite.python.schema_py_generated import ModelT
@@ -25,8 +27,8 @@ def tensors_by_name(subgraph: SubGraphT, names: list) -> list:
 
 def cut_subgraph(
     subgraph: SubGraphT,
-    input_tensor_names: Optional[list],
-    output_tensor_names: Optional[list],
+    input_tensor_names: list | None,
+    output_tensor_names: list | None,
 ) -> None:
     """Change the global inputs and outputs of a graph to the provided named tensors."""
     if input_tensor_names is not None:
@@ -131,11 +133,11 @@ def filter_relabel(src_subgraph: SubGraphT, relabel_filter: set) -> tuple:
 
 
 def cut_model(
-    model_file: str,
-    input_names: Optional[list],
-    output_names: Optional[list],
+    model_file: str | Path,
+    input_names: list | None,
+    output_names: list | None,
     subgraph_index: int,
-    output_file: str,
+    output_file: str | Path,
 ) -> None:
     """Cut subgraphs and simplify a given model."""
     model = load_fb(model_file)
diff --git a/src/mlia/nn/rewrite/core/graph_edit/record.py b/src/mlia/nn/rewrite/core/graph_edit/record.py
index 90f3db8..f85433d 100644
--- a/src/mlia/nn/rewrite/core/graph_edit/record.py
+++ b/src/mlia/nn/rewrite/core/graph_edit/record.py
@@ -6,6 +6,7 @@ from __future__ import annotations
 
 import math
 import os
+from contextlib import ExitStack
 from pathlib import Path
 
 import tensorflow as tf
@@ -15,11 +16,22 @@ from mlia.nn.rewrite.core.utils.numpy_tfrecord import numpytf_count
 from mlia.nn.rewrite.core.utils.numpy_tfrecord import numpytf_read
 from mlia.nn.rewrite.core.utils.numpy_tfrecord import NumpyTFWriter
 from mlia.nn.rewrite.core.utils.parallel import ParallelTFLiteModel
+from mlia.nn.tensorflow.config import NameToTensorMap
 
 os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
 tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
 
 
+DEQUANT_SUFFIX = "_dequant"
+
+
+def dequantized_path(filename: str | Path) -> Path:
+    """Append the de-quantization suffix to the given filename."""
+    path = Path(filename)
+    path = Path(path.parent, f"{path.stem}{DEQUANT_SUFFIX}{path.suffix}")
+    return path
+
+
 def record_model(
     input_filename: str | Path,
     model_filename: str | Path,
@@ -28,11 +40,14 @@ def record_model(
     show_progress: bool = False,
     num_procs: int = 1,
     num_threads: int = 0,
+    dequantize_output: bool = False,
 ) -> None:
     """Model recorder.
 
     num_procs: 0 => detect real cores on system
     num_threads: 0 => TFLite impl. specific setting, usually 3
+
+    dequantize: True => de-quantize the recorded output before saving
     """
     model = ParallelTFLiteModel(model_filename, num_procs, num_threads, batch_size)
     if not batch_size:
@@ -51,22 +66,38 @@ def record_model(
         dataset = dataset.batch(batch_size, drop_remainder=False)
         total = int(math.ceil(total / batch_size))
 
-    with NumpyTFWriter(output_filename) as writer:
-        for _, named_x in enumerate(
-            track(dataset.as_numpy_iterator(), total=total, disable=not show_progress)
-        ):
-            named_y = model(named_x)
+    with ExitStack() as stack:
+        writer = stack.enter_context(NumpyTFWriter(output_filename))
+        writer_dequant = None
+        if dequantize_output:
+            dequant_path = dequantized_path(output_filename)
+            writer_dequant = stack.enter_context(NumpyTFWriter(dequant_path))
+
+        def write(writer: NumpyTFWriter, data: NameToTensorMap) -> None:
+            """Write the data using the given NumpyTFWriter instance."""
             if batch_size > 1:
                 for i in range(batch_size):
                     # Expand the batches and recreate each dict as a
                     # batch-size 1 item for the tfrec output
                     recreated_dict = {
                         k: v[i : i + 1]  # noqa: E203
-                        for k, v in named_y.items()
+                        for k, v in data.items()
                         if i < v.shape[0]
                     }
                     if recreated_dict:
                         writer.write(recreated_dict)
             else:
-                writer.write(named_y)
+                writer.write(data)
+
+        for _, named_x in enumerate(
+            track(dataset.as_numpy_iterator(), total=total, disable=not show_progress)
+        ):
+            named_y = model(named_x)
+            write(writer, named_y)
+
+            if dequantize_output:
+                assert writer_dequant
+                named_y_dequant = model.dequantize_outputs(named_y)
+                write(writer_dequant, named_y_dequant)
+
         model.close()
diff --git a/src/mlia/nn/rewrite/core/train.py b/src/mlia/nn/rewrite/core/train.py
index 82af747..6345f07 100644
--- a/src/mlia/nn/rewrite/core/train.py
+++ b/src/mlia/nn/rewrite/core/train.py
@@ -22,9 +22,11 @@ from typing import Literal
 
 import numpy as np
 import tensorflow as tf
+import tensorflow_model_optimization as tfmot
 from numpy.random import Generator
 
 from mlia.nn.rewrite.core.extract import extract
+from mlia.nn.rewrite.core.extract import ExtractPaths
 from mlia.nn.rewrite.core.graph_edit.diff import diff_stats
 from mlia.nn.rewrite.core.graph_edit.join import join_models
 from mlia.nn.rewrite.core.graph_edit.record import record_model
@@ -34,6 +36,7 @@ from mlia.nn.rewrite.core.utils.parallel import ParallelTFLiteModel
 from mlia.nn.tensorflow.config import TFLiteModel
 from mlia.nn.tensorflow.tflite_graph import load_fb
 from mlia.nn.tensorflow.tflite_graph import save_fb
+from mlia.nn.tensorflow.utils import get_tflite_converter
 from mlia.utils.logging import log_action
 
 
@@ -91,6 +94,7 @@ def train(
             input_tfrec,
             input_tensors,
             output_tensors,
+            dequantize_output=True,
         )
     else:
         unmodified_model_dir = None
@@ -106,6 +110,7 @@ def train(
             output_tensors,
             num_procs=train_params.num_procs,
             num_threads=train_params.num_threads,
+            dequantize_output=True,
         )
 
         tflite_filenames = train_in_dir(
@@ -160,7 +165,10 @@ def train(
 
 
 def eval_in_dir(
-    target_dir: str, new_part: str, num_procs: int = 1, num_threads: int = 0
+    target_dir: str,
+    new_part: str,
+    num_procs: int = 1,
+    num_threads: int = 0,
 ) -> tuple:
     """Evaluate a model in a given directory."""
     model_input_path = Path(target_dir, "input_orig.tfrec")
@@ -168,12 +176,12 @@ def eval_in_dir(
     model_input = (
         model_input_path
         if model_input_path.exists()
-        else Path(target_dir, "input.tfrec")
+        else ExtractPaths.tfrec.input(target_dir, False)
     )
     output = (
         model_output_path
         if model_output_path.exists()
-        else Path(target_dir, "output.tfrec")
+        else ExtractPaths.tfrec.output(target_dir, False)
     )
 
     with tempfile.TemporaryDirectory() as tmp_dir:
@@ -194,8 +202,8 @@ def join_in_dir(model_dir: str, new_part: str, output_model: str) -> None:
     """Join two models in a given directory."""
     with tempfile.TemporaryDirectory() as tmp_dir:
         new_end = Path(tmp_dir, "new_end.tflite")
-        join_models(new_part, Path(model_dir, "end.tflite"), new_end)
-        join_models(Path(model_dir, "start.tflite"), new_end, output_model)
+        join_models(new_part, ExtractPaths.tflite.end(model_dir), new_end)
+        join_models(ExtractPaths.tflite.start(model_dir), new_end, output_model)
 
 
 def _get_io_tensors(model: TFLiteModel) -> tuple[str, str]:
@@ -244,7 +252,9 @@ def set_up_data_pipeline(
 
     input_name, output_name = _get_io_tensors(teacher)
 
-    input_filename = Path(train_dir, "input.tfrec")
+    model_is_quantized = replace.is_tensor_quantized(name=input_name)
+
+    input_filename = ExtractPaths.tfrec.input(train_dir, model_is_quantized)
     total = numpytf_count(str(input_filename))
     dict_inputs = numpytf_read(str(input_filename))
 
@@ -264,13 +274,13 @@ def set_up_data_pipeline(
     if any(augmentations):
         # Map the teacher inputs here because the augmentation stage passes these
         # through a TFLite model to get the outputs
-        teacher_outputs = numpytf_read(str(Path(teacher_dir, "input.tfrec"))).map(
-            lambda d: tf.squeeze(d[input_name], axis=0)
-        )
+        teacher_outputs = numpytf_read(
+            str(ExtractPaths.tfrec.input(teacher_dir, model_is_quantized))
+        ).map(lambda d: tf.squeeze(d[input_name], axis=0))
     else:
-        teacher_outputs = numpytf_read(str(Path(teacher_dir, "output.tfrec"))).map(
-            lambda d: tf.squeeze(d[output_name], axis=0)
-        )
+        teacher_outputs = numpytf_read(
+            str(ExtractPaths.tfrec.output(teacher_dir, model_is_quantized))
+        ).map(lambda d: tf.squeeze(d[output_name], axis=0))
 
     dataset = tf.data.Dataset.zip((inputs, teacher_outputs))
     if epochs > 1:
@@ -285,7 +295,23 @@ def set_up_data_pipeline(
         ) -> tuple:
             """Return results of train and teach based on augmentations."""
             augmented_train = augment_train({input_name: train})[input_name]
-            augmented_teach = teacher(augment_teacher({input_name: teach}))[output_name]
+            # If augmentation of the input is enabled, we need to re-generate
+            # the reference output by running the augmented data through the
+            # teacher model.
+            if model_is_quantized:
+                # If the input model is quantized we have to additionally
+                # - quantize the augmented data before running it through the
+                #   (quantized) teacher model
+                # - de-quantize the output for the training.
+                augmented_teach = teacher.dequantize_outputs(
+                    teacher(
+                        teacher.quantize_inputs(augment_teacher({input_name: teach}))
+                    )
+                )[output_name]
+            else:
+                augmented_teach = teacher(augment_teacher({input_name: teach}))[
+                    output_name
+                ]
             return (augmented_train, augmented_teach)
 
         dataset = dataset.map(
@@ -329,15 +355,20 @@ def train_in_dir(
     """
     teacher_dir = baseline_dir if baseline_dir else train_dir
     teacher = ParallelTFLiteModel(
-        f"{teacher_dir}/replace.tflite",
+        ExtractPaths.tflite.replace(teacher_dir),
         train_params.num_procs,
         train_params.num_threads,
         batch_size=train_params.batch_size,
     )
-    replace = TFLiteModel(f"{train_dir}/replace.tflite")
+    replace = TFLiteModel(ExtractPaths.tflite.replace(train_dir))
 
     input_name, output_name = _get_io_tensors(teacher)
 
+    model_is_quantized = replace.is_tensor_quantized(name=input_name)
+
+    if model_is_quantized:
+        replace.check_datatypes(np.int8)
+
     dataset = set_up_data_pipeline(
         teacher,
         replace,
@@ -354,6 +385,8 @@ def train_in_dir(
 
     optimizer = tf.keras.optimizers.Nadam(learning_rate=train_params.learning_rate)
     loss_fn = tf.keras.losses.MeanSquaredError()
+    if model_is_quantized:
+        model = tfmot.quantization.keras.quantize_model(model)
     model.compile(optimizer=optimizer, loss=loss_fn, metrics=["mae"])
 
     logger.info(model.summary())
@@ -432,6 +465,7 @@ def train_in_dir(
                 replace.shape_from_name[input_name],
                 output_name,
                 replace.shape_from_name[output_name],
+                model_is_quantized,
             )
             output_filenames.append(checkpoint_filename)
 
@@ -446,6 +480,7 @@ def save_as_tflite(
     input_shape: list,
     output_name: str,
     output_shape: list,
+    model_is_quantized: bool = False,
 ) -> None:
     """Save Keras model as TFLite file."""
 
@@ -464,7 +499,7 @@ def save_as_tflite(
             keras_model.input.set_shape(orig_shape)
 
     with fixed_input(keras_model, input_shape) as fixed_model:
-        converter = tf.lite.TFLiteConverter.from_keras_model(fixed_model)
+        converter = get_tflite_converter(fixed_model, quantized=model_is_quantized)
     tflite_model = converter.convert()
 
     with open(filename, "wb") as file: