QAT example for TF backend #3164
Conversation
github-actions
bot
added
the
documentation
andrey-churkin
force-pushed
the
ac/qat_tf_example
branch
from
a621227 to
80c597a
Compare
andrey-churkin
changed the title
| # (default: 300 samples) of the calibration dataset. | ||
|
|
||
| calibration_dataset = nncf.Dataset(val_dataset, transform_fn) | ||
| tf_quantized_model = nncf.quantize(tf_model, calibration_dataset) |
There was a problem hiding this comment.
nncf.quantize applies the strip transformation for the returned tf_qunaized_model. This will lead to incorrect results during training. I suggest to remove applying the strip transformation from nncf.quantize and apply this one before converting model to OpenVINO.
github-actions
bot
added
NNCF TF
andrey-churkin
force-pushed
the
ac/qat_tf_example
branch
from
f1f5493 to
c1faf40
Compare
|
@daniil-lyakhov , please also review it |
|
NNCF/nightly/test_examples : 642 |
| @@ -0,0 +1,200 @@ | |||
| # Copyright (c) 2024 Intel Corporation | |||
There was a problem hiding this comment.
| # Copyright (c) 2024 Intel Corporation | |
| # Copyright (c) 2025 Intel Corporation |
| target_width=padded_center_crop_size, | ||
| ) | ||
|
|
||
| image = tf.compat.v1.image.resize( |
There was a problem hiding this comment.
Can we use https://www.tensorflow.org/api_docs/python/tf/image/resize instead of tf.compat.v1.image.resize? Looks like unique param align_corners=False does not do anything with the image and it safe to migrate to the newest API
| image = tf.image.convert_image_dtype(image, tf.float32) | ||
|
|
||
| label = tf.one_hot(label, DATASET_CLASSES) | ||
|
|
||
| return image, label |
There was a problem hiding this comment.
| image = tf.image.convert_image_dtype(image, tf.float32) | |
| label = tf.one_hot(label, DATASET_CLASSES) | |
| return image, label | |
| image = tf.image.convert_image_dtype(image, tf.float32) | |
| label = tf.one_hot(label, DATASET_CLASSES) | |
| return image, label |
| int8_ir_path = ROOT / "mobilenet_v2_int8.xml" | ||
| ov.save_model(ov_quantized_model, int8_ir_path) | ||
| print(f"[2/7] Save INT8 model: {int8_ir_path}") | ||
| int8_model_size = get_model_size(int8_ir_path) | ||
|
|
||
| print("[3/7] Benchmark FP32 model:") | ||
| fp32_fps = run_benchmark(fp32_ir_path, shape=[1, 224, 224, 3]) | ||
| print("[4/7] Benchmark INT8 model:") | ||
| int8_fps = run_benchmark(int8_ir_path, shape=[1, 224, 224, 3]) | ||
|
|
||
| print("[5/7] Validate OpenVINO FP32 model:") | ||
| fp32_top1 = validate(ov_model, val_dataset) | ||
| print(f"Accuracy @ top1: {fp32_top1:.3f}") | ||
|
|
||
| print("[6/7] Validate OpenVINO INT8 model:") | ||
| int8_top1 = validate(ov_quantized_model, val_dataset) | ||
| print(f"Accuracy @ top1: {int8_top1:.3f}") | ||
|
|
||
| print("[7/7] Report:") | ||
| print(f"Accuracy drop: {fp32_top1 - int8_top1:.3f}") | ||
| print(f"Model compression rate: {fp32_model_size / int8_model_size:.3f}") | ||
| # https://docs.openvino.ai/latest/openvino_docs_optimization_guide_dldt_optimization_guide.html | ||
| print(f"Performance speed up (throughput mode): {int8_fps / fp32_fps:.3f}") |
There was a problem hiding this comment.
Similar Torch quantization aware training example outputs a table with results. Maybe we can align the examples and use the same table format here IMHO
https://github.com/openvinotoolkit/nncf/blob/develop/examples/quantization_aware_training/torch/resnet18/main.py#L338-L355
| Implementation of the nncf.strip() function for the TF backend | ||
| """ |
There was a problem hiding this comment.
Please fill the docstring with parameters/output value
Changes
Add simple QAT example for TF backend
Reason for changes
Ref: 158980
Related tickets
Ref: 158980
Tests