|
|
|
@ -1054,11 +1054,10 @@ class Exporter: |
|
|
|
|
|
|
|
|
|
@try_export |
|
|
|
|
def export_mct(self, prefix=colorstr("Sony MCT:")): |
|
|
|
|
check_requirements(["mct-nightly", "sony-custom-layers[torch]"]) |
|
|
|
|
check_requirements(["model_compression_toolkit==2.1.0", "sony-custom-layers[torch]"]) |
|
|
|
|
import model_compression_toolkit as mct |
|
|
|
|
import onnx |
|
|
|
|
from model_compression_toolkit.core import BitWidthConfig |
|
|
|
|
from model_compression_toolkit.core.common.network_editors import NodeNameScopeFilter |
|
|
|
|
import subprocess |
|
|
|
|
from model_compression_toolkit.core.pytorch.pytorch_device_config import get_working_device, set_working_device |
|
|
|
|
from sony_custom_layers.pytorch.object_detection.nms import multiclass_nms |
|
|
|
|
|
|
|
|
@ -1109,25 +1108,12 @@ class Exporter: |
|
|
|
|
yield [img] |
|
|
|
|
|
|
|
|
|
tpc = mct.get_target_platform_capabilities( |
|
|
|
|
fw_name="pytorch", target_platform_name="imx500", target_platform_version="v3" |
|
|
|
|
) |
|
|
|
|
|
|
|
|
|
# Configure MCT manually for specific layers |
|
|
|
|
bit_cfg = BitWidthConfig() |
|
|
|
|
bit_cfg.set_manual_activation_bit_width( |
|
|
|
|
[ |
|
|
|
|
NodeNameScopeFilter("mul"), |
|
|
|
|
NodeNameScopeFilter("sub"), |
|
|
|
|
NodeNameScopeFilter("add_6"), |
|
|
|
|
NodeNameScopeFilter("cat_17"), |
|
|
|
|
], |
|
|
|
|
16, |
|
|
|
|
fw_name="pytorch", target_platform_name="imx500", target_platform_version="v1" |
|
|
|
|
) |
|
|
|
|
|
|
|
|
|
config = mct.core.CoreConfig( |
|
|
|
|
mixed_precision_config=mct.core.MixedPrecisionQuantizationConfig(num_of_images=10), |
|
|
|
|
quantization_config=mct.core.QuantizationConfig(concat_threshold_update=True), |
|
|
|
|
bit_width_config=bit_cfg, |
|
|
|
|
quantization_config=mct.core.QuantizationConfig(concat_threshold_update=True) |
|
|
|
|
) |
|
|
|
|
|
|
|
|
|
resource_utilization = mct.core.ResourceUtilization(weights_memory=3146176 * 0.76) |
|
|
|
@ -1141,7 +1127,6 @@ class Exporter: |
|
|
|
|
core_config=config, |
|
|
|
|
target_platform_capabilities=tpc, |
|
|
|
|
) |
|
|
|
|
print("Quantized model is ready") |
|
|
|
|
|
|
|
|
|
else: |
|
|
|
|
gptq_config = mct.gptq.get_pytorch_gptq_config(n_epochs=1000, use_hessian_based_weights=False) |
|
|
|
@ -1157,8 +1142,6 @@ class Exporter: |
|
|
|
|
target_platform_capabilities=tpc, |
|
|
|
|
) |
|
|
|
|
|
|
|
|
|
print("Quantized-PTQ model is ready") |
|
|
|
|
|
|
|
|
|
if self.args.nms: |
|
|
|
|
# Define PostProcess params |
|
|
|
|
score_threshold = 0.001 |
|
|
|
@ -1171,7 +1154,7 @@ class Exporter: |
|
|
|
|
iou_threshold=iou_threshold, |
|
|
|
|
max_detections=max_detections, |
|
|
|
|
).to(device=get_working_device()) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
f = Path(str(self.file).replace(self.file.suffix, "_mct_model.onnx")) # js dir |
|
|
|
|
mct.exporter.pytorch_export_model(model=quant_model, save_model_path=f, repr_dataset=representative_dataset_gen) |
|
|
|
|
|
|
|
|
@ -1181,6 +1164,19 @@ class Exporter: |
|
|
|
|
meta.key, meta.value = k, str(v) |
|
|
|
|
|
|
|
|
|
onnx.save(model_onnx, f) |
|
|
|
|
|
|
|
|
|
if not LINUX: |
|
|
|
|
LOGGER.warning(f"{prefix} WARNING ⚠️ MCT imx500-converter is only supported on Linux.") |
|
|
|
|
else: |
|
|
|
|
check_requirements("imx500-converter[pt]==3.14.1") |
|
|
|
|
try: |
|
|
|
|
subprocess.run(["java", "--version"], check=True) |
|
|
|
|
except FileNotFoundError: |
|
|
|
|
LOGGER.error("Java 17 is required for the imx500 conversion. \n Please install Java with: \n sudo apt install openjdk-17-jdk openjdk-17-jre") |
|
|
|
|
return None |
|
|
|
|
|
|
|
|
|
subprocess.run(["imxconv-pt", "-i", "yolov8n_mct_model.onnx", "-o", "yolov8n_imx500_model"], check=True) |
|
|
|
|
|
|
|
|
|
return f, None |
|
|
|
|
|
|
|
|
|
def _add_tflite_metadata(self, file): |
|
|
|
|