PTQ of Fast R-CNN crashes in PyTorch backend

[alexsu52]

Discussed in https://github.com/openvinotoolkit/nncf/discussions/2547

^{Originally posted by **MinGiSa** March 5, 2024} I've been working on converting Torch models into OpenVINO models recently. I was able to easily do this for the existing classification models in torchvision using the tutorials in openvino_notebooks. However, when I tried to convert the detection models, I encountered errors when i do "quantized_model = nncf.quantize(torchModel, calibrationDataset)". I'm not sure how to solve this. Could you help me? Below is my code and my model is from torchvision / fasterrcnn_resnet50_fpn_v2 import sys import os os.environ["PYTHONUTF8"] = str(1) from pathlib import Path from typing import List, Tuple import nncf import openvino as ov import torch import torchvision.transforms as transforms import distutils.command.build_ext import numpy as np import openvino.runtime as ovr import time from DTT_dataLoader import customDatasetDTT from DTT_utils import createModel, validateDTT, getModelSize if sys.platform == "win32": VS_INSTALL_DIR = r"C:/Program Files (x86)/Microsoft Visual Studio" cl_paths = sorted(list(Path(VS_INSTALL_DIR).glob("**/Hostx86/x64/cl.exe"))) if len(cl_paths) == 0: raise ValueError( "Cannot find Visual Studio. This notebook requires C++. If you installed " "a C++ compiler, please add the directory that contains cl.exe to " "`os.environ['PATH']`" ) else: cl_path = cl_paths[-1] vs_dir = str(cl_path.parent) os.environ["PATH"] += f"{os.pathsep}{vs_dir}" d = distutils.core.Distribution() b = distutils.command.build_ext.build_ext(d) b.finalize_options() os.environ["LIB"] = os.pathsep.join(b.library_dirs) def transformDTT_fn(dataItem): images, _ = dataItem return images torch_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(f"Using {torch_device} device") DATASET_FOLDER = - OUTPUT_FOLDER = - TASK_TYPE = - BASE_MODEL = - XML_PATH = - PRETRAINED_MODEL_PATH = - IMAGE_SIZE = [1024, 1024] NUM_CLASSES = 4 BATCH_SIZE = 1 VALIDATE_TRIGGER = False Path(OUTPUT_FOLDER).mkdir(exist_ok=True) openVinoFp32Path = Path(OUTPUT_FOLDER) / Path(BASE_MODEL + "DTT_fp32").with_suffix(".xml") openVinoInt8Path = Path(OUTPUT_FOLDER) / Path(BASE_MODEL + "DTT_int8").with_suffix(".xml") torchModel = createModel(TASK_TYPE, BASE_MODEL, PRETRAINED_MODEL_PATH, NUM_CLASSES, IMAGE_SIZE[0], IMAGE_SIZE[1]) normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) trainTransform = transforms.Compose([transforms.Resize(IMAGE_SIZE), transforms.ToTensor(), normalize]) validTransform = transforms.Compose([transforms.Resize(IMAGE_SIZE), transforms.ToTensor(), normalize]) testTransform = transforms.Compose([transforms.Resize(IMAGE_SIZE), transforms.ToTensor(), normalize]) trainDataset = customDatasetDTT(xmlFile=XML_PATH, mode='TRAINING', imageFolder=DATASET_FOLDER, transform=trainTransform) validDataset = customDatasetDTT(xmlFile=XML_PATH, mode='VALIDATION', imageFolder=DATASET_FOLDER, transform=validTransform) testDataset = customDatasetDTT(xmlFile=XML_PATH, mode='TEST', imageFolder=DATASET_FOLDER, transform=testTransform) trainLoader = torch.utils.data.DataLoader( trainDataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=0, pin_memory=True, sampler=None, ) valLoader = torch.utils.data.DataLoader( validDataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=0, pin_memory=True, ) testLoader = torch.utils.data.DataLoader( testDataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=0, pin_memory=True, ) calibrationDataset = nncf.Dataset(trainLoader, transformDTT_fn) quantized_model = nncf.quantize(torchModel, calibrationDataset) ################################################ and here is error code --------------------------------------------------------------------------- RuntimeError Traceback (most recent call last) Cell In[33], line 1 ----> 1 quantized_model = nncf.quantize(torchModel, calibrationDataset) File c:\Users\sa\anaconda3\envs\openVino\lib\site-packages\nncf\quantization\quantize_model.py:146, in quantize(model, calibration_dataset, mode, preset, target_device, subset_size, fast_bias_correction, model_type, ignored_scope, advanced_parameters) 143 if backend == BackendType.TORCH: 144 from nncf.torch.quantization.quantize_model import quantize_impl --> 146 return quantize_impl( 147 model=model, 148 calibration_dataset=calibration_dataset, 149 mode=mode, 150 preset=preset, 151 target_device=target_device, 152 subset_size=subset_size, 153 fast_bias_correction=fast_bias_correction, 154 model_type=model_type, 155 ignored_scope=ignored_scope, 156 advanced_parameters=advanced_parameters, 157 ) 159 raise RuntimeError(f"Unsupported type of backend: {backend}") File c:\Users\sa\anaconda3\envs\openVino\lib\site-packages\nncf\torch\quantization\quantize_model.py:67, in quantize_impl(model, calibration_dataset, mode, preset, target_device, subset_size, fast_bias_correction, model_type, ignored_scope, advanced_parameters) 55 nncf_network = wrap_model(copied_model.eval(), example_input) 57 quantization_algorithm = PostTrainingQuantization( 58 preset=preset, 59 target_device=target_device, (...) 64 advanced_parameters=advanced_parameters, 65 ) ---> 67 quantized_model = quantization_algorithm.apply( 68 nncf_network, nncf_network.nncf.get_graph(), dataset=calibration_dataset 69 ) 71 quantized_model.nncf.disable_dynamic_graph_building() 73 return quantized_model File c:\Users\sa\anaconda3\envs\openVino\lib\site-packages\nncf\quantization\algorithms\post_training\algorithm.py:112, in PostTrainingQuantization.apply(self, model, graph, statistic_points, dataset) 109 if statistic_points: 110 step_index_to_statistics = {0: statistic_points} --> 112 return self._pipeline.run_from_step(model, dataset, graph, 0, step_index_to_statistics) File c:\Users\sa\anaconda3\envs\openVino\lib\site-packages\nncf\quantization\algorithms\pipeline.py:164, in Pipeline.run_from_step(self, model, dataset, graph, start_step_index, step_index_to_statistics) 161 step_statistics = collect_statistics(statistic_points, step_model, step_graph, dataset) 163 # Run current pipeline step --> 164 step_model = self.run_step(step_index, step_statistics, step_model, step_graph) 166 step_graph = None # We should rebuild the graph for the next pipeline step 168 return step_model File c:\Users\sa\anaconda3\envs\openVino\lib\site-packages\nncf\quantization\algorithms\pipeline.py:117, in Pipeline.run_step(self, step_index, step_statistics, model, graph) 115 pipeline_step = pipeline_steps[step_index] 116 for algorithm in pipeline_step[:-1]: --> 117 current_model = algorithm.apply(current_model, current_graph, step_statistics) 118 current_graph = NNCFGraphFactory.create(current_model) 119 current_model = pipeline_step[-1].apply(current_model, current_graph, step_statistics) File c:\Users\sa\anaconda3\envs\openVino\lib\site-packages\nncf\quantization\algorithms\min_max\algorithm.py:834, in MinMaxQuantization.apply(self, model, graph, statistic_points, dataset) 832 statistics = tensor_collector.get_statistics() 833 if statistics.min_values is None or statistics.max_values is None: --> 834 raise RuntimeError(f"Statistics were not collected for the node {target_node_name}") 835 if self._mode is not None: 836 destination_type = self._quantization_params[quant_group].destination_type RuntimeError: Statistics were not collected for the node FasterRCNN/RoIHeads[roi_heads]/__sub___4

[avitial]

Ref. 138570