NotImplementedError: (Unimplemented) Delete Weight Dequant Linear Op Pass is not supported for per-channel quantization

[nyh3174680732]

问题描述 / Problem Description

模型在自动压缩之前可以正常运行，自动压缩之后开始报错，错误大概是这个错误表明你在尝试使用不支持的操作。根据错误消息，可能是在尝试使用Delete Weight Dequant Linear Op Pass来处理权重去量化线性操作，但该操作不支持通道量化，不知道怎么解决 完全按照https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/detection#4%E9%A2%84%E6%B5%8B%E9%83%A8%E7%BD%B2教程走的。

运行环境 / Runtime Environment

• Paddle:PaddlePaddle 2.5.2

复现代码 / Reproduction Code

完整报错 / Complete Error Message

#predict.py
%cd ~
!python predict.py /home/aistudio/work/data/val/test.txt result.json
/home/aistudio
Model Arch: DETR
--transform op: Resize
--transform op: NormalizeImage
--transform op: Permute
Traceback (most recent call last):
File "/home/aistudio/predict.py", line 221, in
main(infer_txt, result_path, det_model_path, threshold)
File "/home/aistudio/predict.py", line 204, in main
detector = Detector(pred_config, det_model_path)
File "/home/aistudio/predict.py", line 126, in init
self.predictor, self.config = load_predictor(model_dir)
File "/home/aistudio/predict.py", line 78, in load_predictor
predictor = create_predictor(config)
NotImplementedError: (Unimplemented) Delete Weight Dequant Linear Op Pass is not supported for per-channel quantization (at ../paddle/fluid/framework/ir/delete_weight_dequant_linear_op_pass.cc:130)

可能解决方案 / Possible solutions

附件 / Appendix

[nyh3174680732]

predict.py如下

[nyh3174680732]

import os
import time
import sys
sys.path.append('PaddleDetection')
parent_path = os.path.abspath(os.path.join(file, *(['..'] * 2)))
sys.path.insert(0, parent_path)
import json
import yaml
from functools import reduce
import multiprocessing
from concurrent.futures import ThreadPoolExecutor, as_completed

from PIL import Image
import cv2
import numpy as np
import paddle
from paddle.inference import Config
from paddle.inference import create_predictor
from deploy.python.preprocess import preprocess, Resize, NormalizeImage, Permute, PadStride
from deploy.python.utils import argsparser, Timer, get_current_memory_mb

class PredictConfig():
"""set config of preprocess, postprocess and visualize
Args:
model_dir (str): root path of model.yml
"""

def __init__(self, model_dir):
    # parsing Yaml config for Preprocess
    deploy_file = os.path.join(model_dir, 'infer_cfg.yml')
    with open(deploy_file) as f:
        yml_conf = yaml.safe_load(f)
    self.arch = yml_conf['arch']
    self.preprocess_infos = yml_conf['Preprocess']
    self.min_subgraph_size = yml_conf['min_subgraph_size']
    self.labels = yml_conf['label_list']
    self.mask = False
    self.use_dynamic_shape = yml_conf['use_dynamic_shape']
    if 'mask' in yml_conf:
        self.mask = yml_conf['mask']
    self.tracker = None
    if 'tracker' in yml_conf:
        self.tracker = yml_conf['tracker']
    if 'NMS' in yml_conf:
        self.nms = yml_conf['NMS']
    if 'fpn_stride' in yml_conf:
        self.fpn_stride = yml_conf['fpn_stride']
    self.print_config()

def print_config(self):
    print('%s: %s' % ('Model Arch', self.arch))
    for op_info in self.preprocess_infos:
        print('--%s: %s' % ('transform op', op_info['type']))

def get_test_images(infer_file):
with open(infer_file, 'r') as f:
dirs = f.readlines()
images = [eval(repr(dir.replace('\n', ''))).replace('\', '/') for dir in dirs]
assert len(images) > 0, "no image found in {}".format(infer_file)
return images

def load_predictor(model_dir):
config = Config(
os.path.join(model_dir, 'model.pdmodel'),
os.path.join(model_dir, 'model.pdiparams'))
# initial GPU memory(M), device ID
config.enable_use_gpu(2000, 0)
# optimize graph and fuse op
config.switch_ir_optim(True)
# disable print log when predict
config.disable_glog_info()
# enable shared memory
config.enable_memory_optim()
# disable feed, fetch OP, needed by zero_copy_run
config.switch_use_feed_fetch_ops(False)
predictor = create_predictor(config)
return predictor, config

def create_inputs(imgs, im_info):
inputs = {}

im_shape = [np.array((e['im_shape'], )).astype('float32') for e in im_info]
scale_factor = [np.array((e['scale_factor'], )).astype('float32') for e in im_info]

origin_scale_factor = np.concatenate(scale_factor, axis=0)

imgs_shape = [[e.shape[1], e.shape[2]] for e in imgs]
max_shape_h = max([e[0] for e in imgs_shape])
max_shape_w = max([e[1] for e in imgs_shape])
padding_imgs = []
padding_imgs_shape = []
padding_imgs_scale = []
for img in imgs:
    im_c, im_h, im_w = img.shape[:]
    padding_im = np.zeros(
        (im_c, max_shape_h, max_shape_w), dtype=np.float32)
    padding_im[:, :im_h, :im_w] = np.array(img, dtype=np.float32)
    padding_imgs.append(padding_im)
    padding_imgs_shape.append(
        np.array([max_shape_h, max_shape_w]).astype('float32'))
    rescale = [float(max_shape_h) / float(im_h), float(max_shape_w) / float(im_w)]
    padding_imgs_scale.append(np.array(rescale).astype('float32'))
inputs['image'] = np.stack(padding_imgs, axis=0)
inputs['im_shape'] = np.stack(padding_imgs_shape, axis=0)
inputs['scale_factor'] = origin_scale_factor
return inputs

class Detector(object):

def __init__(self,
             pred_config,
             model_dir,
             device='CPU',
             run_mode='paddle',
             batch_size=1,
             trt_min_shape=1,
             trt_max_shape=1280,
             trt_opt_shape=640,
             trt_calib_mode=False,
             cpu_threads=1,
             enable_mkldnn=False):
    self.pred_config = pred_config
    self.predictor, self.config = load_predictor(model_dir)
    self.det_times = Timer()
    self.cpu_mem, self.gpu_mem, self.gpu_util = 0, 0, 0
    self.preprocess_ops = self.get_ops()

def get_ops(self):
    preprocess_ops = []
    for op_info in self.pred_config.preprocess_infos:
        new_op_info = op_info.copy()
        op_type = new_op_info.pop('type')
        preprocess_ops.append(eval(op_type)(**new_op_info))
    return preprocess_ops

def predict(self, inputs):
    # preprocess
    input_names = self.predictor.get_input_names()
    for i in range(len(input_names)):
        input_tensor = self.predictor.get_input_handle(input_names[i])
        input_tensor.copy_from_cpu(inputs[input_names[i]])

    np_boxes, np_boxes_num = [], []

    # model_prediction
    self.predictor.run()
    np_boxes.clear()
    np_boxes_num.clear()
    output_names = self.predictor.get_output_names()
    num_outs = int(len(output_names) / 2)

    for out_idx in range(num_outs):
        np_boxes.append(
            self.predictor.get_output_handle(output_names[out_idx])
            .copy_to_cpu())
        np_boxes_num.append(
            self.predictor.get_output_handle(output_names[
                out_idx + num_outs]).copy_to_cpu())

    np_boxes, np_boxes_num = np.array(np_boxes[0]), np.array(np_boxes_num[0])
    return dict(boxes=np_boxes, boxes_num=np_boxes_num)

def process_image(detector, im_path, threshold):
c_result = []
image_id = os.path.basename(im_path).split('.')[0]
im, im_info = preprocess(im_path, detector.preprocess_ops)
inputs = create_inputs([im], [im_info])
det_results = detector.predict(inputs)

im_bboxes_num = det_results['boxes_num'][0]
if im_bboxes_num > 0:
    bbox_results = det_results['boxes'][0:im_bboxes_num, 2:]
    id_results = det_results['boxes'][0:im_bboxes_num, 0]
    score_results = det_results['boxes'][0:im_bboxes_num, 1]

    for idx in range(im_bboxes_num):
        if float(score_results[idx]) >= threshold:
            c_result.append({"image_id": image_id,
                             "type": int(id_results[idx]) + 1,
                             "x": float(bbox_results[idx][0]),
                             "y": float(bbox_results[idx][1]),
                             "width": float(bbox_results[idx][2]) - float(bbox_results[idx][0]),
                             "height": float(bbox_results[idx][3]) - float(bbox_results[idx][1]),
                             "segmentation": []})
return c_result

def predict_image(detector, image_list, result_path, threshold):
c_results = {"result": []}

with ThreadPoolExecutor(max_workers=2) as executor:
    future_to_image = {executor.submit(process_image, detector, image, threshold): image for image in image_list}
    for future in as_completed(future_to_image):
        c_results["result"].extend(future.result())

# 写文件
with open(result_path, 'w') as ft:
    json.dump(c_results, ft)

def main(infer_txt, result_path, det_model_path, threshold):
pred_config = PredictConfig(det_model_path)
detector = Detector(pred_config, det_model_path)

# predict from image
img_list = get_test_images(infer_txt)
predict_image(detector, img_list, result_path, threshold)

if name == 'main':
start_time = time.time()
det_model_path = "/home/aistudio/PaddleSlim/example/auto_compression/detection/output"
#det_model_path = "/home/aistudio/PaddleDetection/output_inference/rtdetr_r101vd_6x_coco"

threshold = 0.5

paddle.enable_static()
infer_txt = sys.argv[1]
result_path = sys.argv[2]

main(infer_txt, result_path, det_model_path, threshold)
print('total time:', time.time() - start_time)
print('FPS:',len(get_test_images(infer_txt))/(time.time() - start_time))

[GreatV]

This is a Paddle framework issue. You can try updating your PaddlePaddle version or submit an issue to the paddlepaddle/paddle repository.

[nyh3174680732]

This is a Paddle framework issue. You can try updating your PaddlePaddle version or submit an issue to the paddlepaddle/paddle repository.
升级到PaddlePaddle 2.6.1最新版本之后又变成这个错误了
/home/aistudio
Model Arch: DETR
--transform op: Resize
--transform op: NormalizeImage
--transform op: Permute
Traceback (most recent call last):
File "/home/aistudio/predict.py", line 221, in
main(infer_txt, result_path, det_model_path, threshold)
File "/home/aistudio/predict.py", line 208, in main
predict_image(detector, img_list, result_path, threshold)
File "/home/aistudio/predict.py", line 196, in predict_image
c_results["result"].extend(future.result())
File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/concurrent/futures/_base.py", line 451, in result
return self.__get_result()
File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/concurrent/futures/_base.py", line 403, in __get_result
raise self._exception
File "/opt/conda/envs/python35-paddle120-env/lib/python3.10/concurrent/futures/thread.py", line 58, in run
result = self.fn(*self.args, **self.kwargs)
File "/home/aistudio/predict.py", line 171, in process_image
det_results = detector.predict(inputs)
File "/home/aistudio/predict.py", line 149, in predict
self.predictor.run()
ValueError: (InvalidArgument) The type of data we are trying to retrieve (float32) does not match the type of data (int8) currently contained in the container.
[Hint: Expected dtype() == phi::CppTypeToDataType::Type(), but received dtype():3 != phi::CppTypeToDataType::Type():10.] (at /paddle/paddle/phi/core/dense_tensor.cc:171)
[operator < fused_fc_elementwise_layernorm > error]