Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Export semantic model fails

See original GitHub issue

Hello.

Full reproducible code.

import copy
from detectron2 import model_zoo
from detectron2.config import CfgNode
from detectron2.data import build_detection_test_loader
from d2go.export.api import convert_and_export_predictor
from d2go.utils.testing.data_loader_helper import create_fake_detection_data_loader
from d2go.export.d2_meta_arch import patch_d2_meta_arch
from d2go.runner import create_runner


configs = ["COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml","Misc/semantic_R_50_FPN_1x.yaml"]
use_config = 0

cfg = model_zoo.get_config(configs[use_config])
cfg.set_new_allowed(True)
cfg.RCNN_PREPARE_FOR_EXPORT = "default_rcnn_prepare_for_export"
cfg.QUANTIZATION = CfgNode()
cfg.QUANTIZATION.SILICON_QAT = CfgNode({"ENABLED": True})
cfg.QUANTIZATION.QAT = CfgNode({"ENABLED": True})
cfg.MODEL_EMA = CfgNode({"ENABLED": True})
cfg.MODEL.FROZEN_LAYER_REG_EXP = []
cfg.QUANTIZATION.BACKEND = "qnnpack"

runner = create_runner("d2go.runner.GeneralizedRCNNRunner")
if use_config == 0:
    pytorch_model = model_zoo.get(configs[use_config], trained=True) # The trained=True is just to verify that the model is working
else:
    pytorch_model = model_zoo.get(configs[use_config])
    
pytorch_model.cpu()
pytorch_model.eval()

patch_d2_meta_arch()
with create_fake_detection_data_loader(224, 320, is_train=False) as data_loader:
    predictor_path = convert_and_export_predictor(
            cfg,
            copy.deepcopy(pytorch_model),
            "torchscript@tracing",
           './',
            data_loader,
        )

### Uncomment to test that model works. ###

#from d2go.utils.demo_predictor import DemoPredictor
#import numpy as np
#import matplotlib.pyplot as plt
#from PIL import Image
#from detectron2.utils.visualizer import Visualizer
#from detectron2.data import MetadataCatalog
#from mobile_cv.predictor.api import create_predictor

#im = np.uint8(Image.open("./000000439715.jpg")) # Some image, for instance one from coco.

#for model in [pytorch_model, "./torchscript@tracing"]:
#    if isinstance(model, str): # We load the torchscript model
#        model = create_predictor(model)

#    predictor = DemoPredictor(model)
#    outputs = predictor(im)

#    if cfg.MODEL.META_ARCHITECTURE =="SemanticSegmentor":
        # We expect to get nonsens since the model isnt trained.
        # However the only thing that matters is that the model 
        # and the torchscript model are agreeing
#        result = np.argmax(outputs["sem_seg"].to("cpu").numpy(),axis=0)
#        plt.imshow(result)
#        plt.show()
#    else:
#        v = Visualizer(im[:, :, ::-1], MetadataCatalog.get("coco_2017_train"))
#        out = v.draw_instance_predictions(outputs["instances"].to("cpu"))
#        plt.imshow(out.get_image()[:, :, ::-1])
#        plt.show()

You can uncomment the last part to check that both of the mask_rcnn models works.

Changing use_config = 0 to use_config = 1 does the same thing, but for a semantic model in stead of an instance model. Doing this causes an AssertitionError:

File “/home/peter/anaconda3/lib/python3.8/site-packages/d2go/export/api.py”, line 122, in convert_and_export_predictor return export_predictor(cfg, pytorch_model, predictor_type, output_dir, data_loader) File “/home/peter/anaconda3/lib/python3.8/site-packages/d2go/export/api.py”, line 151, in export_predictor return default_export_predictor( File “/home/peter/anaconda3/lib/python3.8/site-packages/d2go/export/api.py”, line 197, in default_export_predictor assert hasattr(pytorch_model, “prepare_for_export”), pytorch_model AssertionError: SemanticSegmentor

It would be very nice to be able to use d2go to export semantic models as well.