Unable to export and run on android the exported custom class trained model using mask_rcnn_fbnetv3a_C4.yaml

[VaibhavAsare]

Hello Everyone,
I have trained a custom class model using mask_rcnn_fbnetv3a_C4.yaml and exported the model demo begineer notebook in the repoistry but the model is not working in android.
How to solve this issue ?? and
how to check the exported model is working fine in our working environment ???

[wat3rBro]

could you provide more context?

[VaibhavAsare]

I have trained a custom class model using mask_rcnn_fbnetv3a_C4.yaml and exported the model using the following code

import copy
from detectron2.data import build_detection_test_loader
from d2go.export.exporter import convert_and_export_predictor

from d2go.export.d2_meta_arch import patch_d2_meta_arch

from d2go.runner import GeneralizedRCNNRunner
from d2go.runner import Detectron2GoRunner
import os
import logging

disable all the warnings

previous_level = logging.root.manager.disable
logging.disable(logging.INFO)

patch_d2_meta_arch()

torch.backends.quantized.engine = 'qnnpack'
cfg.merge_from_file("/home/ubuntu/d2go/codes/trail11/config1.yml")
cfg.QUANTIZATION.BACKEND = 'qnnpack'
runner = GeneralizedRCNNRunner()
model = runner.build_model(cfg, eval_only=True)
model.cpu()

datasets = cfg.DATASETS.TRAIN[0]

cfg.QUANTIZATION.BACKEND = 'qnnpack'
data_loader = runner.build_detection_test_loader(cfg, datasets)

predictor_path = convert_and_export_predictor(
copy.deepcopy(cfg),
copy.deepcopy(model),
"torchscript_int8@tracing",
'./',
data_loader
)

recover the logging level

logging.disable(previous_level)

f = open('config2.yml', 'w')
f.write(cfg.dump())
f.close()

#for android
from typing import List, Dict
import torch

class Wrapper(torch.nn.Module):
def init(self, model):
super().init()
self.model = model
coco_idx_list = [1, 2, 3, 4, 5, 6]

    self.coco_idx = torch.tensor(coco_idx_list)

def forward(self, inputs: List[torch.Tensor]):
    x = inputs[0].unsqueeze(0) * 255
    scale = 320.0 / min(x.shape[-2], x.shape[-1])
    x = torch.nn.functional.interpolate(x, scale_factor=scale, mode="bilinear", align_corners=True, recompute_scale_factor=True)
    out = self.model(x[0])
    res : Dict[str, torch.Tensor] = {}
    res["boxes"] = out[0] / scale
    res["labels"] = torch.index_select(self.coco_idx, 0, out[1])
    res["scores"] = out[2]
    return inputs, [res]

orig_model = torch.jit.load(os.path.join(predictor_path, "model.jit"))
wrapped_model = Wrapper(orig_model)
scripted_model = torch.jit.script(wrapped_model)
scripted_model.save("d2go.pt")

metrics = runner.do_test(cfg, model)

print(metrics)

These two model which I have exported is working fine while inferencing in python notebook but the .pt model for android is not working how to resolve this issue???

[dipu0]

did you find any way out to run custom model in android??

[labs10]

You've to optimize the model and use .ptl format to use it in android.
Add this to your code this will work.

from torch.utils.mobile_optimizer import optimize_for_mobile
optimized_model = optimize_for_mobile(orig_model)
optimized_model._save_for_lite_interpreter(predictor_path+"d2go.ptl")
#Pt model
orig_model.save(predictor_path+"d2go.pt")

[VaibhavAsare]

Hello @labs10 , I have implmented the same solution provided by you, but its giving me the error
I have shared the error message in the picture

[labs12]

@VaibhavAsare d2go provides mobile optimized tracing, use
predictor_path = convert_and_export_predictor( copy.deepcopy(cfg), copy.deepcopy(model), "torchscript_mobile_int8@tracing", './', data_loader ) .
This will provide weights on .ptl format you can use this on android.

[VaibhavAsare]

@labs10 This also not working in android. Is This issue with instance segmentation using mask_rcnn_fbnetv3a_C4.yaml only???

[nkhlS141]

I have trained a custom class model using mask_rcnn_fbnetv3a_C4.yaml and exported the model using the following code

import copy from detectron2.data import build_detection_test_loader from d2go.export.exporter import convert_and_export_predictor

from d2go.export.d2_meta_arch import patch_d2_meta_arch

from d2go.runner import GeneralizedRCNNRunner from d2go.runner import Detectron2GoRunner import os import logging

disable all the warnings

previous_level = logging.root.manager.disable logging.disable(logging.INFO)

patch_d2_meta_arch()

torch.backends.quantized.engine = 'qnnpack' cfg.merge_from_file("/home/ubuntu/d2go/codes/trail11/config1.yml") cfg.QUANTIZATION.BACKEND = 'qnnpack' runner = GeneralizedRCNNRunner() model = runner.build_model(cfg, eval_only=True) model.cpu()

datasets = cfg.DATASETS.TRAIN[0]

cfg.QUANTIZATION.BACKEND = 'qnnpack' data_loader = runner.build_detection_test_loader(cfg, datasets)

predictor_path = convert_and_export_predictor( copy.deepcopy(cfg), copy.deepcopy(model), "torchscript_int8@tracing", './', data_loader )

recover the logging level

logging.disable(previous_level)

f = open('config2.yml', 'w') f.write(cfg.dump()) f.close()

#for android from typing import List, Dict import torch

class Wrapper(torch.nn.Module): def init(self, model): super().init() self.model = model coco_idx_list = [1, 2, 3, 4, 5, 6]

    self.coco_idx = torch.tensor(coco_idx_list)

def forward(self, inputs: List[torch.Tensor]):
    x = inputs[0].unsqueeze(0) * 255
    scale = 320.0 / min(x.shape[-2], x.shape[-1])
    x = torch.nn.functional.interpolate(x, scale_factor=scale, mode="bilinear", align_corners=True, recompute_scale_factor=True)
    out = self.model(x[0])
    res : Dict[str, torch.Tensor] = {}
    res["boxes"] = out[0] / scale
    res["labels"] = torch.index_select(self.coco_idx, 0, out[1])
    res["scores"] = out[2]
    return inputs, [res]

orig_model = torch.jit.load(os.path.join(predictor_path, "model.jit")) wrapped_model = Wrapper(orig_model) scripted_model = torch.jit.script(wrapped_model) scripted_model.save("d2go.pt")

metrics = runner.do_test(cfg, model)

print(metrics)

These two model which I have exported is working fine while inferencing in python notebook but the .pt model for android is not working how to resolve this issue???

@VaibhavAsare I think the forward function must be changed since it is a mask-rcnn model. Ideally a fastrcnn model has three outputs and maskrcnn unlike fastrcnn has an extra output i.e masks
So the ideal change in forward func must be
res["boxes"] = out[0] / scale
res["labels"] = torch.index_select(self.coco_idx, 0, out[1])
res["masks"] = out[2]
res["scores"] = out[3]

No?

[nkhlS141]

@VaibhavAsare please let me know if this works.

[VaibhavAsare]

@nkhlS141 No it is working.
have you find any other way to fix this issue?