Deployment on GroveAIV2 failure at the invoke stage

[EPG6]

I successfully trained a custom model using the sscma library and exported it into the vela.tflite format but once I upload it to groveAIV2 it fails when the invoke function is called (The error is 'MODEL' does not exist). Here is my config file timm_classify.py:

from mmengine.config import read_base

with read_base():
    from .._base_.default_runtime import *
    from .._base_.schedules.schedule_1x import *

from sscma.models import CrossEntropyLoss, Mixup, CutMix, ImageClassifier, MobileNetv2, GlobalAveragePooling, LinearClsHead
from sscma.infer import CustomInferencer

from sscma.datasets.transforms import (
    Resize,
    PackInputs,
    LoadImageFromFile,
    Pad,
    HSVRandomAug,
    RandomFlip,
    Mosaic,
)

from sscma.datasets import ImageNet, ClsDataPreprocessor
from sscma.engine import DetVisualizationHook
from sscma.visualization import UniversalVisualizer
from mmengine.optim import LinearLR, MultiStepLR
from sscma.evaluation import Accuracy

from torch.nn import ReLU6, BCEWithLogitsLoss, ReLU
from torch.optim import Adam, SGD


dataset_type = ImageNet
data_root='dataset/simulation'
train_data = 'train'
val_data = 'val'
metainfo = {'classes' : ['left', 'right', 'forward']}

height = 640
width = 640
imgsz = (width, height)

downsample_factor = (8,)

# TRAIN
batch = 12
workers = 2
persistent_workers = True
widen_factor=0.35

val_batch = 16
val_workers = 2

lr = 0.02
epochs = 50

weight_decay = 0.0005
momentum = 0.95

default_hooks = dict(visualization=dict(type=DetVisualizationHook, score_thr=0.8))
visualizer = dict(type=UniversalVisualizer)

# model settings
data_preprocessor = dict(
    type=ClsDataPreprocessor,
    num_classes=3,
    # RGB format normalization parameters
    mean=[123.675, 116.28, 103.53],
    std=[58.395, 57.12, 57.375],
    # convert image from BGR to RGB
    to_rgb=True,
)


model = dict(
    data_preprocessor=data_preprocessor,
    type=ImageClassifier,
    backbone=dict(type=MobileNetv2, gray_input=False, widen_factor=widen_factor, out_indices=(2,), rep=True),
    neck=dict(type=GlobalAveragePooling),
    head=dict(
        type=LinearClsHead,
        in_channels=32,
        loss=dict(type=CrossEntropyLoss, loss_weight=1.0),
        num_classes=3,
    ),
    train_cfg=dict(
        augments=[dict(type=Mixup, alpha=0.8), dict(type=CutMix, alpha=1.0)]
    ),
)

deploy = dict(
    type=CustomInferencer,
)

imdecode_backend = "cv2"

pre_transform = [
    dict(
        type=LoadImageFromFile,
        imdecode_backend=imdecode_backend
    )
]

train_pipeline = [
    *pre_transform,
     dict(type=RandomFlip, prob=0.5),
    dict(
        type=PackInputs,
    ),
]

test_pipeline = [
    *pre_transform,
    dict(
        type=PackInputs,
    ),
]

train_dataloader = dict(
    batch_size=batch,
    num_workers=workers,
    persistent_workers=persistent_workers,
    drop_last=False,
    dataset=dict(
        type=dataset_type,
        data_root=data_root,
        split=train_data,
        pipeline=train_pipeline,
        metainfo=metainfo,
    ),
)
val_dataloader = dict(
    batch_size=val_batch,
    num_workers=val_workers,
    persistent_workers=persistent_workers,
    drop_last=False,
    dataset=dict(
        type=dataset_type,
        data_root=data_root,
        split=val_data,
        pipeline=test_pipeline,
        metainfo=metainfo,
    ),
)
test_dataloader = val_dataloader

find_unused_parameters = True

optim_wrapper = dict(
    optimizer=dict(
        # type=Adam, lr=lr, betas=momentum, weight_decay=weight_decay, eps=1e-7
        type=SGD,
        lr=lr,
        momentum=momentum,
        weight_decay=weight_decay,
    ),
)

# evaluator
val_evaluator = dict(type=Accuracy)
test_evaluator = val_evaluator

train_cfg = dict(by_epoch=True, max_epochs=epochs)

# learning policy
param_scheduler = [
    dict(type=LinearLR, begin=0, end=30, start_factor=0.001, by_epoch=False),  # warm-up
    dict(
        type=MultiStepLR,
        begin=1,
        end=50,
        milestones=[15, 30, 45],
        gamma=0.3,
        by_epoch=True,
    ),
]

[baorepo]

We haven't seen the steps and logs for your model's quantization, conversion and deployment. If you could provide them, we might be able to assist you.

[EPG6]

I have attached the relevant logs to this comment.

20250717_141109.txt
20250717_134501.txt
20250717_112624.txt

I used the command python3 tools/train.py configs/cls/timm_classify.py to train the model. Then I exported it using the command python3 tools/export.py configs/cls/timm_classify.py work_dirs/timm_classify/epoch_50.pth --format vela --image_path $(pwd)/dataset/simulation/val then I uploaded the resulting model_int8_vela.tflite into groveAiV2 using the Sensecraft website. The model was uploaded successfully but it was not working because the invoke method failed. Which steps are my missing?

[iChizer0]

Hi @EPG6, I need to know the structure of the converted vela model to locate the actual problem, if it is not convenient to share the weight directly, you can view the model structure on Netron and export the structure as a picture to send us. In addition, we note that in your configs, 640 image sizes is used, but the device does not have enough memory to support such large input size, we recommend that you reconfirm the input image size of the model (we generally use 192 or 224), thank you.

[EPG6]

Does this file suffice?

[iChizer0]

This model graph looks fine, I suspect it is a problem with the Vela converter, you can try using the previously verified version of the Vela converter:

python3 -m pip install -U ethos-u-vela==3.12.0

If the problem persists after the replacement, you may need to share the graph information of converted vela model, and we will further investigate the problem, thank you.

[EPG6]

It still does not work. The vela converter functions properly when I train a fomo model but not when I used my custom config file (code in my first comment above) to train a model for a classification task. When I run the command python3 tools/test.py configs/cls/timm_classify.py work_dirs/timm_classify/epoch_50_int8_vela.tflite it returns:

Traceback (most recent call last):
  File "/home/epgk2023/ModelAssistant/tools/test.py", line 153, in <module>
    main()
  File "/home/epgk2023/ModelAssistant/sscma/utils/lazy_import.py", line 69, in wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/tools/test.py", line 111, in main
    model = MODELS.build(cfg.deploy)
            ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/mmengine/registry/registry.py", line 570, in build
    return self.build_func(cfg, *args, **kwargs, registry=self)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/mmengine/registry/build_functions.py", line 232, in build_model_from_cfg
    return build_from_cfg(cfg, registry, default_args)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/mmengine/registry/build_functions.py", line 121, in build_from_cfg
    obj = obj_cls(**args)  # type: ignore
          ^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/mmengine/infer/infer.py", line 169, in __init__
    raise ValueError(
ValueError: If model is None, the weights must be specified since the config needs to be loaded from the weights

In addition to that these are the output csv files from the vela converter:

epoch_50_int8_summary_My_Sys_Cfg.csv

epoch_50_int8_per-layer.csv

[baorepo]

@EPG6 python3 tools/test.py Only TFLITE can be tested, not VALA TFLITE. If the Vela tool is functioning properly, the aforementioned errors will not occur through the normal conversion process. Could you share your original TFLite file instead of the Vala TFLite one?

[EPG6]

This is the structure of my TFLITE file

[baorepo]

From this structure, no problem can be identified. @EPG6

[EPG6]

The vela converter fails though. It fails to properly convert some of the layers from the model. How can I share my model files here?

When I run python3 tools/test.py on my tflite file I get.

python3 tools/test.py     configs/cls/timm_classify.py     work_dirs/timm_classify/epoch_50_int8.tflite
2025-07-23 00:02:08.944785: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
E0000 00:00:1753254128.963816 1298995 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
E0000 00:00:1753254128.969619 1298995 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
W0000 00:00:1753254128.984138 1298995 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
W0000 00:00:1753254128.984157 1298995 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
W0000 00:00:1753254128.984159 1298995 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
W0000 00:00:1753254128.984162 1298995 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
2025-07-23 00:02:08.988440: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
Traceback (most recent call last):
  File "/home/epgk2023/ModelAssistant/tools/test.py", line 153, in <module>
    main()
  File "/home/epgk2023/ModelAssistant/sscma/utils/lazy_import.py", line 69, in wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/tools/test.py", line 111, in main
    model = MODELS.build(cfg.deploy)
            ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/mmengine/registry/registry.py", line 570, in build
    return self.build_func(cfg, *args, **kwargs, registry=self)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/mmengine/registry/build_functions.py", line 232, in build_model_from_cfg
    return build_from_cfg(cfg, registry, default_args)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/mmengine/registry/build_functions.py", line 121, in build_from_cfg
    obj = obj_cls(**args)  # type: ignore
          ^^^^^^^^^^^^^^^
  File "/home/epgk2023/ModelAssistant/mmengine/infer/infer.py", line 169, in __init__
    raise ValueError(
ValueError: If model is None, the weights must be specified since the config needs to be loaded from the weights

[baorepo]

@EPG6

[EPG6]

epoch_50_int8.zip

It does not support tflite format so I send it as a zip file

[iChizer0]

Hi @EPG6, based on the file you provided, I can confirm that this is not a Vela converter issue, it's because the exported Int8 model was not successfully quantized, and there is nothing Vela can do about it. We will check the quantization export script, if it's convenient, please share the following logs of how you ran the quantization export script, thank you.

[iChizer0]

I have just verified and fixed an issue with the export script. The problem was caused by a logic error in the script's generation of the representative dataset. Specifically, when there were no images in the top-level directory of the specified image_path, the model could not be quantized correctly.

You can save this patch:

diff --git a/requirements.txt b/requirements.txt
index d5348b5..d80c287 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -23,3 +23,4 @@ scikit-learn
 timm
 psutil
 regex
+matplotlib
diff --git a/tools/export.py b/tools/export.py
index 3b9274f..471e8dd 100644
--- a/tools/export.py
+++ b/tools/export.py
@@ -93,18 +93,46 @@ def parse_args():
     return args
 
 
+def find_and_sample_images(folder_path, limit=10000, sample_size=100):
+    import random
+
+    image_extensions = {'.jpg', '.jpeg', '.png', '.gif', '.bmp', '.tiff', '.webp'}
+    image_files = []
+
+    if not os.path.isdir(folder_path):
+        raise ValueError(f"The provided image path '{folder_path}' is not a valid directory.")
+
+    for root, _, files in os.walk(folder_path):
+        for file in files:
+            if os.path.splitext(file)[1].lower() in image_extensions:
+                image_files.append(os.path.join(root, file))
+                if len(image_files) >= limit:
+                    break
+        if len(image_files) >= limit:
+            break
+
+    found = len(image_files)
+    if found < sample_size:
+        if found == 0:
+            raise ValueError(f"No images found in the directory '{folder_path}'.")
+        print(f"Warning: Found only {found} images, which is less than the requested sample size of {sample_size}.")
+        sample_size = found
+
+    random_sample = random.sample(image_files, sample_size)
+
+    return random_sample
+
+
 def generate_input(images_path, img_shape):
     import cv2
 
     res = []
-    datasets = [
-        osp.join(images_path, i) for i in os.listdir(images_path) if i.endswith(".jpg")
-    ]
-    for ps in datasets[:100]:
-        img = cv2.imread(ps)
+    datasets = find_and_sample_images(images_path, limit=10000, sample_size=100)
+    for img_path in datasets[:100]:
+        img = cv2.imread(img_path)
         img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255
         img = cv2.resize(img, (img_shape[0], img_shape[1]))
-        res.append(img)
+        res.append(img.astype(np.float32))
     return np.asarray(res)
 
 
@@ -232,6 +260,7 @@ def main():
         runner.test_evaluator.metrics.append(DumpResults(out_file_path=args.out))
 
 
+@lazy_import("ai_edge_litert", install_only=True)
 @lazy_import("onnx2tf", install_only=True)
 @lazy_import("tf-keras", install_only=True)
 @lazy_import("onnx-graphsurgeon", install_only=True)
@@ -360,11 +389,7 @@ def export_tflite(onnx_path: str, img_shape, img_path):
     converter = tf.lite.TFLiteConverter.from_saved_model(osp.dirname(onnx_path))
 
     def representative_dataset():
-        datasets = [
-            osp.join(img_path, i)
-            for i in os.listdir(img_path)
-            if i.lower().endswith((".jpg", ".jpeg", ".png"))
-        ]
+        datasets = find_and_sample_images(img_path, limit=10000, sample_size=300)
         for ps in tqdm(datasets[:300]):
             img = cv2.imread(ps)
             img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255

And apply:

git apply <path to the patch>

Then, try exporting again:

python3 tools/export.py <your configs> <your trained weights> --format vela --image_path dataset/simulation