nanodet keras model fixed to default 80 coco classes

[alexander-sony]

Issue Type

Bug

Source

source

MCT Version

1.10.0

OS Platform and Distribution

Ubuntu 20.04

Python version

3.10.12

Describe the issue

re: tutorials/mct_model_garden/models_keras/nanodet/nanodet_keras_model.py

nanodet_keras_model.py can only handle 80 classes

Can be made more general by replacing the value of 80 with a class parameter
- see def nanodet_box_decoding()

Moreover, in nanodet_plus_head(n, feat_channels=128)
h[idx] = Conv2D(112, 1, name='head.gfl_cls.' + str(idx))(h[idx])
should probably be replaced with
h[idx] = Conv2D(32+num_of_classes, 1, name='head.gfl_cls.' + str(idx))(h[idx])
where num_of_classes corresponds to a new input parameter (default is 80 as in coco)

Expected behaviour

expected to be able to load pth file with a different number of classes than in coco (80)

Code to reproduce the issue

# see https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/keras/ptq/example_keras_nanodet_plus.ipynb
PRETRAINED_WEIGHTS_FILE = <pth file with num of classes != 80>
# Parameters for nanodet-plus-m-1.5x_416
INPUT_RESOLUTION = 416
INPUT_SHAPE = (INPUT_RESOLUTION, INPUT_RESOLUTION, 3)
SCALE_FACTOR = 1.5
BOTTLENECK_RATIO = 0.5
FEATURE_CHANNELS = 128

pretrained_weights = torch.load(PRETRAINED_WEIGHTS_FILE, map_location=torch.device('cpu'))['state_dict']
model = nanodet_plus_m(INPUT_SHAPE, SCALE_FACTOR, BOTTLENECK_RATIO, FEATURE_CHANNELS)
load_state_dict(model, state_dict_torch=pretrained_weights)  # ValueError
...

Log output

(tested with 5 classes instead of 80)
...torch2keras_weights_translation.py", line 15

ValueError: Cannot assign value to variable ' head.gfl_cls.0/kernel:0': Shape mismatch.The variable shape (1, 1, 128, 112), and the assigned value shape (1, 1, 128, 37) are incompatible

[Idan-BenAmi]

Nanodet keras model in MCT Mode Garden supports configurable number of classes #1000