Converting custom tensorflow model to .kmodel

[mhueppe]

Hey,
I want to convert a custom tf model into .kmodel format to run it on the k210 chip.
For testing I wanted to start really simple with 1 input neuron, and one output neuron (here the model trained to return 1 if the input is >0.5 and 0 elsewhen).
However, after saving it to tf.lite format I can't get it to convert using ncc. I currently use the command:

./ncc compile k210_model.tflite k210_model.kmodel --dataset X_train.npy --input-type uint8 --output-type uint8 --inference-type uint8.

However, that does not seem to work. It crashes during quantization. How would I need to format the dataset to work properly?

The code base is the following:

`
import tensorflow as tf
import numpy as np

X_train = np.random.rand(10000, 1) # 1000 samples, 1 feature
y_train = (X_train > 0.5).astype(int) # 1 if >0.5, else 0

model = tf.keras.Sequential([
tf.keras.layers.Input(shape=(1,)),
tf.keras.layers.Dense(1, activation='sigmoid') # Output layer for binary classification
])

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

model.fit(X_train, y_train, epochs=100, batch_size=32, verbose=1)

test_samples = np.array([[0.3], [0.7]])
print(model.predict(test_samples)) # Expect outputs close to [0, 1]

np.save("mnt/data/X_train.npy", X_train)

converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
tflite_model = converter.convert()

with open("model.tflite", "wb") as f:
f.write(tflite_model)

`

The error I receive is the following:

1. Import graph...
2. Optimize Pass 1...
3. Optimize Pass 2...
4. Quantize...
   4.1. Add quantization checkpoints...
   4.2. Get activation ranges...
   Plan buffers...
   Fatal: Invalid dataset, should contain one file at least

However, even when I try to change the path the error maintains. I.e. only giving the directory instead of the file for example.

[curioyang]

@mhueppe
--dataset X_train.npy
You can only give a dir path of images instead of npy.

[mhueppe]

So can you only run comptrr vision models on the k210? Additionally, for images what would the dir structure look like? Is there any way to run a time series model on it?

[curioyang]

images dir such as this. --dataset /path/to/MINIImageNet_20

tree -L 2 MINIImageNet_20
MINIImageNet_20
├── ILSVRC2012_val_00035000.JPEG
├── ILSVRC2012_val_00035001.JPEG
├── ILSVRC2012_val_00035002.JPEG
├── ILSVRC2012_val_00035003.JPEG
├── ILSVRC2012_val_00035004.JPEG
├── ILSVRC2012_val_00035005.JPEG
├── ILSVRC2012_val_00035006.JPEG
├── ILSVRC2012_val_00035007.JPEG
├── ILSVRC2012_val_00035008.JPEG
├── ILSVRC2012_val_00035009.JPEG
├── ILSVRC2012_val_00035010.JPEG
├── ILSVRC2012_val_00035011.JPEG
├── ILSVRC2012_val_00035012.JPEG
├── ILSVRC2012_val_00035013.JPEG
├── ILSVRC2012_val_00035014.JPEG
├── ILSVRC2012_val_00035015.JPEG
├── ILSVRC2012_val_00035016.JPEG
├── ILSVRC2012_val_00035017.JPEG
├── ILSVRC2012_val_00035018.JPEG
└── ILSVRC2012_val_00035019.JPEG

It's difficult to run a time series model on K210.