tflite: grayscale shape using pycoral with PIL

[peepo]

for: image = Image.open(image_file).convert('RGB').resize(size, Image.ANTIALIAS) ValueError: could not broadcast input array from shape (64,64) into shape (64,64,1) else image = Image.open(image_file).resize(size, Image.ANTIALIAS) ValueError: could not broadcast input array from shape (64,64,3) into shape (64,64,1)

I'd like 64,64,1

using only: from pycoral.utils.edgetpu import make_interpreter from pycoral.adapters import common import PIL from PIL import Image

Full error:

Traceback (most recent call last): File "edge.py", line 16, in common.set_input(interpreter, image) File "/usr/lib/python3/dist-packages/pycoral/adapters/common.py", line 75, in set_input input_tensor(interpreter)[:, :] = data ValueError: could not broadcast input array from shape (64,64) into shape (64,64,1)

Apologies if this is working as intended, I did not find a way when creating .tflite with numpy to create 3D array such as 1,64,64 as that also produces an error conv2d requires 4D minimum.

thanks!

I am planning to use monochrome sensor, and would prefer not to convert to RGB.

[peepo]

I tried tf.io.decode_image & tf.io.decode_bmp, ie : img = tf.io.read_file('apple64.png') img = tf.io.decode_image(img, channels=0) tf.expand_dims(img, -1).shape.as_list() model_1 = Sequential()

without success: ValueError: could not broadcast input array from shape (64,64) into shape (64,64,1) https://discuss.tensorflow.org/t/loading-16-grayscale-png-in-tf-lite/5461

[peepo]

Clarification: img = img[:, :, np.newaxis] or img = np.expand_dims(img, axis=(0, -1)) are also similar, the tensorflow works fine, it's when one tries to use the .tflite file created that this error arises. I did not find a reduced testcase or other example.