jschw

angle = [0, math.pi/4, math.pi*2/3, math.pi*4/3]
j=0

for i in angle:
    j=j+1
    x_parabola, y_parabola = createParabola2(centre=[-3+j,-4+j], rotation=i)

jschw

def createParabola(centre, rotation, a=2.0, b=2.0, c=2.0):
    x = np.linspace(0, 7, 100)

    y_parabola = a * x**2 + b * x + c
    x_parabola = 2*x

    if rotation is not None:
        x_parabola, y_parabola = rotateCoordinates(x_parabola, y_parabola, rotation) 
    
    x_parabola = x_parabola + centre[0]

jschw

if pmax > threshold:
    # Max confidence is over threshold
    # Select correct trash type
    if max_index==0 or max_index==3:
        # Papier
        trash_type = 1
        
    elif max_index==2 or max_index==4 or max_index==5:
        # Gelber Sack
        trash_type = 2

jschw

i=0
for val in plist:
  plist[i] = val/img_num_mshot
  i=i+1

jschw

for i in range(img_num_mshot):
     img = sensor.snapshot()
     lcd.display(img)

     # Feed forward
     fmap = kpu.forward(task, img)
     plist_tmp = fmap[:]

     # add values to predicition vector
     i=0

jschw

img = sensor.snapshot()
fmap = kpu.forward(task, img)

plist = fmap[:]
pmax = max(plist)
labelText = ''
unknown = False

if pmax > threshold:
    max_index=plist.index(pmax)

jschw

# Number of images to take in multi shot mode
img_num_mshot = 10

# Image logging
#img_logging = False
img_log_path = '/sd/saved_img'
# Note: Logging is only available in mode 0 and 1

# Path to model file
kmodel_path = '/sd/models/trash/trash.kmodel'

jschw

# Set operation mode for classification script
# 0 -> continuous/test , 1 -> single shot , 2 -> multi shot
op_mode = 0
# Image logging
img_logging = False

# Execute main script
exec(open('/sd/models/trash/trash_main_program.py').read())

jschw

def save_tflite(model, path: str):
    # KPU V4 - nncase >= 0.2.0
    #converter = tf.lite.TFLiteConverter.from_keras_model(model)
    #tflite_model = converter.convert()
    
    # KPU V3 - nncase = 0.1.0rc5
    #model.save(path + 'weights.h5', include_optimizer=False) # <== Save model without callback flag in mode.fit

    converter = tf.compat.v1.lite.TFLiteConverter.from_keras_model_file(path + 'weights.h5')

jschw

# Create callback and save the model with best (=highest) validation accuracy
checkpoint_cb = tf.keras.callbacks.ModelCheckpoint(
                filepath=savepath + 'weights.h5',
                monitor='val_accuracy',
                save_best_only=True
                )

# Create callback for saving training step info to file
csvlogger_cb = CSVLogger(savepath + 'finetune_training_log.csv', append=True, separator=';')