def predict_label_from_image_file(image_file, model: ImageModel):
try:
result = model.predict_from_file(path=image_file)
return result.labels[0]
except Exception as e:
print(f"Problem predicting image from file: {e}")
return None, None
def predict_image_url(url, model: ImageModel, row):
label, confidence = '', ''
try:
result = model.predict_from_url(url=url)
label, confidence = result.labels[0]
except Exception as e:
print(f"Problem predicting image from url: {e}")
return label, confidence, row
import time
from lobe import ImageModel
#from lobe.signature import Signature
from lobe.signature import ImageClassificationSignature
import lobe
from app_settings import AppSettings
modelTier = "tier1"
_app_settings = AppSettings()
sigpath = _app_settings.get_SignaturePath(modelTier)
#sig:Signature = Signature(sigpath)
sig: ImageClassificationSignature = ImageClassificationSignature(sigpath)
#sig:Signature = Signature('./tf_models_lite/tier1/signature.json')
model: ImageModel = ImageModel.load_from_signature(sig)
#test_model: ImageModel = ImageModel.load_from_signature(Signature('./tf_models_lite/sig.json'))
#model_not_a_bee: ImageModel = ImageModel.load()
'''
The TFClassify class (Tensor Flow Classifier), takes a TensorFlow model and allows you
to pass multiple images to it via the addImage() or addImages() methods. It then
returns the predicted classification of the images as a DICT array
{
'image': '<image_path_sent_to_classifier>',
'prediction': 'output_prediction_from_TensorFlow'>
}
'''
class TFClassify:
#!/usr/bin/env python
from lobe import ImageModel
model = ImageModel.load('path/to/exported/model')
# Predict from an image file
result = model.predict_from_file('path/to/file.jpg')
# Predict from an image url
result = model.predict_from_url('http://path/to/file.jpg')
# Predict from Pillow image
from PIL import Image
img = Image.open('path/to/file.jpg')
result = model.predict(img)
# Print top prediction
print("Top prediction:", result.prediction)
# Print all classes
for label, prop in result.labels:
print(f"{label}: {prop*100}%")
def predict_folder(img_dir, model_dir, progress_hook=None):
"""
Run your model on a directory of images. This will also go through any images in existing subdirectories.
Move each image into a subdirectory structure based on the prediction -- the predicted label
becomes the directory name where the image goes.
:param img_dir: the filepath to your directory of images.
:param model_dir: path to the Lobe Tensorflow SavedModel export.
:param progress_hook: an optional function that will be run with progress_hook(currentProgress, totalProgress) when progress updates.
"""
print(f"Predicting {img_dir}")
img_dir = os.path.abspath(img_dir)
# if this a .txt file, don't treat the first row as a header. Otherwise, use the first row for header column names.
if not os.path.isdir(img_dir):
raise ValueError(
f"Please specify a directory to images. Found {img_dir}")
num_items = sum(len(files) for _, _, files in os.walk(img_dir))
print(f"Predicting {num_items} items...")
# load the model
print("Loading model...")
model = ImageModel.load(model_path=model_dir)
print("Model loaded!")
# iterate over the rows and predict the label
curr_progress = 0
no_labels = 0
with tqdm(total=num_items) as pbar:
with ThreadPoolExecutor() as executor:
model_futures = []
# make our prediction jobs
for root, _, files in os.walk(img_dir):
for filename in files:
image_file = os.path.abspath(os.path.join(root, filename))
model_futures.append(
(executor.submit(predict_label_from_image_file,
image_file=image_file,
model=model), image_file))
for future, img_file in model_futures:
label = future.result()
if label == '':
no_labels += 1
filename = os.path.split(img_file)[-1]
name, ext = os.path.splitext(filename)
# move the file
dest_dir = os.path.join(img_dir, label)
os.makedirs(dest_dir, exist_ok=True)
dest_file = os.path.abspath(os.path.join(dest_dir, filename))
# only move if the destination is different than the file
if dest_file != img_file:
try:
# rename the file if there is a conflict
rename_idx = 0
while os.path.exists(dest_file):
new_name = f'{name}_{rename_idx}{ext}'
dest_file = os.path.abspath(
os.path.join(dest_dir, new_name))
rename_idx += 1
shutil.move(img_file, dest_file)
except Exception as e:
print(f"Problem moving file: {e}")
pbar.update(1)
if progress_hook:
curr_progress += 1
progress_hook(curr_progress, num_items)
print(f"Done! Number of images without predicted labels: {no_labels}")
#import Pi GPIO library button class
from gpiozero import LED
from time import sleep
#Import Lobe python library
from lobe import ImageModel
# Define LEDs and GPIO pin numbers
# --> Change GPIO pins as needed
red_led = LED(17)
yellow_led = LED(27)
green_led = LED(22)
# Load Lobe TF model
# --> Change model path
model = ImageModel.load('/path/to/model')
# Run photo through Lobe TF model and get prediction results
# --> Change image path
result = model.predict_from_file('/path/to/image/image.jpg')
# Function that takes in a single string, label, and turns on corresponding LED
# --> Change label1, label2, and label3 to reflect your Lobe model labels.
# Note: Labels are case sensitive!
def ledSelect(label):
print(label)
if label == "label1":
yellow_led.on()
sleep(5)
if label == "label2":
green_led.on()
from time import sleep
from lobe import ImageModel
button = Button(21)
yellow_led = LED(17)
blue_led = LED(27)
green_led = LED(22)
red_led = LED(23)
pink_led = LED(24)
white_led = LED(10)
camera = PiCamera()
model = ImageModel.load('model/')
# Take Photo
def take_photo():
sleep(2)
print("Starting classify process")
camera.start_preview(alpha=200)
camera.rotation = 270
camera.capture('images/predict.png')
camera.stop_preview()
sleep(1)
def ledOff():
yellow_led.off()
blue_led.off()
# Import Pi Camera library
from picamera import PiCamera
from time import sleep
#Import Lobe python library
from lobe import ImageModel
# Create a camera object
camera = PiCamera()
# Load Lobe TF model
# --> Change model path
model = ImageModel.load('/home/pi/model')
if __name__ == '__main__':
# Start the camera preview, make slightly transparent to see any python output
# Note: preview only shows if you have a monitor connected directly to the Pi
camera.start_preview(alpha=200)
# Pi Foundation recommends waiting 2s for light adjustment
sleep(5)
# Optional image rotation for camera
# --> Change or comment out as needed
camera.rotation = 180
#Input image file path here
# --> Change image path as needed
camera.capture('/home/pi/Documents/image.jpg')
#Stop camera
camera.stop_preview()
# Run photo through Lobe TF model and get prediction results
result = model.predict_from_file('/home/pi/Documents/image.jpg')
def main():
global jingle_count
model = ImageModel.load('~/model')
# Check if there is a folder to keep the retraining data, if it there isn't make it
if (not os.path.exists('./retraining_data')):
os.mkdir('./retraining_data')
with picamera.PiCamera(resolution=(224, 224), framerate=30) as camera:
stream = io.BytesIO()
camera.start_preview()
# Camera warm-up time
time.sleep(2)
label = ''
while True:
stream.seek(0)
camera.annotate_text = None
camera.capture(stream, format='jpeg')
camera.annotate_text = label
img = Image.open(stream)
result = model.predict(img)
label = result.prediction
confidence = result.labels[0][1]
camera.annotate_text = label
print(f'\rLabel: {label} | Confidence: {confidence*100: .2f}%',
end='',
flush=True)
# Check if the current label is package and that the label has changed since last tine the code ran
if label == LABEL_CAT:
# Make Servo Jingle Keys
jingle_keys()
elif label == LABEL_MULTI_CAT:
jingle_keys(True)
elif label == LABEL_NOTHING:
jingle_count = 0
time.sleep(0.5)
inputs = get_inputs()
# Check if the joystick is pushed up
if (Input.UP in inputs):
color_fill(GREEN, 0)
# Check if there is a folder to keep the retraining data, if it there isn't make it
if (not os.path.exists(f'./retraining_data/{label}')):
os.mkdir(f'./retraining_data/{label}')
# Remove the text annotation
camera.annotate_text = None
# File name
name = datetime.now()
# Save the current frame
camera.capture(
os.path.join(
f'./retraining_data/{label}',
f'{datetime.now().strftime("%Y-%m-%d_%H:%M:%S")}.jpg'))
color_fill(OFF, 0)
# Check if the joystick is pushed down
elif (Input.DOWN in inputs or Input.BUTTON in inputs):
color_fill(RED, 0)
# Remove the text annotation
camera.annotate_text = None
# Save the current frame to the top level retraining directory
camera.capture(
os.path.join(
f'./retraining_data',
f'{datetime.now().strftime("%Y-%m-%d_%H:%M:%S")}.jpg'))
color_fill(OFF, 0)
def main(check, interval, save_path, model_path):
# Confirm that the provided save path and model paths are valid
try:
path_save = validate_path(save_path)
path_model = validate_path(model_path)
except Exception as err:
print(err)
exit()
# Load Lobe model
model = ImageModel.load(path_model)
with picamera.PiCamera(resolution=(224, 224), framerate=30) as camera:
# Start camera preview
stream = io.BytesIO()
camera.start_preview()
# Camera warm-up time
time.sleep(2)
while True:
# Start stream at the first byte
stream.seek(0)
time_stamp = datetime.now().strftime('%Y-%m-%d_%H:%M:%S')
# Clear the last prediction text
# and capture an image from the camera
camera.annotate_text = None
camera.capture(stream, format='jpeg')
img = Image.open(stream)
# Run inference on the image
result = model.predict(img)
label = result.prediction
confidence = result.labels[0][1]
# Add label text to camera preview
camera.annotate_text = f"{label}\n{confidence}\n{time_stamp}"
# if the image was predicted interesting, save it to the provided path
# and wait the interval set for interesting images
if label == Labels.INTERESTING:
save_filename = path_save.joinpath(f"{time_stamp}.jpg")
img.save(save_filename)
time.sleep(interval)
# if the image was predicted uninteresting, save it to a sub-directory
# (for use to make the interesting/not-interesting model better)
# and wait the interval set for uninteresting images
elif label == Labels.NOT_INTERESTING:
save_filename = path_save.joinpath('uninteresting').joinpath(f"un-{time_stamp}.jpg")
img.save(save_filename)
time.sleep(check)
# if some otherlabel is predicted, there's a problem with the model
# print something out and exit execution
else:
print(f"Unexpected result: {label}")
exit()
def predict_dataset(filepath, model_dir, url_col=None, progress_hook=None):
"""
Given a file with urls to images, predict the given SavedModel on the image and write the label
and confidene back to the file.
:param filepath: path to a valid txt or csv file with image urls to download.
:param model_dir: path to the Lobe Tensorflow SavedModel export.
:param url_col: if this is a csv, the column header name for the urls to download.
:param progress_hook: an optional function that will be run with progress_hook(currentProgress, totalProgress) when progress updates.
"""
print(f"Predicting {filepath}")
filepath = os.path.abspath(filepath)
filename, ext = _name_and_extension(filepath)
# read the file
# if this a .txt file, don't treat the first row as a header. Otherwise, use the first row for header column names.
if ext != '.xlsx':
csv = pd.read_csv(filepath, header=None if ext == '.txt' else 0)
else:
csv = pd.read_excel(filepath, header=0)
if ext in ['.csv', '.xlsx'] and not url_col:
raise ValueError(f"Please specify an image url column for the csv.")
url_col_idx = 0
if url_col:
try:
url_col_idx = list(csv.columns).index(url_col)
except ValueError:
raise ValueError(
f"Image url column {url_col} not found in csv headers {csv.columns}"
)
num_items = len(csv)
print(f"Predicting {num_items} items...")
# load the model
print("Loading model...")
model = ImageModel.load(model_path=model_dir)
print("Model loaded!")
# create our output csv
fname, ext = os.path.splitext(filepath)
out_file = f"{fname}_predictions.csv"
with open(out_file, 'w', encoding="utf-8", newline='') as f:
# our header names from the pandas columns
writer = csv_writer(f)
writer.writerow([
*[str(col) if not pd.isna(col) else '' for col in csv.columns],
'label', 'confidence'
])
# iterate over the rows and predict the label
with tqdm(total=len(csv)) as pbar:
with ThreadPoolExecutor() as executor:
model_futures = []
# make our prediction jobs
for i, row in enumerate(csv.itertuples(index=False)):
url = row[url_col_idx]
model_futures.append(
executor.submit(predict_image_url,
url=url,
model=model,
row=row))
# write the results from the predict (this should go in order of the futures)
for i, future in enumerate(model_futures):
label, confidence, row = future.result()
with open(out_file, 'a', encoding="utf-8", newline='') as f:
writer = csv_writer(f)
writer.writerow([
*[str(col) if not pd.isna(col) else '' for col in row],
label, confidence
])
pbar.update(1)
if progress_hook:
progress_hook(i + 1, len(csv))
#!/usr/bin/python
from time import sleep
from lobe import ImageModel
import subprocess
import tflite_runtime.interpreter
# Learning mostly from https://blog.paperspace.com/tensorflow-lite-raspberry-pi/
# and https://github.com/microsoft/TrashClassifier.
# Took some code from my old project: https://github.com/aHagouel/MauiBot4000/blob/master/src/bot.py
# SHoutout Lobe team for helping debug my execution environment.
model_folder = "/home/pi/Development/hungry-hungry-fishos/utilities/model/"
# Load Lobe.ai TF model. Requires TF Lite & Lobe installation. Please check out readme for more info.
model = ImageModel.load(model_folder)
def take_picture(
file_path='/home/pi/Development/hungry-hungry-fishos/current_state/last_picture'
):
command = "fswebcam -S 2 -r 980x540 --no-banner " + file_path + '.jpg'
process = subprocess.call(command.split(), stdout=subprocess.PIPE)
return file_path + '.jpg'
while True:
photo_path = take_picture()
# Run photo through Lobe TF model
result = model.predict_from_file(photo_path)
print(result)
if (result == "Hungry"):
def on_connect(client, userdata, flags, rc):
print("Connected with result code "+str(rc))
def on_message(client, userdata, msg):
print(msg.topic+" "+str(msg.payload))
client = mqtt.Client()
client.on_connect = on_connect
client.on_message = on_message
# connect to MQTT broker
client.connect("localhost", 1883, 60)
client.loop_start()
model = ImageModel.load(model_dir)
while True:
ret, image_np = cap.read()
image_rgb = cv2.cvtColor(image_np, cv2.COLOR_BGR2RGB)
im_pil = Image.fromarray(image_rgb)
result = model.predict(im_pil)
client.publish("birdcam/detections", json.dumps(result.__dict__), qos=0, retain=False)
print(result.prediction)
#comment the line below if you don't want the webcam preview
cv2.imshow('object detection', image_np)
time.sleep(1)
if cv2.waitKey(25) & 0xFF == ord('q'):
break
cap.release()
white_led = PWMLED(24) # camera state LED
general_ir = Button(23)
recycling_ir = Button(22)
servo = Servo(19)
servo.value = 0 # initate servo to mid point
# count that records row number for google sheets writing
global count
count = 133 # change to current position of last entry
camera = PiCamera()
# Load Lobe TF model
model = ImageModel.load('/home/pi/Lobe')
# run after button press
def get_data(count):
white_led.blink(0.1, 0.1)
sleep(1)
print("Pressed")
white_led.on()
# Start the camera preview
camera.start_preview(alpha=200)
# wait 0.5s for light adjustment
sleep(0.5)
# image rotation for camera
camera.rotation = 90
# new image file path