example_liveprediction_detail.py

from keras.models import load_model
import tensorflow as tf
import numpy as np
from vggish_input import waveform_to_examples
import ubicoustics
import pyaudio
from pathlib import Path
import time
import argparse
import wget
import os
from reprint import output
from helpers import Interpolator, ratio_to_db, dbFS, rangemap

# thresholds
PREDICTION_THRES = 0.8 # confidence
DBLEVEL_THRES = -40 # dB

# Variables
FORMAT = pyaudio.paInt16
CHANNELS = 1
RATE = 16000
CHUNK = RATE
MICROPHONES_DESCRIPTION = []
FPS = 60.0
OUTPUT_LINES = 33

###########################
# Model download
###########################
def download_model(url,output):
    return wget.download(url,output)

###########################
# Check Microphone
###########################
print("=====")
print("1 / 2: Checking Microphones... ")
print("=====")

import microphones
desc, mics, indices = microphones.list_microphones()
if (len(mics) == 0):
    print("Error: No microphone found.")
    exit()

#############
# Read Command Line Args
#############
MICROPHONE_INDEX = indices[0]
parser = argparse.ArgumentParser()
parser.add_argument("-m", "--mic", help="Select which microphone / input device to use")
args = parser.parse_args()
try:
    if args.mic:
        MICROPHONE_INDEX = int(args.mic)
        print("User selected mic: %d" % MICROPHONE_INDEX)
    else:
        mic_in = input("Select microphone [%d]: " % MICROPHONE_INDEX).strip()
        if (mic_in!=''):
            MICROPHONE_INDEX = int(mic_in)
except:
    print("Invalid microphone")
    exit()

# Find description that matches the mic index
mic_desc = ""
for k in range(len(indices)):
    i = indices[k]
    if (i==MICROPHONE_INDEX):
        mic_desc = mics[k]
print("Using mic: %s" % mic_desc)

###########################
# Download model, if it doesn't exist
###########################
MODEL_URL = "https://www.dropbox.com/s/cq1d7uqg0l28211/example_model.hdf5?dl=1"
MODEL_PATH = "models/example_model.hdf5"
print("=====")
print("2 / 2: Checking model... ")
print("=====")
model_filename = "models/example_model.hdf5"
ubicoustics_model = Path(model_filename)
if (not ubicoustics_model.is_file()):
    print("Downloading example_model.hdf5 [867MB]: ")
    download_model(MODEL_URL, MODEL_PATH)

##############################
# Load Deep Learning Model
##############################
print("Using deep learning model: %s" % (model_filename))
model = load_model(model_filename)
graph = tf.get_default_graph()
context = ubicoustics.everything

label = dict()
for k in range(len(context)):
    label[k] = context[k]

##############################
# Setup Audio Callback
##############################
output_lines = []*OUTPUT_LINES
audio_rms = 0
candidate = ("-",0.0)

# Prediction Interpolators
interpolators = []
for k in range(31):
    interpolators.append(Interpolator())

# Audio Input Callback
def audio_samples(in_data, frame_count, time_info, status_flags):
    global graph
    global output_lines
    global interpolators
    global audio_rms
    global candidate
    np_wav = np.fromstring(in_data, dtype=np.int16) / 32768.0 # Convert to [-1.0, +1.0]

    # Compute RMS and convert to dB
    rms = np.sqrt(np.mean(np_wav**2))
    db = dbFS(rms)
    interp = interpolators[30]
    interp.animate(interp.end, db, 1.0)

    # Make Predictions
    x = waveform_to_examples(np_wav, RATE)
    predictions = []
    with graph.as_default():
        if x.shape[0] != 0:
            x = x.reshape(len(x), 96, 64, 1)
            pred = model.predict(x)
            predictions.append(pred)

        for prediction in predictions:
            m = np.argmax(prediction[0])
            candidate = (ubicoustics.to_human_labels[label[m]],prediction[0,m])
            num_classes = len(prediction[0])
            for k in range(num_classes):
                interp = interpolators[k]
                prev = interp.end
                interp.animate(prev,prediction[0,k],1.0)
    return (in_data, pyaudio.paContinue)

##############################
# Main Execution
##############################
while(1):
    ##############################
    # Setup Audio
    ##############################
    p = pyaudio.PyAudio()
    stream = p.open(format=FORMAT, channels=CHANNELS, rate=RATE, input=True, frames_per_buffer=CHUNK, stream_callback=audio_samples, input_device_index=MICROPHONE_INDEX)

    ##############################
    # Start Non-Blocking Stream
    ##############################
    os.system('cls' if os.name == 'nt' else 'clear')
    print("# Live Prediction Using Microphone: %s" % (mic_desc))
    stream.start_stream()
    while stream.is_active():
        with output(initial_len=OUTPUT_LINES, interval=0) as output_lines:
            while True:
                time.sleep(1.0/FPS) # 60fps
                for k in range(30):
                    interp = interpolators[k]
                    val = interp.update()
                    bar = ["|"] * int((val*100.0))
                    output_lines[k] = "%20s: %.2f %s" % (ubicoustics.to_human_labels[label[k]], val, "".join(bar))

                # dB Levels
                interp = interpolators[30]
                db = interp.update()
                val = rangemap(db, -50, 0, 0, 100)
                bar = ["|"] * min(100,int((val)))
                output_lines[30] = "%20s: %.1fdB [%s " % ("Audio Level", db, "".join(bar))

                # Display Thresholds
                output_lines[31] = "%20s: confidence = %.2f, db_level = %.1f" % ("Thresholds", PREDICTION_THRES, DBLEVEL_THRES)

                # Final Prediction
                pred = "-"
                event,conf = candidate
                if (conf > PREDICTION_THRES and db > DBLEVEL_THRES):
                    pred = event
                output_lines[32] = "%20s: %s" % ("Prediction", pred.upper())