akadmin/openwakeword
1
0
Fork 0
mirror of https://github.com/dscripka/openWakeWord.git synced 2026-08-27 18:17:20 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

Updated local demo, changes how pre-trained models are loaded

Browse source
This commit is contained in:
dscripka 2022-11-21 11:56:43 -05:00
parent 802901c5f8
commit 17586c3d8f
5 changed files with 60 additions and 66 deletions
Download patch file Download diff file Expand all files Collapse all files

27
examples/detect_from_microphone.py
View file

@ -27,11 +27,8 @@ mic_stream = sounddevice.InputStream(
dtype = np.int16,
)
# Load openwakeword model(s)
model_name = "alexa_v5"
model = Model(
wakeword_model_paths=[os.path.join("../", "openwakeword", "resources", "models", model_name + ".onnx")],
)
# Load pre-trained openwakeword models
owwModel = Model()
# Run capture loop, checking for hotwords
if __name__ == "__main__":
@ -39,20 +36,26 @@ if __name__ == "__main__":
mic_stream.start()
# Create a prediction buffer
prediction_buffer = [0]*30
while True:
# Get audio
audio, overflowed = mic_stream.read(1280)
audio = audio.squeeze()
# Feed to openWakeWord model
prediction = model.predict(audio)
prediction_buffer = prediction_buffer[1:] + [round(prediction[model_name], 2)]
# Plot predictions in graph
prediction = owwModel.predict(audio)
# Get predictions from prediction buffers and plot
plt.cld()
plt.clt()
plt.plot(prediction_buffer)
for mdl in owwModel.prediction_buffer.keys():
# Plot scores in graph
scores = list(owwModel.prediction_buffer[mdl])
plt.plot(scores)
# Plot text showing name of model with scores >= 0.5 (default threshold)
if max(scores) >= 0.5:
plt.text(mdl, 15, 0.9, alignment="center", color = "blue", style="bold")
plt.ylim(0,1)
plt.show()
plt.sleep(0.005)
plt.sleep(0.005)

17
openwakeword/__init__.py
View file

@ -1,3 +1,20 @@
import os
from openwakeword.model import Model
__all__ = ['Model', ]
models = {
"alexa": {
"model_path": os.path.join(os.path.dirname(os.path.abspath(__file__)), "resources/models/alexa_v5.onnx")
},
"hey_mycroft": {
"model_path": os.path.join(os.path.dirname(os.path.abspath(__file__)), "resources/models/hey_mycroft_v1.onnx")
},
"timer": {
"model_path": os.path.join(os.path.dirname(os.path.abspath(__file__)), "resources/models/timer_v1.onnx")
}
}
def get_pretrained_model_paths():
return [models[i]["model_path"] for i in models.keys()]

7
openwakeword/data.py
View file

@ -483,6 +483,7 @@ class mmap_batch_generator:
return np.vstack(X), np.array(y)
# Function to remove empty rows from the end of a mmap array
def trim_mmap(mmap_path):
"""
@ -502,11 +503,11 @@ def trim_mmap(mmap_path):
i -= 1
N_new = mmap_file1.shape[0] + i + 1
# Create new mmap_file and copy over data in batches
output_file2 = mmap_path.strip(".npy") + "2.npy"
mmap_file2 = open_memmap(output_file2, mode='w+', dtype=np.float32,
shape=(N_new, mmap_file1.shape[1], mmap_file1.shape[2]))
shape=(N_new, mmap_file1.shape[1], mmap_file1.shape[2]))
for i in tqdm(range(0, mmap_file1.shape[0], 1024), total=mmap_file1.shape[0]//1024):
if i + 1024 > N_new:
@ -520,4 +521,4 @@ def trim_mmap(mmap_path):
os.remove(mmap_path)
# Rename new mmap file to match original
os.rename(output_file2, mmap_path)
os.rename(output_file2, mmap_path)

24
openwakeword/model.py
View file

@ -15,9 +15,9 @@
# Imports
import numpy as np
import onnxruntime as ort
import openwakeword
from openwakeword.utils import AudioFeatures
import statistics
import wave
import os
import json
@ -34,12 +34,13 @@ class Model():
The main model class for openWakeWord. Creates a model object with the shared audio pre-processer
and for arbitrarily many custom wake word/wake phrase models.
"""
def __init__(self, wakeword_model_paths: List[str], **kwargs):
def __init__(self, wakeword_model_paths: List[str] = [], **kwargs):
"""
Initialize the openWakeWord model object.
Args:
wakeword_model_paths (List[str]): A list of paths of ONNX models to load into the openWakeWord model object
wakeword_model_paths (List[str]): A list of paths of ONNX models to load into the openWakeWord model object.
If not provided, will load all of the pre-trained models.
"""
# Initialize the ONNX models and store them
@ -47,14 +48,20 @@ class Model():
sessionOptions.inter_op_num_threads = 1
sessionOptions.intra_op_num_threads = 1
# Get model paths for pre-trained models if user doesn't provide models to load
if wakeword_model_paths == []:
wakeword_model_paths = openwakeword.get_pretrained_model_paths()
wakeword_model_names = list(openwakeword.models.keys())
else:
wakeword_model_names = [os.path.basename(i.strip(".onnx")) for i in wakeword_model_paths]
# Create attributes to store models and metadat
self.models = {}
self.model_inputs = {}
self.model_outputs = {}
self.class_mapping = {}
self.model_input_names = {}
for mdl_path in wakeword_model_paths:
mdl_name = mdl_path.split(os.path.sep)[-1].strip(".onnx")
for mdl_path, mdl_name in zip(wakeword_model_paths, wakeword_model_names):
self.models[mdl_name] = ort.InferenceSession(mdl_path, sess_options=sessionOptions,
providers=["CPUExecutionProvider"])
self.model_inputs[mdl_name] = self.models[mdl_name].get_inputs()[0].shape[1]
@ -96,10 +103,11 @@ class Model():
x (Union[ndarray, List]): The input audio data to predict on with the models. Must be 1280
samples of 16khz, 16-bit audio data.
patience (dict): How many consecutive frames (of 1280 samples or 80 ms) above the threshold that must
be observed before the current frame will be returned as non-zero.
be observed before the current frame will be returned as non-zero.
Must be provided as an a dictionary where the keys are the
model names and the values are the number of frames. Can reduce false-positive detections at the
cost of a lower true-positive rate. By default, this behavior is disabled.
model names and the values are the number of frames. Can reduce false-positive
detections at the cost of a lower true-positive rate.
By default, this behavior is disabled.
threshold (dict): The threshold values to use when the `patience` behavior is enabled.
Must be provided as an a dictionary where the keys are the
model names and the values are the thresholds.

51
tests/test_models.py
View file

@ -33,30 +33,24 @@ import numpy as np
from pathlib import Path
import collections
# Define models and corresponding files for testing
test_dict = {
"hey_mycroft_v1": ["hey_mycroft_v1_test.wav"],
"alexa_v5": ["alexa_v5_test.wav"]
}
# Tests
class TestModels:
def test_models(self):
# Load model
models = [str(i) for i in Path(
os.path.join("openwakeword", "resources", "models")
).glob("**/*.onnx")
if "embedding" not in str(i) and "melspec" not in str(i)]
owwModel = openwakeword.Model(
wakeword_model_paths=models,
)
owwModel = openwakeword.Model()
# Get clips for each model (assumes that test clips will have the model name in the filename)
test_dict = {}
for mdl_name in owwModel.models.keys():
all_clips = [str(i) for i in Path(os.path.join("tests", "data")).glob("*.wav")]
test_dict[mdl_name] = [i for i in all_clips if mdl_name in i]
# Predict
for model, clips in test_dict.items():
for clip in clips:
# Get predictions for reach frame in the clip
predictions = owwModel.predict_clip(os.path.join("tests", "data", clip))
predictions = owwModel.predict_clip(clip)
owwModel.reset() # reset after each clip to ensure independent results
# Make predictions dictionary flatter
@ -70,35 +64,6 @@ class TestModels:
else:
assert max(predictions_flat[key]) < 0.5
def test_models_with_median_smooth(self):
# Load models
models = [str(i) for i in Path(
os.path.join("openwakeword", "resources", "models")
).glob("**/*.onnx")
if "embedding" not in str(i) and "melspec" not in str(i)]
owwModel = openwakeword.Model(
wakeword_model_paths=models,
)
# Predict with median smooth
for model, clips in test_dict.items():
for clip in clips:
# Get predictions for reach frame in the clip
predictions = owwModel.predict_clip(os.path.join("tests", "data", clip), median_smooth=True)
owwModel.reset() # reset after each clip to ensure independent results
# Make predictions dictionary flatter
predictions_flat = collections.defaultdict(list)
[predictions_flat[key].append(i[key]) for i in predictions for key in i.keys()]
# Check scores against default threshold (0.5)
for key in predictions_flat.keys():
if key in clip:
assert max(predictions_flat[key]) >= 0.5
else:
print(key, clip)
assert max(predictions_flat[key]) < 0.5
def test_models_with_timing(self):
models = [str(i) for i in Path(
os.path.join("openwakeword", "resources", "models")