Skip to content

Commit

Permalink
Merge branch 'master' of https://github.com/BirdVox/birdvoxdetect
Browse files Browse the repository at this point in the history
  • Loading branch information
Vincent Lostanlen committed Nov 24, 2020
2 parents 4f1854f + 5434f4d commit 1902206
Show file tree
Hide file tree
Showing 3 changed files with 177 additions and 59 deletions.
209 changes: 152 additions & 57 deletions birdvoxdetect/core.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@
from contextlib import redirect_stderr
import datetime
import h5py
import hashlib
import joblib
import json
import librosa
Expand All @@ -11,10 +12,14 @@
import operator
import os
import pandas as pd
import platform
import scipy
import scipy.signal
import sklearn
import socket
import soundfile as sf
import sys
import time
import traceback
import warnings

Expand Down Expand Up @@ -49,6 +54,7 @@ def process_file(
export_context=False,
export_faults=False,
export_logger=False,
predict_proba=False,
threshold=50.0,
suffix="",
clip_duration=1.0,
Expand All @@ -60,6 +66,8 @@ def process_file(
custom_objects=None,
bva_threshold=0.5,
):
# Record local time. This will eventually serve to measure elapsed time.
start_time = time.time()

# Create output_dir if necessary.
if output_dir is not None:
Expand Down Expand Up @@ -221,17 +229,53 @@ def process_file(
df.to_csv(checklist_path,index=False)

# Initialize fault log as a Pandas DataFrame.
faultlist_path = get_output_path(
filepath, suffix + "faults.csv", output_dir=output_dir
)
faultlist_df_columns = [
"Start (hh:mm:ss)",
"Stop (hh:mm:ss)",
"Fault confidence (%)",
]
faultlist_df = pd.DataFrame(columns=faultlist_df_columns)
if export_faults:
faultlist_path = get_output_path(
filepath, suffix + "faults.csv", output_dir=output_dir
faultlist_df.to_csv(
faultlist_path, columns=faultlist_df_columns, index=False)

# Initialize JSON output.
if predict_proba:
json_path = get_output_path(
filepath, suffix + "proba.json", output_dir
)
faultlist_df_columns = [
"Start (hh:mm:ss)",
"Stop (hh:mm:ss)",
"Fault confidence (%)",
]
faultlist_df = pd.DataFrame(columns=faultlist_df_columns)
faultlist_df.to_csv(faultlist_path, columns=faultlist_df_columns, index=False)
# Get MD5 hash.
hash_md5 = hashlib.md5()
with open(filepath, "rb") as fhandle:
for chunk in iter(lambda: fhandle.read(4096), b""):
hash_md5.update(chunk)
json_metadata = {
"file_name": os.path.basename(filepath),
"file_path": os.path.abspath(filepath),
"audio_duration": librosa.get_duration(filename=filepath),
"audio_md5_checksum": hash_md5.hexdigest(),
"birdvoxdetect_threshold": threshold,
"birdvoxactivate_threshold": bva_threshold,
"classifier_name": classifier_name,
"detector_name": detector_name,
"hostname": socket.gethostname(),
"machine_time": datetime.datetime.now().astimezone().isoformat(),
"package_versions": {
module.__name__: module.__version__ for module in modules
},
"platform_machine": platform.machine(),
"platform_processor": platform.processor(),
"platform_release": platform.release(),
"platform_system": platform.system(),
"platform_version": platform.version(),
"sys_version": sys.version
}
with open(json_path, "w") as f:
json.dump({"metadata": json_metadata, "taxonomy": taxonomy}, f)
json_dicts = []

# Create directory of output clips.
if export_clips:
Expand Down Expand Up @@ -321,15 +365,15 @@ def process_file(
has_sensor_fault = False

# Add first row to sensor fault log.
faultlist_df = faultlist_df.append(
{
"Start (hh:mm:ss)": seconds_to_hhmmss(0.0),
"Stop (hh:mm:ss)": seconds_to_hhmmss(queue_length * chunk_duration),
"Fault confidence (%)": int(sensor_fault_probability * 100),
},
ignore_index=True,
)
if export_faults:
faultlist_df = faultlist_df.append(
{
"Start (hh:mm:ss)": seconds_to_hhmmss(0.0),
"Stop (hh:mm:ss)": seconds_to_hhmmss(queue_length * chunk_duration),
"Fault confidence (%)": int(sensor_fault_probability * 100),
},
ignore_index=True,
)
faultlist_df.to_csv(
faultlist_path, columns=faultlist_df_columns, index=False
)
Expand Down Expand Up @@ -381,12 +425,18 @@ def process_file(
n_peaks = len(chunk_timestamps)

# Classify species.
chunk_df = pd.DataFrame(
map(
lambda x: classify_species(classifier, chunk_pcen, x, taxonomy),
th_peak_locs,
)
)
rows = []
for peak_id, th_peak_loc in enumerate(th_peak_locs):
row, json_dict = classify_species(
classifier, chunk_pcen, th_peak_loc, taxonomy)
rows.append(row)
if predict_proba:
chunk_timestamp = chunk_timestamps[peak_id]
json_dict["Time (s)"] = float(chunk_timestamp)
json_dict["Time (hh:mm:ss)"] = seconds_to_hhmmss(chunk_timestamp)
json_dict["Confidence (%)"] = float(th_peak_confidences[peak_id])
json_dicts.append(json_dict)
chunk_df = pd.DataFrame(rows)

# Count flight calls.
if n_peaks>0:
Expand All @@ -403,7 +453,7 @@ def process_file(
logger.info("Number of flight calls in current chunk: 0")
logger.info("")

# Export timestamps.
# Export checklist.
chunk_hhmmss = list(map(seconds_to_hhmmss, chunk_timestamps))
chunk_df["Time (hh:mm:ss)"] = event_hhmmss + chunk_hhmmss
chunk_df["Confidence (%)"] = th_peak_confidences
Expand All @@ -413,6 +463,15 @@ def process_file(
df = df.append(chunk_df)
df.to_csv(checklist_path, columns=df_columns, index=False)

# Export probabilities as JSON file.
if predict_proba:
with open(json_path, "w") as f:
json.dump({
"events": json_dicts,
"metadata": json_metadata,
"taxonomy": taxonomy
}, f)

# Export clips.
if export_clips and len(df)>0:
chunk_zip = zip(
Expand Down Expand Up @@ -487,23 +546,23 @@ def process_file(
)[0][1]

# Add row to sensor fault log.
has_sensor_fault = sensor_fault_probability > bva_threshold
faultlist_df = faultlist_df.append(
{
"Start (hh:mm:ss)": seconds_to_hhmmss(chunk_id * chunk_duration),
"Stop (hh:mm:ss)": seconds_to_hhmmss(
(chunk_id + 1) * chunk_duration
),
"Fault confidence (%)": int(sensor_fault_probability * 100),
},
ignore_index=True,
)
if export_faults:
faultlist_df = faultlist_df.append(
{
"Start (hh:mm:ss)": seconds_to_hhmmss(chunk_id * chunk_duration),
"Stop (hh:mm:ss)": seconds_to_hhmmss(
(chunk_id + 1) * chunk_duration
),
"Fault confidence (%)": int(sensor_fault_probability * 100),
},
ignore_index=True,
)
faultlist_df.to_csv(
faultlist_path, columns=faultlist_df_columns, index=False
)

# If probability of sensor fault is above threshold, exclude chunk.
has_sensor_fault = (sensor_fault_probability > bva_threshold)
if has_sensor_fault:
logger.info(
"Probability of sensor fault: {:5.2f}%".format(
Expand Down Expand Up @@ -565,12 +624,18 @@ def process_file(
n_peaks = len(chunk_timestamps)

# Classify species.
chunk_df = pd.DataFrame(
map(
lambda x: classify_species(classifier, chunk_pcen, x, taxonomy),
th_peak_locs,
)
)
rows = []
for peak_id, th_peak_loc in enumerate(th_peak_locs):
row, json_dict = classify_species(
classifier, chunk_pcen, th_peak_loc, taxonomy)
rows.append(row)
if predict_proba:
chunk_timestamp = chunk_timestamps[peak_id]
json_dict["Time (s)"] = float(chunk_timestamp),
json_dict["Time (hh:mm:ss)"] = seconds_to_hhmmss(chunk_timestamp)
json_dict["Confidence (%)"] = float(th_peak_confidences[peak_id]),
json_dicts.append(json_dict)
chunk_df = pd.DataFrame(rows)

# Count flight calls.
if n_peaks>0:
Expand All @@ -587,7 +652,7 @@ def process_file(
logger.info("Number of flight calls in current chunk: 0")
logger.info("")

# Export timestamps.
# Export checklist.
chunk_hhmmss = list(map(seconds_to_hhmmss, chunk_timestamps))
chunk_df["Time (hh:mm:ss)"] = event_hhmmss + chunk_hhmmss
chunk_df["Confidence (%)"] = th_peak_confidences
Expand All @@ -597,6 +662,15 @@ def process_file(
df = df.append(chunk_df)
df.to_csv(checklist_path, columns=df_columns, index=False)

# Export probabilities as JSON file.
if predict_proba:
with open(json_path, "w") as f:
json.dump({
"events": json_dicts,
"metadata": json_metadata,
"taxonomy": taxonomy
}, f)

# Export clips.
if export_clips and len(df)>0:
chunk_zip = zip(
Expand Down Expand Up @@ -639,7 +713,7 @@ def process_file(
# unstable with files shorter than 30 minutes, which is why we issue a
# warning. Also, we do not try to detect sensor faults in files shorter than
# 30 minutes.
if (n_chunks > 1) and export_faults:
if (n_chunks > 1):
faultlist_df = faultlist_df.append(
{
"Start (hh:mm:ss)": seconds_to_hhmmss(chunk_id * chunk_duration),
Expand All @@ -648,7 +722,9 @@ def process_file(
},
ignore_index=True,
)
faultlist_df.to_csv(faultlist_path, columns=faultlist_df_columns, index=False)
if export_faults:
faultlist_df.to_csv(
faultlist_path, columns=faultlist_df_columns, index=False)

if (n_chunks > 1) and has_sensor_fault:
logger.info(
Expand Down Expand Up @@ -753,12 +829,18 @@ def process_file(
n_peaks = len(chunk_timestamps)

# Classify species.
chunk_df = pd.DataFrame(
map(
lambda x: classify_species(classifier, chunk_pcen, x, taxonomy),
th_peak_locs,
)
)
rows = []
for peak_id, th_peak_loc in enumerate(th_peak_locs):
row, json_dict = classify_species(
classifier, chunk_pcen, th_peak_loc, taxonomy)
rows.append(row)
if predict_proba:
chunk_timestamp = chunk_timestamps[peak_id]
json_dict["Time (s)"] = float(chunk_timestamp)
json_dict["Time (hh:mm:ss)"] = seconds_to_hhmmss(chunk_timestamp)
json_dict["Confidence (%)"] = float(th_peak_confidences[peak_id])
json_dicts.append(json_dict)
chunk_df = pd.DataFrame(rows)

# Count flight calls.
if n_peaks>0:
Expand All @@ -775,7 +857,7 @@ def process_file(
logger.info("Number of flight calls in current chunk: 0")
logger.info("")

# Export timestamps.
# Export checklist.
chunk_hhmmss = list(map(seconds_to_hhmmss, chunk_timestamps))
chunk_df["Time (hh:mm:ss)"] = event_hhmmss + chunk_hhmmss
chunk_df["Confidence (%)"] = th_peak_confidences
Expand All @@ -785,6 +867,18 @@ def process_file(
df = df.append(chunk_df)
df.to_csv(checklist_path, columns=df_columns, index=False)

# Export probabilities as JSON file.
if predict_proba:
with open(json_path, "w") as f:
json_faultlist = faultlist_df.to_json(orient="index")
json_metadata["elapsed_time"] = time.time() - start_time
json.dump({
"events": json_dicts,
"metadata": json_metadata,
"sensor_faults": json.loads(json_faultlist),
"taxonomy": taxonomy
}, f)

# Export clips.
if export_clips and len(df)>0:
chunk_zip = zip(
Expand Down Expand Up @@ -944,23 +1038,24 @@ def classify_species(classifier, chunk_pcen, th_peak_loc, taxonomy):
bvc_prediction = birdvoxclassify.predict(pcen_clip, classifier=classifier)

# Format prediction
prediction = birdvoxclassify.format_pred(bvc_prediction, taxonomy=taxonomy)
formatted_prediction = birdvoxclassify.format_pred(
bvc_prediction, taxonomy=taxonomy)

# Get prediction levels.
pred_levels = list(prediction.keys())
pred_levels = list(formatted_prediction.keys())

# Loop over taxonomical levels.
argmax_prediction = {}
for pred_level in pred_levels:
# List probabilities
prob_dict = {
k: prediction[pred_level][k]["probability"]
for k in prediction[pred_level]
k: formatted_prediction[pred_level][k]["probability"]
for k in formatted_prediction[pred_level]
}

# Extract class of maximum probability
argmax_taxon = max(prob_dict.items(), key=operator.itemgetter(1))[0]
argmax_dict = prediction[pred_level][argmax_taxon]
argmax_dict = formatted_prediction[pred_level][argmax_taxon]

if (pred_level=="coarse"):
argmax_prediction["Order"] = "other"
Expand All @@ -975,7 +1070,7 @@ def classify_species(classifier, chunk_pcen, th_peak_loc, taxonomy):
alias = aliases["species_4letter_code"]
argmax_prediction["Species (4-letter code)"] = alias

return argmax_prediction
return argmax_prediction, formatted_prediction


def compute_pcen(audio, sr):
Expand Down
10 changes: 8 additions & 2 deletions tests/test_cli.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -183,12 +183,18 @@ def test_main():
def test_script_main(capsys):
# Duplicate regression test from test_run just to hit coverage
tempdir = tempfile.mkdtemp()
with patch("sys.argv", ["birdvoxdetect", POSITIVE_PATH, "--output-dir", tempdir]):
with patch(
"sys.argv",
[
"birdvoxdetect", POSITIVE_PATH,
"--output-dir", tempdir,
"--suffix", "cli",
"--export-clips"]):
import birdvoxdetect.__main__

# Check output file created
outfile = os.path.join(
tempdir, "fd79e55d-d3a3-4083-aba1-4f00b545c3d6_checklist.csv"
tempdir, "fd79e55d-d3a3-4083-aba1-4f00b545c3d6_cli_checklist.csv"
)
assert os.path.isfile(outfile)
shutil.rmtree(tempdir)
Expand Down
Loading

0 comments on commit 1902206

Please sign in to comment.