piranaware_version3 / src /processing.py
ashandilgith's picture
pushing tested revised files for HF compatibility
ab65fad
Raw History Blame Contribute Delete
3.6 kB
import numpy as np
import os
import shutil
import json
import tensorflow as tf
from src.preprocess import audio_to_spectrograms
from src.model import build_autoencoder
from src.storage import upload_file, download_file
# --- CRITICAL FIX FOR CLOUD DEPLOYMENT ---
# We use /tmp because it is the ONLY folder guaranteed to be writable in Docker.
TEMP_DIR = "/tmp/temp_models"
# Create the directory if it doesn't exist
if not os.path.exists(TEMP_DIR):
os.makedirs(TEMP_DIR)
def train_mode_cloud(audio_path, mode_name, boat_id):
# 1. Preprocess
# audio_path is now likely "/tmp/input.wav" passed from app.py
X_train = audio_to_spectrograms(audio_path)
if X_train is None: return "❌ Audio too short (min 1 sec)."
# 2. Train
autoencoder = build_autoencoder(X_train.shape[1:])
autoencoder.fit(X_train, X_train, epochs=40, batch_size=4, verbose=0)
# 3. Calculate Threshold (Dynamic Safety Margin)
reconstructions = autoencoder.predict(X_train)
mse = np.mean(np.power(X_train - reconstructions, 2), axis=(1, 2, 3))
# Using 2 Standard Deviations (Industry Standard for Strict Detection)
threshold = float(np.mean(mse) + (2 * np.std(mse)))
# 4. Save Locally (TO /tmp)
model_filename = f"{mode_name}_model.h5"
meta_filename = f"{mode_name}_meta.json"
# These paths now point to /tmp/temp_models/...
local_model_path = os.path.join(TEMP_DIR, model_filename)
local_meta_path = os.path.join(TEMP_DIR, meta_filename)
# Save to the temporary linux folder
autoencoder.save(local_model_path, save_format='h5', include_optimizer=False)
with open(local_meta_path, 'w') as f:
json.dump({"threshold": threshold}, f)
# 5. Upload to Google Cloud
u1 = upload_file(local_model_path, boat_id, model_filename)
u2 = upload_file(local_meta_path, boat_id, meta_filename)
if u1 and u2:
return f"✅ Calibrated {mode_name.upper()} | Threshold: {threshold:.5f}"
else:
return "⚠️ Trained locally, but Cloud Upload Failed."
def predict_health_cloud(audio_path, mode_name, boat_id):
model_filename = f"{mode_name}_model.h5"
meta_filename = f"{mode_name}_meta.json"
local_model_path = os.path.join(TEMP_DIR, model_filename)
local_meta_path = os.path.join(TEMP_DIR, meta_filename)
# 1. Download from Cloud to /tmp
d1 = download_file(boat_id, model_filename, local_model_path)
d2 = download_file(boat_id, meta_filename, local_meta_path)
if not (d1 and d2):
return f"⚠️ No trained model found in cloud for Boat: {boat_id} (Mode: {mode_name})"
# 2. Load
with open(local_meta_path, 'r') as f:
threshold = json.load(f)["threshold"]
model = tf.keras.models.load_model(local_model_path, compile=False)
# 3. Predict
X_test = audio_to_spectrograms(audio_path)
if X_test is None: return "Error: Audio too short."
reconstructions = model.predict(X_test)
mse = np.mean(np.power(X_test - reconstructions, 2), axis=(1, 2, 3))
# 4. Analysis
anomalies = np.sum(mse > threshold)
health_score = 100 * (1 - (anomalies / len(mse)))
# 5. Telemetry
avg_error = np.mean(mse)
max_error = np.max(mse)
status = "🟢 HEALTHY" if health_score > 85 else "🔴 ANOMALY DETECTED"
return f"""
{status}
Confidence Score: {health_score:.1f}%
--- TECHNICAL TELEMETRY ---
Threshold Limit : {threshold:.5f}
Your Avg Error : {avg_error:.5f}
Your Max Error : {max_error:.5f}
Anomalous Secs : {anomalies} / {len(mse)}
"""