Upload 5 files

app.py

@@ -0,0 +1,71 @@
+import gradio as gr
+import torch
+import librosa
+import numpy as np
+import noisereduce as nr  # <-- CRITICAL ADDITION
+from model import MachineSoundCNN  # Make sure this matches your model class name!
+
+# 1. Define your exact classes
+CLASSES = [
+    "Machine 1 Normal", "Machine 1 Abnormal",
+    "Machine 2 Normal", "Machine 2 Abnormal",
+    "Machine 3 Normal", "Machine 3 Abnormal"
+]
+
+# 2. Load the Model (Forcing CPU for the Hugging Face Free Tier)
+device = torch.device('cpu')
+model = MachineSoundCNN() # Initialize EL sir's architecture
+model.load_state_dict(torch.load('best_model.pth', map_location=device))
+model.eval() # Set to evaluation mode (freezes BatchNorm/Dropout)
+
+# 3. Preprocessing & Inference Function
+def predict_machine_sound(audio_path):
+    if audio_path is None:
+        return "Please upload an audio file."
+
+    # Load audio at exactly 16kHz
+    y, sr = librosa.load(audio_path, sr=16000)
+
+    # --- CRITICAL FIX: NOISE REDUCTION ---
+    # Replicating your specific 0.5-second noise profile logic
+    profile_length = int(0.5 * sr)
+    if len(y) > profile_length:
+        noise_profile = y[:profile_length]
+        y_clean = nr.reduce_noise(y=y, sr=sr, y_noise=noise_profile, prop_decrease=0.8)
+    else:
+        y_clean = nr.reduce_noise(y=y, sr=sr, y_noise=y, prop_decrease=0.5)
+        
+    # Clip signal to [-1.0, 1.0] just like your training pipeline
+    y_clean = np.clip(y_clean, -1.0, 1.0)
+    # -------------------------------------
+
+    # Generate Mel Spectrogram (128 bins) using the CLEAN audio
+    mel_spec = librosa.feature.melspectrogram(y=y_clean, sr=sr, n_mels=128)
+    mel_spec_db = librosa.power_to_db(mel_spec, ref=np.max)
+
+    # Reshape for the CNN: (Batch, Channel, Height, Time) -> (1, 1, 128, time_steps)
+    input_tensor = torch.tensor(mel_spec_db, dtype=torch.float32).unsqueeze(0).unsqueeze(0)
+
+    # Forward Pass through the CNN
+    with torch.no_grad():
+        outputs = model(input_tensor)
+        # Convert raw output numbers into percentages (0.0 to 1.0)
+        probabilities = torch.nn.functional.softmax(outputs[0], dim=0)
+
+    # Format output for the Gradio UI: a dictionary of {Class Name: Probability}
+    result = {CLASSES[i]: float(probabilities[i]) for i in range(len(CLASSES))}
+    return result
+
+# 4. Build the Web Interface
+interface = gr.Interface(
+    fn=predict_machine_sound,
+    inputs=gr.Audio(type="filepath", label="Upload Machine Audio (.wav)"),
+    outputs=gr.Label(num_top_classes=6, label="CNN Prediction Confidence"),
+    title="Industrial Machine Sound Anomaly Detector",
+    description="Upload an audio clip of an industrial machine. The Custom CNN will analyze the Mel Spectrogram and predict if it is operating normally or failing.",
+    allow_flagging="never"
+)
+
+# 5. Launch the App
+if __name__ == "__main__":
+    interface.launch()

best_model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1a1f9eb14f2604bd2493cd2130b198a8d6ee4d032154bc9a9922eb94950f930
+size 9970018

config.py

@@ -0,0 +1,204 @@
+"""
+===============================================================================
+config.py — Central Configuration for the Machine Sound Classification Pipeline
+===============================================================================
+
+PURPOSE:
+    This file is the SINGLE SOURCE OF TRUTH for every hyperparameter,
+    file path, and constant used across the project. Every team member
+    should import values from here instead of hard-coding numbers.
+
+    If you need to change a value (e.g. sample rate, n_mels, batch size),
+    change it HERE and it propagates everywhere automatically.
+
+OWNER: Shared (everyone imports from here)
+===============================================================================
+"""
+
+import os
+
+# =============================================================================
+# 1. PATH CONFIGURATION
+# =============================================================================
+# Root directory of the project (where this file lives)
+PROJECT_ROOT = os.path.dirname(os.path.abspath(__file__))
+
+# Directory where the examiner places test .wav files
+# CRITICAL: infer.py reads from this exact folder
+DATA_DIR = os.path.join(PROJECT_ROOT, "data")
+
+# Directory for the training dataset (organized by class)
+TRAIN_DATA_DIR = os.path.join(PROJECT_ROOT, "train_data")
+
+PROCESSED_DATA_DIR = os.path.join(PROJECT_ROOT, "processed_features")
+
+# Directory where trained model checkpoints are saved
+CHECKPOINT_DIR = os.path.join(PROJECT_ROOT, "checkpoints")
+
+# Output files required by the submission
+RESULTS_FILE = os.path.join(PROJECT_ROOT, "results.txt")
+TIME_FILE = os.path.join(PROJECT_ROOT, "time.txt")
+
+# Directory for split metadata (train/val/test file lists)
+SPLITS_DIR = os.path.join(PROJECT_ROOT, "splits")
+
+
+# =============================================================================
+# 2. AUDIO PREPROCESSING CONSTANTS
+# =============================================================================
+# Target sampling rate — all audio is resampled to this before processing.
+# 16 kHz is standard for machine sound analysis; Nyquist limit = 8 kHz.
+# Owner: JSON (resampling.py)
+TARGET_SR = 16000
+
+# Silence removal threshold in dB.
+# Frames quieter than this (relative to peak) are considered silence.
+# 20 dB is a safe starting point for factory recordings.
+# Owner: EL sir (silence_removal.py)
+SILENCE_TOP_DB = 40
+
+# Noise reduction — duration (in seconds) of the noise profile sample.
+# We estimate the noise floor from the first N seconds of each clip.
+# Owner: EL sir (noise_reduction.py)
+NOISE_PROFILE_DURATION = 0.5  # seconds
+
+
+# =============================================================================
+# 3. MEL SPECTROGRAM PARAMETERS
+# =============================================================================
+# These MUST be agreed upon by sala7 (feature extraction) and EL sir (CNN input).
+# Changing n_mels here changes the CNN input height automatically.
+# Owner: sala7 (mel_spectrogram.py) — coordinated with EL sir
+
+# Number of mel filter banks (height of the spectrogram "image")
+N_MELS = 128
+
+# FFT window size — 1024 samples @ 16 kHz = 64 ms analysis window.
+# For machine fault detection, 64 ms captures one full rotation of many motors.
+N_FFT = 1024
+
+# Hop length — 512 samples = 50% overlap between consecutive frames.
+# Good time resolution without quadrupling computation.
+HOP_LENGTH = 512
+
+# Maximum frequency for the mel filterbank.
+# At 16 kHz sampling rate, Nyquist = 8 kHz, so fmax = 8000.
+FMAX = 8000
+
+# Power for the mel spectrogram (2.0 = power spectrogram)
+MEL_POWER = 2.0
+
+
+# =============================================================================
+# 4. FIXED-SIZE TENSOR PARAMETERS
+# =============================================================================
+# After silence removal, audio clips vary in length. We must pad/trim
+# spectrograms to a uniform time dimension for batching.
+#
+# CALCULATION:
+#   Original audio ≈ 11 seconds → 11 * 16000 = 176,000 samples
+#   Time frames = ceil(176000 / 512) = 344 frames (full 11s)
+#   After silence trimming, clips are shorter → 256 frames ≈ 8.2 seconds
+#   is a reasonable target that captures the machine sound while
+#   discarding silence.
+#
+
+# Final input shape to the CNN: (batch, channels, n_mels, time_frames)
+# channels = 1 (grayscale spectrogram)
+CNN_INPUT_CHANNELS = 1
+
+
+# =============================================================================
+# 5. DATA SPLIT RATIOS
+# =============================================================================
+# Stratified split ratios — every class appears in every split at the same proportion.
+# Owner: JSON (splits.py)
+TRAIN_RATIO = 0.70
+VAL_RATIO = 0.15
+TEST_RATIO = 0.15
+
+# Random seed for reproducibility across all random operations
+RANDOM_SEED = 42
+
+
+# =============================================================================
+# 6. MODEL ARCHITECTURE PARAMETERS
+# =============================================================================
+# Number of output classes:
+#   0 = Machine 1 Normal, 1 = Machine 1 Abnormal,
+#   2 = Machine 2 Normal, 3 = Machine 2 Abnormal,
+#   4 = Machine 3 Normal, 5 = Machine 3 Abnormal
+# Owner: EL sir (cnn.py)
+NUM_CLASSES = 6
+
+# Convolutional layer filter counts (depth progression)
+# Each successive layer doubles the filters to capture more complex patterns.
+CNN_FILTERS = [32, 64, 128, 256]
+
+# Kernel size for all Conv2D layers
+CNN_KERNEL_SIZE = 3
+
+# Padding for Conv2D layers (1 = 'same' padding with kernel_size=3)
+CNN_PADDING = 1
+
+# Pool size for MaxPool2d layers
+CNN_POOL_SIZE = 2
+
+# Output size of AdaptiveAvgPool2d before the classifier head
+# This makes the model accept any time-length input gracefully.
+ADAPTIVE_POOL_OUTPUT = (4, 4)
+
+
+# =============================================================================
+# 7. TRAINING HYPERPARAMETERS
+# =============================================================================
+# Owner: Osama (trainer.py)
+
+# Optimizer: AdamW (corrects weight decay application vs vanilla Adam)
+LEARNING_RATE = 1e-3
+WEIGHT_DECAY = 1e-4
+
+# Batch size for training DataLoader
+BATCH_SIZE = 64
+
+# Maximum number of training epochs
+MAX_EPOCHS = 100
+
+# Early stopping — stop if val loss doesn't improve for this many epochs
+EARLY_STOPPING_PATIENCE = 10
+
+# Learning rate scheduler — cosine annealing
+LR_SCHEDULER_T_MAX = MAX_EPOCHS  # period of the cosine cycle
+
+# Number of DataLoader workers for parallel data loading
+NUM_WORKERS = 8
+
+
+# =============================================================================
+# 8. AUGMENTATION PARAMETERS
+# =============================================================================
+# Owner: sala7 (augmentation.py)
+# These are applied ONLY during training (not val/test).
+
+# SpecAugment: number of frequency bands to mask
+FREQ_MASK_PARAM = 20
+
+# SpecAugment: number of time steps to mask
+TIME_MASK_PARAM = 30
+
+# Gaussian noise injection — standard deviation
+NOISE_STD = 0.005
+
+# Probability of applying each augmentation
+AUGMENT_PROB = 0.5
+
+
+# =============================================================================
+# 9. INFERENCE PARAMETERS
+# =============================================================================
+# Path to the best saved model checkpoint (used by infer.py)
+BEST_MODEL_PATH = os.path.join(CHECKPOINT_DIR, "best_model.pth")
+
+# Device selection for inference (auto-detect GPU)
+import torch
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model.py

@@ -0,0 +1,128 @@
+"""
+===============================================================================
+model/cnn.py — CNN Architecture for Machine Sound Classification
+===============================================================================
+
+OWNER: EL sir
+"""
+
+import torch
+import torch.nn as nn
+from config import (
+    CNN_INPUT_CHANNELS,
+    NUM_CLASSES,
+    CNN_FILTERS,
+    CNN_KERNEL_SIZE,
+    CNN_PADDING,
+    CNN_POOL_SIZE,
+    ADAPTIVE_POOL_OUTPUT,
+)
+
+class ConvBlock(nn.Module):
+    """
+    A single convolutional block: Conv2d → BatchNorm → ReLU → MaxPool2d.
+    """
+    def __init__(self, in_channels, out_channels):
+        super(ConvBlock, self).__init__()
+        self.conv = nn.Conv2d(
+            in_channels, out_channels,
+            kernel_size=CNN_KERNEL_SIZE,
+            padding=CNN_PADDING,
+            bias=False  # Bias is redundant when followed by BatchNorm
+        )
+        self.bn = nn.BatchNorm2d(out_channels)
+        self.relu = nn.ReLU(inplace=True)
+        self.pool = nn.MaxPool2d(CNN_POOL_SIZE)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        x = self.relu(x)
+        x = self.pool(x)
+        return x
+
+
+class MachineSoundCNN(nn.Module):
+    """
+    Custom CNN for 6-class machine sound classification from mel spectrograms.
+    """
+    def __init__(self, num_classes=NUM_CLASSES):
+        super(MachineSoundCNN, self).__init__()
+
+        # ---------------------------------------------------------------------
+        # 1. Feature Extractor (Dynamic Conv Blocks)
+        # ---------------------------------------------------------------------
+        layers = []
+        in_channels = CNN_INPUT_CHANNELS
+        
+        # Dynamically build blocks based on config.py (e.g., 32 -> 64 -> 128 -> 256)
+        for out_channels in CNN_FILTERS:
+            layers.append(ConvBlock(in_channels, out_channels))
+            in_channels = out_channels  # The output of this layer is the input to the next
+            
+        self.features = nn.Sequential(*layers)
+        
+        # ---------------------------------------------------------------------
+        # 2. Adaptive Pooling
+        # ---------------------------------------------------------------------
+        # This squashes whatever time dimension is left into a fixed (4, 4) grid
+        self.adaptive_pool = nn.AdaptiveAvgPool2d(ADAPTIVE_POOL_OUTPUT)
+
+        # ---------------------------------------------------------------------
+        # 3. Classifier Head (Fully Connected)
+        # ---------------------------------------------------------------------
+        # Calculate flattened size: Last filter size (256) * height (4) * width (4) = 4096
+        flattened_size = CNN_FILTERS[-1] * ADAPTIVE_POOL_OUTPUT[0] * ADAPTIVE_POOL_OUTPUT[1]
+
+        self.classifier = nn.Sequential(
+            nn.Flatten(),
+            nn.Dropout(p=0.5),                   # Regularization: Prevent overfitting
+            nn.Linear(flattened_size, 512),      # Hidden layer to compress features
+            nn.ReLU(inplace=True),
+            nn.Dropout(p=0.5),                   # Regularization
+            nn.Linear(512, num_classes)          # Output raw logits (6 classes)
+        )
+
+    def forward(self, x):
+        # Extract visual features from the spectrogram
+        x = self.features(x)
+        # Pool them to a fixed mathematical size
+        x = self.adaptive_pool(x)
+        # Make the final classification
+        x = self.classifier(x)
+        return x
+
+
+# =============================================================================
+# SHAPE SANITY CHECK
+# =============================================================================
+if __name__ == "__main__":
+    print("=" * 60)
+    print("CNN Architecture Shape Sanity Check")
+    print("=" * 60)
+
+    model = MachineSoundCNN()
+
+    # Create a dummy batch based on JSON's exact tensor output (batch=4, ch=1, mels=128, time=281)
+    dummy_input = torch.randn(4, 1, 128, 281)
+    print(f"\nInput shape:  {dummy_input.shape}")
+
+    # Trace through the architecture
+    x = dummy_input
+    for i, block in enumerate(model.features):
+        x = block(x)
+        print(f"After Conv Block {i+1} ({CNN_FILTERS[i]} filters): {x.shape}")
+
+    x = model.adaptive_pool(x)
+    print(f"After AdaptivePool:  {x.shape}")
+
+    x = model.classifier(x)
+    print(f"After Classifier:    {x.shape}")
+
+    print(f"\n✓ Output shape is correct: {x.shape} (batch=4, classes={NUM_CLASSES})")
+
+    # Count parameters
+    total_params = sum(p.numel() for p in model.parameters())
+    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    print(f"\nTotal parameters:     {total_params:,}")
+    print("=" * 60)

requirements.txt

@@ -0,0 +1,33 @@
+# =============================================================================
+# requirements.txt
+# =============================================================================
+
+gradio==3.36.0
+
+# Core deep learning
+torch==2.9.1
+torchaudio==2.9.1
+torchvision==0.24.1
+
+# Audio processing
+librosa==0.10.1
+soundfile==0.12.1
+resampy==0.4.3 #added this for kaiser window in resampling
+
+# Noise reduction
+noisereduce==3.0.2
+
+# Numerical / scientific
+numpy==1.26.4
+scipy==1.12.0
+
+# Data handling
+scikit-learn==1.4.1.post1
+pandas==2.2.1
+
+# Visualization
+matplotlib==3.8.3
+seaborn==0.13.2
+
+# Progress bars
+tqdm==4.66.2