Merge pull request #1 from VFluger/fix_dependencies

README.md

@@ -16,9 +16,11 @@ Classifies the dance style that best accompanies a provided song. Users record o
 
 ## Getting Started
 
-1. Download dependencies: `conda env create --file environment.yml`
-2. Open environment: `conda activate dancer-net`
-3. Start the demo application: `python app.py`
+1. Clone this repo: `git clone https://github.com/Waidhoferj/dance-classifier`
+2. Download git LSF files: `git lfs pull`
+3. Download dependencies: `conda env create --file environment.yml`
+4. Open environment: `conda activate dancer-classifier`
+5. Start the demo application: `python app.py`
 
 ## Training
 

TODO.md

@@ -1,20 +0,0 @@
-- ✅ Ensure app.py audio input sounds like training data
-- ✅ Use a huggingface transformer with the dataset
-- Verify that the training spectrogram matches the predict spectrogram
-- Count number of example misses in dataset loading
-- Verify windowing and jitter params in Song Dataset
-- Create an attention-based network
-- ✅ Increase parameter count in network
-- Verify that labels really match what is on the music4dance site
-- ✅ Read the Medium series about audio DL
-- double check \_rectify_duration
-- ✅ Filter out songs that have only one vote
-- ✅ Download songs from [Best Ballroom](https://www.youtube.com/channel/UC0bYSnzAFMwPiEjmVsrvmRg)
-
-- ✅ fix nan values
-- Try higher mels (224) and more ffts (2048)
-- Verify random sample of dataset outputs by hand.
-
-- Train with non music data and add a non music category
-- Add back class weights
-- Add back multi label classification

app.py

@@ -85,9 +85,10 @@ class DancePredictor:
         if waveform.ndim == 1:
             waveform = np.stack([waveform, waveform]).T
         waveform = torch.from_numpy(waveform.T)
-        waveform = torchaudio.functional.apply_codec(
-            waveform, sample_rate, "wav", channels_first=True
-        )
+        # Convert to proper format instead of using deprecated apply_codec
+        # The apply_codec was mainly used for format conversion, but since we're already
+        # working with tensor data, we can skip this step
+        waveform = waveform.float()
 
         waveform = torchaudio.functional.resample(
             waveform, sample_rate, self.resample_frequency

environment.yml

@@ -9,7 +9,9 @@ dependencies:
   - pytorch
   - torchaudio
   - librosa
-  - numpy
+  - numpy<2
+  - sounddevice
+  - gradio
   - pandas
   - bs4
   - requests

requirements.txt

@@ -1,3 +1,4 @@
+sounddevice
 torch
 torchaudio
 pytorch-lightning