Fixing Github workflow, and setting .yml files up for configuring experiments.

pseeth · pseeth · commit 6bd0646bfc1e · 2020-09-28T02:28:02.000-07:00
diff --git a/.github/workflows/main.yml b/.github/workflows/main.yml
@@ -40,7 +40,7 @@ jobs:
 
     - name: Build the book
       run: |
-        python common/download.py
+        python -m common.data --run.cmd='download'
         jupyter-book build book/
 
     # Push the book's HTML to github-pages
diff --git a/README.md b/README.md
@@ -20,6 +20,17 @@ python -m common.data --run.cmd='download'
 jb build --all book/
 ```
 
+## Running an experiment
+
+To run a basic mask estimation experiment with a Chimera network,
+do the following:
+
+```
+python -m common.data --run.cmd='prepare_musdb'
+python -m common.exp.chimera --run.cmd='train'
+python -m common.exp.chimera --run.cmd='evaluate'
+```
+
 ## Questions? Comments? Typos? Bugs? Issues?
 
 Open a github issue [here](https://github.com/source-separation/tutorial/issues/new)
diff --git a/common/data.py b/common/data.py
@@ -287,9 +287,9 @@ def __init__(
             'source_time': ('uniform', 0, MAX_SOURCE_TIME),
             'event_time': ('const', 0),
             'event_duration': ('const', duration),
-            'snr': snr,
-            'pitch_shift': pitch_shift,
-            'time_stretch': time_stretch,
+            'snr': tuple(snr),
+            'pitch_shift': tuple(pitch_shift),
+            'time_stretch': tuple(time_stretch),
         }
         self.fg_path = fg_path
         self.sample_rate = sample_rate
diff --git a/common/exp/chimera.py b/common/exp/chimera.py
@@ -13,8 +13,8 @@
 def train(
     args,
     seed : int = 0,
-    num_epochs : int = 1,
-    epoch_length : int = 1,
+    num_epochs : int = 100,
+    epoch_length : int = 1000,
     lr : float = 1e-3,
     batch_size : int = 1,
     dpcl_weight : float = .75,
@@ -128,6 +128,7 @@ def evaluate(
     output_folder : str = './results',
     num_workers : int = 1,
 ):
+    output_folder = Path(output_folder) 
     stft_params, sample_rate = data.signal()
     # Output of net is always in alphabetical order
     labels = ['bass', 'drums', 'other', 'vocals']
@@ -140,6 +141,8 @@ def evaluate(
 
     _device = utils.device()
     separator = models.deep_mask_estimation(_device)
+    
+    utils.plot_metrics(separator, 'l1_loss', output_folder / 'metrics.png')
 
     pbar = tqdm.tqdm(musdb)
     for item in pbar:
@@ -158,7 +161,7 @@ def evaluate(
             json.dump(scores, f, indent=4)
         break
 
-    output_file = Path(output_folder) / 'report_card.txt'    
+    output_file = output_folder / 'report_card.txt'    
 
     json_files = glob.glob(f"{output_folder}/*.json")
     if not json_files:
diff --git a/common/exp/conf/base.yml b/common/exp/conf/base.yml
@@ -0,0 +1,45 @@
+autoclip.percentile: 10
+
+evaluate.folder: data/test
+evaluate.num_workers: 1
+evaluate.output_folder: ./results
+
+logger.level: info
+
+mixer.coherent_prob: 0.25
+mixer.master_label: vocals
+mixer.n_channels: 1
+mixer.num_mixtures: 10
+mixer.pitch_shift: [uniform, -2, 2]
+mixer.ref_db: [-30, -10]
+mixer.sample_rate: 44100
+mixer.snr: [uniform, -5, 5]
+mixer.source_file: [choose, []]
+mixer.time_stretch: [uniform, .9, 1.1]
+
+train/mixer.fg_path: data/train
+train/mixer.duration: 4.0
+train/transform.excerpt_length: 4.0
+
+val/mixer.duration: 10.0
+val/transform.excerpt_length: 10.0
+
+val/mixer.coherent_prob: 1.0
+val.mixer.pitch_shift: [const, 0]
+val/mixer.time_stretch: [const, 1.0]
+val/mixer.fg_path: data/valid
+val/mixer.num_mixtures: 100
+
+device.use: cuda
+
+early_stopping.cumulative_delta: 0
+early_stopping.epochs: 30
+early_stopping.min_delta: 0.0
+
+patience.epochs: 5
+patience.factor: 0.5
+patience.mode: min
+patience.verbose: false
+
+transform.audio_only: false
+transform.mask_type: msa
diff --git a/common/exp/conf/chimera.yml b/common/exp/conf/chimera.yml
@@ -0,0 +1,34 @@
+$include: 
+  - common/conf/exp/base.yml
+
+build.bidirectional: 1
+build.dropout: 0.3
+build.embedding_activation:
+- sigmoid
+- unit_norm
+build.embedding_size: 20
+build.hidden_size: 300
+build.mask_activation:
+- sigmoid
+build.num_audio_channels: 1
+build.num_layers: 4
+build.num_sources: 4
+build.rnn_type: lstm
+
+deep_mask_estimation.mask_type: soft
+deep_mask_estimation.model_path: checkpoints/best.model.pth
+
+signal.hop_length: 512
+signal.sample_rate: 44100
+signal.window_length: 2048
+signal.window_type: sqrt_hann
+
+train.batch_size: 32
+train.dpcl_weight: 0.75
+train.epoch_length: 1000
+train.lr: 0.001
+train.mi_weight: 0.25
+train.num_epochs: 100
+train.num_workers: 4
+train.output_folder: .
+train.seed: 0
diff --git a/common/utils.py b/common/utils.py
@@ -2,7 +2,7 @@
 import json
 import sys
 import torch
-
+import matplotlib.pyplot as plt
 from . import argbind
 
 @argbind.bind_to_parser()
@@ -69,3 +69,19 @@ def pprint(data):
                 f"Source time : {obs.value['source_time']} \n"
             )
             logging.info('\n' + desc)
+
+
+def plot_metrics(separator, key, output_path):
+    data = separator.metadata['trainer.state.epoch_history']
+    plt.figure(figsize=(5, 4))
+
+    plt.subplot(111)
+    plt.plot(data[f'validation/{key}'], label='val')
+    plt.plot(data[f'train/{key}'], label='train')
+    plt.xlabel('Epoch')
+    plt.ylabel('Loss')
+    plt.title('Loss')
+    plt.legend()
+    plt.tight_layout()
+
+    plt.savefig(output_path)
