Feature/log time sample (#97)

* log information per sample in artifacts (audio and spec)
* bugfix in default train run

soundbay/conf/data/defaults.yaml

@@ -12,6 +12,7 @@ data:
   train_dataset:
     _target_: soundbay.data.ClassifierDataset
     data_path: './tests/assets/data/'
+    path_hierarchy: 0
     mode: train
     metadata_path: './tests/assets/annotations/sample_annotations.csv'
     augmentations_p: 0.8
@@ -25,6 +26,7 @@ data:
   val_dataset:
     _target_: soundbay.data.ClassifierDataset
     data_path: './tests/assets/data'
+    path_hierarchy: 0
     mode: val
     metadata_path: './tests/assets/annotations/sample_annotations.csv'
     augmentations_p: 0

soundbay/data.py

@@ -239,7 +239,7 @@ def __getitem__(self, idx):
 
         if self.mode == "train" or self.mode == "val":
             label = self.metadata["label"][idx]
-            return audio_processed, label, audio_raw, idx
+            return audio_processed, label, audio_raw, {"idx": idx, "begin_time": begin_time, "org_file": Path(path_to_file).stem}
 
         elif self.mode == "test":
             return audio_processed

soundbay/trainers.py

@@ -98,11 +98,11 @@ def train_epoch(self, epoch):
                 break
 
             self.model.zero_grad()
-            audio, label, raw_wav, idx = batch
-            audio, label  = audio.to(self.device), label.to(self.device)
+            audio, label, raw_wav, meta = batch
+            audio, label = audio.to(self.device), label.to(self.device)
 
             if (it == 0) and (not self.debug) and ((epoch % 5) == 0):
-                self.logger.upload_artifacts(audio, label, raw_wav, idx, sample_rate=self.train_dataloader.dataset.sample_rate, flag='train')
+                self.logger.upload_artifacts(audio, label, raw_wav, meta, sample_rate=self.train_dataloader.dataset.sample_rate, flag='train')
 
             # estimate and calc losses
             estimated_label = self.model(audio)
@@ -137,10 +137,10 @@ def eval_epoch(self, epoch: int, datatset_name: str = None):
             for it, batch in tqdm(enumerate(dataloader), desc=datatset_name):
                 if it == 3 and self.debug:
                     break
-                audio, label, raw_wav, idx = batch
+                audio, label, raw_wav, meta = batch
                 audio, label = audio.to(self.device), label.to(self.device)
                 if (it == 0) and (not self.debug) and ((epoch % 5) == 0):
-                    self.logger.upload_artifacts(audio, label, raw_wav, idx, sample_rate=self.train_dataloader.dataset.sample_rate, flag=datatset_name)
+                    self.logger.upload_artifacts(audio, label, raw_wav, meta, sample_rate=self.train_dataloader.dataset.sample_rate, flag=datatset_name)
 
                 # estimate and calc losses
                 estimated_label = self.model(audio)

soundbay/utils/logging.py

@@ -153,11 +153,12 @@ def calc_metrics(self, epoch: int, mode: str = 'train', label_names: Optional[Li
         self.label_list = []
         self.pred_proba_list = []
 
-    def upload_artifacts(self, audio: torch.Tensor, label: torch.Tensor, raw_wav: torch.Tensor, idx: torch.Tensor, sample_rate: int=16000, flag: str='train'):
+    def upload_artifacts(self, audio: torch.Tensor, label: torch.Tensor, raw_wav: torch.Tensor, meta: dict, sample_rate: int=16000, flag: str='train'):
         """upload algorithm artifacts to W&B during training session"""
         volume = 50
         matplotlib.use('Agg')
-        idx = idx.detach().cpu().numpy()
+        idx = meta['idx'].detach().cpu().numpy()
+        meta['begin_time'] = meta['begin_time'].detach().cpu().numpy()
         label = label.detach().cpu().numpy()
 
         if audio.shape[0] > self.upload_artifacts_limit:
@@ -170,7 +171,7 @@ def upload_artifacts(self, audio: torch.Tensor, label: torch.Tensor, raw_wav: to
 
         artifact_wav = torch.squeeze(raw_wav).detach().cpu().numpy()
         artifact_wav = artifact_wav / np.expand_dims(np.abs(artifact_wav).max(axis=1) + 1e-8, 1) * 0.5  # gain -6dB
-        list_of_wavs_objects = [wandb.Audio(data_or_path=wav, caption=f'label_{lab}_{ind}_train', sample_rate=sample_rate) for wav, ind, lab in zip(artifact_wav,idx, label)]
+        list_of_wavs_objects = [wandb.Audio(data_or_path=wav, caption=f'{flag}_label{lab}_i{ind}_{round(b_t/sample_rate,3)}sec_{f_n}', sample_rate=sample_rate) for wav, ind, lab, b_t, f_n in zip(artifact_wav,idx, label, meta['begin_time'], meta['org_file'])]
 
         # Spectrograms batch
         artifact_spec = torch.squeeze(audio).detach().cpu().numpy()
@@ -180,7 +181,7 @@ def upload_artifacts(self, audio: torch.Tensor, label: torch.Tensor, raw_wav: to
             specs.append(librosa.display.specshow(artifact_spec[artifact_id,...], ax=ax[1]))
             plt.close('all')
             del ax
-        list_of_specs_objects = [wandb.Image(data_or_path=spec, caption=f'label_{lab}_{ind}_train') for spec, ind, lab in zip(specs,idx, label)]
+        list_of_specs_objects = [wandb.Image(data_or_path=spec, caption=f'{flag}_label{lab}_i{ind}_{round(b_t/sample_rate,2)}sec_{f_n}') for spec, ind, lab, b_t, f_n in zip(specs,idx, label, meta['begin_time'], meta['org_file'])]
         log_wavs = {f'First batch {flag} original wavs': list_of_wavs_objects}
         log_specs = {f'First batch {flag} augmented spectrogram\'s': list_of_specs_objects}