Combine train/valid loops

scripts/train.py

@@ -17,7 +17,7 @@
 import sys
 import torchaudio
 import torchinfo
-from contextlib import contextmanager
+from contextlib import nullcontext
 import sklearn.preprocessing
 from torch.autograd import Variable
 from torch.utils.tensorboard import SummaryWriter
@@ -32,56 +32,36 @@
 tqdm.monitor_interval = 0
 
 
-def train(args, unmix, encoder, device, train_sampler, criterion, optimizer):
+def loop(args, unmix, encoder, device, sampler, criterion, optimizer, train=True):
     # unpack encoder object
     nsgt, _, cnorm = encoder
 
     losses = utils.AverageMeter()
-    unmix.train()
-    pbar = tqdm.tqdm(train_sampler, disable=args.quiet)
-
-    for x, y_bass, y_vocals, y_other, y_drums in pbar:
-        pbar.set_description("Training batch")
-        x, y_bass, y_vocals, y_other, y_drums = x.to(device), y_bass.to(device), y_vocals.to(device), y_other.to(device), y_drums.to(device)
-        optimizer.zero_grad()
-        X = nsgt(x)
-        Xmag = cnorm(X)
-        Ymag_bass = cnorm(nsgt(y_bass))
-        Ymag_vocals = cnorm(nsgt(y_vocals))
-        Ymag_drums = cnorm(nsgt(y_drums))
-        Ymag_other = cnorm(nsgt(y_other))
-
-        loss = criterion(
-            *unmix(Xmag), # forward call to unmix returns bass, vocals, other, drums
-            *(Ymag_bass, Ymag_vocals, Ymag_other, Ymag_drums),
-            *(y_bass, y_vocals, y_other, y_drums),
-            X,
-            x.shape[-1]
-        )
-
-        loss.backward()
-        optimizer.step()
-        losses.update(loss.item(), x.size(1))
-
-    return losses.avg
 
+    cm = None
+    name = ''
+    if train:
+        unmix.train()
+        cm = nullcontext
+        name = 'Train'
+    else:
+        unmix.eval()
+        cm = torch.no_grad
+        name = 'Validation'
 
-def valid(args, unmix, encoder, device, valid_sampler, criterion):
-    # unpack encoder object
-    nsgt, _, cnorm = encoder
+    pbar = tqdm.tqdm(sampler, disable=args.quiet)
 
-    losses = utils.AverageMeter()
-    unmix.eval()
-
-    with torch.no_grad():
-        pbar = tqdm.tqdm(valid_sampler, disable=args.quiet)
+    with cm():
         for x, y_bass, y_vocals, y_other, y_drums in pbar:
-            pbar.set_description("Validation batch")
+            pbar.set_description(f"{name} batch")
+
             x, y_bass, y_vocals, y_other, y_drums = x.to(device), y_bass.to(device), y_vocals.to(device), y_other.to(device), y_drums.to(device)
 
+            if train:
+                optimizer.zero_grad()
+
             X = nsgt(x)
             Xmag = cnorm(X)
-
             Ymag_bass = cnorm(nsgt(y_bass))
             Ymag_vocals = cnorm(nsgt(y_vocals))
             Ymag_drums = cnorm(nsgt(y_drums))
@@ -95,8 +75,13 @@ def valid(args, unmix, encoder, device, valid_sampler, criterion):
                 x.shape[-1]
             )
 
+            if train:
+                loss.backward()
+                optimizer.step()
+
             losses.update(loss.item(), x.size(1))
-        return losses.avg
+
+    return losses.avg
 
 
 def get_statistics(args, encoder, dataset, time_blocks):
@@ -426,8 +411,8 @@ def kill_tboard():
     for epoch in t:
         t.set_description("Training Epoch")
         end = time.time()
-        train_loss = train(args, unmix, encoder, device, train_sampler, criterion, optimizer)
-        valid_loss = valid(args, unmix, encoder, device, valid_sampler, criterion)
+        train_loss = loop(args, unmix, encoder, device, train_sampler, criterion, optimizer, train=True)
+        valid_loss = loop(args, unmix, encoder, device, valid_sampler, criterion, None, train=False)
 
         scheduler.step(valid_loss)
         train_losses.append(train_loss)