From e20a989c4163507ad4c86fb53cf84c8526706686 Mon Sep 17 00:00:00 2001
From: Laurent Fainsin <laurentfainsin@protonmail.com>
Date: Thu, 30 Jun 2022 21:26:12 +0200
Subject: [PATCH] feat: added some training metrics

Former-commit-id: 05f9c9c44914cc2ef3443791e71d2f9ca893751b [formerly 0cec2187f92c52328f07cca65ebaeb7ab3641ac3]
Former-commit-id: 4749dbea8fd69798e08e34a91156d3111c7504a8
---
 src/train.py       | 26 +++++++++++++++++++-------
 src/unet/blocks.py |  2 +-
 2 files changed, 20 insertions(+), 8 deletions(-)

diff --git a/src/train.py b/src/train.py
index c717652..5feb01c 100644
--- a/src/train.py
+++ b/src/train.py
@@ -26,7 +26,7 @@ def main():
             DIR_VALID_IMG="/home/lilian/data_disk/lfainsin/smoltrain2017/",
             DIR_SPHERE_IMG="/home/lilian/data_disk/lfainsin/spheres/Images/",
             DIR_SPHERE_MASK="/home/lilian/data_disk/lfainsin/spheres/Masks/",
-            FEATURES=[16, 32, 64, 128],
+            FEATURES=[64, 128, 256, 512],
             N_CHANNELS=3,
             N_CLASSES=1,
             AMP=True,
@@ -35,7 +35,7 @@ def main():
             DEVICE="cuda",
             WORKERS=8,
             EPOCHS=5,
-            BATCH_SIZE=64,
+            BATCH_SIZE=16,
             LEARNING_RATE=1e-4,
             IMG_SIZE=512,
             SPHERES=5,
@@ -105,11 +105,15 @@ def main():
     )
 
     # 4. Set up the optimizer, the loss, the learning rate scheduler and the loss scaling for amp
-    optimizer = torch.optim.Adam(net.parameters(), lr=wandb.config.LEARNING_RATE)
+    optimizer = torch.optim.RMSprop(net.parameters(), lr=wandb.config.LEARNING_RATE, weight_decay=1e-8, momentum=0.9)
     scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, "max", patience=2)
     grad_scaler = torch.cuda.amp.GradScaler(enabled=wandb.config.AMP)
     criterion = torch.nn.BCEWithLogitsLoss()
 
+    # accuracy stuff
+    mse = torch.nn.MSELoss()
+    mae = torch.nn.L1Loss()
+
     # save model.pth
     torch.save(net.state_dict(), "checkpoints/model-0.pth")
     artifact = wandb.Artifact("pth", type="model")
@@ -151,7 +155,7 @@ def main():
                     # forward
                     with torch.cuda.amp.autocast(enabled=wandb.config.AMP):
                         pred_masks = net(images)
-                        train_loss = criterion(true_masks, pred_masks)
+                        train_loss = criterion(pred_masks, true_masks)
 
                     # backward
                     optimizer.zero_grad(set_to_none=True)
@@ -159,21 +163,29 @@ def main():
                     grad_scaler.step(optimizer)
                     grad_scaler.update()
 
+                    # compute metrics
+                    accuracy = (true_masks == pred_masks).float().mean()
+                    mse = torch.nn.functional.mse_loss(pred_masks, true_masks)
+                    mae = torch.nn.functional.l1_loss(pred_masks, true_masks)
+
                     # update tqdm progress bar
                     pbar.update(images.shape[0])
                     pbar.set_postfix(**{"loss": train_loss.item()})
 
-                    # log training metrics
+                    # log metrics
                     wandb.log(
                         {
                             "train/epoch": epoch - 1 + step / len(train_loader),
-                            "train/train_loss": train_loss,
+                            "train/accuracy": accuracy,
+                            "train/loss": train_loss,
+                            "train/mse": mse,
+                            "train/mae": mae,
                         }
                     )
 
                 # Evaluation round
                 val_score = evaluate(net, val_loader, device)
-                # scheduler.step(val_score)
+                scheduler.step(val_score)
 
                 # log validation metrics
                 wandb.log(
diff --git a/src/unet/blocks.py b/src/unet/blocks.py
index b5b9267..b1f7d7f 100644
--- a/src/unet/blocks.py
+++ b/src/unet/blocks.py
@@ -72,7 +72,7 @@ class OutConv(nn.Module):
 
         self.conv = nn.Sequential(
             nn.Conv2d(in_channels, out_channels, kernel_size=1),
-            nn.Sigmoid(),
+            # nn.Sigmoid(),
         )
 
     def forward(self, x):