From 82f4acabe24e5171c40afa2939a4777ba87bcc30 Mon Sep 17 00:00:00 2001
From: Gustaf Rydholm <gustaf.rydholm@gmail.com>
Date: Sun, 8 Aug 2021 21:43:39 +0200
Subject: Add training of VQGAN

---
 text_recognizer/criterions/vqgan_loss.py | 23 ++++++++---------------
 1 file changed, 8 insertions(+), 15 deletions(-)

(limited to 'text_recognizer/criterions')

diff --git a/text_recognizer/criterions/vqgan_loss.py b/text_recognizer/criterions/vqgan_loss.py
index 8bb568f..87f0f1c 100644
--- a/text_recognizer/criterions/vqgan_loss.py
+++ b/text_recognizer/criterions/vqgan_loss.py
@@ -1,5 +1,5 @@
 """VQGAN loss for PyTorch Lightning."""
-from typing import Dict
+from typing import Dict, Optional
 from click.types import Tuple
 
 import torch
@@ -40,9 +40,9 @@ class VQGANLoss(nn.Module):
         vq_loss: Tensor,
         optimizer_idx: int,
         stage: str,
-    ) -> Tuple[Tensor, Dict[str, Tensor]]:
+    ) -> Optional[Tuple]:
         """Calculates the VQGAN loss."""
-        rec_loss = self.reconstruction_loss(
+        rec_loss: Tensor = self.reconstruction_loss(
             data.contiguous(), reconstructions.contiguous()
         )
 
@@ -51,13 +51,13 @@ class VQGANLoss(nn.Module):
             logits_fake = self.discriminator(reconstructions.contiguous())
             g_loss = -torch.mean(logits_fake)
 
-            loss = (
+            loss: Tensor = (
                 rec_loss
                 + self.discriminator_weight * g_loss
                 + self.vq_loss_weight * vq_loss
             )
             log = {
-                f"{stage}/loss": loss,
+                f"{stage}/total_loss": loss,
                 f"{stage}/vq_loss": vq_loss,
                 f"{stage}/rec_loss": rec_loss,
                 f"{stage}/g_loss": g_loss,
@@ -68,18 +68,11 @@ class VQGANLoss(nn.Module):
             logits_fake = self.discriminator(reconstructions.contiguous().detach())
             logits_real = self.discriminator(data.contiguous().detach())
 
-            d_loss = self.adversarial_loss(
+            d_loss = self.discriminator_weight * self.adversarial_loss(
                 logits_real=logits_real, logits_fake=logits_fake
             )
-            loss = (
-                rec_loss
-                + self.discriminator_weight * d_loss
-                + self.vq_loss_weight * vq_loss
-            )
+
             log = {
-                f"{stage}/loss": loss,
-                f"{stage}/vq_loss": vq_loss,
-                f"{stage}/rec_loss": rec_loss,
                 f"{stage}/d_loss": d_loss,
             }
-            return loss, log
+            return d_loss, log
-- 
cgit v1.2.3-70-g09d2