Add char htr experiment, rename from ocr to htr, vqvae loss collapses

2 files changed, 4 insertions, 9 deletions

diff --git a/text_recognizer/models/vqvae.py b/text_recognizer/models/vqvae.py
index 5890fd9..7f79b78 100644
--- a/text_recognizer/models/vqvae.py
+++ b/text_recognizer/models/vqvae.py
@@ -1,11 +1,8 @@
 """PyTorch Lightning model for base Transformers."""
-from typing import Any, Dict, Union, Tuple, Type
+from typing import Tuple
 
 import attr
-from omegaconf import DictConfig
-from torch import nn
 from torch import Tensor
-import wandb
 
 from text_recognizer.models.base import BaseLitModel
 
@@ -25,7 +22,7 @@ class VQVAELitModel(BaseLitModel):
         data, _ = batch
         reconstructions, vq_loss = self(data)
         loss = self.loss_fn(reconstructions, data)
-        loss += self.latent_loss_weight * vq_loss
+        loss = loss + self.latent_loss_weight * vq_loss
         self.log("train/loss", loss)
         return loss
 
@@ -34,7 +31,7 @@ class VQVAELitModel(BaseLitModel):
         data, _ = batch
         reconstructions, vq_loss = self(data)
         loss = self.loss_fn(reconstructions, data)
-        loss += self.latent_loss_weight * vq_loss
+        loss = loss + self.latent_loss_weight * vq_loss
         self.log("val/loss", loss, prog_bar=True)
 
     def test_step(self, batch: Tuple[Tensor, Tensor], batch_idx: int) -> None:
@@ -42,5 +39,5 @@ class VQVAELitModel(BaseLitModel):
         data, _ = batch
         reconstructions, vq_loss = self(data)
         loss = self.loss_fn(reconstructions, data)
-        loss += self.latent_loss_weight * vq_loss
+        loss = loss + self.latent_loss_weight * vq_loss
         self.log("test/loss", loss)
diff --git a/text_recognizer/networks/vqvae/quantizer.py b/text_recognizer/networks/vqvae/quantizer.py
index 5e0b602..1b59e78 100644
--- a/text_recognizer/networks/vqvae/quantizer.py
+++ b/text_recognizer/networks/vqvae/quantizer.py
@@ -83,8 +83,6 @@ class VectorQuantizer(nn.Module):
     def compute_ema(self, one_hot_encoding: Tensor, latent: Tensor) -> None:
         batch_cluster_size = one_hot_encoding.sum(axis=0)
         batch_embedding_avg = (latent.t() @ one_hot_encoding).t()
-        print(batch_cluster_size.shape)
-        print(self.embedding._cluster_size.shape)
         self.embedding._cluster_size.data.mul_(self.decay).add_(
             batch_cluster_size, alpha=1 - self.decay
         )