Some changes to the tensorboard code and hypernetwork support

1 files changed, 27 insertions, 18 deletions

diff --git a/modules/textual_inversion/textual_inversion.py b/modules/textual_inversion/textual_inversion.py
index ec8176bf..b1dc2596 100644
--- a/modules/textual_inversion/textual_inversion.py
+++ b/modules/textual_inversion/textual_inversion.py
@@ -201,19 +201,30 @@ def write_loss(log_directory, filename, step, epoch_len, values):
             **values,

         })

 

+def tensorboard_setup(log_directory):

+    os.makedirs(os.path.join(log_directory, "tensorboard"), exist_ok=True)

+    return SummaryWriter(

+            log_dir=os.path.join(log_directory, "tensorboard"),

+            flush_secs=shared.opts.training_tensorboard_flush_every)

+

+def tensorboard_add(tensorboard_writer, loss, global_step, step, learn_rate, epoch_num):

+    tensorboard_add_scaler(tensorboard_writer, "Loss/train", loss, global_step)

+    tensorboard_add_scaler(tensorboard_writer, f"Loss/train/epoch-{epoch_num}", loss, step)

+    tensorboard_add_scaler(tensorboard_writer, "Learn rate/train", learn_rate, global_step)

+    tensorboard_add_scaler(tensorboard_writer, f"Learn rate/train/epoch-{epoch_num}", learn_rate, step)

+

 def tensorboard_add_scaler(tensorboard_writer, tag, value, step):

-    if shared.opts.training_enable_tensorboard:

-        tensorboard_writer.add_scalar(tag=tag, 

-            scalar_value=value, global_step=step)

+    tensorboard_writer.add_scalar(tag=tag, 

+        scalar_value=value, global_step=step)

 

 def tensorboard_add_image(tensorboard_writer, tag, pil_image, step):

-    if shared.opts.training_enable_tensorboard:

-        # Convert a pil image to a torch tensor

-        img_tensor = torch.as_tensor(np.array(pil_image, copy=True))

-        img_tensor = img_tensor.view(pil_image.size[1], pil_image.size[0], len(pil_image.getbands()))

-        img_tensor = img_tensor.permute((2, 0, 1))

+    # Convert a pil image to a torch tensor

+    img_tensor = torch.as_tensor(np.array(pil_image, copy=True))

+    img_tensor = img_tensor.view(pil_image.size[1], pil_image.size[0], 

+        len(pil_image.getbands()))

+    img_tensor = img_tensor.permute((2, 0, 1))

                 

-        tensorboard_writer.add_image(tag, img_tensor, global_step=step)

+    tensorboard_writer.add_image(tag, img_tensor, global_step=step)

 

 def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):

     assert embedding_name, 'embedding not selected'

@@ -268,10 +279,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
     optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)

 

     if shared.opts.training_enable_tensorboard:

-        os.makedirs(os.path.join(log_directory, "tensorboard"), exist_ok=True)

-        tensorboard_writer = SummaryWriter(

-                log_dir=os.path.join(log_directory, "tensorboard"),

-                flush_secs=shared.opts.training_tensorboard_flush_every)

+        tensorboard_writer = tensorboard_setup(log_directory)

 

     pbar = tqdm.tqdm(enumerate(ds), total=steps-initial_step)

     for i, entries in pbar:

@@ -308,10 +316,8 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
             embedding_yet_to_be_embedded = True

 

         if shared.opts.training_enable_tensorboard:

-            tensorboard_add_scaler(tensorboard_writer, "Loss/train", losses.mean(), embedding.step)

-            tensorboard_add_scaler(tensorboard_writer, f"Loss/train/epoch-{epoch_num}", losses.mean(), epoch_step)

-            tensorboard_add_scaler(tensorboard_writer, "Learn rate/train", scheduler.learn_rate, embedding.step)

-            tensorboard_add_scaler(tensorboard_writer, f"Learn rate/train/epoch-{epoch_num}", scheduler.learn_rate, epoch_step)

+            tensorboard_add(tensorboard_writer, loss=losses.mean(), global_step=embedding.step, 

+                step=epoch_step, learn_rate=scheduler.learn_rate, epoch_num=epoch_num)

 

         write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, len(ds), {

             "loss": f"{losses.mean():.7f}",

@@ -377,7 +383,10 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
                 embedding_yet_to_be_embedded = False

 

             image.save(last_saved_image)

-            tensorboard_add_image(tensorboard_writer, f"Validation at epoch {epoch_num}", image, embedding.step)

+

+            if shared.opts.training_enable_tensorboard and shared.opts.training_tensorboard_save_images:

+                tensorboard_add_image(tensorboard_writer, f"Validation at epoch {epoch_num}", 

+                    image, embedding.step)

 

             last_saved_image += f", prompt: {preview_text}"