diff options
| -rw-r--r-- | modules/hypernetworks/hypernetwork.py | 13 | ||||
| -rw-r--r-- | modules/prompt_parser.py | 7 | ||||
| -rw-r--r-- | modules/sd_hijack_clip.py | 17 | ||||
| -rw-r--r-- | modules/shared.py | 2 |
4 files changed, 27 insertions, 12 deletions
diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index bbd1f673..c963fc40 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -561,6 +561,7 @@ def train_hypernetwork(id_task, hypernetwork_name, learn_rate, batch_size, gradi _loss_step = 0 #internal
# size = len(ds.indexes)
# loss_dict = defaultdict(lambda : deque(maxlen = 1024))
+ loss_logging = deque(maxlen=len(ds) * 3) # this should be configurable parameter, this is 3 * epoch(dataset size)
# losses = torch.zeros((size,))
# previous_mean_losses = [0]
# previous_mean_loss = 0
@@ -610,7 +611,7 @@ def train_hypernetwork(id_task, hypernetwork_name, learn_rate, batch_size, gradi # go back until we reach gradient accumulation steps
if (j + 1) % gradient_step != 0:
continue
-
+ loss_logging.append(_loss_step)
if clip_grad:
clip_grad(weights, clip_grad_sched.learn_rate)
@@ -644,7 +645,7 @@ def train_hypernetwork(id_task, hypernetwork_name, learn_rate, batch_size, gradi if shared.opts.training_enable_tensorboard:
epoch_num = hypernetwork.step // len(ds)
epoch_step = hypernetwork.step - (epoch_num * len(ds)) + 1
- mean_loss = sum(sum(x) for x in loss_dict.values()) / sum(len(x) for x in loss_dict.values())
+ mean_loss = sum(loss_logging) / len(loss_logging)
textual_inversion.tensorboard_add(tensorboard_writer, loss=mean_loss, global_step=hypernetwork.step, step=epoch_step, learn_rate=scheduler.learn_rate, epoch_num=epoch_num)
textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, steps_per_epoch, {
@@ -688,9 +689,6 @@ def train_hypernetwork(id_task, hypernetwork_name, learn_rate, batch_size, gradi processed = processing.process_images(p)
image = processed.images[0] if len(processed.images) > 0 else None
-
- if shared.opts.training_enable_tensorboard and shared.opts.training_tensorboard_save_images:
- textual_inversion.tensorboard_add_image(tensorboard_writer, f"Validation at epoch {epoch_num}", image, hypernetwork.step)
if unload:
shared.sd_model.cond_stage_model.to(devices.cpu)
@@ -701,7 +699,10 @@ def train_hypernetwork(id_task, hypernetwork_name, learn_rate, batch_size, gradi hypernetwork.train()
if image is not None:
shared.state.assign_current_image(image)
-
+ if shared.opts.training_enable_tensorboard and shared.opts.training_tensorboard_save_images:
+ textual_inversion.tensorboard_add_image(tensorboard_writer,
+ f"Validation at epoch {epoch_num}", image,
+ hypernetwork.step)
last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
last_saved_image += f", prompt: {preview_text}"
diff --git a/modules/prompt_parser.py b/modules/prompt_parser.py index 870218db..69665372 100644 --- a/modules/prompt_parser.py +++ b/modules/prompt_parser.py @@ -274,6 +274,7 @@ re_attention = re.compile(r""" :
""", re.X)
+re_break = re.compile(r"\s*\bBREAK\b\s*", re.S)
def parse_prompt_attention(text):
"""
@@ -339,7 +340,11 @@ def parse_prompt_attention(text): elif text == ']' and len(square_brackets) > 0:
multiply_range(square_brackets.pop(), square_bracket_multiplier)
else:
- res.append([text, 1.0])
+ parts = re.split(re_break, text)
+ for i, part in enumerate(parts):
+ if i > 0:
+ res.append(["BREAK", -1])
+ res.append([part, 1.0])
for pos in round_brackets:
multiply_range(pos, round_bracket_multiplier)
diff --git a/modules/sd_hijack_clip.py b/modules/sd_hijack_clip.py index 852afc66..9fa5c5c5 100644 --- a/modules/sd_hijack_clip.py +++ b/modules/sd_hijack_clip.py @@ -96,13 +96,18 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module): token_count = 0
last_comma = -1
- def next_chunk():
- """puts current chunk into the list of results and produces the next one - empty"""
+ def next_chunk(is_last=False):
+ """puts current chunk into the list of results and produces the next one - empty;
+ if is_last is true, tokens <end-of-text> tokens at the end won't add to token_count"""
nonlocal token_count
nonlocal last_comma
nonlocal chunk
- token_count += len(chunk.tokens)
+ if is_last:
+ token_count += len(chunk.tokens)
+ else:
+ token_count += self.chunk_length
+
to_add = self.chunk_length - len(chunk.tokens)
if to_add > 0:
chunk.tokens += [self.id_end] * to_add
@@ -116,6 +121,10 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module): chunk = PromptChunk()
for tokens, (text, weight) in zip(tokenized, parsed):
+ if text == 'BREAK' and weight == -1:
+ next_chunk()
+ continue
+
position = 0
while position < len(tokens):
token = tokens[position]
@@ -159,7 +168,7 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module): position += embedding_length_in_tokens
if len(chunk.tokens) > 0 or len(chunks) == 0:
- next_chunk()
+ next_chunk(is_last=True)
return chunks, token_count
diff --git a/modules/shared.py b/modules/shared.py index 3483db1c..06810435 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -228,7 +228,7 @@ class State: if not parallel_processing_allowed:
return
- if self.sampling_step - self.current_image_sampling_step >= opts.show_progress_every_n_steps and opts.live_previews_enable:
+ if self.sampling_step - self.current_image_sampling_step >= opts.show_progress_every_n_steps and opts.live_previews_enable and opts.show_progress_every_n_steps != -1:
self.do_set_current_image()
def do_set_current_image(self):
|