fix: how print training resume logs.

examples/custom_diffusion/train_custom_diffusion.py

@@ -1075,30 +1075,30 @@ def main(args):
                 f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
             )
             args.resume_from_checkpoint = None
+            initial_global_step = 0
         else:
             accelerator.print(f"Resuming from checkpoint {path}")
             accelerator.load_state(os.path.join(args.output_dir, path))
             global_step = int(path.split("-")[1])
 
-            resume_global_step = global_step * args.gradient_accumulation_steps
+            initial_global_step = global_step
             first_epoch = global_step // num_update_steps_per_epoch
-            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
-
-    # Only show the progress bar once on each machine.
-    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
-    progress_bar.set_description("Steps")
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
 
     for epoch in range(first_epoch, args.num_train_epochs):
         unet.train()
         if args.modifier_token is not None:
             text_encoder.train()
         for step, batch in enumerate(train_dataloader):
-            # Skip steps until we reach the resumed step
-            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
-                if step % args.gradient_accumulation_steps == 0:
-                    progress_bar.update(1)
-                continue
-
             with accelerator.accumulate(unet), accelerator.accumulate(text_encoder):
                 # Convert images to latent space
                 latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()

examples/dreambooth/train_dreambooth.py

@@ -1178,30 +1178,30 @@ def compute_text_embeddings(prompt):
                 f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
             )
             args.resume_from_checkpoint = None
+            initial_global_step = 0
         else:
             accelerator.print(f"Resuming from checkpoint {path}")
             accelerator.load_state(os.path.join(args.output_dir, path))
             global_step = int(path.split("-")[1])
 
-            resume_global_step = global_step * args.gradient_accumulation_steps
+            initial_global_step = global_step
             first_epoch = global_step // num_update_steps_per_epoch
-            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
-
-    # Only show the progress bar once on each machine.
-    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
-    progress_bar.set_description("Steps")
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
 
     for epoch in range(first_epoch, args.num_train_epochs):
         unet.train()
         if args.train_text_encoder:
             text_encoder.train()
         for step, batch in enumerate(train_dataloader):
-            # Skip steps until we reach the resumed step
-            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
-                if step % args.gradient_accumulation_steps == 0:
-                    progress_bar.update(1)
-                continue
-
             with accelerator.accumulate(unet):
                 pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
 

examples/dreambooth/train_dreambooth_lora.py

@@ -1108,30 +1108,30 @@ def compute_text_embeddings(prompt):
                 f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
             )
             args.resume_from_checkpoint = None
+            initial_global_step = 0
         else:
             accelerator.print(f"Resuming from checkpoint {path}")
             accelerator.load_state(os.path.join(args.output_dir, path))
             global_step = int(path.split("-")[1])
 
-            resume_global_step = global_step * args.gradient_accumulation_steps
+            initial_global_step = global_step
             first_epoch = global_step // num_update_steps_per_epoch
-            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
-
-    # Only show the progress bar once on each machine.
-    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
-    progress_bar.set_description("Steps")
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
 
     for epoch in range(first_epoch, args.num_train_epochs):
         unet.train()
         if args.train_text_encoder:
             text_encoder.train()
         for step, batch in enumerate(train_dataloader):
-            # Skip steps until we reach the resumed step
-            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
-                if step % args.gradient_accumulation_steps == 0:
-                    progress_bar.update(1)
-                continue
-
             with accelerator.accumulate(unet):
                 pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
 

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -1048,31 +1048,32 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                 f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
             )
             args.resume_from_checkpoint = None
+            initial_global_step = 0
         else:
             accelerator.print(f"Resuming from checkpoint {path}")
             accelerator.load_state(os.path.join(args.output_dir, path))
             global_step = int(path.split("-")[1])
 
-            resume_global_step = global_step * args.gradient_accumulation_steps
+            initial_global_step = global_step
             first_epoch = global_step // num_update_steps_per_epoch
-            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
 
-    # Only show the progress bar once on each machine.
-    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
-    progress_bar.set_description("Steps")
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
 
     for epoch in range(first_epoch, args.num_train_epochs):
         unet.train()
         if args.train_text_encoder:
             text_encoder_one.train()
             text_encoder_two.train()
         for step, batch in enumerate(train_dataloader):
-            # Skip steps until we reach the resumed step
-            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
-                if step % args.gradient_accumulation_steps == 0:
-                    progress_bar.update(1)
-                continue
-
             with accelerator.accumulate(unet):
                 pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
 

examples/instruct_pix2pix/train_instruct_pix2pix_sdxl.py

@@ -726,6 +726,9 @@ def preprocess_images(examples):
     text_encoder_1.requires_grad_(False)
     text_encoder_2.requires_grad_(False)
 
+    # Set UNet to trainable.
+    unet.train()
+
     # Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
     def encode_prompt(text_encoders, tokenizers, prompt):
         prompt_embeds_list = []
@@ -933,29 +936,28 @@ def collate_fn(examples):
                 f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
             )
             args.resume_from_checkpoint = None
+            initial_global_step = 0
         else:
             accelerator.print(f"Resuming from checkpoint {path}")
             accelerator.load_state(os.path.join(args.output_dir, path))
             global_step = int(path.split("-")[1])
 
-            resume_global_step = global_step * args.gradient_accumulation_steps
+            initial_global_step = global_step
             first_epoch = global_step // num_update_steps_per_epoch
-            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
-
-    # Only show the progress bar once on each machine.
-    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
-    progress_bar.set_description("Steps")
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
 
     for epoch in range(first_epoch, args.num_train_epochs):
-        unet.train()
         train_loss = 0.0
         for step, batch in enumerate(train_dataloader):
-            # Skip steps until we reach the resumed step
-            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
-                if step % args.gradient_accumulation_steps == 0:
-                    progress_bar.update(1)
-                continue
-
             with accelerator.accumulate(unet):
                 # We want to learn the denoising process w.r.t the edited images which
                 # are conditioned on the original image (which was edited) and the edit instruction.

examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py

@@ -512,6 +512,9 @@ def deepspeed_zero_init_disabled_context_manager():
     vae.requires_grad_(False)
     image_encoder.requires_grad_(False)
 
+    # Set unet to trainable.
+    unet.train()
+
     # Create EMA for the unet.
     if args.use_ema:
         ema_unet = UNet2DConditionModel.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="unet")
@@ -727,27 +730,28 @@ def collate_fn(examples):
                 f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
             )
             args.resume_from_checkpoint = None
+            initial_global_step = 0
         else:
             accelerator.print(f"Resuming from checkpoint {path}")
             accelerator.load_state(os.path.join(args.output_dir, path))
             global_step = int(path.split("-")[1])
 
-            resume_global_step = global_step * args.gradient_accumulation_steps
+            initial_global_step = global_step
             first_epoch = global_step // num_update_steps_per_epoch
-            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
 
-    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
-    progress_bar.set_description("Steps")
     for epoch in range(first_epoch, args.num_train_epochs):
-        unet.train()
         train_loss = 0.0
         for step, batch in enumerate(train_dataloader):
-            # Skip steps until we reach the resumed step
-            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
-                if step % args.gradient_accumulation_steps == 0:
-                    progress_bar.update(1)
-                continue
-
             with accelerator.accumulate(unet):
                 # Convert images to latent space
                 images = batch["pixel_values"].to(weight_dtype)

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py

@@ -579,29 +579,29 @@ def collate_fn(examples):
                 f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
             )
             args.resume_from_checkpoint = None
+            initial_global_step = 0
         else:
             accelerator.print(f"Resuming from checkpoint {path}")
             accelerator.load_state(os.path.join(args.output_dir, path))
             global_step = int(path.split("-")[1])
 
-            resume_global_step = global_step * args.gradient_accumulation_steps
+            initial_global_step = global_step
             first_epoch = global_step // num_update_steps_per_epoch
-            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
+    else:
+        initial_global_step = 0
 
-    # Only show the progress bar once on each machine.
-    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
-    progress_bar.set_description("Steps")
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
 
     for epoch in range(first_epoch, args.num_train_epochs):
         unet.train()
         train_loss = 0.0
         for step, batch in enumerate(train_dataloader):
-            # Skip steps until we reach the resumed step
-            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
-                if step % args.gradient_accumulation_steps == 0:
-                    progress_bar.update(1)
-                continue
-
             with accelerator.accumulate(unet):
                 # Convert images to latent space
                 images = batch["pixel_values"].to(weight_dtype)

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py

@@ -595,30 +595,33 @@ def collate_fn(examples):
                 f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
             )
             args.resume_from_checkpoint = None
+            initial_global_step = 0
         else:
             accelerator.print(f"Resuming from checkpoint {path}")
             accelerator.load_state(os.path.join(args.output_dir, path))
             global_step = int(path.split("-")[1])
 
-            resume_global_step = global_step * args.gradient_accumulation_steps
+            initial_global_step = global_step
             first_epoch = global_step // num_update_steps_per_epoch
-            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
 
-    # Only show the progress bar once on each machine.
-    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
-    progress_bar.set_description("Steps")
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
+
     clip_mean = clip_mean.to(weight_dtype).to(accelerator.device)
     clip_std = clip_std.to(weight_dtype).to(accelerator.device)
+
     for epoch in range(first_epoch, args.num_train_epochs):
         prior.train()
         train_loss = 0.0
         for step, batch in enumerate(train_dataloader):
-            # Skip steps until we reach the resumed step
-            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
-                if step % args.gradient_accumulation_steps == 0:
-                    progress_bar.update(1)
-                continue
-
             with accelerator.accumulate(prior):
                 # Convert images to latent space
                 text_input_ids, text_mask, clip_images = (