[dreambooth] allow fine-tuning text encoder (#883)

* allow fine-tuning text encoder

* fix a few things

* update readme

examples/dreambooth/README.md

@@ -160,6 +160,39 @@ accelerate launch train_dreambooth.py \
   --mixed_precision=fp16
 ```
 
+### Fine-tune text encoder with the UNet.
+
+The script also allows to fine-tune the `text_encoder` along with the `unet`. It's been observed experimentally that fine-tuning `text_encoder` gives much better results especially on faces. 
+Pass the `--train_text_encoder` argument to the script to enable training `text_encoder`.
+
+___Note: Training text encoder requires more memory, with this option the training won't fit on 16GB GPU. It needs at least 24GB VRAM.___
+
+```bash
+export MODEL_NAME="CompVis/stable-diffusion-v1-4"
+export INSTANCE_DIR="path-to-instance-images"
+export CLASS_DIR="path-to-class-images"
+export OUTPUT_DIR="path-to-save-model"
+
+accelerate launch train_dreambooth.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --train_text_encoder \
+  --instance_data_dir=$INSTANCE_DIR \
+  --class_data_dir=$CLASS_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --with_prior_preservation --prior_loss_weight=1.0 \
+  --instance_prompt="a photo of sks dog" \
+  --class_prompt="a photo of dog" \
+  --resolution=512 \
+  --train_batch_size=1 \
+  --use_8bit_adam
+  --gradient_checkpointing \
+  --learning_rate=2e-6 \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --num_class_images=200 \
+  --max_train_steps=800
+```
+
 ## Inference
 
 Once you have trained a model using above command, the inference can be done simply using the `StableDiffusionPipeline`. Make sure to include the `identifier`(e.g. sks in above example) in your prompt.

examples/dreambooth/train_dreambooth.py

@@ -1,4 +1,5 @@
 import argparse
+import itertools
 import math
 import os
 from pathlib import Path
@@ -100,6 +101,7 @@ def parse_args():
     parser.add_argument(
         "--center_crop", action="store_true", help="Whether to center crop images before resizing to resolution"
     )
+    parser.add_argument("--train_text_encoder", action="store_true", help="Whether to train the text encoder")
     parser.add_argument(
         "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
     )
@@ -320,6 +322,15 @@ def main():
         logging_dir=logging_dir,
     )
 
+    # Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate
+    # This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models.
+    # TODO (patil-suraj): Remove this check when gradient accumulation with two models is enabled in accelerate.
+    if args.train_text_encoder and args.gradient_accumulation_steps > 1 and accelerator.num_processes > 1:
+        raise ValueError(
+            "Gradient accumulation is not supported when training the text encoder in distributed training. "
+            "Please set gradient_accumulation_steps to 1. This feature will be supported in the future."
+        )
+
     if args.seed is not None:
         set_seed(args.seed)
 
@@ -385,8 +396,14 @@ def main():
     vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae")
     unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet")
 
+    vae.requires_grad_(False)
+    if not args.train_text_encoder:
+        text_encoder.requires_grad_(False)
+
     if args.gradient_checkpointing:
         unet.enable_gradient_checkpointing()
+        if args.train_text_encoder:
+            text_encoder.gradient_checkpointing_enable()
 
     if args.scale_lr:
         args.learning_rate = (
@@ -406,8 +423,11 @@ def main():
     else:
         optimizer_class = torch.optim.AdamW
 
+    params_to_optimize = (
+        itertools.chain(unet.parameters(), text_encoder.parameters()) if args.train_text_encoder else unet.parameters()
+    )
     optimizer = optimizer_class(
-        unet.parameters(),  # only optimize unet
+        params_to_optimize,
         lr=args.learning_rate,
         betas=(args.adam_beta1, args.adam_beta2),
         weight_decay=args.adam_weight_decay,
@@ -467,9 +487,14 @@ def main():
         num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
     )
 
-    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        unet, optimizer, train_dataloader, lr_scheduler
-    )
+    if args.train_text_encoder:
+        unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            unet, text_encoder, optimizer, train_dataloader, lr_scheduler
+        )
+    else:
+        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            unet, optimizer, train_dataloader, lr_scheduler
+        )
 
     weight_dtype = torch.float32
     if args.mixed_precision == "fp16":
@@ -480,8 +505,9 @@ def main():
     # Move text_encode and vae to gpu.
     # For mixed precision training we cast the text_encoder and vae weights to half-precision
     # as these models are only used for inference, keeping weights in full precision is not required.
-    text_encoder.to(accelerator.device, dtype=weight_dtype)
     vae.to(accelerator.device, dtype=weight_dtype)
+    if not args.train_text_encoder:
+        text_encoder.to(accelerator.device, dtype=weight_dtype)
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
@@ -516,9 +542,8 @@ def main():
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
                 # Convert images to latent space
-                with torch.no_grad():
-                    latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
-                    latents = latents * 0.18215
+                latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
+                latents = latents * 0.18215
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(latents)
@@ -532,8 +557,7 @@ def main():
                 noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
 
                 # Get the text embedding for conditioning
-                with torch.no_grad():
-                    encoder_hidden_states = text_encoder(batch["input_ids"])[0]
+                encoder_hidden_states = text_encoder(batch["input_ids"])[0]
 
                 # Predict the noise residual
                 noise_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
@@ -556,7 +580,12 @@ def main():
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients:
-                    accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
+                    params_to_clip = (
+                        itertools.chain(unet.parameters(), text_encoder.parameters())
+                        if args.train_text_encoder
+                        else unet.parameters()
+                    )
+                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()
@@ -578,7 +607,9 @@ def main():
     # Create the pipeline using using the trained modules and save it.
     if accelerator.is_main_process:
         pipeline = StableDiffusionPipeline.from_pretrained(
-            args.pretrained_model_name_or_path, unet=accelerator.unwrap_model(unet)
+            args.pretrained_model_name_or_path,
+            unet=accelerator.unwrap_model(unet),
+            text_encoder=accelerator.unwrap_model(text_encoder),
         )
         pipeline.save_pretrained(args.output_dir)