changes for freezing and upscaling

shreyazomato · Sep 30, 2022 · 31ab82d · 31ab82d
1 parent 0eafc35
commit 31ab82d
Show file tree

Hide file tree

Showing 3 changed files with 196 additions and 10 deletions.
diff --git a/configs/stable-diffusion/upscaling_256.yaml b/configs/stable-diffusion/upscaling_256.yaml
@@ -0,0 +1,181 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentUpscaleDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 32
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: "hybrid-adm"
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    low_scale_key: "lr"
+
+    low_scale_config:
+      target: ldm.modules.encoders.modules.LowScaleEncoder
+      params:
+        scale_factor: 0.18215
+        linear_start: 0.00085
+        linear_end: 0.0120
+        timesteps: 1000
+        max_noise_level: 100
+        output_size: null
+        model_config:
+          target: ldm.models.autoencoder.AutoencoderKL
+          params:
+            embed_dim: 4
+            monitor: val/rec_loss
+            ckpt_path: "models/first_stage_models/kl-f8/model.ckpt"
+            ddconfig:
+              double_z: true
+              z_channels: 4
+              resolution: 256
+              in_channels: 3
+              out_ch: 3
+              ch: 128
+              ch_mult:
+                - 1
+                - 2
+                - 4
+                - 4
+              num_res_blocks: 2
+              attn_resolutions: [ ]
+              dropout: 0.0
+            lossconfig:
+              target: torch.nn.Identity
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 10000 ] # NOTE for resuming. use 10000 if starting from scratch
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1. ]
+        f_min: [ 1. ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        num_classes: 1000
+        image_size: 16 # unused
+        in_channels: 8
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: True
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ckpt_path: "models/first_stage_models/kl-f8/model.ckpt"
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
+
+
+data:
+  target: ldm.data.laion.WebDataModuleFromConfig
+  params:
+    tar_base: "/mnt/data_rome/laion/improved_aesthetics_6plus/ims"
+    batch_size: 50
+    num_workers: 8
+    multinode: True
+    train:
+      shards: '{00000..01209}.tar'
+      shuffle: 10000
+      image_key: jpg
+      image_transforms:
+      - target: torchvision.transforms.Resize
+        params:
+          size: 256
+          interpolation: 3
+      - target: torchvision.transforms.RandomCrop
+        params:
+          size: 256
+      postprocess:
+        target: ldm.data.laion.AddLR
+        params:
+          factor: 4
+          output_size: 256
+
+    # NOTE use enough shards to avoid empty validation loops in workers
+    validation:
+      shards: '{00000..00012}.tar'
+      shuffle: 0
+      image_key: jpg
+      image_transforms:
+      - target: torchvision.transforms.Resize
+        params:
+          size: 256
+          interpolation: 3
+      - target: torchvision.transforms.CenterCrop
+        params:
+          size: 256
+      postprocess:
+        target: ldm.data.laion.AddLR
+        params:
+          factor: 1
+          output_size: 256
+
+
+lightning:
+  find_unused_parameters: False
+
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 20
+        max_images: 4
+        increase_log_steps: False
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          inpaint: False
+          plot_progressive_rows: False
+          plot_diffusion_rows: False
+          N: 4
+          unconditional_guidance_scale: 3.0
+          unconditional_guidance_label: [""]
+
+  trainer:
+    benchmark: True
+    val_check_interval: 5000000  # really sorry
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
diff --git a/ldm/data/laion.py b/ldm/data/laion.py
@@ -219,10 +219,12 @@ def test_dataloader(self):
 
 
 from ldm.modules.image_degradation import degradation_fn_bsr_light
+import cv2
 
 class AddLR(object):
-    def __init__(self, factor):
+    def __init__(self, factor, output_size):
         self.factor = factor
+        self.output_size = output_size
 
     def pt2np(self, x):
         x = ((x+1.0)*127.5).clamp(0, 255).to(dtype=torch.uint8).detach().cpu().numpy()
@@ -236,6 +238,7 @@ def __call__(self, sample):
         # sample['jpg'] is tensor hwc in [-1, 1] at this point
         x = self.pt2np(sample['jpg'])
         x = degradation_fn_bsr_light(x, sf=self.factor)['image']
+        x = cv2.resize(x, (self.output_size, self.output_size), interpolation=2)
         x = self.np2pt(x)
         sample['lr'] = x
         return sample

diff --git a/ldm/models/diffusion/ddpm.py b/ldm/models/diffusion/ddpm.py
@@ -26,6 +26,7 @@
 from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
 from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
 from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.modules.attention import CrossAttention
 
 
 __conditioning_keys__ = {'concat': 'c_concat',
@@ -1396,17 +1397,18 @@ def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=
 
     def configure_optimizers(self):
         lr = self.learning_rate
-        if self.unet_trainable:
+        params = []
+        if self.unet_trainable == "attn":
+            print("Training only unet attention layers")
+            for n, m in self.model.named_modules():
+                if isinstance(m, CrossAttention) and n.endswith('attn2'):
+                    params.extend(m.parameters())
+        elif self.unet_trainable is True or self.unet_trainable == "all":
+            print("Training the full unet")
             params = list(self.model.parameters())
         else:
-            params = []
-
-        # TODO allow certain parts trainables
-        # from ldm.modules.attention import CrossAttention
-        # for n, m in self.model.named_modules():
-            # if isinstance(m, CrossAttention) and n.endswith('attn2'):
-                # params.extend(m.parameters())
-        # END JP
+            raise ValueError(f"Unrecognised setting for unet_trainable: {self.unet_trainable}")
+
         if self.cond_stage_trainable:
             print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
             params = params + list(self.cond_stage_model.parameters())