add options for scaling

configs/stable-diffusion/txt2img-2B-clip-encoder-high-res-512-dev.yaml

@@ -0,0 +1,149 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    #ckpt_path: "/home/mchorse/stable-diffusion-ckpts/256pretrain-2022-06-09.ckpt"
+    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: 64
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 10000 ]
+        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:
+        image_size: 64 # unused
+        in_channels: 4
+        out_channels: 4
+        model_channels: 352
+        attention_resolutions: [ 8, 4, 2 ]
+        num_res_blocks: [ 2, 2, 2, 6 ]
+        channel_mult: [ 1, 2, 4, 4 ]
+        disable_self_attentions: [ True, True, True, False ]  # converts the self-attention to a cross-attention layer if true
+        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
+        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: main.DataModuleFromConfig
+  params:
+    batch_size: 1
+    num_workers: 4
+    wrap: false
+    train:
+      target: ldm.data.dummy.DummyData
+      params:
+        length: 20000
+        size: [512, 512, 3]
+    validation:
+      target: ldm.data.dummy.DummyData
+      params:
+        length: 10000
+        size: [512, 512, 3]
+
+#data:
+#  target: ldm.data.laion.WebDataModuleFromConfig
+#  params:
+#    tar_base: "pipe:aws s3 cp s3://s-datasets/laion-high-resolution/"
+#    batch_size: 4
+#    num_workers: 4
+#    multinode: True
+#    train:
+#      shards: '{00000..17279}.tar -'
+#      shuffle: 10000
+#      image_key: jpg
+#      image_transforms:
+#      - target: torchvision.transforms.Resize
+#        params:
+#          size: 512
+#          interpolation: 3
+#      - target: torchvision.transforms.RandomCrop
+#        params:
+#          size: 512
+#
+#    # NOTE use enough shards to avoid empty validation loops in workers
+#    validation:
+#      shards: '{17280..17535}.tar -'
+#      shuffle: 0
+#      image_key: jpg
+#      image_transforms:
+#      - target: torchvision.transforms.Resize
+#        params:
+#          size: 512
+#          interpolation: 3
+#      - target: torchvision.transforms.CenterCrop
+#        params:
+#          size: 512
+
+
+lightning:
+  callbacks:
+    image_logger:
+      target: main.ImageLogger
+      params:
+        batch_frequency: 5000
+        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:
+    #replace_sampler_ddp: False
+    benchmark: True
+    val_check_interval: 1000 # TODO: 1e10 # really sorry
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 2

ldm/modules/attention.py

@@ -194,9 +194,12 @@ class CrossAttention(nn.Module):
 
 
 class BasicTransformerBlock(nn.Module):
-    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True):
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
+                 disable_self_attn=False):
         super().__init__()
-        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout)  # is a self-attention
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+                                    context_dim=context_dim if self.disable_self_attn else None)  # is a self-attention if not self.disable_self_attn
         self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
         self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
                                     heads=n_heads, dim_head=d_head, dropout=dropout)  # is self-attn if context is none
@@ -209,7 +212,7 @@ class BasicTransformerBlock(nn.Module):
         return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
 
     def _forward(self, x, context=None):
-        x = self.attn1(self.norm1(x)) + x
+        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x
         x = self.attn2(self.norm2(x), context=context) + x
         x = self.ff(self.norm3(x)) + x
         return x
@@ -224,7 +227,8 @@ class SpatialTransformer(nn.Module):
     Finally, reshape to image
     """
     def __init__(self, in_channels, n_heads, d_head,
-                 depth=1, dropout=0., context_dim=None):
+                 depth=1, dropout=0., context_dim=None,
+                 disable_self_attn=False):
         super().__init__()
         self.in_channels = in_channels
         inner_dim = n_heads * d_head
@@ -237,7 +241,8 @@ class SpatialTransformer(nn.Module):
                                  padding=0)
 
         self.transformer_blocks = nn.ModuleList(
-            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim,
+                                   disable_self_attn=disable_self_attn)
                 for d in range(depth)]
         )
 

ldm/modules/diffusionmodules/openaimodel.py

@@ -18,6 +18,7 @@ from ldm.modules.diffusionmodules.util import (
     timestep_embedding,
 )
 from ldm.modules.attention import SpatialTransformer
+from ldm.util import exists
 
 
 # dummy replace
@@ -466,6 +467,7 @@ class UNetModel(nn.Module):
         context_dim=None,                 # custom transformer support
         n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
         legacy=True,
+        disable_self_attentions=None
     ):
         super().__init__()
         if use_spatial_transformer:
@@ -490,7 +492,18 @@ class UNetModel(nn.Module):
         self.in_channels = in_channels
         self.model_channels = model_channels
         self.out_channels = out_channels
-        self.num_res_blocks = num_res_blocks
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+        #self.num_res_blocks = num_res_blocks
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+
         self.attention_resolutions = attention_resolutions
         self.dropout = dropout
         self.channel_mult = channel_mult
@@ -525,7 +538,7 @@ class UNetModel(nn.Module):
         ch = model_channels
         ds = 1
         for level, mult in enumerate(channel_mult):
-            for _ in range(num_res_blocks):
+            for _ in range(self.num_res_blocks[level]):
                 layers = [
                     ResBlock(
                         ch,
@@ -547,6 +560,10 @@ class UNetModel(nn.Module):
                     if legacy:
                         #num_heads = 1
                         dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
                     layers.append(
                         AttentionBlock(
                             ch,
@@ -555,7 +572,8 @@ class UNetModel(nn.Module):
                             num_head_channels=dim_head,
                             use_new_attention_order=use_new_attention_order,
                         ) if not use_spatial_transformer else SpatialTransformer(
-                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                            disable_self_attn=disabled_sa
                         )
                     )
                 self.input_blocks.append(TimestepEmbedSequential(*layers))
@@ -609,7 +627,7 @@ class UNetModel(nn.Module):
                 num_heads=num_heads,
                 num_head_channels=dim_head,
                 use_new_attention_order=use_new_attention_order,
-            ) if not use_spatial_transformer else SpatialTransformer(
+            ) if not use_spatial_transformer else SpatialTransformer(  # always uses a self-attn
                             ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
                         ),
             ResBlock(
@@ -625,7 +643,7 @@ class UNetModel(nn.Module):
 
         self.output_blocks = nn.ModuleList([])
         for level, mult in list(enumerate(channel_mult))[::-1]:
-            for i in range(num_res_blocks + 1):
+            for i in range(self.num_res_blocks[level] + 1):
                 ich = input_block_chans.pop()
                 layers = [
                     ResBlock(
@@ -648,6 +666,10 @@ class UNetModel(nn.Module):
                     if legacy:
                         #num_heads = 1
                         dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
                     layers.append(
                         AttentionBlock(
                             ch,
@@ -656,10 +678,11 @@ class UNetModel(nn.Module):
                             num_head_channels=dim_head,
                             use_new_attention_order=use_new_attention_order,
                         ) if not use_spatial_transformer else SpatialTransformer(
-                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
+                            disable_self_attn=disabled_sa
                         )
                     )
-                if level and i == num_res_blocks:
+                if level and i == self.num_res_blocks[level]:
                     out_ch = ch
                     layers.append(
                         ResBlock(