resolve more issues

codes/models/audio/tts/unet_diffusion_tts9.py

@@ -157,6 +157,7 @@ class DiffusionTts(nn.Module):
             kernel_size=3,
             scale_factor=2,
             time_embed_dim_multiplier=4,
+            freeze_main_net=False,
             efficient_convs=True,  # Uses kernels with width of 1 in several places rather than 3.
             use_scale_shift_norm=True,
             # Parameters for regularization.
@@ -188,6 +189,7 @@ class DiffusionTts(nn.Module):
         self.unconditioned_percentage = unconditioned_percentage
         self.enable_fp16 = use_fp16
         self.alignment_size = 2 ** (len(channel_mult)+1)
+        self.freeze_main_net = freeze_main_net
         padding = 1 if kernel_size == 3 else 2
         down_kernel = 1 if efficient_convs else 3
 
@@ -379,7 +381,17 @@ class DiffusionTts(nn.Module):
             zero_module(conv_nd(dims, model_channels, out_channels, kernel_size, padding=padding)),
         )
 
+        if self.freeze_main_net:
+            mains = [self.time_embed, self.contextual_embedder, self.conditioning_conv, self.unconditioned_embedding, self.conditioning_timestep_integrator,
+                     self.input_blocks, self.middle_block, self.output_blocks, self.out]
+            for m in mains:
+                for p in m.parameters():
+                    p.requires_grad = False
+                    p.DO_NOT_TRAIN = True
+
     def get_grad_norm_parameter_groups(self):
+        if self.freeze_main_net:
+            return {}
         groups = {
             'minicoder': list(self.contextual_embedder.parameters()),
             'input_blocks': list(self.input_blocks.parameters()),

codes/models/audio/tts/unet_diffusion_tts_flat.py

@@ -80,6 +80,7 @@ class DiffusionTtsFlat(nn.Module):
         self.contextual_embedder = nn.Sequential(nn.Conv1d(in_channels,model_channels,3,padding=1,stride=2),
                                                  CheckpointedXTransformerEncoder(
                                                      needs_permute=True,
+                                                     checkpoint=False,  # This is repeatedly executed for many conditioning signals, which is incompatible with checkpointing & DDP.
                                                      max_seq_len=-1,
                                                      use_pos_emb=False,
                                                      attn_layers=Encoder(
@@ -166,18 +167,16 @@ class DiffusionTtsFlat(nn.Module):
             conds = []
             for j in range(speech_conditioning_input.shape[1]):
                 conds.append(self.contextual_embedder(speech_conditioning_input[:, j]))
-            conds = torch.stack(conds, dim=1)
+            conds = torch.cat(conds, dim=-1)
-            cond_emb = conds.mean(dim=1)
+            cond_emb = conds.mean(dim=-1).unsqueeze(-1)
 
-            if len(cond_emb.shape) == 3:  # Just take the first element.
-                cond_emb = cond_emb[:, :, 0]
             if is_latent(aligned_conditioning):
                 code_emb = self.latent_converter(aligned_conditioning)
                 unused_params.extend(list(self.code_converter.parameters()))
             else:
                 code_emb = self.code_converter(aligned_conditioning)
                 unused_params.extend(list(self.latent_converter.parameters()))
-            cond_emb_spread = cond_emb.unsqueeze(-1).repeat(1, 1, code_emb.shape[-1])
+            cond_emb_spread = cond_emb.repeat(1, 1, code_emb.shape[-1])
             code_emb = self.conditioning_conv(torch.cat([cond_emb_spread, code_emb], dim=1))
 
         # Mask out the conditioning branch for whole batch elements, implementing something similar to classifier-free guidance.