Unverified Commit ef02f1c2 authored by Phil Wang's avatar Phil Wang Committed by GitHub
Browse files

Merge pull request #21 from eonglints/depth-and-heads-options 

Depth and heads options for initializing transformers
parents 06eecbdb e67d4661

audiolm_pytorch/audiolm_pytorch.py

+13 −6
Original line number Diff line number Diff line
@@ -283,6 +283,7 @@ class Transformer(nn.Module): 
        *,

        dim,

        depth,

        heads,

        dim_context = None,

        cross_attend = False,

        grad_shrink_alpha = 0.1,
@@ -293,12 +294,12 @@ class Transformer(nn.Module): 


        self.layers = nn.ModuleList([])



        self.rel_pos_bias = RelativePositionBias()

        self.rel_pos_bias = RelativePositionBias(heads = heads)



        for _ in range(depth):

            self.layers.append(nn.ModuleList([

                Attention(dim = dim, causal = True, **kwargs),

                Attention(dim = dim, dim_context = dim_context, num_null_kv = 1, norm_context = True, **kwargs) if cross_attend else None,

                Attention(dim = dim, heads = heads, causal = True, **kwargs),

                Attention(dim = dim, heads = heads, dim_context = dim_context, num_null_kv = 1, norm_context = True, **kwargs) if cross_attend else None,

                FeedForward(dim = dim)

            ]))
@@ -337,6 +338,8 @@ class SemanticTransformer(nn.Module): 
        self,

        *,

        dim,

        depth,

        heads,

        num_semantic_tokens,

        t5_name = DEFAULT_T5_NAME,

        has_condition = False,
@@ -358,7 +361,7 @@ class SemanticTransformer(nn.Module): 
        self.eos_id = num_semantic_tokens

        self.pad_id = pad_id



        self.transformer = Transformer(dim = dim, dim_context = get_encoded_dim(t5_name), cross_attend = has_condition, grad_shrink_alpha = grad_shrink_alpha, **kwargs)

        self.transformer = Transformer(dim = dim, depth = depth, heads = heads, dim_context = get_encoded_dim(t5_name), cross_attend = has_condition, grad_shrink_alpha = grad_shrink_alpha, **kwargs)

        self.to_logits = nn.Linear(dim, num_semantic_tokens + 1)



    @property
@@ -428,6 +431,8 @@ class CoarseTransformer(nn.Module): 
        codebook_size,

        num_coarse_quantizers,

        dim,

        depth,

        heads,

        num_semantic_tokens,

        t5_name = DEFAULT_T5_NAME,

        has_condition = False,
@@ -450,7 +455,7 @@ class CoarseTransformer(nn.Module): 
        codebook_size_with_eos = codebook_size + 1

        self.coarse_embedding = nn.Embedding(num_coarse_quantizers * codebook_size_with_eos, dim)



        self.transformer = Transformer(dim = dim, dim_context = get_encoded_dim(t5_name), cross_attend = has_condition, grad_shrink_alpha = grad_shrink_alpha, **kwargs)

        self.transformer = Transformer(dim = dim, depth = depth, heads = heads, dim_context = get_encoded_dim(t5_name), cross_attend = has_condition, grad_shrink_alpha = grad_shrink_alpha, **kwargs)



        self.codebook_size = codebook_size

        self.num_coarse_quantizers = num_coarse_quantizers
@@ -574,6 +579,8 @@ class FineTransformer(nn.Module): 
        num_fine_quantizers,

        codebook_size,

        dim,

        depth,

        heads,

        t5_name = DEFAULT_T5_NAME,

        has_condition = False,

        cond_drop_prob = 0.5,
@@ -597,7 +604,7 @@ class FineTransformer(nn.Module): 


        self.eos_id = codebook_size



        self.transformer = Transformer(dim = dim, dim_context = get_encoded_dim(t5_name), cross_attend = has_condition, grad_shrink_alpha = grad_shrink_alpha, **kwargs)

        self.transformer = Transformer(dim = dim, depth = depth, heads = heads, dim_context = get_encoded_dim(t5_name), cross_attend = has_condition, grad_shrink_alpha = grad_shrink_alpha, **kwargs)



        self.codebook_size = codebook_size

        self.num_coarse_quantizers = num_coarse_quantizers