- Notifications
You must be signed in to change notification settings - Fork 5.9k
/
Copy pathconvert_music_spectrogram_to_diffusers.py
203 lines (158 loc) · 10 KB
/
convert_music_spectrogram_to_diffusers.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
#!/usr/bin/env python3
importargparse
importos
importjaxasjnp
importnumpyasonp
importtorch
importtorch.nnasnn
frommusic_spectrogram_diffusionimportinference
fromt5ximportcheckpoints
fromdiffusersimportDDPMScheduler, OnnxRuntimeModel, SpectrogramDiffusionPipeline
fromdiffusers.pipelines.spectrogram_diffusionimportSpectrogramContEncoder, SpectrogramNotesEncoder, T5FilmDecoder
MODEL="base_with_context"
defload_notes_encoder(weights, model):
model.token_embedder.weight=nn.Parameter(torch.Tensor(weights["token_embedder"]["embedding"]))
model.position_encoding.weight=nn.Parameter(torch.Tensor(weights["Embed_0"]["embedding"]), requires_grad=False)
forlyr_num, lyrinenumerate(model.encoders):
ly_weight=weights[f"layers_{lyr_num}"]
lyr.layer[0].layer_norm.weight=nn.Parameter(torch.Tensor(ly_weight["pre_attention_layer_norm"]["scale"]))
attention_weights=ly_weight["attention"]
lyr.layer[0].SelfAttention.q.weight=nn.Parameter(torch.Tensor(attention_weights["query"]["kernel"].T))
lyr.layer[0].SelfAttention.k.weight=nn.Parameter(torch.Tensor(attention_weights["key"]["kernel"].T))
lyr.layer[0].SelfAttention.v.weight=nn.Parameter(torch.Tensor(attention_weights["value"]["kernel"].T))
lyr.layer[0].SelfAttention.o.weight=nn.Parameter(torch.Tensor(attention_weights["out"]["kernel"].T))
lyr.layer[1].layer_norm.weight=nn.Parameter(torch.Tensor(ly_weight["pre_mlp_layer_norm"]["scale"]))
lyr.layer[1].DenseReluDense.wi_0.weight=nn.Parameter(torch.Tensor(ly_weight["mlp"]["wi_0"]["kernel"].T))
lyr.layer[1].DenseReluDense.wi_1.weight=nn.Parameter(torch.Tensor(ly_weight["mlp"]["wi_1"]["kernel"].T))
lyr.layer[1].DenseReluDense.wo.weight=nn.Parameter(torch.Tensor(ly_weight["mlp"]["wo"]["kernel"].T))
model.layer_norm.weight=nn.Parameter(torch.Tensor(weights["encoder_norm"]["scale"]))
returnmodel
defload_continuous_encoder(weights, model):
model.input_proj.weight=nn.Parameter(torch.Tensor(weights["input_proj"]["kernel"].T))
model.position_encoding.weight=nn.Parameter(torch.Tensor(weights["Embed_0"]["embedding"]), requires_grad=False)
forlyr_num, lyrinenumerate(model.encoders):
ly_weight=weights[f"layers_{lyr_num}"]
attention_weights=ly_weight["attention"]
lyr.layer[0].SelfAttention.q.weight=nn.Parameter(torch.Tensor(attention_weights["query"]["kernel"].T))
lyr.layer[0].SelfAttention.k.weight=nn.Parameter(torch.Tensor(attention_weights["key"]["kernel"].T))
lyr.layer[0].SelfAttention.v.weight=nn.Parameter(torch.Tensor(attention_weights["value"]["kernel"].T))
lyr.layer[0].SelfAttention.o.weight=nn.Parameter(torch.Tensor(attention_weights["out"]["kernel"].T))
lyr.layer[0].layer_norm.weight=nn.Parameter(torch.Tensor(ly_weight["pre_attention_layer_norm"]["scale"]))
lyr.layer[1].DenseReluDense.wi_0.weight=nn.Parameter(torch.Tensor(ly_weight["mlp"]["wi_0"]["kernel"].T))
lyr.layer[1].DenseReluDense.wi_1.weight=nn.Parameter(torch.Tensor(ly_weight["mlp"]["wi_1"]["kernel"].T))
lyr.layer[1].DenseReluDense.wo.weight=nn.Parameter(torch.Tensor(ly_weight["mlp"]["wo"]["kernel"].T))
lyr.layer[1].layer_norm.weight=nn.Parameter(torch.Tensor(ly_weight["pre_mlp_layer_norm"]["scale"]))
model.layer_norm.weight=nn.Parameter(torch.Tensor(weights["encoder_norm"]["scale"]))
returnmodel
defload_decoder(weights, model):
model.conditioning_emb[0].weight=nn.Parameter(torch.Tensor(weights["time_emb_dense0"]["kernel"].T))
model.conditioning_emb[2].weight=nn.Parameter(torch.Tensor(weights["time_emb_dense1"]["kernel"].T))
model.position_encoding.weight=nn.Parameter(torch.Tensor(weights["Embed_0"]["embedding"]), requires_grad=False)
model.continuous_inputs_projection.weight=nn.Parameter(
torch.Tensor(weights["continuous_inputs_projection"]["kernel"].T)
)
forlyr_num, lyrinenumerate(model.decoders):
ly_weight=weights[f"layers_{lyr_num}"]
lyr.layer[0].layer_norm.weight=nn.Parameter(
torch.Tensor(ly_weight["pre_self_attention_layer_norm"]["scale"])
)
lyr.layer[0].FiLMLayer.scale_bias.weight=nn.Parameter(
torch.Tensor(ly_weight["FiLMLayer_0"]["DenseGeneral_0"]["kernel"].T)
)
attention_weights=ly_weight["self_attention"]
lyr.layer[0].attention.to_q.weight=nn.Parameter(torch.Tensor(attention_weights["query"]["kernel"].T))
lyr.layer[0].attention.to_k.weight=nn.Parameter(torch.Tensor(attention_weights["key"]["kernel"].T))
lyr.layer[0].attention.to_v.weight=nn.Parameter(torch.Tensor(attention_weights["value"]["kernel"].T))
lyr.layer[0].attention.to_out[0].weight=nn.Parameter(torch.Tensor(attention_weights["out"]["kernel"].T))
attention_weights=ly_weight["MultiHeadDotProductAttention_0"]
lyr.layer[1].attention.to_q.weight=nn.Parameter(torch.Tensor(attention_weights["query"]["kernel"].T))
lyr.layer[1].attention.to_k.weight=nn.Parameter(torch.Tensor(attention_weights["key"]["kernel"].T))
lyr.layer[1].attention.to_v.weight=nn.Parameter(torch.Tensor(attention_weights["value"]["kernel"].T))
lyr.layer[1].attention.to_out[0].weight=nn.Parameter(torch.Tensor(attention_weights["out"]["kernel"].T))
lyr.layer[1].layer_norm.weight=nn.Parameter(
torch.Tensor(ly_weight["pre_cross_attention_layer_norm"]["scale"])
)
lyr.layer[2].layer_norm.weight=nn.Parameter(torch.Tensor(ly_weight["pre_mlp_layer_norm"]["scale"]))
lyr.layer[2].film.scale_bias.weight=nn.Parameter(
torch.Tensor(ly_weight["FiLMLayer_1"]["DenseGeneral_0"]["kernel"].T)
)
lyr.layer[2].DenseReluDense.wi_0.weight=nn.Parameter(torch.Tensor(ly_weight["mlp"]["wi_0"]["kernel"].T))
lyr.layer[2].DenseReluDense.wi_1.weight=nn.Parameter(torch.Tensor(ly_weight["mlp"]["wi_1"]["kernel"].T))
lyr.layer[2].DenseReluDense.wo.weight=nn.Parameter(torch.Tensor(ly_weight["mlp"]["wo"]["kernel"].T))
model.decoder_norm.weight=nn.Parameter(torch.Tensor(weights["decoder_norm"]["scale"]))
model.spec_out.weight=nn.Parameter(torch.Tensor(weights["spec_out_dense"]["kernel"].T))
returnmodel
defmain(args):
t5_checkpoint=checkpoints.load_t5x_checkpoint(args.checkpoint_path)
t5_checkpoint=jnp.tree_util.tree_map(onp.array, t5_checkpoint)
gin_overrides= [
"from __gin__ import dynamic_registration",
"from music_spectrogram_diffusion.models.diffusion import diffusion_utils",
"diffusion_utils.ClassifierFreeGuidanceConfig.eval_condition_weight = 2.0",
"diffusion_utils.DiffusionConfig.classifier_free_guidance = @diffusion_utils.ClassifierFreeGuidanceConfig()",
]
gin_file=os.path.join(args.checkpoint_path, "..", "config.gin")
gin_config=inference.parse_training_gin_file(gin_file, gin_overrides)
synth_model=inference.InferenceModel(args.checkpoint_path, gin_config)
scheduler=DDPMScheduler(beta_schedule="squaredcos_cap_v2", variance_type="fixed_large")
notes_encoder=SpectrogramNotesEncoder(
max_length=synth_model.sequence_length["inputs"],
vocab_size=synth_model.model.module.config.vocab_size,
d_model=synth_model.model.module.config.emb_dim,
dropout_rate=synth_model.model.module.config.dropout_rate,
num_layers=synth_model.model.module.config.num_encoder_layers,
num_heads=synth_model.model.module.config.num_heads,
d_kv=synth_model.model.module.config.head_dim,
d_ff=synth_model.model.module.config.mlp_dim,
feed_forward_proj="gated-gelu",
)
continuous_encoder=SpectrogramContEncoder(
input_dims=synth_model.audio_codec.n_dims,
targets_context_length=synth_model.sequence_length["targets_context"],
d_model=synth_model.model.module.config.emb_dim,
dropout_rate=synth_model.model.module.config.dropout_rate,
num_layers=synth_model.model.module.config.num_encoder_layers,
num_heads=synth_model.model.module.config.num_heads,
d_kv=synth_model.model.module.config.head_dim,
d_ff=synth_model.model.module.config.mlp_dim,
feed_forward_proj="gated-gelu",
)
decoder=T5FilmDecoder(
input_dims=synth_model.audio_codec.n_dims,
targets_length=synth_model.sequence_length["targets_context"],
max_decoder_noise_time=synth_model.model.module.config.max_decoder_noise_time,
d_model=synth_model.model.module.config.emb_dim,
num_layers=synth_model.model.module.config.num_decoder_layers,
num_heads=synth_model.model.module.config.num_heads,
d_kv=synth_model.model.module.config.head_dim,
d_ff=synth_model.model.module.config.mlp_dim,
dropout_rate=synth_model.model.module.config.dropout_rate,
)
notes_encoder=load_notes_encoder(t5_checkpoint["target"]["token_encoder"], notes_encoder)
continuous_encoder=load_continuous_encoder(t5_checkpoint["target"]["continuous_encoder"], continuous_encoder)
decoder=load_decoder(t5_checkpoint["target"]["decoder"], decoder)
melgan=OnnxRuntimeModel.from_pretrained("kashif/soundstream_mel_decoder")
pipe=SpectrogramDiffusionPipeline(
notes_encoder=notes_encoder,
continuous_encoder=continuous_encoder,
decoder=decoder,
scheduler=scheduler,
melgan=melgan,
)
ifargs.save:
pipe.save_pretrained(args.output_path)
if__name__=="__main__":
parser=argparse.ArgumentParser()
parser.add_argument("--output_path", default=None, type=str, required=True, help="Path to the converted model.")
parser.add_argument(
"--save", default=True, type=bool, required=False, help="Whether to save the converted model or not."
)
parser.add_argument(
"--checkpoint_path",
default=f"{MODEL}/checkpoint_500000",
type=str,
required=False,
help="Path to the original jax model checkpoint.",
)
args=parser.parse_args()
main(args)
