deepanwayx commited on
Commit
2c55b55
1 Parent(s): 09e9063

update app

Browse files
Files changed (2) hide show
  1. .ipynb_checkpoints/app-checkpoint.py +283 -0
  2. app.py +16 -25
.ipynb_checkpoints/app-checkpoint.py ADDED
@@ -0,0 +1,283 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ import json
3
+ import torch
4
+ import wavio
5
+ import numpy as np
6
+ from tqdm import tqdm
7
+ from huggingface_hub import snapshot_download
8
+
9
+ from audioldm.audio.stft import TacotronSTFT
10
+ from audioldm.variational_autoencoder import AutoencoderKL
11
+
12
+ from transformers import AutoTokenizer, T5ForConditionalGeneration
13
+ from modelling_deberta_v2 import DebertaV2ForTokenClassificationRegression
14
+
15
+ import sys
16
+ sys.path.insert(0, "diffusers/src")
17
+
18
+ from diffusers import DDPMScheduler
19
+ from models import MusicAudioDiffusion
20
+
21
+ from gradio import Markdown
22
+
23
+ class MusicFeaturePredictor:
24
+ def __init__(self, path, device="cuda:0", cache_dir=None, local_files_only=False):
25
+ self.beats_tokenizer = AutoTokenizer.from_pretrained(
26
+ "microsoft/deberta-v3-large",
27
+ use_fast=False,
28
+ cache_dir=cache_dir,
29
+ local_files_only=local_files_only,
30
+ )
31
+ self.beats_model = DebertaV2ForTokenClassificationRegression.from_pretrained(
32
+ "microsoft/deberta-v3-large",
33
+ cache_dir=cache_dir,
34
+ local_files_only=local_files_only,
35
+ )
36
+ self.beats_model.eval()
37
+ self.beats_model.to(device)
38
+
39
+ beats_ckpt = f"{path}/beats/microsoft-deberta-v3-large.pt"
40
+ beats_weight = torch.load(beats_ckpt, map_location="cpu")
41
+ self.beats_model.load_state_dict(beats_weight)
42
+
43
+ self.chords_tokenizer = AutoTokenizer.from_pretrained(
44
+ "google/flan-t5-large",
45
+ cache_dir=cache_dir,
46
+ local_files_only=local_files_only,
47
+ )
48
+ self.chords_model = T5ForConditionalGeneration.from_pretrained(
49
+ "google/flan-t5-large",
50
+ cache_dir=cache_dir,
51
+ local_files_only=local_files_only,
52
+ )
53
+ self.chords_model.eval()
54
+ self.chords_model.to(device)
55
+
56
+ chords_ckpt = f"{path}/chords/flan-t5-large.bin"
57
+ chords_weight = torch.load(chords_ckpt, map_location="cpu")
58
+ self.chords_model.load_state_dict(chords_weight)
59
+
60
+ def generate_beats(self, prompt):
61
+ tokenized = self.beats_tokenizer(
62
+ prompt, max_length=512, padding=True, truncation=True, return_tensors="pt"
63
+ )
64
+ tokenized = {k: v.to(self.beats_model.device) for k, v in tokenized.items()}
65
+
66
+ with torch.no_grad():
67
+ out = self.beats_model(**tokenized)
68
+
69
+ max_beat = (
70
+ 1 + torch.argmax(out["logits"][:, 0, :], -1).detach().cpu().numpy()
71
+ ).tolist()[0]
72
+ intervals = (
73
+ out["values"][:, :, 0]
74
+ .detach()
75
+ .cpu()
76
+ .numpy()
77
+ .astype("float32")
78
+ .round(4)
79
+ .tolist()
80
+ )
81
+
82
+ intervals = np.cumsum(intervals)
83
+ predicted_beats_times = []
84
+ for t in intervals:
85
+ if t < 10:
86
+ predicted_beats_times.append(round(t, 2))
87
+ else:
88
+ break
89
+ predicted_beats_times = list(np.array(predicted_beats_times)[:50])
90
+
91
+ if len(predicted_beats_times) == 0:
92
+ predicted_beats = [[], []]
93
+ else:
94
+ beat_counts = []
95
+ for i in range(len(predicted_beats_times)):
96
+ beat_counts.append(float(1.0 + np.mod(i, max_beat)))
97
+ predicted_beats = [[predicted_beats_times, beat_counts]]
98
+
99
+ return max_beat, predicted_beats_times, predicted_beats
100
+
101
+ def generate(self, prompt):
102
+ max_beat, predicted_beats_times, predicted_beats = self.generate_beats(prompt)
103
+
104
+ chords_prompt = "Caption: {} \\n Timestamps: {} \\n Max Beat: {}".format(
105
+ prompt,
106
+ " , ".join([str(round(t, 2)) for t in predicted_beats_times]),
107
+ max_beat,
108
+ )
109
+
110
+ tokenized = self.chords_tokenizer(
111
+ chords_prompt,
112
+ max_length=512,
113
+ padding=True,
114
+ truncation=True,
115
+ return_tensors="pt",
116
+ )
117
+ tokenized = {k: v.to(self.chords_model.device) for k, v in tokenized.items()}
118
+
119
+ generated_chords = self.chords_model.generate(
120
+ input_ids=tokenized["input_ids"],
121
+ attention_mask=tokenized["attention_mask"],
122
+ min_length=8,
123
+ max_length=128,
124
+ num_beams=5,
125
+ early_stopping=True,
126
+ num_return_sequences=1,
127
+ )
128
+
129
+ generated_chords = self.chords_tokenizer.decode(
130
+ generated_chords[0],
131
+ skip_special_tokens=True,
132
+ clean_up_tokenization_spaces=True,
133
+ ).split(" n ")
134
+
135
+ predicted_chords, predicted_chords_times = [], []
136
+ for item in generated_chords:
137
+ c, ct = item.split(" at ")
138
+ predicted_chords.append(c)
139
+ predicted_chords_times.append(float(ct))
140
+
141
+ return predicted_beats, predicted_chords, predicted_chords_times
142
+
143
+
144
+ class Mustango:
145
+ def __init__(
146
+ self,
147
+ name="declare-lab/mustango",
148
+ device="cuda:0",
149
+ cache_dir=None,
150
+ local_files_only=False,
151
+ ):
152
+ path = snapshot_download(repo_id=name, cache_dir=cache_dir)
153
+
154
+ self.music_model = MusicFeaturePredictor(
155
+ path, device, cache_dir=cache_dir, local_files_only=local_files_only
156
+ )
157
+
158
+ vae_config = json.load(open(f"{path}/configs/vae_config.json"))
159
+ stft_config = json.load(open(f"{path}/configs/stft_config.json"))
160
+ main_config = json.load(open(f"{path}/configs/main_config.json"))
161
+
162
+ self.vae = AutoencoderKL(**vae_config).to(device)
163
+ self.stft = TacotronSTFT(**stft_config).to(device)
164
+ self.model = MusicAudioDiffusion(
165
+ main_config["text_encoder_name"],
166
+ main_config["scheduler_name"],
167
+ unet_model_config_path=f"{path}/configs/music_diffusion_model_config.json",
168
+ ).to(device)
169
+ self.model.device = device
170
+
171
+ vae_weights = torch.load(
172
+ f"{path}/vae/pytorch_model_vae.bin", map_location=device
173
+ )
174
+ stft_weights = torch.load(
175
+ f"{path}/stft/pytorch_model_stft.bin", map_location=device
176
+ )
177
+ main_weights = torch.load(
178
+ f"{path}/ldm/pytorch_model_ldm.bin", map_location=device
179
+ )
180
+
181
+ self.vae.load_state_dict(vae_weights)
182
+ self.stft.load_state_dict(stft_weights)
183
+ self.model.load_state_dict(main_weights)
184
+
185
+ print("Successfully loaded checkpoint from:", name)
186
+
187
+ self.vae.eval()
188
+ self.stft.eval()
189
+ self.model.eval()
190
+
191
+ self.scheduler = DDPMScheduler.from_pretrained(
192
+ main_config["scheduler_name"], subfolder="scheduler"
193
+ )
194
+
195
+ def generate(self, prompt, steps=100, guidance=3, samples=1, disable_progress=True):
196
+ """Genrate music for a single prompt string."""
197
+
198
+ with torch.no_grad():
199
+ beats, chords, chords_times = self.music_model.generate(prompt)
200
+ latents = self.model.inference(
201
+ [prompt],
202
+ beats,
203
+ [chords],
204
+ [chords_times],
205
+ self.scheduler,
206
+ steps,
207
+ guidance,
208
+ samples,
209
+ disable_progress,
210
+ )
211
+ mel = self.vae.decode_first_stage(latents)
212
+ wave = self.vae.decode_to_waveform(mel)
213
+
214
+ return wave[0]
215
+
216
+
217
+ # Initialize Mustango
218
+ if torch.cuda.is_available():
219
+ mustango = Mustango()
220
+ else:
221
+ mustango = Mustango(device="cpu")
222
+
223
+ # output_wave = mustango.generate("This techno song features a synth lead playing the main melody.", 5, 3, disable_progress=False)
224
+
225
+ def gradio_generate(prompt, steps, guidance):
226
+ output_wave = mustango.generate(prompt, steps, guidance)
227
+ # output_filename = f"{prompt.replace(' ', '_')}_{steps}_{guidance}"[:250] + ".wav"
228
+ output_filename = "temp.wav"
229
+ wavio.write(output_filename, output_wave, rate=16000, sampwidth=2)
230
+
231
+ return output_filename
232
+
233
+
234
+ title="Mustango: Toward Controllable Text-to-Music Generation"
235
+ description_text = """
236
+ <p><a href="https://huggingface.co/spaces/declare-lab/mustango/blob/main/app.py?duplicate=true"> <img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a> For faster inference without waiting in queue, you may duplicate the space and upgrade to a GPU in the settings. <br/><br/>
237
+ Generate music using Mustango by providing a text prompt.
238
+ <br/><br/> This is the demo for Mustango for controllable text to music generation: <a href="https://arxiv.org/abs/2311.08355">Read our paper.</a>
239
+ <p/>
240
+ """
241
+ #description_text = ""
242
+ # Gradio input and output components
243
+ input_text = gr.Textbox(lines=2, label="Prompt")
244
+ output_audio = gr.Audio(label="Generated Music", type="filepath")
245
+ denoising_steps = gr.Slider(minimum=100, maximum=200, value=100, step=1, label="Steps", interactive=True)
246
+ guidance_scale = gr.Slider(minimum=1, maximum=10, value=3, step=0.1, label="Guidance Scale", interactive=True)
247
+
248
+ # CSS styling for the Duplicate button
249
+ css = '''
250
+ #duplicate-button {
251
+ margin: auto;
252
+ color: white;
253
+ background: #1565c0;
254
+ border-radius: 100vh;
255
+ }
256
+ '''
257
+
258
+ # Gradio interface
259
+ gr_interface = gr.Interface(
260
+ fn=gradio_generate,
261
+ inputs=[input_text, denoising_steps, guidance_scale],
262
+ outputs=[output_audio],
263
+ description=description_text,
264
+ allow_flagging=False,
265
+ examples=[
266
+ ["This techno song features a synth lead playing the main melody. This is accompanied by programmed percussion playing a simple kick focused beat. The hi-hat is accented in an open position on the 3-and count of every bar. The synth plays the bass part with a voicing that sounds like a cello. This techno song can be played in a club. The chord sequence is Gm, A7, Eb, Bb, C, F, Gm. The beat counts to 2. The tempo of this song is 128.0 beats per minute. The key of this song is G minor.", 100, 3],
267
+ ["This is a new age piece. There is a flute playing the main melody with a lot of staccato notes. The rhythmic background consists of a medium tempo electronic drum beat with percussive elements all over the spectrum. There is a playful atmosphere to the piece. This piece can be used in the soundtrack of a children's TV show or an advertisement jingle.", 100, 3],
268
+ ["The song is an instrumental. The song is in medium tempo with a classical guitar playing a lilting melody in accompaniment style. The song is emotional and romantic. The song is a romantic instrumental song. The chord sequence is Gm, F6, Ebm. The time signature is 4/4. This song is in Adagio. The key of this song is G minor.", 100, 3],
269
+ ["This folk song features a female voice singing the main melody. This is accompanied by a tabla playing the percussion. A guitar strums chords. For most parts of the song, only one chord is played. At the last bar, a different chord is played. This song has minimal instruments. This song has a story-telling mood. This song can be played in a village scene in an Indian movie. The chord sequence is Bbm, Ab. The beat is 3. The tempo of this song is Allegro. The key of this song is Bb minor.", 100, 3],
270
+ ["This is a live performance of a classical music piece. There is an orchestra performing the piece with a violin lead playing the main melody. The atmosphere is sentimental and heart-touching. This piece could be playing in the background at a classy restaurant. The chord progression in this song is Am7, Gm, Dm, A7, Dm. The beat is 3. This song is in Largo. The key of this song is D minor.", 100, 3],
271
+ ["This is a techno piece with drums and beats and a leading melody. A synth plays chords. The music kicks off with a powerful and relentless drumbeat. Over the pounding beats, a leading melody emerges. In the middle of the song, a flock of seagulls flies over the venue and make loud bird sounds. It has strong danceability and can be played in a club. The tempo is 120 bpm. The chords played by the synth are Am, Cm, Dm, Gm.", 100, 3],
272
+ ],
273
+ cache_examples=True,
274
+ )
275
+
276
+ with gr.Blocks(css=css) as demo:
277
+ title=gr.HTML(f"<h1><center>{title}</center></h1>")
278
+ dupe = gr.DuplicateButton(value="Duplicate Space for private use", elem_id="duplicate-button")
279
+ gr_interface.render()
280
+
281
+
282
+ # Launch Gradio app
283
+ demo.queue().launch()
app.py CHANGED
@@ -144,13 +144,12 @@ class MusicFeaturePredictor:
144
  class Mustango:
145
  def __init__(
146
  self,
 
147
  device="cuda:0",
148
  cache_dir=None,
149
  local_files_only=False,
150
  ):
151
-
152
- path = snapshot_download(repo_id="declare-lab/mustango", cache_dir=cache_dir)
153
- pretrained_path = snapshot_download(repo_id="declare-lab/mustango-pretrained", cache_dir=cache_dir)
154
 
155
  self.music_model = MusicFeaturePredictor(
156
  path, device, cache_dir=cache_dir, local_files_only=local_files_only
@@ -175,16 +174,15 @@ class Mustango:
175
  stft_weights = torch.load(
176
  f"{path}/stft/pytorch_model_stft.bin", map_location=device
177
  )
178
- self.main_weights = torch.load(
179
  f"{path}/ldm/pytorch_model_ldm.bin", map_location=device
180
  )
181
-
182
- self.main_weights_pretrained = torch.load(
183
- f"{pretrained_path}/ldm/pytorch_model_ldm.bin", map_location=device
184
- )
185
 
186
  self.vae.load_state_dict(vae_weights)
187
  self.stft.load_state_dict(stft_weights)
 
 
 
188
 
189
  self.vae.eval()
190
  self.stft.eval()
@@ -194,13 +192,8 @@ class Mustango:
194
  main_config["scheduler_name"], subfolder="scheduler"
195
  )
196
 
197
- def generate(self, model_name, prompt, steps=100, guidance=3, samples=1, disable_progress=True):
198
  """Genrate music for a single prompt string."""
199
-
200
- if model_name == "declare-lab/mustango":
201
- self.model.load_state_dict(self.main_weights)
202
- else:
203
- self.model.load_state_dict(self.main_weights_pretrained)
204
 
205
  with torch.no_grad():
206
  beats, chords, chords_times = self.music_model.generate(prompt)
@@ -229,8 +222,8 @@ else:
229
 
230
  # output_wave = mustango.generate("This techno song features a synth lead playing the main melody.", 5, 3, disable_progress=False)
231
 
232
- def gradio_generate(model_name, prompt, steps, guidance):
233
- output_wave = mustango.generate(model_name, prompt, steps, guidance)
234
  # output_filename = f"{prompt.replace(' ', '_')}_{steps}_{guidance}"[:250] + ".wav"
235
  output_filename = "temp.wav"
236
  wavio.write(output_filename, output_wave, rate=16000, sampwidth=2)
@@ -251,8 +244,6 @@ input_text = gr.Textbox(lines=2, label="Prompt")
251
  output_audio = gr.Audio(label="Generated Music", type="filepath")
252
  denoising_steps = gr.Slider(minimum=100, maximum=200, value=100, step=1, label="Steps", interactive=True)
253
  guidance_scale = gr.Slider(minimum=1, maximum=10, value=3, step=0.1, label="Guidance Scale", interactive=True)
254
- model_name = gr.Radio(["declare-lab/mustango", "declare-lab/mustango-pretrained",], label="Choose a model type", value="declare-lab/mustango", type="value")
255
-
256
 
257
  # CSS styling for the Duplicate button
258
  css = '''
@@ -267,17 +258,17 @@ css = '''
267
  # Gradio interface
268
  gr_interface = gr.Interface(
269
  fn=gradio_generate,
270
- inputs=[model_name, input_text, denoising_steps, guidance_scale],
271
  outputs=[output_audio],
272
  description=description_text,
273
  allow_flagging=False,
274
  examples=[
275
- ["declare-lab/mustango", "This techno song features a synth lead playing the main melody. This is accompanied by programmed percussion playing a simple kick focused beat. The hi-hat is accented in an open position on the 3-and count of every bar. The synth plays the bass part with a voicing that sounds like a cello. This techno song can be played in a club. The chord sequence is Gm, A7, Eb, Bb, C, F, Gm. The beat counts to 2. The tempo of this song is 128.0 beats per minute. The key of this song is G minor.", 200, 3],
276
- ["declare-lab/mustango", "This is a new age piece. There is a flute playing the main melody with a lot of staccato notes. The rhythmic background consists of a medium tempo electronic drum beat with percussive elements all over the spectrum. There is a playful atmosphere to the piece. This piece can be used in the soundtrack of a children's TV show or an advertisement jingle.", 200, 3],
277
- ["declare-lab/mustango", "The song is an instrumental. The song is in medium tempo with a classical guitar playing a lilting melody in accompaniment style. The song is emotional and romantic. The song is a romantic instrumental song. The chord sequence is Gm, F6, Ebm. The time signature is 4/4. This song is in Adagio. The key of this song is G minor.", 200, 3],
278
- ["declare-lab/mustango", "This folk song features a female voice singing the main melody. This is accompanied by a tabla playing the percussion. A guitar strums chords. For most parts of the song, only one chord is played. At the last bar, a different chord is played. This song has minimal instruments. This song has a story-telling mood. This song can be played in a village scene in an Indian movie. The chord sequence is Bbm, Ab. The beat is 3. The tempo of this song is Allegro. The key of this song is Bb minor.", 200, 3],
279
- ["declare-lab/mustango", "This is a live performance of a classical music piece. There is an orchestra performing the piece with a violin lead playing the main melody. The atmosphere is sentimental and heart-touching. This piece could be playing in the background at a classy restaurant. The chord progression in this song is Am7, Gm, Dm, A7, Dm. The beat is 3. This song is in Largo. The key of this song is D minor.", 200, 3],
280
- ["declare-lab/mustango", "This is a techno piece with drums and beats and a leading melody. A synth plays chords. The music kicks off with a powerful and relentless drumbeat. Over the pounding beats, a leading melody emerges. In the middle of the song, a flock of seagulls flies over the venue and make loud bird sounds. It has strong danceability and can be played in a club. The tempo is 120 bpm. The chords played by the synth are Am, Cm, Dm, Gm.", 200, 3],
281
  ],
282
  cache_examples=True,
283
  )
 
144
  class Mustango:
145
  def __init__(
146
  self,
147
+ name="declare-lab/mustango",
148
  device="cuda:0",
149
  cache_dir=None,
150
  local_files_only=False,
151
  ):
152
+ path = snapshot_download(repo_id=name, cache_dir=cache_dir)
 
 
153
 
154
  self.music_model = MusicFeaturePredictor(
155
  path, device, cache_dir=cache_dir, local_files_only=local_files_only
 
174
  stft_weights = torch.load(
175
  f"{path}/stft/pytorch_model_stft.bin", map_location=device
176
  )
177
+ main_weights = torch.load(
178
  f"{path}/ldm/pytorch_model_ldm.bin", map_location=device
179
  )
 
 
 
 
180
 
181
  self.vae.load_state_dict(vae_weights)
182
  self.stft.load_state_dict(stft_weights)
183
+ self.model.load_state_dict(main_weights)
184
+
185
+ print("Successfully loaded checkpoint from:", name)
186
 
187
  self.vae.eval()
188
  self.stft.eval()
 
192
  main_config["scheduler_name"], subfolder="scheduler"
193
  )
194
 
195
+ def generate(self, prompt, steps=100, guidance=3, samples=1, disable_progress=True):
196
  """Genrate music for a single prompt string."""
 
 
 
 
 
197
 
198
  with torch.no_grad():
199
  beats, chords, chords_times = self.music_model.generate(prompt)
 
222
 
223
  # output_wave = mustango.generate("This techno song features a synth lead playing the main melody.", 5, 3, disable_progress=False)
224
 
225
+ def gradio_generate(prompt, steps, guidance):
226
+ output_wave = mustango.generate(prompt, steps, guidance)
227
  # output_filename = f"{prompt.replace(' ', '_')}_{steps}_{guidance}"[:250] + ".wav"
228
  output_filename = "temp.wav"
229
  wavio.write(output_filename, output_wave, rate=16000, sampwidth=2)
 
244
  output_audio = gr.Audio(label="Generated Music", type="filepath")
245
  denoising_steps = gr.Slider(minimum=100, maximum=200, value=100, step=1, label="Steps", interactive=True)
246
  guidance_scale = gr.Slider(minimum=1, maximum=10, value=3, step=0.1, label="Guidance Scale", interactive=True)
 
 
247
 
248
  # CSS styling for the Duplicate button
249
  css = '''
 
258
  # Gradio interface
259
  gr_interface = gr.Interface(
260
  fn=gradio_generate,
261
+ inputs=[input_text, denoising_steps, guidance_scale],
262
  outputs=[output_audio],
263
  description=description_text,
264
  allow_flagging=False,
265
  examples=[
266
+ ["This techno song features a synth lead playing the main melody. This is accompanied by programmed percussion playing a simple kick focused beat. The hi-hat is accented in an open position on the 3-and count of every bar. The synth plays the bass part with a voicing that sounds like a cello. This techno song can be played in a club. The chord sequence is Gm, A7, Eb, Bb, C, F, Gm. The beat counts to 2. The tempo of this song is 128.0 beats per minute. The key of this song is G minor.", 100, 3],
267
+ ["This is a new age piece. There is a flute playing the main melody with a lot of staccato notes. The rhythmic background consists of a medium tempo electronic drum beat with percussive elements all over the spectrum. There is a playful atmosphere to the piece. This piece can be used in the soundtrack of a children's TV show or an advertisement jingle.", 100, 3],
268
+ ["The song is an instrumental. The song is in medium tempo with a classical guitar playing a lilting melody in accompaniment style. The song is emotional and romantic. The song is a romantic instrumental song. The chord sequence is Gm, F6, Ebm. The time signature is 4/4. This song is in Adagio. The key of this song is G minor.", 100, 3],
269
+ ["This folk song features a female voice singing the main melody. This is accompanied by a tabla playing the percussion. A guitar strums chords. For most parts of the song, only one chord is played. At the last bar, a different chord is played. This song has minimal instruments. This song has a story-telling mood. This song can be played in a village scene in an Indian movie. The chord sequence is Bbm, Ab. The beat is 3. The tempo of this song is Allegro. The key of this song is Bb minor.", 100, 3],
270
+ ["This is a live performance of a classical music piece. There is an orchestra performing the piece with a violin lead playing the main melody. The atmosphere is sentimental and heart-touching. This piece could be playing in the background at a classy restaurant. The chord progression in this song is Am7, Gm, Dm, A7, Dm. The beat is 3. This song is in Largo. The key of this song is D minor.", 100, 3],
271
+ ["This is a techno piece with drums and beats and a leading melody. A synth plays chords. The music kicks off with a powerful and relentless drumbeat. Over the pounding beats, a leading melody emerges. In the middle of the song, a flock of seagulls flies over the venue and make loud bird sounds. It has strong danceability and can be played in a club. The tempo is 120 bpm. The chords played by the synth are Am, Cm, Dm, Gm.", 100, 3],
272
  ],
273
  cache_examples=True,
274
  )