Update inference.py
Browse files- inference.py +388 -388
inference.py
CHANGED
@@ -1,388 +1,388 @@
|
|
1 |
-
import os
|
2 |
-
import time
|
3 |
-
import torch
|
4 |
-
import re
|
5 |
-
import difflib
|
6 |
-
from utils import *
|
7 |
-
from config import *
|
8 |
-
from transformers import GPT2Config
|
9 |
-
from abctoolkit.utils import Exclaim_re, Quote_re, SquareBracket_re, Barline_regexPattern
|
10 |
-
from abctoolkit.transpose import Note_list, Pitch_sign_list
|
11 |
-
from abctoolkit.duration import calculate_bartext_duration
|
12 |
-
import requests
|
13 |
-
import torch
|
14 |
-
from huggingface_hub import hf_hub_download
|
15 |
-
import logging
|
16 |
-
|
17 |
-
# Setup logging
|
18 |
-
logging.basicConfig(level=logging.INFO)
|
19 |
-
logger = logging.getLogger(__name__)
|
20 |
-
|
21 |
-
Note_list = Note_list + ['z', 'x']
|
22 |
-
|
23 |
-
if torch.cuda.is_available():
|
24 |
-
device = torch.device("cuda")
|
25 |
-
else:
|
26 |
-
device = torch.device("cpu")
|
27 |
-
|
28 |
-
patchilizer = Patchilizer()
|
29 |
-
|
30 |
-
patch_config = GPT2Config(num_hidden_layers=PATCH_NUM_LAYERS,
|
31 |
-
max_length=PATCH_LENGTH,
|
32 |
-
max_position_embeddings=PATCH_LENGTH,
|
33 |
-
n_embd=HIDDEN_SIZE,
|
34 |
-
num_attention_heads=HIDDEN_SIZE // 64,
|
35 |
-
vocab_size=1)
|
36 |
-
byte_config = GPT2Config(num_hidden_layers=CHAR_NUM_LAYERS,
|
37 |
-
max_length=PATCH_SIZE + 1,
|
38 |
-
max_position_embeddings=PATCH_SIZE + 1,
|
39 |
-
hidden_size=HIDDEN_SIZE,
|
40 |
-
num_attention_heads=HIDDEN_SIZE // 64,
|
41 |
-
vocab_size=128)
|
42 |
-
|
43 |
-
model = NotaGenLMHeadModel(encoder_config=patch_config, decoder_config=byte_config).to(device)
|
44 |
-
|
45 |
-
|
46 |
-
def download_model_weights():
|
47 |
-
weights_path = "weights_notagenx_p_size_16_p_length_1024_p_layers_20_h_size_1280.pth"
|
48 |
-
local_weights_path = os.path.join(os.getcwd(), weights_path)
|
49 |
-
|
50 |
-
# Check if weights already exist locally
|
51 |
-
if os.path.exists(local_weights_path):
|
52 |
-
logger.info(f"Model weights already exist at {local_weights_path}")
|
53 |
-
return local_weights_path
|
54 |
-
|
55 |
-
logger.info("Downloading model weights from HuggingFace Hub...")
|
56 |
-
try:
|
57 |
-
# Download from HuggingFace
|
58 |
-
downloaded_path = hf_hub_download(
|
59 |
-
repo_id="ElectricAlexis/NotaGen",
|
60 |
-
filename=weights_path,
|
61 |
-
local_dir=os.getcwd(),
|
62 |
-
local_dir_use_symlinks=False
|
63 |
-
)
|
64 |
-
logger.info(f"Model weights downloaded successfully to {downloaded_path}")
|
65 |
-
return downloaded_path
|
66 |
-
except Exception as e:
|
67 |
-
logger.error(f"Error downloading model weights: {str(e)}")
|
68 |
-
raise RuntimeError(f"Failed to download model weights: {str(e)}")
|
69 |
-
|
70 |
-
|
71 |
-
def prepare_model_for_kbit_training(model, use_gradient_checkpointing=True):
|
72 |
-
"""
|
73 |
-
Prepare model for k-bit training.
|
74 |
-
Features include:
|
75 |
-
1. Convert model to mixed precision (FP16).
|
76 |
-
2. Disable unnecessary gradient computations.
|
77 |
-
3. Enable gradient checkpointing (optional).
|
78 |
-
"""
|
79 |
-
# Convert model to mixed precision
|
80 |
-
model = model.to(dtype=torch.float16)
|
81 |
-
|
82 |
-
# Disable gradients for embedding layers
|
83 |
-
for param in model.parameters():
|
84 |
-
if param.dtype == torch.float32:
|
85 |
-
param.requires_grad = False
|
86 |
-
|
87 |
-
# Enable gradient checkpointing
|
88 |
-
if use_gradient_checkpointing:
|
89 |
-
model.gradient_checkpointing_enable()
|
90 |
-
|
91 |
-
return model
|
92 |
-
|
93 |
-
|
94 |
-
model = prepare_model_for_kbit_training(
|
95 |
-
model,
|
96 |
-
use_gradient_checkpointing=False
|
97 |
-
)
|
98 |
-
|
99 |
-
print("Parameter Number: " + str(sum(p.numel() for p in model.parameters() if p.requires_grad)))
|
100 |
-
|
101 |
-
# Download weights at startup
|
102 |
-
model_weights_path = download_model_weights()
|
103 |
-
checkpoint = torch.load(model_weights_path, map_location=torch.device(device))
|
104 |
-
model.load_state_dict(checkpoint['model'], strict=False)
|
105 |
-
|
106 |
-
model = model.to(device)
|
107 |
-
model.eval()
|
108 |
-
|
109 |
-
|
110 |
-
def postprocess_inst_names(abc_text):
|
111 |
-
with open('standard_inst_names.txt', 'r', encoding='utf-8') as f:
|
112 |
-
standard_instruments_list = [line.strip() for line in f if line.strip()]
|
113 |
-
|
114 |
-
with open('instrument_mapping.json', 'r', encoding='utf-8') as f:
|
115 |
-
instrument_mapping = json.load(f)
|
116 |
-
|
117 |
-
abc_lines = abc_text.split('\n')
|
118 |
-
abc_lines = list(filter(None, abc_lines))
|
119 |
-
abc_lines = [line + '\n' for line in abc_lines]
|
120 |
-
|
121 |
-
for i, line in enumerate(abc_lines):
|
122 |
-
if line.startswith('V:') and 'nm=' in line:
|
123 |
-
match = re.search(r'nm="([^"]*)"', line)
|
124 |
-
if match:
|
125 |
-
inst_name = match.group(1)
|
126 |
-
|
127 |
-
# Check if the instrument name is already standard
|
128 |
-
if inst_name in standard_instruments_list:
|
129 |
-
continue
|
130 |
-
|
131 |
-
# Find the most similar key in instrument_mapping
|
132 |
-
matching_key = difflib.get_close_matches(inst_name, list(instrument_mapping.keys()), n=1, cutoff=0.6)
|
133 |
-
|
134 |
-
if matching_key:
|
135 |
-
# Replace the instrument name with the standardized version
|
136 |
-
replacement = instrument_mapping[matching_key[0]]
|
137 |
-
new_line = line.replace(f'nm="{inst_name}"', f'nm="{replacement}"')
|
138 |
-
abc_lines[i] = new_line
|
139 |
-
|
140 |
-
# Combine the lines back into a single string
|
141 |
-
processed_abc_text = ''.join(abc_lines)
|
142 |
-
return processed_abc_text
|
143 |
-
|
144 |
-
|
145 |
-
def complete_brackets(s):
|
146 |
-
stack = []
|
147 |
-
bracket_map = {'{': '}', '[': ']', '(': ')'}
|
148 |
-
|
149 |
-
# Iterate through each character, handle bracket matching
|
150 |
-
for char in s:
|
151 |
-
if char in bracket_map:
|
152 |
-
stack.append(char)
|
153 |
-
elif char in bracket_map.values():
|
154 |
-
# Find the corresponding left bracket
|
155 |
-
for key, value in bracket_map.items():
|
156 |
-
if value == char:
|
157 |
-
if stack and stack[-1] == key:
|
158 |
-
stack.pop()
|
159 |
-
break # Found matching right bracket, process next character
|
160 |
-
|
161 |
-
# Complete missing right brackets (in reverse order of remaining left brackets in stack)
|
162 |
-
completion = ''.join(bracket_map[c] for c in reversed(stack))
|
163 |
-
return s + completion
|
164 |
-
|
165 |
-
|
166 |
-
def rest_unreduce(abc_lines):
|
167 |
-
tunebody_index = None
|
168 |
-
for i in range(len(abc_lines)):
|
169 |
-
if abc_lines[i].startswith('%%score'):
|
170 |
-
abc_lines[i] = complete_brackets(abc_lines[i])
|
171 |
-
if '[V:' in abc_lines[i]:
|
172 |
-
tunebody_index = i
|
173 |
-
break
|
174 |
-
|
175 |
-
metadata_lines = abc_lines[: tunebody_index]
|
176 |
-
tunebody_lines = abc_lines[tunebody_index:]
|
177 |
-
|
178 |
-
part_symbol_list = []
|
179 |
-
voice_group_list = []
|
180 |
-
for line in metadata_lines:
|
181 |
-
if line.startswith('%%score'):
|
182 |
-
for round_bracket_match in re.findall(r'\((.*?)\)', line):
|
183 |
-
voice_group_list.append(round_bracket_match.split())
|
184 |
-
existed_voices = [item for sublist in voice_group_list for item in sublist]
|
185 |
-
if line.startswith('V:'):
|
186 |
-
symbol = line.split()[0]
|
187 |
-
part_symbol_list.append(symbol)
|
188 |
-
if symbol[2:] not in existed_voices:
|
189 |
-
voice_group_list.append([symbol[2:]])
|
190 |
-
z_symbol_list = [] # voices that use z as rest
|
191 |
-
x_symbol_list = [] # voices that use x as rest
|
192 |
-
for voice_group in voice_group_list:
|
193 |
-
z_symbol_list.append('V:' + voice_group[0])
|
194 |
-
for j in range(1, len(voice_group)):
|
195 |
-
x_symbol_list.append('V:' + voice_group[j])
|
196 |
-
|
197 |
-
part_symbol_list.sort(key=lambda x: int(x[2:]))
|
198 |
-
|
199 |
-
unreduced_tunebody_lines = []
|
200 |
-
|
201 |
-
for i, line in enumerate(tunebody_lines):
|
202 |
-
unreduced_line = ''
|
203 |
-
|
204 |
-
line = re.sub(r'^\[r:[^\]]*\]', '', line)
|
205 |
-
|
206 |
-
pattern = r'\[V:(\d+)\](.*?)(?=\[V:|$)'
|
207 |
-
matches = re.findall(pattern, line)
|
208 |
-
|
209 |
-
line_bar_dict = {}
|
210 |
-
for match in matches:
|
211 |
-
key = f'V:{match[0]}'
|
212 |
-
value = match[1]
|
213 |
-
line_bar_dict[key] = value
|
214 |
-
|
215 |
-
# calculate duration and collect barline
|
216 |
-
dur_dict = {}
|
217 |
-
for symbol, bartext in line_bar_dict.items():
|
218 |
-
right_barline = ''.join(re.split(Barline_regexPattern, bartext)[-2:])
|
219 |
-
bartext = bartext[:-len(right_barline)]
|
220 |
-
try:
|
221 |
-
bar_dur = calculate_bartext_duration(bartext)
|
222 |
-
except:
|
223 |
-
bar_dur = None
|
224 |
-
if bar_dur is not None:
|
225 |
-
if bar_dur not in dur_dict.keys():
|
226 |
-
dur_dict[bar_dur] = 1
|
227 |
-
else:
|
228 |
-
dur_dict[bar_dur] += 1
|
229 |
-
|
230 |
-
try:
|
231 |
-
ref_dur = max(dur_dict, key=dur_dict.get)
|
232 |
-
except:
|
233 |
-
pass # use last ref_dur
|
234 |
-
|
235 |
-
if i == 0:
|
236 |
-
prefix_left_barline = line.split('[V:')[0]
|
237 |
-
else:
|
238 |
-
prefix_left_barline = ''
|
239 |
-
|
240 |
-
for symbol in part_symbol_list:
|
241 |
-
if symbol in line_bar_dict.keys():
|
242 |
-
symbol_bartext = line_bar_dict[symbol]
|
243 |
-
else:
|
244 |
-
if symbol in z_symbol_list:
|
245 |
-
symbol_bartext = prefix_left_barline + 'z' + str(ref_dur) + right_barline
|
246 |
-
elif symbol in x_symbol_list:
|
247 |
-
symbol_bartext = prefix_left_barline + 'x' + str(ref_dur) + right_barline
|
248 |
-
unreduced_line += '[' + symbol + ']' + symbol_bartext
|
249 |
-
|
250 |
-
unreduced_tunebody_lines.append(unreduced_line + '\n')
|
251 |
-
|
252 |
-
unreduced_lines = metadata_lines + unreduced_tunebody_lines
|
253 |
-
|
254 |
-
return unreduced_lines
|
255 |
-
|
256 |
-
|
257 |
-
def inference_patch(period, composer, instrumentation):
|
258 |
-
prompt_lines = [
|
259 |
-
'%' + period + '\n',
|
260 |
-
'%' + composer + '\n',
|
261 |
-
'%' + instrumentation + '\n']
|
262 |
-
|
263 |
-
while True:
|
264 |
-
|
265 |
-
failure_flag = False
|
266 |
-
|
267 |
-
bos_patch = [patchilizer.bos_token_id] * (PATCH_SIZE - 1) + [patchilizer.eos_token_id]
|
268 |
-
|
269 |
-
start_time = time.time()
|
270 |
-
|
271 |
-
prompt_patches = patchilizer.patchilize_metadata(prompt_lines)
|
272 |
-
byte_list = list(''.join(prompt_lines))
|
273 |
-
context_tunebody_byte_list = []
|
274 |
-
metadata_byte_list = []
|
275 |
-
|
276 |
-
print(''.join(byte_list), end='')
|
277 |
-
|
278 |
-
prompt_patches = [[ord(c) for c in patch] + [patchilizer.special_token_id] * (PATCH_SIZE - len(patch)) for patch
|
279 |
-
in prompt_patches]
|
280 |
-
prompt_patches.insert(0, bos_patch)
|
281 |
-
|
282 |
-
input_patches = torch.tensor(prompt_patches, device=device).reshape(1, -1)
|
283 |
-
|
284 |
-
end_flag = False
|
285 |
-
cut_index = None
|
286 |
-
|
287 |
-
tunebody_flag = False
|
288 |
-
|
289 |
-
with torch.inference_mode():
|
290 |
-
|
291 |
-
while True:
|
292 |
-
with torch.autocast(device_type='cuda', dtype=torch.float16):
|
293 |
-
predicted_patch = model.generate(input_patches.unsqueeze(0),
|
294 |
-
top_k=TOP_K,
|
295 |
-
top_p=TOP_P,
|
296 |
-
temperature=TEMPERATURE)
|
297 |
-
if not tunebody_flag and patchilizer.decode([predicted_patch]).startswith(
|
298 |
-
'[r:'): # 初次进入tunebody,必须以[r:0/开头
|
299 |
-
tunebody_flag = True
|
300 |
-
r0_patch = torch.tensor([ord(c) for c in '[r:0/']).unsqueeze(0).to(device)
|
301 |
-
temp_input_patches = torch.concat([input_patches, r0_patch], axis=-1)
|
302 |
-
predicted_patch = model.generate(temp_input_patches.unsqueeze(0),
|
303 |
-
top_k=TOP_K,
|
304 |
-
top_p=TOP_P,
|
305 |
-
temperature=TEMPERATURE)
|
306 |
-
predicted_patch = [ord(c) for c in '[r:0/'] + predicted_patch
|
307 |
-
if predicted_patch[0] == patchilizer.bos_token_id and predicted_patch[1] == patchilizer.eos_token_id:
|
308 |
-
end_flag = True
|
309 |
-
break
|
310 |
-
next_patch = patchilizer.decode([predicted_patch])
|
311 |
-
|
312 |
-
for char in next_patch:
|
313 |
-
byte_list.append(char)
|
314 |
-
if tunebody_flag:
|
315 |
-
context_tunebody_byte_list.append(char)
|
316 |
-
else:
|
317 |
-
metadata_byte_list.append(char)
|
318 |
-
print(char, end='')
|
319 |
-
|
320 |
-
patch_end_flag = False
|
321 |
-
for j in range(len(predicted_patch)):
|
322 |
-
if patch_end_flag:
|
323 |
-
predicted_patch[j] = patchilizer.special_token_id
|
324 |
-
if predicted_patch[j] == patchilizer.eos_token_id:
|
325 |
-
patch_end_flag = True
|
326 |
-
|
327 |
-
predicted_patch = torch.tensor([predicted_patch], device=device) # (1, 16)
|
328 |
-
input_patches = torch.cat([input_patches, predicted_patch], dim=1) # (1, 16 * patch_len)
|
329 |
-
|
330 |
-
if len(byte_list) > 102400:
|
331 |
-
failure_flag = True
|
332 |
-
break
|
333 |
-
if time.time() - start_time > 10 * 60:
|
334 |
-
failure_flag = True
|
335 |
-
break
|
336 |
-
|
337 |
-
if input_patches.shape[1] >= PATCH_LENGTH * PATCH_SIZE and not end_flag:
|
338 |
-
print('Stream generating...')
|
339 |
-
|
340 |
-
metadata = ''.join(metadata_byte_list)
|
341 |
-
context_tunebody = ''.join(context_tunebody_byte_list)
|
342 |
-
|
343 |
-
if '\n' not in context_tunebody:
|
344 |
-
break # Generated content is all metadata, abandon
|
345 |
-
|
346 |
-
context_tunebody_lines = context_tunebody.strip().split('\n')
|
347 |
-
|
348 |
-
if not context_tunebody.endswith('\n'):
|
349 |
-
context_tunebody_lines = [context_tunebody_lines[i] + '\n' for i in
|
350 |
-
range(len(context_tunebody_lines) - 1)] + [context_tunebody_lines[-1]]
|
351 |
-
else:
|
352 |
-
context_tunebody_lines = [context_tunebody_lines[i] + '\n' for i in
|
353 |
-
range(len(context_tunebody_lines))]
|
354 |
-
|
355 |
-
cut_index = len(context_tunebody_lines) // 2
|
356 |
-
abc_code_slice = metadata + ''.join(context_tunebody_lines[-cut_index:])
|
357 |
-
|
358 |
-
input_patches = patchilizer.encode_generate(abc_code_slice)
|
359 |
-
|
360 |
-
input_patches = [item for sublist in input_patches for item in sublist]
|
361 |
-
input_patches = torch.tensor([input_patches], device=device)
|
362 |
-
input_patches = input_patches.reshape(1, -1)
|
363 |
-
|
364 |
-
context_tunebody_byte_list = list(''.join(context_tunebody_lines[-cut_index:]))
|
365 |
-
|
366 |
-
if not failure_flag:
|
367 |
-
abc_text = ''.join(byte_list)
|
368 |
-
|
369 |
-
# unreduce
|
370 |
-
abc_lines = abc_text.split('\n')
|
371 |
-
abc_lines = list(filter(None, abc_lines))
|
372 |
-
abc_lines = [line + '\n' for line in abc_lines]
|
373 |
-
try:
|
374 |
-
unreduced_abc_lines = rest_unreduce(abc_lines)
|
375 |
-
except:
|
376 |
-
failure_flag = True
|
377 |
-
pass
|
378 |
-
else:
|
379 |
-
unreduced_abc_lines = [line for line in unreduced_abc_lines if
|
380 |
-
not (line.startswith('%') and not line.startswith('%%'))]
|
381 |
-
unreduced_abc_lines = ['X:1\n'] + unreduced_abc_lines
|
382 |
-
unreduced_abc_text = ''.join(unreduced_abc_lines)
|
383 |
-
return unreduced_abc_text
|
384 |
-
|
385 |
-
|
386 |
-
if __name__ == '__main__':
|
387 |
-
inference_patch('Classical', 'Beethoven, Ludwig van', 'Orchestral')
|
388 |
-
|
|
|
1 |
+
import os
|
2 |
+
import time
|
3 |
+
import torch
|
4 |
+
import re
|
5 |
+
import difflib
|
6 |
+
from utils import *
|
7 |
+
from config import *
|
8 |
+
from transformers import GPT2Config
|
9 |
+
from abctoolkit.utils import Exclaim_re, Quote_re, SquareBracket_re, Barline_regexPattern
|
10 |
+
from abctoolkit.transpose import Note_list, Pitch_sign_list
|
11 |
+
from abctoolkit.duration import calculate_bartext_duration
|
12 |
+
import requests
|
13 |
+
import torch
|
14 |
+
from huggingface_hub import hf_hub_download
|
15 |
+
import logging
|
16 |
+
|
17 |
+
# Setup logging
|
18 |
+
logging.basicConfig(level=logging.INFO)
|
19 |
+
logger = logging.getLogger(__name__)
|
20 |
+
|
21 |
+
Note_list = Note_list + ['z', 'x']
|
22 |
+
|
23 |
+
if torch.cuda.is_available():
|
24 |
+
device = torch.device("cuda")
|
25 |
+
else:
|
26 |
+
device = torch.device("cpu")
|
27 |
+
|
28 |
+
patchilizer = Patchilizer()
|
29 |
+
|
30 |
+
patch_config = GPT2Config(num_hidden_layers=PATCH_NUM_LAYERS,
|
31 |
+
max_length=PATCH_LENGTH,
|
32 |
+
max_position_embeddings=PATCH_LENGTH,
|
33 |
+
n_embd=HIDDEN_SIZE,
|
34 |
+
num_attention_heads=HIDDEN_SIZE // 64,
|
35 |
+
vocab_size=1)
|
36 |
+
byte_config = GPT2Config(num_hidden_layers=CHAR_NUM_LAYERS,
|
37 |
+
max_length=PATCH_SIZE + 1,
|
38 |
+
max_position_embeddings=PATCH_SIZE + 1,
|
39 |
+
hidden_size=HIDDEN_SIZE,
|
40 |
+
num_attention_heads=HIDDEN_SIZE // 64,
|
41 |
+
vocab_size=128)
|
42 |
+
|
43 |
+
model = NotaGenLMHeadModel(encoder_config=patch_config, decoder_config=byte_config).to(device)
|
44 |
+
|
45 |
+
|
46 |
+
def download_model_weights():
|
47 |
+
weights_path = "weights_notagenx_p_size_16_p_length_1024_p_layers_20_h_size_1280.pth"
|
48 |
+
local_weights_path = os.path.join(os.getcwd(), weights_path)
|
49 |
+
|
50 |
+
# Check if weights already exist locally
|
51 |
+
if os.path.exists(local_weights_path):
|
52 |
+
logger.info(f"Model weights already exist at {local_weights_path}")
|
53 |
+
return local_weights_path
|
54 |
+
|
55 |
+
logger.info("Downloading model weights from HuggingFace Hub...")
|
56 |
+
try:
|
57 |
+
# Download from HuggingFace
|
58 |
+
downloaded_path = hf_hub_download(
|
59 |
+
repo_id="ElectricAlexis/NotaGen",
|
60 |
+
filename=weights_path,
|
61 |
+
local_dir=os.getcwd(),
|
62 |
+
local_dir_use_symlinks=False
|
63 |
+
)
|
64 |
+
logger.info(f"Model weights downloaded successfully to {downloaded_path}")
|
65 |
+
return downloaded_path
|
66 |
+
except Exception as e:
|
67 |
+
logger.error(f"Error downloading model weights: {str(e)}")
|
68 |
+
raise RuntimeError(f"Failed to download model weights: {str(e)}")
|
69 |
+
|
70 |
+
|
71 |
+
def prepare_model_for_kbit_training(model, use_gradient_checkpointing=True):
|
72 |
+
"""
|
73 |
+
Prepare model for k-bit training.
|
74 |
+
Features include:
|
75 |
+
1. Convert model to mixed precision (FP16).
|
76 |
+
2. Disable unnecessary gradient computations.
|
77 |
+
3. Enable gradient checkpointing (optional).
|
78 |
+
"""
|
79 |
+
# Convert model to mixed precision
|
80 |
+
model = model.to(dtype=torch.float16)
|
81 |
+
|
82 |
+
# Disable gradients for embedding layers
|
83 |
+
for param in model.parameters():
|
84 |
+
if param.dtype == torch.float32:
|
85 |
+
param.requires_grad = False
|
86 |
+
|
87 |
+
# Enable gradient checkpointing
|
88 |
+
if use_gradient_checkpointing:
|
89 |
+
model.gradient_checkpointing_enable()
|
90 |
+
|
91 |
+
return model
|
92 |
+
|
93 |
+
|
94 |
+
model = prepare_model_for_kbit_training(
|
95 |
+
model,
|
96 |
+
use_gradient_checkpointing=False
|
97 |
+
)
|
98 |
+
|
99 |
+
print("Parameter Number: " + str(sum(p.numel() for p in model.parameters() if p.requires_grad)))
|
100 |
+
|
101 |
+
# Download weights at startup
|
102 |
+
model_weights_path = download_model_weights()
|
103 |
+
checkpoint = torch.load(model_weights_path, weights_only=True, map_location=torch.device(device))
|
104 |
+
model.load_state_dict(checkpoint['model'], strict=False)
|
105 |
+
|
106 |
+
model = model.to(device)
|
107 |
+
model.eval()
|
108 |
+
|
109 |
+
|
110 |
+
def postprocess_inst_names(abc_text):
|
111 |
+
with open('standard_inst_names.txt', 'r', encoding='utf-8') as f:
|
112 |
+
standard_instruments_list = [line.strip() for line in f if line.strip()]
|
113 |
+
|
114 |
+
with open('instrument_mapping.json', 'r', encoding='utf-8') as f:
|
115 |
+
instrument_mapping = json.load(f)
|
116 |
+
|
117 |
+
abc_lines = abc_text.split('\n')
|
118 |
+
abc_lines = list(filter(None, abc_lines))
|
119 |
+
abc_lines = [line + '\n' for line in abc_lines]
|
120 |
+
|
121 |
+
for i, line in enumerate(abc_lines):
|
122 |
+
if line.startswith('V:') and 'nm=' in line:
|
123 |
+
match = re.search(r'nm="([^"]*)"', line)
|
124 |
+
if match:
|
125 |
+
inst_name = match.group(1)
|
126 |
+
|
127 |
+
# Check if the instrument name is already standard
|
128 |
+
if inst_name in standard_instruments_list:
|
129 |
+
continue
|
130 |
+
|
131 |
+
# Find the most similar key in instrument_mapping
|
132 |
+
matching_key = difflib.get_close_matches(inst_name, list(instrument_mapping.keys()), n=1, cutoff=0.6)
|
133 |
+
|
134 |
+
if matching_key:
|
135 |
+
# Replace the instrument name with the standardized version
|
136 |
+
replacement = instrument_mapping[matching_key[0]]
|
137 |
+
new_line = line.replace(f'nm="{inst_name}"', f'nm="{replacement}"')
|
138 |
+
abc_lines[i] = new_line
|
139 |
+
|
140 |
+
# Combine the lines back into a single string
|
141 |
+
processed_abc_text = ''.join(abc_lines)
|
142 |
+
return processed_abc_text
|
143 |
+
|
144 |
+
|
145 |
+
def complete_brackets(s):
|
146 |
+
stack = []
|
147 |
+
bracket_map = {'{': '}', '[': ']', '(': ')'}
|
148 |
+
|
149 |
+
# Iterate through each character, handle bracket matching
|
150 |
+
for char in s:
|
151 |
+
if char in bracket_map:
|
152 |
+
stack.append(char)
|
153 |
+
elif char in bracket_map.values():
|
154 |
+
# Find the corresponding left bracket
|
155 |
+
for key, value in bracket_map.items():
|
156 |
+
if value == char:
|
157 |
+
if stack and stack[-1] == key:
|
158 |
+
stack.pop()
|
159 |
+
break # Found matching right bracket, process next character
|
160 |
+
|
161 |
+
# Complete missing right brackets (in reverse order of remaining left brackets in stack)
|
162 |
+
completion = ''.join(bracket_map[c] for c in reversed(stack))
|
163 |
+
return s + completion
|
164 |
+
|
165 |
+
|
166 |
+
def rest_unreduce(abc_lines):
|
167 |
+
tunebody_index = None
|
168 |
+
for i in range(len(abc_lines)):
|
169 |
+
if abc_lines[i].startswith('%%score'):
|
170 |
+
abc_lines[i] = complete_brackets(abc_lines[i])
|
171 |
+
if '[V:' in abc_lines[i]:
|
172 |
+
tunebody_index = i
|
173 |
+
break
|
174 |
+
|
175 |
+
metadata_lines = abc_lines[: tunebody_index]
|
176 |
+
tunebody_lines = abc_lines[tunebody_index:]
|
177 |
+
|
178 |
+
part_symbol_list = []
|
179 |
+
voice_group_list = []
|
180 |
+
for line in metadata_lines:
|
181 |
+
if line.startswith('%%score'):
|
182 |
+
for round_bracket_match in re.findall(r'\((.*?)\)', line):
|
183 |
+
voice_group_list.append(round_bracket_match.split())
|
184 |
+
existed_voices = [item for sublist in voice_group_list for item in sublist]
|
185 |
+
if line.startswith('V:'):
|
186 |
+
symbol = line.split()[0]
|
187 |
+
part_symbol_list.append(symbol)
|
188 |
+
if symbol[2:] not in existed_voices:
|
189 |
+
voice_group_list.append([symbol[2:]])
|
190 |
+
z_symbol_list = [] # voices that use z as rest
|
191 |
+
x_symbol_list = [] # voices that use x as rest
|
192 |
+
for voice_group in voice_group_list:
|
193 |
+
z_symbol_list.append('V:' + voice_group[0])
|
194 |
+
for j in range(1, len(voice_group)):
|
195 |
+
x_symbol_list.append('V:' + voice_group[j])
|
196 |
+
|
197 |
+
part_symbol_list.sort(key=lambda x: int(x[2:]))
|
198 |
+
|
199 |
+
unreduced_tunebody_lines = []
|
200 |
+
|
201 |
+
for i, line in enumerate(tunebody_lines):
|
202 |
+
unreduced_line = ''
|
203 |
+
|
204 |
+
line = re.sub(r'^\[r:[^\]]*\]', '', line)
|
205 |
+
|
206 |
+
pattern = r'\[V:(\d+)\](.*?)(?=\[V:|$)'
|
207 |
+
matches = re.findall(pattern, line)
|
208 |
+
|
209 |
+
line_bar_dict = {}
|
210 |
+
for match in matches:
|
211 |
+
key = f'V:{match[0]}'
|
212 |
+
value = match[1]
|
213 |
+
line_bar_dict[key] = value
|
214 |
+
|
215 |
+
# calculate duration and collect barline
|
216 |
+
dur_dict = {}
|
217 |
+
for symbol, bartext in line_bar_dict.items():
|
218 |
+
right_barline = ''.join(re.split(Barline_regexPattern, bartext)[-2:])
|
219 |
+
bartext = bartext[:-len(right_barline)]
|
220 |
+
try:
|
221 |
+
bar_dur = calculate_bartext_duration(bartext)
|
222 |
+
except:
|
223 |
+
bar_dur = None
|
224 |
+
if bar_dur is not None:
|
225 |
+
if bar_dur not in dur_dict.keys():
|
226 |
+
dur_dict[bar_dur] = 1
|
227 |
+
else:
|
228 |
+
dur_dict[bar_dur] += 1
|
229 |
+
|
230 |
+
try:
|
231 |
+
ref_dur = max(dur_dict, key=dur_dict.get)
|
232 |
+
except:
|
233 |
+
pass # use last ref_dur
|
234 |
+
|
235 |
+
if i == 0:
|
236 |
+
prefix_left_barline = line.split('[V:')[0]
|
237 |
+
else:
|
238 |
+
prefix_left_barline = ''
|
239 |
+
|
240 |
+
for symbol in part_symbol_list:
|
241 |
+
if symbol in line_bar_dict.keys():
|
242 |
+
symbol_bartext = line_bar_dict[symbol]
|
243 |
+
else:
|
244 |
+
if symbol in z_symbol_list:
|
245 |
+
symbol_bartext = prefix_left_barline + 'z' + str(ref_dur) + right_barline
|
246 |
+
elif symbol in x_symbol_list:
|
247 |
+
symbol_bartext = prefix_left_barline + 'x' + str(ref_dur) + right_barline
|
248 |
+
unreduced_line += '[' + symbol + ']' + symbol_bartext
|
249 |
+
|
250 |
+
unreduced_tunebody_lines.append(unreduced_line + '\n')
|
251 |
+
|
252 |
+
unreduced_lines = metadata_lines + unreduced_tunebody_lines
|
253 |
+
|
254 |
+
return unreduced_lines
|
255 |
+
|
256 |
+
|
257 |
+
def inference_patch(period, composer, instrumentation):
|
258 |
+
prompt_lines = [
|
259 |
+
'%' + period + '\n',
|
260 |
+
'%' + composer + '\n',
|
261 |
+
'%' + instrumentation + '\n']
|
262 |
+
|
263 |
+
while True:
|
264 |
+
|
265 |
+
failure_flag = False
|
266 |
+
|
267 |
+
bos_patch = [patchilizer.bos_token_id] * (PATCH_SIZE - 1) + [patchilizer.eos_token_id]
|
268 |
+
|
269 |
+
start_time = time.time()
|
270 |
+
|
271 |
+
prompt_patches = patchilizer.patchilize_metadata(prompt_lines)
|
272 |
+
byte_list = list(''.join(prompt_lines))
|
273 |
+
context_tunebody_byte_list = []
|
274 |
+
metadata_byte_list = []
|
275 |
+
|
276 |
+
print(''.join(byte_list), end='')
|
277 |
+
|
278 |
+
prompt_patches = [[ord(c) for c in patch] + [patchilizer.special_token_id] * (PATCH_SIZE - len(patch)) for patch
|
279 |
+
in prompt_patches]
|
280 |
+
prompt_patches.insert(0, bos_patch)
|
281 |
+
|
282 |
+
input_patches = torch.tensor(prompt_patches, device=device).reshape(1, -1)
|
283 |
+
|
284 |
+
end_flag = False
|
285 |
+
cut_index = None
|
286 |
+
|
287 |
+
tunebody_flag = False
|
288 |
+
|
289 |
+
with torch.inference_mode():
|
290 |
+
|
291 |
+
while True:
|
292 |
+
with torch.autocast(device_type='cuda', dtype=torch.float16):
|
293 |
+
predicted_patch = model.generate(input_patches.unsqueeze(0),
|
294 |
+
top_k=TOP_K,
|
295 |
+
top_p=TOP_P,
|
296 |
+
temperature=TEMPERATURE)
|
297 |
+
if not tunebody_flag and patchilizer.decode([predicted_patch]).startswith(
|
298 |
+
'[r:'): # 初次进入tunebody,必须以[r:0/开头
|
299 |
+
tunebody_flag = True
|
300 |
+
r0_patch = torch.tensor([ord(c) for c in '[r:0/']).unsqueeze(0).to(device)
|
301 |
+
temp_input_patches = torch.concat([input_patches, r0_patch], axis=-1)
|
302 |
+
predicted_patch = model.generate(temp_input_patches.unsqueeze(0),
|
303 |
+
top_k=TOP_K,
|
304 |
+
top_p=TOP_P,
|
305 |
+
temperature=TEMPERATURE)
|
306 |
+
predicted_patch = [ord(c) for c in '[r:0/'] + predicted_patch
|
307 |
+
if predicted_patch[0] == patchilizer.bos_token_id and predicted_patch[1] == patchilizer.eos_token_id:
|
308 |
+
end_flag = True
|
309 |
+
break
|
310 |
+
next_patch = patchilizer.decode([predicted_patch])
|
311 |
+
|
312 |
+
for char in next_patch:
|
313 |
+
byte_list.append(char)
|
314 |
+
if tunebody_flag:
|
315 |
+
context_tunebody_byte_list.append(char)
|
316 |
+
else:
|
317 |
+
metadata_byte_list.append(char)
|
318 |
+
print(char, end='')
|
319 |
+
|
320 |
+
patch_end_flag = False
|
321 |
+
for j in range(len(predicted_patch)):
|
322 |
+
if patch_end_flag:
|
323 |
+
predicted_patch[j] = patchilizer.special_token_id
|
324 |
+
if predicted_patch[j] == patchilizer.eos_token_id:
|
325 |
+
patch_end_flag = True
|
326 |
+
|
327 |
+
predicted_patch = torch.tensor([predicted_patch], device=device) # (1, 16)
|
328 |
+
input_patches = torch.cat([input_patches, predicted_patch], dim=1) # (1, 16 * patch_len)
|
329 |
+
|
330 |
+
if len(byte_list) > 102400:
|
331 |
+
failure_flag = True
|
332 |
+
break
|
333 |
+
if time.time() - start_time > 10 * 60:
|
334 |
+
failure_flag = True
|
335 |
+
break
|
336 |
+
|
337 |
+
if input_patches.shape[1] >= PATCH_LENGTH * PATCH_SIZE and not end_flag:
|
338 |
+
print('Stream generating...')
|
339 |
+
|
340 |
+
metadata = ''.join(metadata_byte_list)
|
341 |
+
context_tunebody = ''.join(context_tunebody_byte_list)
|
342 |
+
|
343 |
+
if '\n' not in context_tunebody:
|
344 |
+
break # Generated content is all metadata, abandon
|
345 |
+
|
346 |
+
context_tunebody_lines = context_tunebody.strip().split('\n')
|
347 |
+
|
348 |
+
if not context_tunebody.endswith('\n'):
|
349 |
+
context_tunebody_lines = [context_tunebody_lines[i] + '\n' for i in
|
350 |
+
range(len(context_tunebody_lines) - 1)] + [context_tunebody_lines[-1]]
|
351 |
+
else:
|
352 |
+
context_tunebody_lines = [context_tunebody_lines[i] + '\n' for i in
|
353 |
+
range(len(context_tunebody_lines))]
|
354 |
+
|
355 |
+
cut_index = len(context_tunebody_lines) // 2
|
356 |
+
abc_code_slice = metadata + ''.join(context_tunebody_lines[-cut_index:])
|
357 |
+
|
358 |
+
input_patches = patchilizer.encode_generate(abc_code_slice)
|
359 |
+
|
360 |
+
input_patches = [item for sublist in input_patches for item in sublist]
|
361 |
+
input_patches = torch.tensor([input_patches], device=device)
|
362 |
+
input_patches = input_patches.reshape(1, -1)
|
363 |
+
|
364 |
+
context_tunebody_byte_list = list(''.join(context_tunebody_lines[-cut_index:]))
|
365 |
+
|
366 |
+
if not failure_flag:
|
367 |
+
abc_text = ''.join(byte_list)
|
368 |
+
|
369 |
+
# unreduce
|
370 |
+
abc_lines = abc_text.split('\n')
|
371 |
+
abc_lines = list(filter(None, abc_lines))
|
372 |
+
abc_lines = [line + '\n' for line in abc_lines]
|
373 |
+
try:
|
374 |
+
unreduced_abc_lines = rest_unreduce(abc_lines)
|
375 |
+
except:
|
376 |
+
failure_flag = True
|
377 |
+
pass
|
378 |
+
else:
|
379 |
+
unreduced_abc_lines = [line for line in unreduced_abc_lines if
|
380 |
+
not (line.startswith('%') and not line.startswith('%%'))]
|
381 |
+
unreduced_abc_lines = ['X:1\n'] + unreduced_abc_lines
|
382 |
+
unreduced_abc_text = ''.join(unreduced_abc_lines)
|
383 |
+
return unreduced_abc_text
|
384 |
+
|
385 |
+
|
386 |
+
if __name__ == '__main__':
|
387 |
+
inference_patch('Classical', 'Beethoven, Ludwig van', 'Orchestral')
|
388 |
+
|