-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathtrain_maestro_GPT2.py
536 lines (472 loc) · 22.8 KB
/
train_maestro_GPT2.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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
maestro_v3_midi_url = "https://storage.googleapis.com/magentadata/datasets/maestro/v3.0.0/maestro-v3.0.0-midi.zip"
maestro_v3_meta_url = "https://storage.googleapis.com/magentadata/datasets/maestro/v3.0.0/maestro-v3.0.0.csv"
maestro_v3_json_url = "https://storage.googleapis.com/magentadata/datasets/maestro/v3.0.0/maestro-v3.0.0.json"
# %%
# download and unzip the urls above
# !wget $maestro_v3_midi_url
# !wget $maestro_v3_meta_url
# !wget $maestro_v3_json_url
# !unzip maestro-v3.0.0-midi.zip
# %%
import os
from os.path import join
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from transformers import GPT2LMHeadModel, GPT2Tokenizer, GPT2Config
from torch.optim import AdamW, Adam
from transformers import AdamW, get_linear_schedule_with_warmup
import librosa.display
import librosa
import matplotlib.pyplot as plt
from tqdm import tqdm, trange
from typing import Union
import pretty_midi
import pickle
# maestro_root = r"E:\Datasets\maestro-v3.0.0-midi\maestro-v3.0.0"
maestro_root = r"/home/binxu/Datasets/maestro-v3.0.0"
# visualize the midi file as piano roll
def plot_piano_roll(pm, start_pitch, end_pitch, fs=100):
# Use librosa's specshow function for displaying the piano roll
librosa.display.specshow(pm.get_piano_roll(fs)[start_pitch:end_pitch],
hop_length=1, sr=fs, x_axis='time', y_axis='cqt_note',
fmin=pretty_midi.note_number_to_hz(start_pitch))
# load the metadata
maestro_meta = pd.read_csv(join(maestro_root, "maestro-v3.0.0.csv"))
# %%
# https://notebook.community/craffel/pretty-midi/Tutorial
# Each note has a start time, end time, pitch, and velocity.
#
# Velocity / volume of a note: 1-127
# Pitch of a note: 0-127 C1-G9
# Duration of a note: seconds, float number
# Start time of a note: seconds, float number
# %%
# midi to note sequence
def midi2notes(midi_file: Union[str, pretty_midi.PrettyMIDI]):
if isinstance(midi_file, str):
midi_data = pretty_midi.PrettyMIDI(midi_file)
else:
midi_data = midi_file
note_t_seq = []
note_dt_seq = []
note_duration_seq = []
note_velo_seq = []
note_pitch_seq = []
prev_note_end = 0.0
for i, note_sample in enumerate(midi_data.instruments[0].notes[1:]):
note_dt_seq.append(note_sample.start - prev_note_end)
note_t_seq.append(note_sample.start)
note_duration_seq.append(note_sample.duration)
note_velo_seq.append(note_sample.velocity)
note_pitch_seq.append(note_sample.pitch)
prev_note_end = note_sample.end
note_seq = pd.DataFrame({"t": note_t_seq, "dt": note_dt_seq, "duration": note_duration_seq, "velo": note_velo_seq,
"pitch": note_pitch_seq}) # .to_csv("note_seq.csv", index=False)
return note_seq
midi_file = join(maestro_root, maestro_meta["midi_filename"][0].replace("/", os.path.sep))
note_seq = midi2notes(midi_file)
# %%
class MusicScoreDataset(torch.utils.data.Dataset):
def __init__(self, maestro_root, maestro_meta=None, ):
self.maestro_root = maestro_root
if maestro_meta is None:
self.maestro_meta = pd.read_csv(join(maestro_root, "maestro-v3.0.0.csv"))
else:
self.maestro_meta = maestro_meta
self.dataset = {}
def load_dataset(self):
for i in tqdm(range(len(self.maestro_meta))):
row = self.maestro_meta.iloc[i]
midi_path = join(self.maestro_root, row["midi_filename"].replace("/", os.path.sep))
# load sample midi file
midi_data = pretty_midi.PrettyMIDI(midi_path)
note_seq = midi2notes(midi_data)
self.dataset[i] = (row[["canonical_composer", 'canonical_title', 'year', 'duration']], note_seq)
return
def __len__(self):
return len(self.dataset)
def __getitem__(self, idx):
meta, note_seq = self.dataset[idx]
return meta, note_seq
dataset = MusicScoreDataset(maestro_root, maestro_meta)
# %%
dataset.dataset = pickle.load(open(join(maestro_root, "dataset.pkl"), "rb"))
# %%
# import pickle
# pickle.dump(dataset.dataset, open(join(maestro_root, "dataset.pkl"), "wb"))
# %%
meta, note_seq = dataset[0]
# %%
# %%
from torch.nn.utils.rnn import pad_sequence, pack_padded_sequence, pad_packed_sequence
# from torch.utils.data import DataLoader
# loader = DataLoader(dataset, batch_size=128, collate_fn=batch_sampler, shuffle=True)
# next(iter(loader))
token_num = 128 # 128 pitch classes
BOS_id = token_num
EOS_id = token_num + 1
PAD_ID = EOS_id + 1
MASK_LABEL_ID = - 100
def notedf2tensor(note_df, device="cuda"):
pitch_ids = torch.tensor(note_df.pitch.values).unsqueeze(0).long().to(device)
velo = torch.tensor(note_df.velo.values).unsqueeze(0).float().to(device) / 128.0
dur = torch.tensor(note_df.duration.values).unsqueeze(0).float().to(device)
dt = torch.tensor(note_df.dt.values).unsqueeze(0).float().to(device)
return pitch_ids, velo, dur, dt
def batch_sampler(data_source, batch_idxs, max_seq_len=1024):
batch_pitch = [] # integer value
batch_velo = [] # integer value
batch_dt = [] # float value
batch_duration = [] # float value
label_batch = []
if isinstance(batch_idxs, int):
batch_idxs = range(batch_idxs)
for i in batch_idxs:
meta, note_seq = data_source[i]
# sample note sequence with max_seq_len
if len(note_seq) <= max_seq_len:
note_seq_subsamp = note_seq
else:
seq_start = np.random.randint(0, len(note_seq) - max_seq_len)
note_seq_subsamp = note_seq.iloc[seq_start:seq_start + max_seq_len]
pitch, velo, dur, dt = notedf2tensor(note_seq_subsamp, device="cpu")
batch_pitch.append(pitch[0])
batch_velo.append(velo[0])
batch_dt.append(dt[0])
batch_duration.append(dur[0])
# batch_pitch.append(torch.tensor(note_seq_subsamp["pitch"].values).long())
# batch_velo.append(torch.tensor(note_seq_subsamp["velo"].values).long())
# batch_dt.append(torch.tensor(note_seq_subsamp["dt"].values))
# batch_duration.append(torch.tensor(note_seq_subsamp["duration"].values))
batch_pitch = pad_sequence(batch_pitch, batch_first=True, padding_value=PAD_ID)
batch_velo = pad_sequence(batch_velo, batch_first=True, padding_value=0.5)
batch_dt = pad_sequence(batch_dt, batch_first=True, padding_value=0.0)
batch_duration = pad_sequence(batch_duration, batch_first=True, padding_value=0.0)
# label_batch = pad_sequence(label_batch, batch_first=True, padding_value=MASK_LABEL_ID)
return batch_pitch, batch_velo, batch_dt, batch_duration
batch_pitch, batch_velo, batch_dt, batch_duration = batch_sampler(dataset, range(32), max_seq_len=512)
# %% Submodules for a Music Note Transformer
class NoteReadoutHeads(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.velocity_head = nn.Sequential(
nn.Linear(config.n_embd, config.n_embd),
nn.Tanh(),
nn.Linear(config.n_embd, 1),
)
self.duration_head = nn.Sequential(
nn.Linear(config.n_embd, config.n_embd),
nn.Tanh(),
nn.Linear(config.n_embd, 1),
nn.Softplus(), # duration should be positive
)
self.dt_head = nn.Sequential(
nn.Linear(config.n_embd, config.n_embd),
nn.Tanh(),
nn.Linear(config.n_embd, 1),
)
def forward(self, hidden_states):
velocity = self.velocity_head(hidden_states).squeeze(-1)
duration = self.duration_head(hidden_states).squeeze(-1)
dt = self.dt_head(hidden_states).squeeze(-1)
return velocity, duration, dt
class ScalarEmbedding(nn.Module):
def __init__(self, config, ):
super().__init__()
self.n_embd = config.n_embd
self.weights = nn.Parameter(torch.randn(1, 1, self.n_embd) * np.sqrt(1 / self.n_embd))
def forward(self, x):
return self.weights * x[:, :, None]
# %%
def naive_greedy_decode(model, prompt_ids, prompt_velos, prompt_durations, prompt_dts, max_gen_tokens=500,
temperature=None):
model.eval()
readout.eval()
dt_emb.eval()
duration_emb.eval()
velocity_emb.eval()
pitch_embed = model.get_input_embeddings()(prompt_ids)
velo_embed = velocity_emb(prompt_velos)
dur_embed = duration_emb(prompt_durations)
dt_embed = dt_emb(prompt_dts)
input_embed = pitch_embed + velo_embed + dur_embed + dt_embed
input_embed_full = input_embed.clone()
generated_note_seq = []
with torch.no_grad():
for i in trange(max_gen_tokens):
out = model(inputs_embeds=input_embed_full, output_hidden_states=True)
last_hidden = out.hidden_states[-1]
last_logits = out.logits[:, -1, :]
# pred_velos, pred_durations, pred_dts = readout(last_hidden)
pred_velo, pred_duration, pred_dt = readout(last_hidden[:, -1, :])
if temperature is None:
# greedy decoding
pitch_maxprob = torch.argmax(last_logits, dim=-1)
else:
# Sample from the distribution with temperature
pitch_maxprob = torch.multinomial(F.softmax(last_logits / temperature, dim=-1),
num_samples=1).squeeze(1)
input_embed_new = model.get_input_embeddings()(pitch_maxprob) + \
velocity_emb(pred_velo.unsqueeze(0)) + \
duration_emb(pred_duration.unsqueeze(0)) + \
dt_emb(pred_dt.unsqueeze(0))
generated_note_seq.append((pitch_maxprob, pred_velo, pred_duration, pred_dt))
input_embed_full = torch.concat((input_embed_full, input_embed_new), dim=1)
pitch_gen = torch.stack([pitch for pitch, velo, dur, dt in generated_note_seq], dim=1)
velo_gen = torch.stack([velo for pitch, velo, dur, dt in generated_note_seq], dim=1)
dur_gen = torch.stack([dur for pitch, velo, dur, dt in generated_note_seq], dim=1)
dt_gen = torch.stack([dt for pitch, velo, dur, dt in generated_note_seq], dim=1)
pitch_gen = torch.concat((prompt_ids, pitch_gen), dim=1).cpu()
velo_gen = torch.concat((prompt_velos, velo_gen), dim=1).cpu() # this is 0, 1 normalized
dur_gen = torch.concat((prompt_durations, dur_gen), dim=1).cpu()
dt_gen = torch.concat((prompt_dts, dt_gen), dim=1).cpu()
velo_gen = (velo_gen * 128).long() # convert to integer
return pitch_gen, velo_gen, dur_gen, dt_gen
def tensor2midi(pitch, velo, dur, dt, savedir, filename, instrument="Acoustic Grand Piano"):
generated_midi = pretty_midi.PrettyMIDI()
piano_program = pretty_midi.instrument_name_to_program(instrument)
piano = pretty_midi.Instrument(program=piano_program)
cur_time = 0
for i, pitch in enumerate(pitch):
cur_time = cur_time + dt[i].item()
note = pretty_midi.Note(velocity=min(127, max(0, velo[i].item())),
pitch=min(127, max(0, pitch.item())),
start=cur_time, end=cur_time + dur[i].item())
cur_time = cur_time + dur[i].item()
piano.notes.append(note)
generated_midi.instruments.append(piano)
generated_midi.write(join(savedir, filename))
return generated_midi
# %%
import torch.nn.functional as F
from torch.optim import AdamW
from torch.utils.tensorboard import SummaryWriter
saveroot = "/home/binxu/DL_Projects/Maestro-GPT" #"runs"
# config = GPT2Config(n_embd=128, n_layer=12, n_head=8, n_positions=512, n_ctx=128,
# vocab_size=132, bos_token_id=BOS_id, eos_token_id=EOS_id, )
# config = GPT2Config(n_embd=128, n_layer=18, n_head=8, n_positions=512, n_ctx=128,
# vocab_size=token_num + 4, bos_token_id=BOS_id, eos_token_id=EOS_id, )
config = GPT2Config(n_embd=128, n_layer=36, n_head=8, n_positions=256,
vocab_size=token_num + 4, bos_token_id=BOS_id, eos_token_id=EOS_id, )
model = GPT2LMHeadModel(config)
readout = NoteReadoutHeads(config)
dt_emb = ScalarEmbedding(config).cuda()
duration_emb = ScalarEmbedding(config).cuda()
velocity_emb = ScalarEmbedding(config).cuda()
model.cuda()
readout.cuda()
optimizer = AdamW([*model.parameters(),
*readout.parameters(),
*dt_emb.parameters(),
*duration_emb.parameters(),
*velocity_emb.parameters(),
], lr=10e-4)
scheduler = get_linear_schedule_with_warmup(optimizer, 50, 3000)
# savedir = join(saveroot, "runs_L18")
# savedir = join(saveroot, "runs_L18_lrsched_nexttok")
# savedir = join(saveroot, "runs_L36_ctx256_lrsched_nexttok")
savedir = join(saveroot, "runs_L36_ctx256_lrsched_nexttok_fixeval")
os.makedirs(savedir, exist_ok=True)
os.makedirs(join(savedir, "ckpt"), exist_ok=True)
# %%
save_per_epoch = 10
synth_per_epoch = 10
batch_size = 24
max_seq_len = 256
writer = SummaryWriter(savedir)
for epoch in trange(0, 3000):
for module in [model, readout, dt_emb, duration_emb, velocity_emb]:
module.train()
rand_idx_seq = np.random.permutation(len(dataset))
for i, csr in enumerate(trange(0, len(dataset), batch_size)):
batch_idxs = rand_idx_seq[csr:csr + batch_size]
batch_pitch, batch_velo, batch_dt, batch_duration = batch_sampler(dataset,
batch_idxs, max_seq_len=max_seq_len)
pitch_embed = model.transformer.wte(batch_pitch.cuda())
note_embed = pitch_embed + velocity_emb(batch_velo.cuda()) + \
dt_emb(batch_dt.cuda()) + duration_emb(batch_duration.cuda()) #
# out = model(batch_pitch.cuda(), labels=batch_pitch.cuda(), output_hidden_states=True)
out = model(inputs_embeds=note_embed, labels=batch_pitch.cuda(), output_hidden_states=True)
loss = out.loss
last_hidden = out.hidden_states[-1] # (batch, seq_len, embd)
velocity_pred, duration_pred, dt_pred = readout(last_hidden)
# note this is wrong, need to predict the next note so shift by 1
# loss_velo = F.mse_loss(velocity_pred, batch_velo.cuda())
# loss_duration = F.mse_loss(duration_pred, batch_duration.cuda())
# loss_dt = F.mse_loss(dt_pred, batch_dt.cuda())
# this is correct prediction
loss_velo = F.mse_loss(velocity_pred[:, :-1], batch_velo.cuda()[:, 1:])
loss_duration = F.mse_loss(duration_pred[:, :-1], batch_duration.cuda()[:, 1:])
loss_dt = F.mse_loss(dt_pred[:, :-1], batch_dt.cuda()[:, 1:])
loss += loss_velo + loss_duration + loss_dt
# compute additional loss based on last hidden state
loss.backward()
optimizer.step()
optimizer.zero_grad()
print(
f"epoch{epoch}-step{i:03d} {loss.item():.5f} vel {loss_velo.item():.5f} dur {loss_duration.item():.5f} dt {loss_dt.item():.5f}")
writer.add_scalar("loss", loss.item(), epoch * len(dataset) // batch_size + i)
writer.add_scalar("loss_velo", loss_velo.item(), epoch * len(dataset) // batch_size + i)
writer.add_scalar("loss_duration", loss_duration.item(), epoch * len(dataset) // batch_size + i)
writer.add_scalar("loss_dt", loss_dt.item(), epoch * len(dataset) // batch_size + i)
writer.add_scalar("lr", optimizer.param_groups[0]["lr"], epoch * len(dataset) // batch_size + i)
writer.add_scalar("epoch", epoch, epoch * len(dataset) // batch_size + i)
# writer.add_scalar("grad_norm", torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0),)
scheduler.step()
if epoch % save_per_epoch == 0:
torch.save(model.state_dict(), join(savedir, "ckpt", f"model_{epoch:03d}.pt"))
torch.save(readout.state_dict(), join(savedir, "ckpt", f"readout_{epoch:03d}.pt"))
torch.save(dt_emb.state_dict(), join(savedir, "ckpt", f"dt_emb_{epoch:03d}.pt"))
torch.save(duration_emb.state_dict(), join(savedir, "ckpt", f"duration_emb_{epoch:03d}.pt"))
torch.save(velocity_emb.state_dict(), join(savedir, "ckpt", f"velocity_emb_{epoch:03d}.pt"))
if epoch % synth_per_epoch == 0:
prompt_ids, prompt_velos, prompt_durations, prompt_dts = notedf2tensor(note_seq[:10], device="cuda")
pitch_gen, velo_gen, dur_gen, dt_gen = naive_greedy_decode(model,
prompt_ids, prompt_velos, prompt_durations, prompt_dts,
temperature=.2, max_gen_tokens=max_seq_len - 12)
tensor2midi(pitch_gen[0], velo_gen[0], dur_gen[0], dt_gen[0],
savedir, f"generated_midi_with_tempo_T02_{epoch:03d}.mid")
#%%
torch.save(readout.state_dict(), join(savedir, "ckpt", f"readout_{epoch:03d}.pt"))
torch.save(dt_emb.state_dict(), join(savedir, "ckpt", f"dt_emb_{epoch:03d}.pt"))
torch.save(duration_emb.state_dict(), join(savedir, "ckpt", f"duration_emb_{epoch:03d}.pt"))
torch.save(velocity_emb.state_dict(), join(savedir, "ckpt", f"velocity_emb_{epoch:03d}.pt"))
model.save_pretrained(join(savedir, "ckpt", f"model_{epoch:03d}.pt"))
config.save_pretrained(join(savedir, "ckpt","config.json"))
#%%
# sampling from the model
model.eval()
readout.eval()
dt_emb.eval()
duration_emb.eval()
velocity_emb.eval()
# %%
prompt_ids = torch.tensor(note_seq.pitch[:10]).unsqueeze(0).long().cuda()
answers = model.generate(prompt_ids, max_length=128, do_sample=True,
top_k=0, top_p=0.90, num_return_sequences=3,
bos_token_id=BOS_id, eos_token_id=EOS_id, pad_token_id=PAD_ID,
output_hidden_states=True, return_dict_in_generate=True)
#%% Generate sequence
prompt_ids = torch.tensor(note_seq.pitch[:10]).unsqueeze(0).long().cuda()
answers = model.generate(prompt_ids, max_length=100, do_sample=True,
top_k=0, top_p=0.90, num_return_sequences=3,
bos_token_id=BOS_id, eos_token_id=EOS_id, pad_token_id=PAD_ID,
output_hidden_states=True, return_dict_in_generate=True)
#%%
#%%
# generate sequence with readout and embedding
#%%
# prompt_ids = torch.tensor(note_seq.pitch[:12]).unsqueeze(0).long().cuda()
# prompt_velos = torch.tensor(note_seq.velo[:12]).unsqueeze(0).cuda().float() / 128.0
# prompt_durations = torch.tensor(note_seq.duration[:12]).unsqueeze(0).cuda().float()
# prompt_dts = torch.tensor(note_seq.dt[:12]).unsqueeze(0).cuda().float()
prompt_ids, prompt_velos, prompt_durations, prompt_dts = notedf2tensor(note_seq[:10], device="cuda")
pitch_gen, velo_gen, dur_gen, dt_gen = naive_greedy_decode(model,
prompt_ids, prompt_velos, prompt_durations, prompt_dts,
temperature=.5, max_gen_tokens=500)
tensor2midi(pitch_gen[0], velo_gen[0], dur_gen[0], dt_gen[0],
savedir, "generated_midi_with_tempo_T0_5.mid")
#%%
note_seq_full = answers.sequences[0].cpu()
note_seq_full = note_seq_full[note_seq_full != PAD_ID]
for pitch in note_seq_full:
print(librosa.midi_to_note(pitch), end=" ")
#%% save to midi and play
# create midi file based on the generated sequence
generated_midi = pretty_midi.PrettyMIDI()
piano_program = pretty_midi.instrument_name_to_program('Acoustic Grand Piano')
piano = pretty_midi.Instrument(program=piano_program)
delta = 0.25
for i, pitch in enumerate(note_seq_full):
note = pretty_midi.Note(velocity=100, pitch=pitch.item(), start=i*delta, end=(i+1)*delta)
piano.notes.append(note)
#%%
generated_midi.instruments.append(piano)
generated_midi.write(join(savedir, "generated_midi.mid"))
# %% Old version
# dataset = MusicScoreDataset(maestro_root, maestro_meta)
batch_size = 12
writer = SummaryWriter(savedir)
for epoch in trange(0, 100):
rand_idx_seq = np.random.permutation(len(dataset))
for i, csr in enumerate(trange(0, len(dataset), batch_size)):
batch_idxs = rand_idx_seq[csr:csr + batch_size]
batch_pitch, batch_velo, batch_dt, batch_duration = batch_sampler(dataset,
batch_idxs, max_seq_len=512)
out = model(batch_pitch.cuda(), labels=batch_pitch.cuda(), output_hidden_states=True)
loss = out.loss
last_hidden = out.hidden_states[-1] # (batch, seq_len, embd)
velocity_pred, duration_pred, dt_pred = readout(last_hidden)
# note this is wrong, need to predict the next note so shift by 1
loss_velo = F.mse_loss(velocity_pred, batch_velo.cuda())
loss_duration = F.mse_loss(duration_pred, batch_duration.cuda())
loss_dt = F.mse_loss(dt_pred, batch_dt.cuda())
loss += loss_velo + loss_duration + loss_dt
# compute additional loss based on last hidden state
loss.backward()
optimizer.step()
optimizer.zero_grad()
print(
f"epoch{epoch}-step{i:03d} {loss.item():.5f} vel {loss_velo.item():.5f} dur {loss_duration.item():.5f} dt {loss_dt.item():.5f}")
writer.add_scalar("loss", loss.item(), epoch * len(dataset) // batch_size + i)
writer.add_scalar("loss_velo", loss_velo.item(), epoch * len(dataset) // batch_size + i)
writer.add_scalar("loss_duration", loss_duration.item(), epoch * len(dataset) // batch_size + i)
writer.add_scalar("loss_dt", loss_dt.item(), epoch * len(dataset) // batch_size + i)
writer.add_scalar("lr", optimizer.param_groups[0]["lr"], epoch * len(dataset) // batch_size + i)
writer.add_scalar("epoch", epoch, epoch * len(dataset) // batch_size + i)
# writer.add_scalar("grad_norm", torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0),)
torch.save(model.state_dict(), join(savedir, "ckpt", f"model_{epoch:03d}.pt"))
#%%
#%%# %%
# # load sample midi file
# midi_file = join(maestro_root, maestro_meta["midi_filename"][0].replace("/", os.path.sep))
# midi_data = pretty_midi.PrettyMIDI(midi_file)
#
# print(midi_data.instruments)
# len(midi_data.instruments[0].notes)
# # %%
# # create a new figure
# fig, ax = plt.subplots(figsize=(12, 4))
# # plot the piano roll
# plot_piano_roll(midi_data, 24, 84)
# plt.tight_layout()
# plt.show()
# # %%
# note_dt_dist = []
# note_len_dist = []
# note_velo_dist = []
# note_pitch_dist = []
# for i, note_sample in enumerate(midi_data.instruments[0].notes[1:]):
# note_dt_dist.append(note_sample.start - midi_data.instruments[0].notes[i - 1].end)
# note_len_dist.append(note_sample.duration)
# note_velo_dist.append(note_sample.velocity)
# note_pitch_dist.append(note_sample.pitch)
# # %%
#
# # %%
# plt.subplots(1, 3, figsize=(12, 4))
# plt.subplot(1, 3, 1)
# plt.hist(note_len_dist, bins=5000)
# plt.xlim(0, 3)
# plt.title("Note Length Distribution")
# plt.subplot(1, 3, 2)
# plt.hist(note_velo_dist, bins=100)
# plt.title("Note Velocity Distribution")
# plt.subplot(1, 3, 3)
# plt.hist(note_pitch_dist, bins=100)
# plt.title("Note Pitch Distribution")
# plt.show()
#
# # %%
# print(len(maestro_meta.canonical_composer.unique()))
# print(len(maestro_meta.canonical_title.unique()))
#
# # %%
# notes_str = [librosa.midi_to_note(note.pitch)
# for note in midi_data.instruments[0].notes]
# # %%
# midi_data.time_to_tick(midi_data.instruments[0].notes[-1].end)
# # %%