-
Notifications
You must be signed in to change notification settings - Fork 0
/
transformer101.py
294 lines (244 loc) · 11.3 KB
/
transformer101.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
import torch
import torch.nn as nn
import torch.optim as optim
import torch.utils.data as data
import math
import copy
import pandas as pd
from tqdm import tqdm
from os.path import exists
from os import remove, chdir
import pickle
from torch.utils.tensorboard import SummaryWriter
# from synthesizer import parse_csv, synthesize # functions in synthesizer.ipynb, PrettyMiDI and MuseScore needed
DEVICE = "cuda"
class MultiHeadAttention(nn.Module):
def __init__(self, d_model, num_heads):
super(MultiHeadAttention, self).__init__()
assert d_model % num_heads == 0, "d_model must be divisible by num_heads"
self.d_model = d_model
self.num_heads = num_heads
self.d_k = d_model // num_heads
self.W_q = nn.Linear(d_model, d_model)
self.W_k = nn.Linear(d_model, d_model)
self.W_v = nn.Linear(d_model, d_model)
self.W_o = nn.Linear(d_model, d_model)
def scaled_dot_product_attention(self, Q, K, V, mask=None):
attn_scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(self.d_k)
if mask is not None:
attn_scores = attn_scores.masked_fill(mask == 0, -1e9)
attn_probs = torch.softmax(attn_scores, dim=-1)
output = torch.matmul(attn_probs, V)
return output
def split_heads(self, x):
batch_size, seq_length, d_model = x.size()
return x.view(batch_size, seq_length, self.num_heads, self.d_k).transpose(1, 2)
def combine_heads(self, x):
batch_size, _, seq_length, d_k = x.size()
return x.transpose(1, 2).contiguous().view(batch_size, seq_length, self.d_model)
def forward(self, Q, K, V, mask=None):
Q = self.split_heads(self.W_q(Q))
K = self.split_heads(self.W_k(K))
V = self.split_heads(self.W_v(V))
attn_output = self.scaled_dot_product_attention(Q, K, V, mask)
output = self.W_o(self.combine_heads(attn_output))
return output
class PositionWiseFeedForward(nn.Module):
def __init__(self, d_model, d_ff):
super(PositionWiseFeedForward, self).__init__()
self.fc1 = nn.Linear(d_model, d_ff)
self.fc2 = nn.Linear(d_ff, d_model)
self.relu = nn.ReLU()
def forward(self, x):
return self.fc2(self.relu(self.fc1(x)))
class PositionalEncoding(nn.Module):
def __init__(self, d_model):
super(PositionalEncoding, self).__init__()
self.d_model = d_model
def forward(self, x):
max_len = x.size(1)
pe = torch.zeros(max_len, self.d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, self.d_model, 2).float() * -(math.log(10000.0) / self.d_model))
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0).to(x.device)
return x + pe
class EncoderLayer(nn.Module):
def __init__(self, d_model, num_heads, d_ff, dropout):
super(EncoderLayer, self).__init__()
self.self_attn = MultiHeadAttention(d_model, num_heads)
self.feed_forward = PositionWiseFeedForward(d_model, d_ff)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.dropout = nn.Dropout(dropout)
def forward(self, x, mask):
attn_output = self.self_attn(x, x, x, mask)
x = self.norm1(x + self.dropout(attn_output))
ff_output = self.feed_forward(x)
x = self.norm2(x + self.dropout(ff_output))
return x
class EmbedHead(nn.Module):
def __init__(
self,
input_dim,
inner_dim_1,
inner_dim_2,
out_dim
):
super().__init__()
self.linear1 = nn.Linear(input_dim, inner_dim_1)
self.linear2 = nn.Linear(inner_dim_1, inner_dim_2)
self.linear3 = nn.Linear(inner_dim_2, out_dim)
self.activation_fn = nn.functional.gelu
def forward(self, x):
x = self.linear1(x)
x = self.activation_fn(x)
x = self.linear2(x)
x = self.activation_fn(x)
x = self.linear3(x)
return x
class Transformer(nn.Module):
def __init__(self, src_vocab_size, tgt_vocab_size, d_model, num_heads, num_layers, d_ff, max_seq_length, dropout):
super(Transformer, self).__init__()
self.encoder_embedding = EmbedHead(src_vocab_size, d_model, d_model, d_model)
self.positional_encoding = PositionalEncoding(d_model)
self.encoder_layers = nn.ModuleList([EncoderLayer(d_model, num_heads, d_ff, dropout) for _ in range(num_layers)])
self.fc = nn.Linear(d_model, tgt_vocab_size)
self.dropout = nn.Dropout(dropout)
def generate_mask(self, src):
src_mask = (torch.sum(src, dim=2) > 0).unsqueeze(1).unsqueeze(2)
return src_mask
def forward(self, src):
src_mask = self.generate_mask(src)
src_embedded = self.dropout(self.positional_encoding(self.encoder_embedding(src)))
enc_output = src_embedded
for enc_layer in self.encoder_layers:
enc_output = enc_layer(enc_output, src_mask)
output = self.fc(enc_output)
return torch.sigmoid(output)
if __name__ == "__main__":
src_vocab_size = 12
tgt_vocab_size = 12
d_model = 512
num_heads = 8
num_layers = 4
d_ff = 4096//8
max_seq_length = 2400
dropout = 0.1
batchsize = 8
mode = "train"
writer = SummaryWriter('.log-tf2')
transformer = Transformer(src_vocab_size, tgt_vocab_size, d_model, num_heads, num_layers, d_ff, max_seq_length, dropout).to(DEVICE)
n = 0
for p in transformer.parameters():
n += p.numel()
print(n)
# load data
if exists("trainset_w.pkl") and exists("validset_w.pkl") and exists("testset_w.pkl"):
print("splitted dataset found!")
with open("trainset_w.pkl", "rb") as f:
trainset = pickle.load(f)
with open("validset_w.pkl", "rb") as f:
validset = pickle.load(f)
with open("testset_w.pkl", "rb") as f:
testset = pickle.load(f)
else:
print("splitted dataset (unzip) not found!")
dataset = []
for i in tqdm(range(1, 910)):
idx = f"{i:0>3}"
data_frame = pd.read_csv(f'POP909/POP909/{idx}/melody_chord_1_beat.csv')
data_frame['melody'] = data_frame['melody'].apply(lambda x: [float(n.strip().rstrip('.')) for n in x.strip('[]').split(',') if n.strip()])
data_frame['chord'] = data_frame['chord'].apply(lambda x: [float(n.strip().rstrip('.')) for n in x.strip('[]').split(' ') if n.strip()])
# If you want to extract these columns as lists of lists
melody_list = data_frame['melody'].tolist()
chord_list = data_frame['chord'].tolist()
old_chord = None
weight_list = []
for chord in chord_list:
if chord != old_chord:
weight_list.append(1.)
else:
weight_list.append(0.)
old_chord = chord
# # Define a default value for NaNs
# default_value = 2
src_data = torch.tensor(melody_list)
# Add an additional dimension at the front
tgt_data = torch.tensor(chord_list)
# Add an additional dimension at the front
weights = torch.tensor(weight_list)
dataset.append((idx, src_data, tgt_data, weights))
generator = torch.Generator().manual_seed(0)
trainset, validset, testset = data.random_split(dataset, [709, 100, 100], generator)
with open("trainset_w.pkl", "wb") as f:
pickle.dump(trainset, f)
with open("validset_w.pkl", "wb") as f:
pickle.dump(validset, f)
with open("testset_w.pkl", "wb") as f:
pickle.dump(testset, f)
def collate_fn(batch):
idx, src_data, tgt_data, weights = zip(*batch)
src_data = nn.utils.rnn.pad_sequence(src_data, batch_first=True, padding_value=0.).to(DEVICE)
tgt_data = nn.utils.rnn.pad_sequence(tgt_data, batch_first=True, padding_value=0.).to(DEVICE)
weights = nn.utils.rnn.pad_sequence(weights, batch_first=True, padding_value=0.).to(DEVICE)
return idx, src_data, tgt_data, weights
trainset = data.DataLoader(trainset, batch_size=batchsize, collate_fn=collate_fn)
validset = data.DataLoader(validset, batch_size=1, collate_fn=collate_fn)
testset = data.DataLoader(testset, batch_size=1, collate_fn=collate_fn)
def loss_fn(output, target, weights):
flat_weight = 1 + weights.repeat_interleave(12, 1).flatten()
bce_loss = nn.BCELoss(weight=flat_weight, reduction="mean")(output.contiguous().view(-1), target.contiguous().view(-1))
return bce_loss
if mode == "train":
optimizer = optim.Adam(transformer.parameters(), lr=0.0001, betas=(0.9, 0.98), eps=1e-9)
val_cnt = 0
beat_epoch = 0
best_val_loss = float("inf")
for epoch in range(2000):
transformer.train()
for i, pair_data in enumerate(trainset):
_, src_data, tgt_data, weights = pair_data
optimizer.zero_grad()
output = transformer(src_data)
loss = loss_fn(output, tgt_data, weights)
loss.backward()
optimizer.step()
transformer.eval()
total_val_loss = 0
for i, pair_data in enumerate(validset):
_, src_data, tgt_data, weights = pair_data
output = transformer(src_data).detach()
val_loss = loss_fn(output, tgt_data, weights).detach().item()
total_val_loss += val_loss
total_val_loss /= 100
writer.add_scalar("train_loss", loss.item(), global_step=epoch)
writer.add_scalar("valid_loss", total_val_loss, global_step=epoch)
if total_val_loss < best_val_loss:
if epoch > 0:
remove("model2_best.pt")
best_val_loss = total_val_loss
best_epoch = epoch
val_cnt = 0
print(f"Epoch: {epoch+1}, Train Loss: {loss.item()}, Valid Loss: {total_val_loss} (best)")
torch.save(transformer, f"model2_best.pt")
else:
val_cnt += 1
print(f"Epoch: {epoch+1}, Train Loss: {loss.item()}, Valid Loss: {total_val_loss} ({val_cnt}/50)")
if val_cnt >= 50:
break
if (epoch + 1) % 50 == 0:
torch.save(transformer, f"model2_{epoch + 1}.pt")
elif mode == "eval":
transformer.eval()
transformer.load_state_dict(torch.load("model2_best.pt").state_dict())
# tot_loss = 0
for i, pair_data in tqdm(enumerate(testset)):
idx, src_data, tgt_data, weights = pair_data
idx = idx[0]
output = transformer(src_data, tgt_data).detach()
print(loss_fn(output, tgt_data, weights))
chords = torch.round(output).squeeze().cpu().numpy().tolist()
real_chords, beats, durations = parse_csv(f"POP909/POP909/{idx}/melody_chord_1_beat.csv")
synthesize(f"POP909/POP909/{idx}/{idx}.mid", chords, beats, durations, f"../gen/test_m2/test_{idx}_m2mlp_best_ns.mid", to_mp3=False)