File size: 11,012 Bytes
d8968a6 | 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 | import glob
from collections import OrderedDict, defaultdict
import os
import dill
import wandb
import torch_optimizer as optim_extra
import torch
from torch import optim
from torch.utils.data import ConcatDataset, DataLoader
import torchaudio
import numpy as np
from tqdm import tqdm
from dataloader import (TrainTestDataset,
TrainValTestDataset, collate_fn_padd, spectral_size)
from next_frame_classifier import NextFrameClassifier
from utils import (PrecisionRecallMetric, StatsMeter,
detect_peaks, line, max_min_norm, replicate_first_k_frames)
import dutch_preprocess
from utils import timit_to_leehon_map_MACRO
class Solver(torch.nn.Module):
def __init__(self, cfg):
super().__init__()
hp = cfg
self.hp = hp
self.hparams = hp
self.current_epoch = 0
self.peak_detection_params = defaultdict(lambda: {
"width": None,
"distance": None
})
self.pr = defaultdict(lambda: {
"train": PrecisionRecallMetric(),
"val": PrecisionRecallMetric(),
"test": PrecisionRecallMetric()
})
self.best_l1_dist= defaultdict(lambda: {
"train": (float("inf"), 0),
"val": (float("inf"), 0),
"test": (float("inf"), 0)
})
self.overall_best_l1_dist = 0
self.stats = defaultdict(lambda: {
"train": StatsMeter(),
"val": StatsMeter(),
"test": StatsMeter()
})
self.build_model()
self.prepare_data()
self.configure_optimizers()
def build_model(self):
print("MODEL:")
self.NFC = NextFrameClassifier(self.hp)
line()
def prepare_data(self):
if "timit" in self.hp.data and "buckeye" in self.hp.data:
# joint training on TIMIT + Buckeye (data: timit_buckeye)
t_train, t_val, t_test = TrainTestDataset.get_datasets(path=self.hp.timit_path)
b_train, b_val, b_test = TrainValTestDataset.get_datasets(path=self.hp.buckeye_path, percent=self.hp.buckeye_percent)
train = ConcatDataset([t_train, b_train])
val = ConcatDataset([t_val, b_val])
test = ConcatDataset([t_test, b_test])
train.path = f"TIMIT+Buckeye train ({len(t_train)}+{len(b_train)})"
val.path = f"TIMIT+Buckeye val ({len(t_val)}+{len(b_val)})"
test.path = f"TIMIT+Buckeye test ({len(t_test)}+{len(b_test)})"
elif "timit" in self.hp.data:
train, val, test = TrainTestDataset.get_datasets(path=self.hp.timit_path)
elif "buckeye" in self.hp.data:
train, val, test = TrainValTestDataset.get_datasets(path=self.hp.buckeye_path, percent=self.hp.buckeye_percent)
else:
raise Exception("no such training data!")
self.train_dataset = train
self.valid_dataset = val
self.test_dataset = test
line()
print("DATA:")
print(f"train: {self.train_dataset.path} ({len(self.train_dataset)})")
print(f"valid: {self.valid_dataset.path} ({len(self.valid_dataset)})")
print(f"test: {self.test_dataset.path} ({len(self.test_dataset)})")
line()
_persistent = self.hp.dataloader_n_workers > 0
self.train_loader = DataLoader(self.train_dataset,
batch_size=self.hp.batch_size,
shuffle=True,
collate_fn=collate_fn_padd,
num_workers=self.hp.dataloader_n_workers,
pin_memory=True,
persistent_workers=_persistent)
self.valid_loader = DataLoader(self.valid_dataset,
batch_size=self.hp.batch_size,
shuffle=False,
collate_fn=collate_fn_padd,
num_workers=self.hp.dataloader_n_workers,
pin_memory=True,
persistent_workers=_persistent)
self.test_loader = DataLoader(self.test_dataset,
batch_size=self.hp.batch_size,
shuffle=False,
collate_fn=collate_fn_padd,
num_workers=self.hp.dataloader_n_workers,
pin_memory=True,
persistent_workers=_persistent)
def configure_optimizers(self):
parameters = filter(lambda p: p.requires_grad, self.NFC.parameters())
if self.hp.optimizer == "sgd":
self.optimizer = optim.SGD(parameters, lr=self.hparams.lr, momentum=0.9, weight_decay=5e-4)
elif self.hp.optimizer == "adam":
self.optimizer = optim.Adam(parameters, lr=self.hparams.lr, weight_decay=5e-4)
elif self.hp.optimizer == "adamW":
self.optimizer = optim.AdamW(parameters, lr=self.hparams.lr)
elif self.hp.optimizer == "ranger":
self.optimizer = optim_extra.Ranger(parameters, lr=self.hparams.lr, alpha=0.5, k=6, N_sma_threshhold=5, betas=(.95, 0.999), eps=1e-5, weight_decay=0)
else:
raise Exception("unknown optimizer")
print(f"optimizer: {self.optimizer}")
line()
self.scheduler = optim.lr_scheduler.StepLR(self.optimizer,
step_size=self.hp.lr_anneal_step,
gamma=self.hp.lr_anneal_gamma)
return [self.optimizer]
def train_one_epoch(self):
self.run_epoch("train", self.current_epoch)
return self.stats['nfc_loss']['train'].get_stats()
def validate(self):
self.evaluate("val", self.current_epoch)
return self.stats['nfc_loss']['val'].get_stats()
def run_epoch(self, mode, epoch):
self.current_epoch = epoch
is_train = mode=="train"
loader = getattr(self, f"{mode}_loader")
self.NFC.train(mode=="train")
for batch_i, batch in enumerate(tqdm(loader)):
if is_train:
self.optimizer.zero_grad()
result = self.forward(batch,batch_i,mode)
if is_train:
result['loss'].backward()
self.optimizer.step()
if is_train:
self.scheduler.step()
self.generic_eval_end(mode, epoch)
def evaluate(self, mode, epoch=None):
loader = self.valid_loader if mode == 'val' else self.test_loader
self.NFC.eval()
with torch.no_grad():
for batch_i, batch in enumerate(tqdm(loader)):
self.forward(batch, batch_i, mode)
self.generic_eval_end(mode, epoch if epoch is not None else self.current_epoch)
def test(self):
self.evaluate('test', epoch=-1)
def forward(self, data_batch, batch_i, mode):
loss = 0
# TRAIN
audio, seg, phonemes, length, fname = data_batch
language = "english" # "dutch"
if language == "dutch":
lh39_ph = []
for phoneme_seq in phonemes:
lh39_ph_seq = []
for IFA_ph in phoneme_seq:
print(f"\nINPUT: {IFA_ph}")
output = dutch_preprocess.aligner_pipeline(timit_to_leehon_map_MACRO[IFA_ph.lower()])
lh39_ph_seq.append([x["lh39"] for x in output])
if not output:
print("Results: None")
print(phoneme_seq)
print(f"Dutch IPA to LH39 mapping: {lh39_ph_seq}")
lh39_ph.append(np.hstack(lh39_ph_seq).tolist())
print(f"Dutch IPA to LH39 total: {lh39_ph}")
phonemes = lh39_ph
if mode in ["test", "val"]:
with torch.no_grad():
self.NFC.eval()
preds,original_lengths, probs, frame_labels, _,preds_peaks, w_phi = self.NFC(audio,None,phonemes,length)
else:
self.NFC.train()
preds,original_lengths, probs, frame_labels,_,preds_peaks, w_phi = self.NFC(audio,seg,phonemes,length)
epoch = self.current_epoch
if epoch < 5:
total_loss, ph_loss, loss_nce, sum_mse, w_pos_neg, w_phi = self.NFC.loss_ph(preds,original_lengths, probs, frame_labels, seg,preds_peaks, w_phi, phonemes)
else:
total_loss, ph_loss, loss_nce, sum_mse, w_pos_neg, w_phi = self.NFC.total_loss(preds,original_lengths, probs, frame_labels, seg,preds_peaks, w_phi, phonemes)
# InfoNCE LOSS ABLATION -
# total_loss, ph_loss, loss_nce, sum_mse, w_pos_neg, w_phi = self.NFC.loss_InfoNCE_classic(preds, original_lengths, probs, frame_labels, seg, preds_peaks, w_phi, phonemes)
NFC_loss = total_loss
self.stats['w_pos'][mode].update(torch.tensor(w_pos_neg[0].item()))
self.stats['w_neg'][mode].update(torch.tensor(w_pos_neg[1].item()))
self.stats['w_phi1'][mode].update(torch.tensor(w_phi[0].item()))
self.stats['w_phi2'][mode].update(torch.tensor(w_phi[1].item()))
self.stats['ph_loss'][mode].update(torch.tensor(ph_loss.item()))
self.stats['nce_loss'][mode].update(torch.tensor(loss_nce.item()))
self.stats['softDPmse_loss'][mode].update(torch.tensor(sum_mse.item()))
self.stats['nfc_loss'][mode].update(torch.tensor(NFC_loss.item()))
loss += NFC_loss
loss_key = "loss" if mode == "train" else f"{mode}_loss"
return OrderedDict({
loss_key: loss
})
def generic_eval_end(self, mode, epoch):
metrics = {}
data = self.hp.data
for k, v in self.stats.items():
metrics[f"{mode}_{k}"] = self.stats[k][mode].get_stats()
metrics['epoch'] = epoch+1
metrics['current_lr'] = self.optimizer.param_groups[0]['lr']
line()
# get best_l1_dist from all l1_dist types and all epochs
best_overall_l1_dist = float("inf")
for pred_type, l1_dist in self.best_l1_dist.items():
if l1_dist[mode][0] < best_overall_l1_dist:
best_overall_l1_dist = l1_dist[mode][0]
metrics[f'{mode}_min_l1_dist'] = best_overall_l1_dist
for k, v in metrics.items():
print(f"\t{k:<30} -- {v}")
line()
wandb.log(metrics, step=epoch)
output = OrderedDict({
'log': metrics
})
return output
def get_ckpt_path(self):
# return glob.glob(self.hp.wd + "/*.ckpt")[0]
return glob.glob(os.path.join(self.hp.wd + "/*.ckpt"))[0] |