File size: 5,439 Bytes
3ade64c |
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 |
import sys
import torch
from tqdm import tqdm
from torch.utils.data import DataLoader
from torch.optim import AdamW
from transformers import get_scheduler
import utils_ctc
from models import Swin_CTC, VED
from mydatasets import myDatasetCTC, myDatasetTransformerDecoder
torch.set_float32_matmul_precision('medium')
#################################################################
# Experiment Settings
#################################################################
NUM_EPOCHS = int(sys.argv[0])
LR = float(sys.argv[1])
STRATEGY = str(sys.argv[2])
BATCH_SIZE = int(sys.argv[3])
MODEL_NAME = str(sys.argv[4])
NUM_ACCUMULATION_STEPS = int(sys.argv[5])
print(30*'*')
print("EXPERIMENT PARAMS: ")
print("\tNUM_EPOCHS: ", NUM_EPOCHS)
print("\tLR: ", LR)
print("\tSTRATEGY: ", STRATEGY)
print("\tBATCH_SIZE: ", BATCH_SIZE)
print("\tMODEL_NAME: ", MODEL_NAME)
print("\tNUM_ACCUMULATION_BATCHES: ", NUM_ACCUMULATION_STEPS)
print(30*'*')
#################################################################
# Load Torch Dataset and Create Vocab
#################################################################
l_of_transcrips = []
if MODEL_NAME == "Swin_CTC":
train_dataset = myDatasetCTC(partition="train")
else:
train_dataset = myDatasetTransformerDecoder(partition="train")
l_of_transcrips = train_dataset.label_list
text_to_seq, seq_to_text = utils_ctc.create_char_dicts(l_of_transcrips)
# update dics in datasets
train_dataset.text_to_seq = text_to_seq
train_dataset.seq_to_text = seq_to_text
print("Len dict text_to_seq: ", len(text_to_seq))
print("Len dict seq_to_text: ", len(seq_to_text))
print("Dict text_to_seq: ", (text_to_seq))
print("Dict seq_to_text: ", (seq_to_text))
#################################################################
# Load Model
#################################################################
# Create model
if MODEL_NAME == "Swin_CTC":
model = Swin_CTC(len(text_to_seq))
else:
model = VED()
#################################################################
# Training Settings
#################################################################
device = "cuda:0"
if MODEL_NAME == "Swin_CTC":
mycollate_fn = utils_ctc.custom_collate
else:
mycollate_fn = None
train_dataloader = DataLoader(
train_dataset,
BATCH_SIZE,
shuffle=True,
num_workers=23,
collate_fn=mycollate_fn)
optimizer = AdamW(model.parameters(), lr=LR, betas=(0.9, 0.999), eps=1e-6, weight_decay=0.0)
num_training_steps = NUM_EPOCHS # * len(train_dataloader)
lr_scheduler = get_scheduler(
"linear", optimizer=optimizer, num_warmup_steps=0, num_training_steps=num_training_steps
)
#################################################################
# Frozen Strategies
#################################################################
model.to(device)
model.train()
if MODEL_NAME == "Swin_CTC":
if STRATEGY == "CTC-fclayer":
for name_p,p in model.named_parameters():
p.requires_grad = False
if "projection_V" in name_p:
p.requires_grad = True
print("Train only: ", name_p)
elif STRATEGY == "CTC-Swin":
for name_p,p in model.named_parameters():
p.requires_grad = True
if "projection_V" in name_p:
p.requires_grad = False
print("No train: ", name_p)
else:
for name_p,p in model.named_parameters():
p.requires_grad = True
print("Train all layers")
else:
if STRATEGY == "VED-encoder":
for name_p,p in model.named_parameters():
p.requires_grad = False
if "model.encoder." in name_p:
p.requires_grad = True
print("Train only: ", name_p)
elif STRATEGY == "VED-decoder":
for name_p,p in model.named_parameters():
p.requires_grad = False
if "model.decoder." in name_p:
p.requires_grad = True
print("Train only: ", name_p)
else:
for name_p,p in model.named_parameters():
p.requires_grad = True
print("Train all layers")
def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
print("Params: ", count_parameters(model))
#################################################################
# Training
#################################################################
for epoch in range(NUM_EPOCHS):
epoch_loss = 0
print("Epoch ", epoch)
idx = 0
optimizer.zero_grad(set_to_none=True)
model.train()
with tqdm(iter(train_dataloader), desc="Training set", unit="batch") as tepoch:
for batch in tepoch:
inputs: torch.Tensor = batch["img"].to(device)
labels: torch.Tensor = batch["label"].to(device)
if MODEL_NAME == "Swin_CTC":
target_lengths: torch.Tensor = batch["target_lengths"].to(device)
outputs, loss = model(inputs, labels, target_lengths)
else:
outputs, loss = model(inputs, labels)
loss.backward()
if ((idx + 1) % NUM_ACCUMULATION_STEPS == 0):
optimizer.step()
optimizer.zero_grad(set_to_none=True)
tepoch.set_postfix(loss=loss.data.item())
epoch_loss += loss.data.item()
idx += 1
# Save Final model
torch.save(model.state_dict(), './FINAL_MODEL') |