def run():
# Load Dataset
train_dataset = dataset.autoencoderDataset("bibtex_train.embeddings")
test_dataset = dataset.autoencoderDataset("bibtex_test.embeddings")
# Dataloaders
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size = 1, shuffle=True, num_workers = 4)
test_dataloader = torch.utils.data.DataLoader(test_dataset, batch_size = 1, num_workers = 1)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
autoencoder_model = model.autoEncoder(100) # 100 is the input dimension
autoencoder_model.to(device)
# Creating the optimizer
optimizer = torch.optim.Adam(autoencoder_model.parameters(), lr=1e-3)
for epoch in range(0,10):
training_loss = engine.train(train_dataloader, autoencoder_model, optimizer, device)
print("Epoch: {} Loss: {}".format(epoch+1, training_loss))
# Model evaluation
engine.eval(test_dataloader, autoencoder_model, device)
# Generating the embeddings now
train_set_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size = 1, shuffle = False)
test_set_dataloader = torch.utils.data.DataLoader(test_dataset, batch_size=1, shuffle = False)
embed_list = engine.generate_low_dimensional_embeddings(train_set_dataloader, autoencoder_model, device)
embed_list.extend(engine.generate_low_dimensional_embeddings(test_set_dataloader, autoencoder_model, device))
with open("bibtex_low_dimension.embeddings", mode='w+') as file:
for each_elem in embed_list:
line_to_write = " ".join(str(v) for v in each_elem[0])+'\n'
file.write(line_to_write)
def evaluate(model, dataset, device, load_model=True, load_dataset=True, draw_results=False):
try:
'''
if load_* params. are True then method accepts path for
model, dataset, otherwise objects.
load_dataset sent without '.pkl'
'''
if load_model is True:
model = torch.load(model) # model is expected to be the model path.
elif load_dataset is True:
dataset = load_obj(dataset)
eval(model, dataset, device, draw=draw_results)
except Exception as e:
print(e)
return False
return True
def run_training():
image_files = glob.glob(os.path.join(config.DATA_DIR, '*.png'))
targets_orig = [x.split('/')[-1][:-4] for x in image_files]
targets = [[c for c in x] for x in targets_orig]
targets_flat = [c for clist in targets for c in clist]
label_enc = preprocessing.LabelEncoder()
label_enc.fit(targets_flat)
targets_enc = [label_enc.transform(x) for x in targets]
targets_enc = np.array(targets_enc) + 1
train_imgs, test_imgs, train_targets, test_targets, train_orig_targets, test_orig_targets = model_selection.train_test_split(
image_files, targets_enc, targets_orig, test_size=0.1, random_state=42)
train_dataset = dataset.Classification(image_paths=train_imgs,
targets=train_targets,
resize=(config.IMAGE_HEIGHT,
config.IMAGE_WIDTH))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=config.BATCH_SIZE,
shuffle=True)
test_dataset = dataset.Classification(image_paths=test_imgs,
targets=test_targets,
resize=(config.IMAGE_HEIGHT,
config.IMAGE_WIDTH))
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=config.BATCH_SIZE,
shuffle=False)
model = CaptchaModel(num_chars=len(label_enc.classes_))
model = model.to(config.DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=3e-4)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.8,
patience=5,
verbose=True)
for epoch in range(config.EPOCHS):
train_loss = engine.train(model, train_loader, optimizer)
val_preds, valid_loss = engine.eval(model, test_loader)
print(
f"Epoch: {epoch}: Train loss: {train_loss}, Valid loss: {valid_loss}"
)
valid_cap_preds = []
for vp in val_preds:
current_preds = decode_predictions(vp, label_enc)
valid_cap_preds.extend(current_preds)
print(list(zip(test_orig_targets, valid_cap_preds))[6:11])
def main():
image_paths = glob.glob(os.path.join(config.PATH, "*.png"))
image_paths = [path.replace("\\", "/") for path in image_paths]
targets = [path.split("/")[-1][:-4] for path in image_paths]
targets_listed = [[char for char in target] for target in targets]
targets_flattened = [char for target in targets_listed for char in target]
label_enc = preprocessing.LabelEncoder()
label_enc.fit(targets_flattened)
targets_encoded = np.array(
[label_enc.transform(target) for target in targets_listed])
targets_encoded += 1 # to keep the "0" class for UNK chars
(
train_imgs,
valid_imgs,
train_enc_targets,
valid_enc_targets,
_,
valid_targets,
) = model_selection.train_test_split(image_paths,
targets_encoded,
targets,
test_size=0.1,
random_state=42)
train_loader = make_loader(
mode="train",
image_paths=train_imgs,
targets=train_enc_targets,
size=(config.HEIGHT, config.WIDTH),
resize=True,
)
valid_loader = make_loader(
mode="valid",
image_paths=valid_imgs,
targets=valid_enc_targets,
size=(config.HEIGHT, config.WIDTH),
resize=True,
)
model = OCRModel(num_classes=len(label_enc.classes_),
dropout=config.DROPOUT).to(config.DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=config.LR)
lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
mode="min",
patience=config.PATIENCE,
factor=config.FACTOR,
verbose=True,
)
if config.MODE == "train":
best_loss = float("inf")
for epoch in range(config.EPOCHS):
model.train()
_ = engine.train(model, train_loader, optimizer)
model.eval()
with torch.no_grad():
valid_preds, valid_loss = engine.eval(model, valid_loader)
captcha_preds = []
for preds in valid_preds:
preds_ = decode_predictions(preds, label_enc)
captcha_preds.extend(preds_)
print(f"Epoch: {epoch}")
pprint(list(zip(valid_targets, captcha_preds))[:10])
lr_scheduler.step(valid_loss.avg)
if valid_loss.avg < best_loss:
best_loss = valid_loss.avg
torch.save(model.state_dict(), "model.pt")
else:
model.load_state_dict(
torch.load("./models/model.pt", map_location=config.DEVICE))
model.eval()
with torch.no_grad():
valid_preds, valid_loss = engine.eval(model, valid_loader)
captcha_preds = []
for preds in valid_preds:
preds_ = decode_predictions(preds, label_enc)
captcha_preds.extend(preds_)
pprint(list(zip(valid_targets, captcha_preds))[:10])
return valid_loader, captcha_preds, valid_targets
ecgs, y_true, y_pred = test(
net=net,
test_loader=loader,
device=DEVICE,
batch_size=BATCH_SIZE,
plot_ecg=False,
plot_ecg_windows_size=WINDOWS_SIZE
)
eval(
ecgs=ecgs,
y_true=y_true,
y_pred=y_pred,
labels=[0, 1, 2, 3, 4],
target_names=['none', 'p_wave', 'qrs', 't_wave', 'extrasystole'],
plot_acc=True,
plot_loss=True,
plot_conf_matrix=True,
plot_ecg=True,
plot_ecg_windows_size=WINDOWS_SIZE
)
def run():
trainDataset, testDataset, labelGenerator = utils.loadDataset()
# Making DataLoaders
trainDataLoader = torch.utils.data.DataLoader(
trainDataset,
batch_size=config.TRAIN_BATCH_SIZE,
shuffle=True,
num_workers=4,
pin_memory=True)
testDataLoader = torch.utils.data.DataLoader(
testDataset, batch_size=config.TEST_BATCH_SIZE, num_workers=1)
totalNOsOfLabels = len(labelGenerator.classes_)
device = torch.device(config.DEVICE)
# Defining Model
print("Making model:- ", config.modelName)
citeModel = None
if config.modelName == "BertBase":
citemodel = model.BERTBaseUncased(numOfLabels=totalNOsOfLabels,
dropout=config.DROPOUT)
elif config.modelName == "SciBert":
citemodel = model.SciBertUncased(numOfLabels=totalNOsOfLabels,
dropout=config.DROPOUT)
citemodel.to(device)
param_optimizer = list(citemodel.named_parameters())
'''
There is generally no need to apply L2 penalty (i.e. weight decay) to biases and LayerNorm.weight.
Hence, we have following line.
Update: There is need to apply L2 to LayerNorm.weight as per Google TF implementation so reverting it ;)
'''
no_decay = ["bias", "LayerNorm.bias",
"LayerNorm.weight"] # Removed "LayerNorm.bias",
optimizer_parameters = [
{
"params": [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
"weight_decay":
0.01, # changed this from 0.001 to 0.1
},
{
"params":
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
"weight_decay":
0.0,
}
]
num_train_steps = int(len(trainDataLoader) * config.EPOCHS)
optimizer = AdamW(optimizer_parameters, lr=config.LEARNING_RATE)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=num_train_steps * config.WARMUP_PROPORTION,
num_training_steps=num_train_steps)
if config.dotrain:
print('In Training')
for epoch in range(config.EPOCHS):
trainingLoss = engine.train(trainDataLoader, citemodel, optimizer,
device, scheduler)
print("Epoch: ", epoch, " Loss: ", trainingLoss, '\n')
# Saving the model
os.makedirs(os.path.dirname(config.MODEL_SAVED.format(
config.modelName)),
exist_ok=True)
torch.save(citemodel.state_dict(),
config.MODEL_SAVED.format(config.modelName))
print('Model is saved at: ',
config.MODEL_SAVED.format(config.modelName))
'''
Evaluating the model
'''
print("Loading the model")
#citemodel = model.BERTBaseUncased(*args, **kwargs)
citemodel.load_state_dict(
torch.load(config.MODEL_SAVED.format(config.modelName)))
outputs, targets = engine.eval(testDataLoader, citemodel, device)
# Saving the results with corresponding targets
os.makedirs(os.path.dirname(
config.PREDICTIONS_PATH.format(config.modelName)),
exist_ok=True)
with open(config.PREDICTIONS_PATH.format(config.modelName), 'wb') as f:
pickle.dump(outputs, f) # First saved the predicted outputs
pickle.dump(targets, f) # Then saved the corresponding targets
print('Starting Evaluation...')
utils.metric(outputs, targets)
def run():
df = pd.read_csv('input/IMDB Dataset.csv').sample(frac=1).reset_index(drop=True)
df.sentiment = df.sentiment.apply(
lambda x:
1 if x == 'positive' else 0
)
df_train, df_val = train_test_split(df, test_size=0.1)
train_data = DataLoader.DataLoader(df_train)
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=config.Batch_Size,
num_workers=4,
pin_memory=True,
collate_fn=DataLoader.MyCollate(config.pad_idx)
)
val_data = DataLoader.DataLoader(df_val)
val_loader = torch.utils.data.DataLoader(
val_data,
batch_size=config.Batch_Size,
num_workers=4,
pin_memory=True,
collate_fn=DataLoader.MyCollate(config.pad_idx)
)
if torch.cuda.is_available():
accelarator = 'cuda'
torch.backends.cudnn.benchmark=True
else:
accelarator = 'cpu'
device = torch.device(accelarator)
model = BertSentimentModel.BERTMODEL()
model = model.to(device)
model_param = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_param = [
{'params': [p for n, p in model_param if not any(nd in n for nd in no_decay)], 'weight_decay':0.001},
{'params': [p for n, p in model_param if any(nd in n for nd in no_decay)], 'weight_decay':0.0 }
]
num_training_steps = len(df_train)*config.Epochs//config.Batch_Size
optimizer = transformers.AdamW(optimizer_param, lr=3e-5)
scheduler = transformers.get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=0,
num_training_steps=num_training_steps
)
best_loss = 1e32
best_model = None
print('---------- [INFO] STARTING TRAINING ---------')
for epoch in range(config.Epochs):
train_acc, train_loss = engine.train(
model,
train_loader,
optimizer,
scheduler,
device
)
val_acc, val_loss = engine.eval(
model,
val_loader,
device
)
print(f'EPOCH : {epoch+1}/{config.Epochs}')
print(f'TRIAN_ACC = {train_acc}% | TRAIN LOSS = {train_loss}')
print(f'VAL ACC = {val_acc}% | VAL LOSS {val_loss}')
if best_loss > val_loss:
best_loss = val_loss
best_model = model.state_dict()
torch.save(best_model, config.Model_Path)
def run():
df = pd.read_csv('input/train.csv').sample(frac=1).dropna().reset_index(
drop=True)
df_train, df_val = train_test_split(df,
test_size=0.1,
stratify=df.sentiment.values)
train_data = DataLoader.DataLoader(df_train)
val_data = DataLoader.DataLoader(df_val)
train_loader = torch.utils.data.DataLoader(
train_data,
num_workers=0,
batch_size=CONFIG.Batch_Size,
pin_memory=True,
collate_fn=DataLoader.MyCollate())
val_loader = torch.utils.data.DataLoader(val_data,
num_workers=0,
batch_size=CONFIG.Batch_Size,
pin_memory=True,
collate_fn=DataLoader.MyCollate())
if torch.cuda.is_available():
compute = 'cuda'
torch.backends.cudnn.benchmark = True
else:
compute = 'cpu'
device = torch.device(compute)
device = torch.device('cpu')
model = ROBERTA_MODEL.QnAModel(CONFIG.Dropout)
mode = model.to(device)
parameters = list(model.parameters())
decay_parameter = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimized_parameters = [{
'params': [
p for n, p in model.named_parameters()
if not any(nd in n for nd in decay_parameter)
],
'weight_decay':
0.001
}, {
'params': [
p for n, p in model.named_parameters()
if any(nd in n for nd in decay_parameter)
],
'weight_decay':
0.0
}]
optimizer = transformers.AdamW(optimized_parameters, lr=CONFIG.LR)
num_training_steps = len(df_train) * CONFIG.Epochs / CONFIG.Batch_Size
scheduler = transformers.get_linear_schedule_with_warmup(
optimizer, num_training_steps=num_training_steps, num_warmup_steps=0)
best_loss = 100
best_jaccard = 0
print('-------------[INFO] STARTING TRANING--------------\n')
for epoch in range(CONFIG.Epochs):
train_loss = engine.train(model, train_loader, optimizer, scheduler,
device)
val_loss, jaccard_score = engine.eval(model, val_loader, device)
print(
f'EPOCHS -> {epoch+1}/{CONFIG.Epochs} | TRAINING LOSS = {train_loss} | VAL LOSS = {val_loss} | JACCARD SCORE = {jaccard_score}\n'
)
if jaccard_score > best_jaccard:
best_jaccard = jaccard_score
best_model = model.state_dict()
torch.save(best_model, CONFIG.MODLE_PATH)
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.001
},
{
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
},
]
num_train_steps = int(len(df_train) / config.TRAIN_BATCH_SIZE * config.EPOCHS)
optimizer = AdamW(optimizer_parameters, lr=3e-5)
scheduler = get_linear_schedule_with_warmup(optimizer,
num_warmup_steps=0,
num_training_steps=num_train_steps)
model = nn.DataParallel(model)
best_accuracy = 0
for epoch in range(config.EPOCHS):
print('======== Epoch {} ========'.format(epoch))
engine.train(train_data_loader, model, optimizer, scheduler, device)
outputs, targets = engine.eval(valid_data_loader, model, device)
outputs = np.array(outputs) >= 0.5
accuracy = metrics.accuracy_score(targets, outputs)
print(f"Accuracy Score = {accuracy}")
if accuracy > best_accuracy:
torch.save(model.state_dict(), config.MODEL_PATH)
best_accuracy = accuracy
def run_training():
image_files = glob.glob("../input/dl_captcha/train_images/*.png")
print(image_files.__len__())
targets = [int(i.split("/")[-1].split(".")[0].split("_")[1])
for i in image_files]
print(targets.__len__())
lbl_encoder = preprocessing.LabelEncoder()
lbl_encoder.fit(targets)
enc_targets = np.array(lbl_encoder.transform(targets))
print(len(lbl_encoder.classes_))
train_imgs, val_imgs, train_enc_targets, val_enc_targets, train_orig_targets, val_orig_targets = model_selection.train_test_split(
image_files, enc_targets, targets, test_size=0.1, random_state=42)
train_dataset = dataset.RotnetDataset(
train_imgs, train_enc_targets, resize=(config.IMAGE_HEIGHT, config.IMAGE_WIDTH))
print(len(train_dataset))
val_dataset = dataset.RotnetDataset(
val_imgs, val_enc_targets, resize=(config.IMAGE_HEIGHT, config.IMAGE_WIDTH))
print(len(val_dataset))
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=config.BATCH_SIZE,
shuffle=True, num_workers=config.NUM_WORKERS
)
val_loader = torch.utils.data.DataLoader(
val_dataset, batch_size=config.BATCH_SIZE,
shuffle=False, num_workers=config.NUM_WORKERS
)
model = RotNetclassifier(num_classes=len(lbl_encoder.classes_))
model.to(config.DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=3e-4)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, factor=0.8, patience=5, verbose=True
)
for epoch in range(config.EPOCHS):
train_loss = engine.train(model, train_loader, optimizer)
valid_preds, valid_loss = engine.eval(model, val_loader)
final_val_preds = []
for vp in valid_preds:
current_preds = decode_preds(vp, lbl_encoder)
final_val_preds.extend(current_preds)
combined = list(zip(val_orig_targets, final_val_preds))
print(combined[:20])
accuracy = metrics.accuracy_score(final_val_preds, val_orig_targets)
print(
f"Epoch={epoch}, Train Loss={train_loss}, Test Loss={valid_loss} Accuracy={accuracy}"
)
scheduler.step(valid_loss)
joblib.dump(lbl_encoder, "../input/pickles/lbl_encoder.pkl")
torch.save(model.state_dict(), "../input/pickles/captcha.pth")
def run():
df = pd.read_csv('input/ner_dataset.csv', encoding='latin-1')
df['Sentence #'] = df['Sentence #'].fillna(method='ffill')
if os.path.exists('input/pos_lb.pickle'):
pos_lb = pickle.load(open('input/pos_lb.pickle', 'rb'))
else:
pos_lb = LabelEncoder().fit(df.POS.values)
pickle.dump(pos_lb, open('input/pos_lb.pickle', 'wb'))
df['POS'] = pos_lb.transform(df.POS.values)
pos_pad_idx = pos_lb.transform(['.'])[0]
sentence = df.groupby('Sentence #')['Word'].apply(list).values
pos = df.groupby('Sentence #')['POS'].apply(list).values
print('-------- [INFO] TOKENIZING --------\n')
data = DataLoader.DataLoader(sentence, pos)
data_len = len(data)
indices = np.arange(0, data_len)
valid_len = int(data_len * CONFIG.Valid_split)
train_index = indices[valid_len:]
valid_index = indices[:valid_len]
train_sampler = torch.utils.data.sampler.SubsetRandomSampler(train_index)
valid_sampler = torch.utils.data.sampler.SubsetRandomSampler(valid_index)
if not os.path.exists('input/word_to_idx.pickle'):
pickle.dump(data.vocab.word_to_idx,
open('input/word_to_idx.pickle', 'wb'))
pad_idx = data.vocab.word_to_idx['<PAD>']
train_loader = torch.utils.data.DataLoader(data,
num_workers=4,
batch_size=CONFIG.BATCH_SIZE,
pin_memory=True,
collate_fn=DataLoader.MyCollate(
pad_idx, pos_pad_idx),
sampler=train_sampler)
valid_loader = torch.utils.data.DataLoader(data,
num_workers=4,
batch_size=CONFIG.BATCH_SIZE,
pin_memory=True,
collate_fn=DataLoader.MyCollate(
pad_idx, pos_pad_idx),
sampler=valid_sampler)
vocab_size = len(data.vocab.word_to_idx)
num_pos_class = len(list(pos_lb.classes_))
tag_to_idx = {str(x): num for num, x in enumerate(list(pos_lb.classes_))}
tag_to_idx['start_tag'] = num_pos_class
tag_to_idx['stop_tag'] = num_pos_class + 1
if torch.cuda.is_available():
accelarator = 'cuda'
torch.backends.cudnn.benchmark = True
else:
accelarator = 'cpu'
device = torch.device(accelarator)
model = NERCRFModel.NER(vocab_size=vocab_size,
embed_dims=CONFIG.EMBED_DIMS,
hidden_dims=CONFIG.HIDDEN_DIMS,
num_layers=CONFIG.NUM_HIDDEN_LAYER,
num_classes=len(tag_to_idx),
dropout=CONFIG.DROPOUT,
bidirectional=CONFIG.BIDIRECTIONAL,
tag_to_idx=tag_to_idx,
device=device)
model = model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=CONFIG.LR)
best_loss = 1e32
print('\n-----------[INFO] STARTING TRAINING ----------------\n')
for epoch in range(CONFIG.EPOCHS):
train_loss = engine.train(model, train_loader, optimizer, device)
eval_loss, val_pos_acc = engine.eval(model, valid_loader, device)
print(f'EPOCH -> {epoch+1}/{CONFIG.EPOCHS}')
print(f'TRAIN LOSS = {np.round(train_loss, 5)}')
print(
f'VAL LOSS = {np.round(eval_loss, 5)} | VAL POS ACC = {np.round(val_pos_acc*100, 5)}%'
)
if best_loss > eval_loss:
best_loss = eval_loss
best_model = model.state_dict()
torch.save(best_model, CONFIG.Model_Path)
def run():
df = pd.read_csv(CONFIG.Path).sample(frac=1).reset_index(drop=True)[:10]
mean = (0.5, 0.5, 0.5)
std = (0.5, 0.5, 0.5)
mean = (0.485, 0.456, 0.406)
std = (0.229, 0.224, 0.225)
aug = albumentations.Compose([
albumentations.Normalize(mean, std, always_apply=True),
albumentations.RandomBrightness(),
albumentations.HueSaturationValue(),
albumentations.Resize(224, 224, always_apply=True),
AT.ToTensor()
])
print('-------[INFO] TOKENIZING CAPTIONS -------')
dataset = DataLoader.DataLoader(df, aug)
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(0.1 * dataset_size))
train_indices, val_indices = indices[split:], indices[:split]
train_sampler = torch.utils.data.sampler.SubsetRandomSampler(train_indices)
valid_sampler = torch.utils.data.sampler.SubsetRandomSampler(val_indices)
pickle.dump(dataset.vocab.word_to_idx,
open('model/word_to_idx.pickle', 'wb'))
pickle.dump(dataset.vocab.idx_to_word,
open('model/idx_to_word.pickle', 'wb'))
pad_idx = dataset.vocab.word_to_idx['<PAD>']
train_loader = torch.utils.data.DataLoader(
dataset,
batch_size=CONFIG.BATCH_SIZE,
sampler=train_sampler,
pin_memory=True,
num_workers=8,
collate_fn=DataLoader.MyCollate(pad_idx=pad_idx))
val_loader = torch.utils.data.DataLoader(
dataset,
batch_size=CONFIG.BATCH_SIZE,
sampler=valid_sampler,
pin_memory=True,
num_workers=8,
collate_fn=DataLoader.MyCollate(pad_idx=pad_idx))
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch.backends.cudnn.benchmark = True
model = ImageCaptioningModel.EncoderDecoder(
embedding_dims=CONFIG.embedding_dims,
vocab_size=len(dataset.vocab.word_to_idx),
hidden_dims=CONFIG.hidden_dims,
num_layers=CONFIG.num_layer,
bidirectional=CONFIG.bidirectional,
dropout=CONFIG.dropout)
torch.backends.cudnn.benchmark = True
model = model.to(device)
for name, param in model.encoder.base_model.named_parameters():
if "linear" in name:
param.requires_grad = True
else:
param.requires_grad = False
optimizer = torch.optim.AdamW(model.parameters(), lr=CONFIG.LR)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=2,
threshold=0.001,
mode='min')
best_loss = 1e4
print(
'------------------------------ [INFO] STARTING TRAINING --------------------------------'
)
for epoch in range(CONFIG.EPOCHS):
print(f'-----EPOCH - {epoch+1}/ {CONFIG.EPOCHS} -----')
train_loss = engine.train(model, train_loader, optimizer, device,
pad_idx)
val_loss = engine.eval(model, val_loader, device, pad_idx)
scheduler.step(val_loss)
print(f'Train Loss = {train_loss} | Eval Loss = {val_loss}\n')
if best_loss > val_loss:
best_loss = val_loss
best_model = model.state_dict()
torch.save(best_model, CONFIG.MODEL_PATH)
predict.predict('1.jpg')
def run_training():
# image_files
image_files = glob.glob("../input/train_all_captchas/*.png")
print(image_files[:4])
# targets
targets_orig = [i.split("/")[-1][:-4] for i in image_files]
print(targets_orig[:5])
# creating a list of list for the targets
targets = [[j for j in i] for i in targets_orig]
# flattening the lists
targets_flat = [item for sublists in targets for item in sublists]
# print(targets_flat)
lbl_encoder = preprocessing.LabelEncoder()
lbl_encoder.fit(targets_flat)
enc_targets = [lbl_encoder.transform(x) for x in targets]
# this +1 is to add 1 to all the encoded labels, so that we could use 0 for the unknown values
enc_targets = np.array(enc_targets) + 1
print(len(enc_targets))
print(len(lbl_encoder.classes_))
(
train_imgs,
test_imgs,
train_targets_orig,
test_target_orig,
train_targets,
test_targets,
) = model_selection.train_test_split(image_files,
targets_orig,
enc_targets,
test_size=0.1,
random_state=42)
print(len(train_imgs), len(train_targets))
print(len(test_imgs), len(test_targets))
train_dataset = dataset.ClassificationDataset(
image_paths=train_imgs,
targets=train_targets,
resize=(config.IMAGE_HEIGHT, config.IMAGE_WIDTH),
)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config.BATCH_SIZE,
num_workers=config.NUM_WORKERS,
shuffle=True,
)
test_dataset = dataset.ClassificationDataset(
image_paths=test_imgs,
targets=test_targets,
resize=(config.IMAGE_HEIGHT, config.IMAGE_WIDTH),
)
test_dataloader = torch.utils.data.DataLoader(
test_dataset,
batch_size=config.BATCH_SIZE,
num_workers=config.NUM_WORKERS,
shuffle=False,
)
model = CaptchaModel(num_chars=len(lbl_encoder.classes_))
model.to(config.DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=3e-4)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.8,
patience=5,
verbose=True)
for epoch in range(config.EPOCHS):
train_loss = engine.train(model, train_dataloader, optimizer)
valid_preds, valid_loss = engine.eval(model, test_dataloader)
valid_captcha_preds = []
for vp in valid_preds:
current_preds = decode_predictions(vp, lbl_encoder)
valid_captcha_preds.extend(current_preds)
combined = list(zip(test_target_orig, valid_captcha_preds))
print(combined[:20])
test_dup_rem = [remove_duplicates(c) for c in test_target_orig]
accuracy = metrics.accuracy_score(test_dup_rem, valid_captcha_preds)
print(
f"Epoch={epoch}, Train Loss={train_loss}, Test Loss={valid_loss} Accuracy={accuracy}"
)
scheduler.step(valid_loss)
joblib.dump(
lbl_encoder,
"../input/pickles/tan_pan_oltas_gst_epfo_rc_lbl_encoder.pkl")
torch.save(model.state_dict(),
"../input/pickles/tan_pan_oltas_gst_epfo_rc_model.pth")