def train_new():
df = pd.DataFrame(request.json)
# We use a global here so we can modify the model and list of column names
global model_columns, model
model_columns, model = model_utils.train(df)
joblib.dump(model_columns, model_utils.MODEL_COLUMNS_FILE_NAME)
joblib.dump(model, model_utils.MODEL_FILE_NAME)
return 'Success'
def train():
df = pd.read_csv(TRAINING_FILE_PATH)
# We use a global here so we can modify the model and list of column names
global model_columns, model
model_columns, model = model_utils.train(df)
joblib.dump(model_columns, model_utils.MODEL_COLUMNS_FILE_NAME)
joblib.dump(model, model_utils.MODEL_FILE_NAME)
return 'Success'
def train_model_without_file():
print('training model...wait')
df = pd.read_csv(TRAINING_FILE_PATH)
global model_columns, model
model_columns, model = model_utils.train(df)
#joblib.dump(model_columns, model_utils.MODEL_COLUMNS_FILE_NAME)
#joblib.dump(model, model_utils.MODEL_FILE_NAME)
pickle.dump(model, open(model_utils.MODEL_FILE_NAME, 'wb'))
pickle.dump(model_columns, open(model_utils.MODEL_COLUMNS_FILE_NAME, 'wb'))
return 'success: from train_model_without_file'
def train_model():
print('training model...waiting for data')
try:
df= pd.DataFrame(request.json)
except Exception as e:
return jsonify({'error': str(e), 'trace': traceback.format_exc()})
global model_columns, model
model_columns, model = model_utils.train(df)
pickle.dump(model, open(model_utils.MODEL_FILE_NAME, 'wb'))
pickle.dump(model_columns, open(model_utils.MODEL_COLUMNS_FILE_NAME, 'wb'))
return 'success:from train_model'
def train():
df = pd.read_csv(TRAINING_FILE_PATH)
# We use a global here so we can modify the model and list of column names
global model_columns, model
model_columns, model = model_utils.train(df)
# save the model
# serialize model to JSON
model_json = model.to_json()
with open(model_utils.MODEL_FILE_NAME + ".json", "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
model.save_weights(model_utils.MODEL_FILE_NAME + ".h5")
print("Saved model to disk")
return jsonify({'success': 'Model training successful'})
]
# Define optimizer
optimizer = AdamW(grouped_model_parameters)
# Place model & loss function on GPU
model, criterion = model.to(DEVICE), criterion.to(DEVICE)
# Start actual training, check test loss after each epoch
best_test_loss = float('inf')
for epoch in range(NUM_EPOCHS):
print("EPOCH NO: %d" % (epoch + 1))
train_loss, train_acc = train(model=model,
iterator=train_loader,
criterion=criterion,
optimizer=optimizer,
device=DEVICE,
include_bert_masks=True)
test_loss, test_acc = test(model=model,
iterator=test_loader,
criterion=criterion,
device=DEVICE,
include_bert_masks=True)
if test_loss < best_test_loss:
best_test_loss = test_loss
torch.save(model.state_dict(), 'saved_models/finetuned-bert-model.pt')
print(f'\tTrain Loss: {train_loss:.3f} | Train Accuracy: {train_acc * 100:.2f}%')
print(f'\tTest Loss: {test_loss:.3f} | Test Accuracy: {test_acc * 100:.2f}%')