def test_keep_n_captions():
params = load_parameters()
params['REBUILD_DATASET'] = True
params['DATASET_STORE_PATH'] = './'
ds = build_dataset(params)
len_splits = {'train': 9900, 'val': 100, 'test': 2996}
for splits in [[], None, ['val'], ['val', 'test']]:
keep_n_captions(ds, 1, n=1, set_names=splits)
if splits is not None:
for split in splits:
len_split = len_splits[split]
assert eval('ds.len_' + split) == len_split
assert eval('all(ds.loaded_' + split + ')')
assert len(eval('ds.X_' + split + str([params['INPUTS_IDS_DATASET'][0]]))) == len_split
assert len(eval('ds.Y_' + split + str([params['OUTPUTS_IDS_DATASET'][0]]))) == len_split
if __name__ == '__main__':
pytest.main([__file__])
from keras_wrapper.extra import evaluation
Control_path = 'Control_M7.pred'
list2file(Control_path, Control_predictions)
dataset.setOutput('data/Ross_test.reply',
'test',
type='text',
id='target_text',
pad_on_batch=True,
tokenization='tokenize_basic',
sample_weights=True,
max_text_len=30,
max_words=0)
keep_n_captions(dataset, repeat=1, n=1, set_names=['test'])
metric = 'coco'
#Apply sampling
extra_vars = dict()
extra_vars['tokenize_f'] = eval('dataset.' + 'tokenize_basic')
extra_vars['language'] = params['TRG_LAN']
extra_vars['test'] = dict()
extra_vars['test']['references'] = dataset.extra_variables['test'][
'target_text']
Control_metrics = evaluation.select[metric](pred_list=Control_predictions,
verbose=1,
extra_vars=extra_vars,
split='test')
print("Control:")
pad_on_batch=True,
required=False)
ds.setOutput('data/Cornell_train_reply.en',
'train',
type='text',
id='target_text',
tokenization='tokenize_basic',
build_vocabulary=True,
pad_on_batch=True,
sample_weights=True,
max_text_len=30,
max_words=30000,
min_occ=0)
ds.setOutput('data/Cornell_valid_reply.en',
'val',
type='text',
id='target_text',
pad_on_batch=True,
tokenization='tokenize_basic',
sample_weights=True,
max_text_len=30,
max_words=0)
#ds.merge_vocabularies(['source_text', 'target_text'])
keep_n_captions(ds, repeat=1, n=1, set_names=['val'])
saveDataset(ds, 'query_to_reply')
def start_training(use_gpu):
ds = Dataset('tutorial_dataset', 'tutorial', silence=False)
ds.setOutput(DATA_PATH + "train_y.txt",
'train',
type='text',
id='target_text',
tokenization='tokenize_basic',
build_vocabulary=True,
pad_on_batch=True,
sample_weights=True,
max_text_len=30,
max_words=30000,
min_occ=0)
ds.setOutput(DATA_PATH + "val_y.txt",
'val',
type='text',
id='target_text',
pad_on_batch=True,
tokenization='tokenize_basic',
sample_weights=True,
max_text_len=30,
max_words=0)
ds.setInput(DATA_PATH + "train_x.txt",
'train',
type='text',
id='source_text',
pad_on_batch=True,
tokenization='tokenize_basic',
build_vocabulary=True,
fill='end',
max_text_len=30,
max_words=30000,
min_occ=0)
ds.setInput(DATA_PATH + "val_x.txt",
'val',
type='text',
id='source_text',
pad_on_batch=True,
tokenization='tokenize_basic',
fill='end',
max_text_len=30,
min_occ=0)
ds.setInput(DATA_PATH + "train_y.txt",
'train',
type='text',
id='state_below',
required=False,
tokenization='tokenize_basic',
pad_on_batch=True,
build_vocabulary='target_text',
offset=1,
fill='end',
max_text_len=30,
max_words=30000)
ds.setInput(None, 'val', type='ghost', id='state_below', required=False)
for split, input_text_filename in zip(
['train', 'val'],
[DATA_PATH + "train_x.txt", DATA_PATH + "val_x.txt"]):
ds.setRawInput(input_text_filename,
split,
type='file-name',
id='raw_source_text',
overwrite_split=True)
"""We also need to match the references with the inputs. Since we only have one reference per input sample, we set `repeat=1`."""
keep_n_captions(ds, repeat=1, n=1, set_names=['val'])
"""Finally, we can save our dataset instance for using in other experiments:"""
saveDataset(ds, MODEL_PATH + "/dataset")
"""## 2. Creating and training a Neural Translation Model
Now, we'll create and train a Neural Machine Translation (NMT) model. Since there is a significant number of hyperparameters, we'll use the default ones, specified in the `config.py` file. Note that almost every hardcoded parameter is automatically set from config if we run `main.py `.
We'll create an `'AttentionRNNEncoderDecoder'` (a LSTM encoder-decoder with attention mechanism). Refer to the [`model_zoo.py`](https://github.com/lvapeab/nmt-keras/blob/master/nmt_keras/model_zoo.py) file for other models (e.g. Transformer).
So first, let's import the model and the hyperparameters. We'll also load the dataset we stored in the previous section (not necessary as it is in memory, but as a demonstration):
"""
params = load_parameters()
dataset = loadDataset(MODEL_PATH + "/dataset/Dataset_tutorial_dataset.pkl")
"""Since the number of words in the dataset may be unknown beforehand, we must update the params information according to the dataset instance:"""
params['MODEL_TYPE'] = 'Transformer'
params['USE_CUDNN'] = use_gpu
params['EARLY_STOP'] = True
params['PATIENCE'] = 10
params['SAVE_EACH_EVALUATION'] = True
params['STORE_PATH'] = MODEL_PATH
params['N_LAYERS_ENCODER'] = 2
params['N_LAYERS_DECODER'] = 2
params['N_HEADS'] = 100
params['POS_UNK'] = False # current Transformer model requires this
params[
'ATTEND_ON_OUTPUT'] = True # current Transformer model requires this
params['MODEL_SIZE'] = 100
params['SOURCE_TEXT_EMBEDDING_SIZE'] = 100
params['TARGET_TEXT_EMBEDDING_SIZE'] = 100
params['SKIP_VECTORS_HIDDEN_SIZE'] = 100
params['ENCODER_HIDDEN_SIZE'] = 100
params['DECODER_HIDDEN_SIZE'] = 100
params['APPLY_DETOKENIZATION'] = True
params['LENGTH_PENALTY'] = True
params['LENGTH_NORM_FACTOR'] = 0.8
params['MAX_INPUT_TEXT_LEN'] = 128
params['MAX_OUTPUT_TEXT_LEN'] = 128
params['STOP_METRIC'] = 'perplexity'
params['BEAM_SIZE'] = 20
params['N_GPUS'] = 2
params['START_EVAL_ON_EPOCH'] = 1
params['BATCH_SIZE'] = 128
params['EVAL_EACH'] = 1
params['MAX_EPOCH'] = 100
params['PLOT_EVALULATION'] = True
params['APPLY_DETOKENIZATION'] = True
params['MODE'] = 'training'
params['BEAM_SEARCH'] = True
params['TENSORBOARD'] = True
train_model(params,
load_dataset=MODEL_PATH +
"/dataset/Dataset_tutorial_dataset.pkl")
def start_training(use_gpu):
ds = Dataset('tutorial_dataset', 'tutorial', silence=False)
ds.setOutput(PATH + "train_correct.txt",
'train',
type='text',
id='target_text',
tokenization='tokenize_basic',
build_vocabulary=True,
pad_on_batch=True,
sample_weights=True,
max_text_len=100,
max_words=55000,
min_occ=1)
ds.setOutput(PATH + "validation_correct.txt",
'val',
type='text',
id='target_text',
pad_on_batch=True,
tokenization='tokenize_basic',
sample_weights=True,
max_text_len=100,
max_words=0)
ds.setInput(PATH + "train_error.txt",
'train',
type='text',
id='source_text',
pad_on_batch=True,
tokenization='tokenize_basic',
build_vocabulary=True,
fill='end',
max_text_len=100,
max_words=55000,
min_occ=1)
ds.setInput(PATH + "validation_error.txt",
'val',
type='text',
id='source_text',
pad_on_batch=True,
tokenization='tokenize_basic',
fill='end',
max_text_len=100,
min_occ=1)
"""...and for the 'state_below' data. Note that: 1) The offset flat is set to 1, which means that the text will be shifted to the right 1 position. 2) During sampling time, we won't have this input. Hence, we 'hack' the dataset model by inserting an artificial input, of type 'ghost' for the validation split."""
ds.setInput(PATH + "train_correct.txt",
'train',
type='text',
id='state_below',
required=False,
tokenization='tokenize_basic',
pad_on_batch=True,
build_vocabulary='target_text',
offset=1,
fill='end',
max_text_len=100,
max_words=55000)
ds.setInput(None, 'val', type='ghost', id='state_below', required=False)
"""We can also keep the literal source words (for replacing unknown words)."""
for split, input_text_filename in zip(
['train', 'val'],
[PATH + "train_error.txt", PATH + "validation_error.txt"]):
ds.setRawInput(input_text_filename,
split,
type='file-name',
id='raw_source_text',
overwrite_split=True)
"""We also need to match the references with the inputs. Since we only have one reference per input sample, we set `repeat=1`."""
keep_n_captions(ds, repeat=1, n=1, set_names=['val'])
"""Finally, we can save our dataset instance for using in other experiments:"""
saveDataset(ds, PATH + "dataset")
"""## 2. Creating and training a Neural Translation Model
Now, we'll create and train a Neural Machine Translation (NMT) model. Since there is a significant number of hyperparameters, we'll use the default ones, specified in the `config.py` file. Note that almost every hardcoded parameter is automatically set from config if we run `main.py `.
We'll create an `'AttentionRNNEncoderDecoder'` (a LSTM encoder-decoder with attention mechanism). Refer to the [`model_zoo.py`](https://github.com/lvapeab/nmt-keras/blob/master/nmt_keras/model_zoo.py) file for other models (e.g. Transformer).
So first, let's import the model and the hyperparameters. We'll also load the dataset we stored in the previous section (not necessary as it is in memory, but as a demonstration):
"""
params = load_parameters()
dataset = loadDataset(PATH + "dataset/Dataset_tutorial_dataset.pkl")
"""Since the number of words in the dataset may be unknown beforehand, we must update the params information according to the dataset instance:"""
params['INPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len['source_text']
params['OUTPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len['target_text']
params['USE_CUDNN'] = use_gpu
params['N_GPUS'] = 2
params['MAX_EPOCH'] = 1000
params['EARLY_STOP'] = True
params['PATIENCE'] = 10
params['SAVE_EACH_EVALUATION'] = True
params['STORE_PATH'] = PATH + "model/"
params['BATCH_SIZE'] = 128
params['ATTENTION_MODE'] = "add"
params['N_LAYERS_ENCODER'] = 2
params['N_LAYERS_DECODER'] = 2
params['SOURCE_TEXT_EMBEDDING_SIZE'] = 512
params['TARGET_TEXT_EMBEDDING_SIZE'] = 512
params['SKIP_VECTORS_HIDDEN_SIZE'] = 512
params['ATTENTION_SIZE'] = 512
params['ENCODER_HIDDEN_SIZE'] = 512
params['DECODER_HIDDEN_SIZE'] = 512
params['ENCODER_RNN_TYPE'] = "LSTM"
params['DECODER_RNN_TYPE'] = "ConditionalLSTM"
params['METRICS'] = ['coco']
params['KERAS_METRICS'] = ['perplexity']
params['APPLY_DETOKENIZATION'] = True
params['LENGTH_PENALTY'] = True
params['LENGTH_NORM_FACTOR'] = 1.0
params['BEAM_SIZE'] = 1
params['BEAM_SEARCH'] = True
params['PLOT_EVALUATION'] = True
params['MAX_PLOT_Y'] = 1.
params['MODE'] = 'training'
params['TENSORBOARD'] = True
result = pyfiglet.figlet_format("START TRAINING FROM SCRATCH".format(mode),
font="digital")
print(result)
train_model(params,
load_dataset=os.getcwd() +
"/dataset/Dataset_tutorial_dataset.pkl")