使用Tensor2Tensor和10行代码训练尖端语言翻译神经网络

from tensor2tensorimport problems

# Print all T2T problems to console

problems.available()

t2t-datagen \

  --data_dir=$DATA_DIR \

  --tmp_dir=$TMP_DIR \

  --problem=$PROBLEM

from tensor2tensorimport problems

PROBLEM= '{Tensor2Tensor_problem_name}

TMP_DIR= '{/Tmp_Dir_Path}' # Where data files from internet stored

DATA_DIR= '{/Data_Dir_Path}' # Where pre-prcessed data is stored

# Init problem T2T object the generated training data

t2t_problem= problems.problem(PROBLEM)

t2t_problem.generate_data(DATA_DIR, TMP_DIR)

from tensor2tensor.utilsimport registry

from tensor2tensorimport models

# Print all models in T2T to console

registry.list_models()

t2t-trainer \

  --data_dir=$DATA_DIR \

  --problem=$PROBLEM \

  --model=$MODEL \

  --hparams_set=$HPARAMS \

  --output_dir=$TRAIN_DIR

PROBLEM =  ' translate_enfr_wmt32k_rev '

MODEL =  ' TRANSFORMER '

HPARAMS =  ' transformer_base '

TRAIN_DIR =  '〜/ translator / model_files '

DATA_DIR =  '〜/ translator / fr_en_data '

from tensor2tensor.utils.trainer_libimport create_hparams

# Init Hparams object from T2T Problem

hparams= create_hparams(HPARAMS)

# Make Chngaes to Hparams

hparams.batch_size= 1024

hparams.learning_rate_warmup_steps= 45000

hparams.learning_rate= .4

# Can see all Hparams with code below

print(json.loads(hparams.to_json())

from tensor2tensor.utils.trainer_libimport create_run_config, create_experiment

# Initi Run COnfig for Model Training

RUN_CONFIG= create_run_config(

      model_dir=TRAIN_DIR# Location of where model file is store

      # More Params here in this fucntion for controling how noften to tave checkpoints and more.

)

# Create Tensorflow Experiment Object

tensorflow_exp_fn= create_experiment(

        run_config=RUN_CONFIG,

        hparams=hparams,

        model_name=MODEL,

        problem_name=PROBLEM,

        data_dir=DATA_DIR,

        train_steps=400000,# Total number of train steps for all Epochs

        eval_steps=100 # Number of steps to perform for each evaluation

    )

# Kick off Training

tensorflow_exp_fn.train_and_evaluate()

t2t-decoder \

  --data_dir=$DATA_DIR \

  --problem=$PROBLEM \

  --model=$MODEL \

  --hparams_set=$HPARAMS \

  --output_dir=$TRAIN_DIR \

  --decode_hparams="beam_size=$BEAM_SIZE,alpha=$ALPHA" \

  --decode_from_file=$DECODE_FILE \

  --decode_to_file=translation.en

from tensor2tensor.utils.trainer_libimport create_hparams, registry

from tensor2tensorimport problems

INPUT_TEXT_TO_TRANSLATE= 'Translate this sentence into French'

# Set Tensor2Tensor Arguments

MODEL_DIR_PATH= ~/En_to_Fr_translator'

MODEL= 'transformer'

HPARAMS= 'transformer_big_single_gpu'

T2T_PROBLEM= 'translate_enfr_wmt32k'

hparams= create_hparams(HPARAMS, data_dir=model_dir, problem_name=T2T_PROBLEM)

# Make any changes to default Hparams for model architechture used during training

hparams.batch_size= 1024

hparams.hidden_size= 7*80

hparams.filter_size= 7*80*4

hparams.num_heads= 8

# Load model into Memory

T2T_MODEL= registry.model(MODEL)(hparams, tf.estimator.ModeKeys.PREDICT)

# Init T2T Token Encoder/ Decoders

DATA_ENCODERS= problems.problem(T2T_PROBLEM).feature_encoders(model_dir)

### START USING MODELS

encoded_inputs= encode(INPUT_TEXT_TO_TRANSLATE, DATA_ENCODERS)

model_output= T2T_MODEL.infer(encoded_inputs, beam_size=2)["outputs"]

translated_text_in_french=  decode(model_output, DATA_ENCODERS)

print(translated_text_in_french)

def encode(input_txt, encoders):

    """List of Strings to features dict, ready for inference"""

    encoded_inputs= [encoders["inputs"].encode(x)+ [1]for xin input_txt]

    # pad each input so is they are the same length

    biggest_seq= len(max(encoded_inputs, key=len))

    for i, text_inputin enumerate(encoded_inputs):

        encoded_inputs[i]= text_input+ [0 for xin range(biggest_seq- len(text_input))]

    # Format Input Data For Model

    batched_inputs= tf.reshape(encoded_inputs, [len(encoded_inputs),-1,1])

    return {"inputs": batched_inputs}

def decode(integers, encoders):

    """Decode list of ints to list of strings"""

    # Turn to list to remove EOF mark

    to_decode= list(np.squeeze(integers))

    if isinstance(to_decode[0], np.ndarray):

        to_decode= map(lambda x:list(np.squeeze(x)), to_decode)

    else:

        to_decode= [to_decode]

    # remove <EOF> Tag before decoding

    to_decode= map(lambda x: x[:x.index(1)],filter(lambda x:1 in x, to_decode))

    # Decode and return Translated text

    return [encoders["inputs"].decode(np.squeeze(x))for xin to_decode]