Commit 46e7157a authored by Rudolf Chrispens's avatar Rudolf Chrispens
Browse files

reworked code and bugfixing encountered not solvable bug with callable

parent c191b483

caption-lib/lstm/configuration.py

+1 −0
Original line number Diff line number Diff line
@@ -24,6 +24,7 @@ class default_reader(object): # default reader config 
    verbose = False

    read_raw_files = False  # this will only read files use only for testing

    create_vocab = True

    # current_reader = "basic_text_reader.py"



class default_validator(object):

    validator_on = True

caption-lib/lstm/lstm_train.py

+74 −60
Original line number Diff line number Diff line
@@ -5,6 +5,7 @@ from lstm.configuration import current as config 
import lstm.Readers as reader



from tensorflow.python.ops.rnn_cell import LSTMCell, LSTMStateTuple

from tensorflow.python.framework import constant_op



# def train(config)

# instanziiert das model mit hife der config
@@ -14,6 +15,61 @@ from tensorflow.python.ops.rnn_cell import LSTMCell, LSTMStateTuple 
# speichert das trainierte model mit saver

#   unter einem bestimmtem path



# manually spcifying loop function over time -  to get initial cell state and input to RNN

# normally we would just use dynamic_rnn, but lets get detailed here with raw_rnn



# we define and return these values, no operations occure here

def loop_fn_initial(_decoder_lengths, _eos_step_embedded, _encoder_final_state):

    initial_elements_finished = (0 >= _decoder_lengths)

    # end of sentence

    initial_input = _eos_step_embedded

    # last time steps cell state

    initial_cell_state = _encoder_final_state

    # none

    initial_cell_output = None

    # none

    initial_loop_state = None

    return(initial_elements_finished,

        initial_input,

        initial_cell_state,

        initial_cell_output,

        initial_loop_state)



# attention mechanism --choose which previously generated token to pass as input in the next time step

def loop_fn_transition(time, previous_output, previous_state, previous_loop_state, _embeddings, _W, _b, _decoder_lengths, _pad_step_embedded):

    def get_next_input():

        output_logits = tf.add(tf.matmul(previous_output, _W), _b)

        # this next line is attention

        prediction = tf.argmax(output_logits, axis=1)

        next_inpuut = tf.nn.embedding_lookup(_embeddings, prediction)

        return next_inpuut



    elements_finished = (time >= _decoder_lengths)



    finished = tf.reduce_all(elements_finished)



    input = tf.cond(finished, lambda: _pad_step_embedded, get_next_input)



    state = previous_state

    output = previous_output

    loop_state = None



    return (elements_finished,

        input,

        state,

        output,

        loop_state)



def loop_fn(time, previous_output, previous_state, previous_loop_state, _W, _b, _decoder_lengths, _pad_step_embedded, _eos_step_embedded, _encoder_final_state):

    if previous_state is None:

        assert previous_output is None and previous_state is None

        return loop_fn_initial(_decoder_lengths, _eos_step_embedded, _encoder_final_state)

    else:

        return loop_fn_transition(time, previous_output, previous_state, previous_loop_state, _W, _b, _decoder_lengths, _pad_step_embedded)









def train():

    # TODO split into sub functions

    # this is a tutorial of Siraj Raval!
@@ -52,8 +108,8 @@ def train(): 
    # bidirectional step1

    # encoder_outputs = tf.concat((encoder_fw_outputs, encoder_bw_outputs, 2))



    encoder_final_state_c = tf.concat((encoder_fw_final_state.c, encoder_bw_final_state.c, 1))

    encoder_final_state_h = tf.concat((encoder_fw_final_state.h, encoder_bw_final_state.h, 1))

    encoder_final_state_c = tf.concat((encoder_fw_final_state.c, encoder_bw_final_state.c), 1)

    encoder_final_state_h = tf.concat((encoder_fw_final_state.h, encoder_bw_final_state.h), 1)



    # TF Tuple by LSTM Cells for state size, zerostate and output state

    encoder_final_state = LSTMStateTuple(
@@ -69,21 +125,33 @@ def train(): 
    # output projects

    # weights and biases

    # SOFT ATTENTION

    W = tf.Variable(tf.random_uniform([config.trainer.decoder_hidden_units], config.trainer.vocab_size), -1, 1, dtype = tf.float32)

    b = tf.Variable(tf.zeroes([config.trainer.vocab_size]), dtype = tf.float32)

    W = tf.Variable(tf.random_uniform([config.trainer.decoder_hidden_units, config.trainer.vocab_size], -1, 1), dtype=tf.float32)

    b = tf.Variable(tf.zeros([config.trainer.vocab_size]), dtype=tf.float32)





    # create padded inputs for the decoder from the word embeddings

    # we are telling the program to test a condition, and trigger an error if the condition is false

    assert config.trainer.end_of_sentence == 1 and config.trainer.padding == 0

    eos_time_slice = tf.ones([config.traer.batch_size], dtype = tf.int32, name = 'EOS')

    eos_time_slice = tf.ones([config.trainer.batch_size], dtype = tf.int32, name = 'EOS')

    pad_time_slice = tf.ones([config.trainer.batch_size], dtype = tf.int32, name = 'PAD')

    # retrieves rows of the params tensor. The behaviour is similar to using indexing with arrays in numpy

    eos_step_embedded = tf.nn.embedding_lookup(embeddings, eos_time_slice)

    pad_step_embedded = tf.nn.embedding_lookup(embeddings, pad_time_slice)



    time = constant_op.constant(0, dtype=tf.int32)

    callable_loop_fn = loop_fn(

        time=time,

        previous_output=None,

        previous_state=None,

        previous_loop_state=None,

        _W=W, _b=b,

        _decoder_lengths=decoder_lengths,

        _pad_step_embedded=pad_step_embedded,

        _eos_step_embedded=eos_step_embedded,

        _encoder_final_state=encoder_final_state)



    # using the functions for the attention decoder

    decoder_outputs_ta, decoder_final_state, _ = tf.nn.raw_rnn(decoder_cell, loop_fn)

    decoder_outputs_ta, decoder_final_state, decoder_loop_state = tf.nn.raw_rnn(decoder_cell, callable_loop_fn)

    decoder_outputs = decoder_outputs_ta.stack()



    # unpacks the given dimension of a rank-R tensr into a R-1 tensor
@@ -138,60 +206,6 @@ def train(): 
        print('\n.\n.\n.\n...training interupted')





# manually spcifying loop function over time -  to get initial cell state and input to RNN

# normally we would just use dynamic_rnn, but lets get detailed here with raw_rnn



# we define and return these values, no operations occure here

def loop_fn_initial(_decoder_lengths, _eos_step_embedded, _encoder_final_state):

    initial_elements_finished = (0 >= _decoder_lengths)

    # end of sentence

    initial_input = _eos_step_embedded

    # last time steps cell state

    initial_cell_state = _encoder_final_state

    # none

    initial_cell_output = None

    # none

    initial_loop_state = None

    return(initial_elements_finished,

        initial_input,

        initial_cell_state,

        initial_cell_output,

        initial_loop_state)



# attention mechanism --choose which previously generated token to pass as input in the next time step

def loop_fn_transition(time, previous_output, previous_state, previous_loop_state, _embeddings, _w, _b, _decoder_lengths, _pad_step_embedded):

    def get_next_input():

        output_logits = tf.add(tf.matmul(previous_output, _w), _b)

        # this next line is attention

        prediction = tf.argmax(output_logits, axis=1)

        next_inpuut = tf.nn.embedding_lookup(_embeddings, prediction)

        return next_inpuut



    elements_finished = (time >= _decoder_lengths)



    finished = tf.reduce_all(elements_finished)



    input = tf.cond(finished, lambda: _pad_step_embedded, get_next_input)



    state = previous_state

    output = previous_output

    loop_state = None



    return (elements_finished,

        input,

        state,

        output,

        loop_state)



def loop_fn(time, previous_output, previous_state, previous_loop_state, _w, _b, _decoder_lengths, _pad_step_embedded):

    if previous_state is None:

        assert previous_output is None and previous_state is None

        return loop_fn_initial()

    else:

        return loop_fn_transition(time, previous_output, previous_state, previous_loop_state, _w, _b, _decoder_lengths, _pad_step_embedded)







"""

# Embedding

embedding_encoder = variable_scope.get_variable(