(Depends on PR #9) Add seq2seq, modify RNN to enable seq2seq

configs/default/components/models/recurrent_neural_network.yml

@@ -9,7 +9,15 @@ hidden_size: 100
 
 # Flag informing the model to learn the intial state (h0/c0) (LOADED)
 # When false, (c0/c0) will be initialized as zeros.
-initial_state_trainable: True
+
+# Initial state type:
+#   * Zero (null vector)
+#   * Trainable (xavier initialization, trainable)
+#   * Input (the initial hidden state comes from an input stream)
+initial_state: Trainable
+
+# Wether to include the last hidden state in the outputs
+output_last_state: False
 
 # Type of recurrent cell (LOADED)
 # Options: LSTM | GRU | RNN_TANH | RNN_RELU
@@ -25,9 +33,19 @@ dropout_rate: 0
 # Prediction mode (LOADED)
 # Options: 
 #   * Dense (passes every activation through output layer) |
-#   * Last (passes only the last activation though output layer)
+#   * Last (passes only the last activation though output layer) |
+#   * None (all outputs are discarded)
 prediction_mode: Dense
 
+# Input mode
+# Options:
+#   * Dense (every iteration expects an input)
+#   * Autoregression_First (Autoregression, expects an input for the first iteration)
+#   * Autoregression_None (Autoregression, first input will be a null vector)
+input_mode: Dense
+
+autoregression_length: 42
+
 # If true, output of the last layer will be additionally processed with Log Softmax (LOADED)
 use_logsoftmax: True
 
@@ -39,9 +57,17 @@ streams:
   # Stream containing batch of images (INPUT)
   inputs: inputs
 
+  # Stream containing the inital state of the RNN (INPUT)
+  # The stream will be actually created only if `inital_state: Input`
+  input_state: input_state
+
   # Stream containing predictions (OUTPUT)
   predictions: predictions
 
+  # Stream containing the final output state of the RNN (output)
+  # The stream will be actually created only if `output_last_state: True`
+  output_state: output_state
+
 globals:
   ####################################################################
   # 3. Keymappings of variables that will be RETRIEVED from GLOBALS.

configs/default/components/models/seq2seq_rnn.yml

@@ -0,0 +1,81 @@
+# This file defines the default values for the RNN model.
+
+####################################################################
+# 1. CONFIGURATION PARAMETERS that will be LOADED by the component.
+####################################################################
+
+# Size of the hidden state (LOADED)
+hidden_size: 100
+
+# Flag informing the model to learn the intial state (h0/c0) (LOADED)
+# When false, (c0/c0) will be initialized as zeros.
+
+# Initial state type:
+#   * Zero (null vector)
+#   * Trainable (xavier initialization, trainable)
+#   * Input (the initial hidden state comes from an input stream)
+initial_state: Trainable
+
+# Wether to include the last hidden state in the outputs
+output_last_state: False
+
+# Type of recurrent cell (LOADED)
+# Options: LSTM | GRU | RNN_TANH | RNN_RELU
+cell_type: LSTM
+
+# Number of "stacked" layers (LOADED)
+num_layers: 1
+
+# Dropout rate (LOADED)
+# Default: 0 (means that it is turned off)
+dropout_rate: 0
+
+# Prediction mode (LOADED)
+# Options: 
+#   * Dense (passes every activation through output layer) |
+#   * Last (passes only the last activation though output layer) |
+#   * None (all outputs are discarded)
+prediction_mode: Dense
+
+# Input mode
+# Options:
+#   * Dense (every iteration expects an input)
+#   * Autoregression_First (Autoregression, expects an input for the first iteration)
+#   * Autoregression_None (Autoregression, first input will be a null vector)
+input_mode: Dense
+
+autoregression_length: 50
+
+# If true, output of the last layer will be additionally processed with Log Softmax (LOADED)
+use_logsoftmax: True
+
+streams: 
+  ####################################################################
+  # 2. Keymappings associated with INPUT and OUTPUT streams.
+  ####################################################################
+
+  # Stream containing batch of images (INPUT)
+  inputs: inputs
+
+  # Stream containing predictions (OUTPUT)
+  predictions: predictions
+
+globals:
+  ####################################################################
+  # 3. Keymappings of variables that will be RETRIEVED from GLOBALS.
+  ####################################################################
+
+  # Size of the input (RETRIEVED)
+  input_size: input_size
+
+  # Size of the prediction (RETRIEVED)
+  prediction_size: prediction_size
+
+  ####################################################################
+  # 4. Keymappings associated with GLOBAL variables that will be SET.
+  ####################################################################
+
+  ####################################################################
+  # 5. Keymappings associated with statistics that will be ADDED.
+  ####################################################################
+

configs/vqa_med_2019/c2_classification/default_c2_classification.yml

@@ -17,7 +17,6 @@ training:
     episode_limit: 10000
     epoch_limit: -1
 
-
 # Validation parameters:
 validation:
   problem:

configs/wikitext/wikitext_language_modeling_seq2seq.yml

@@ -0,0 +1,196 @@
+# This pipeline applies seq2seq on wikitext-2 to make word-level prediction.
+# It's been made for test purposes only, as it is doing:
+# [word 0 , ... , word 49] -> [word 1 , ... , word 50] (basically copying most of the input)
+#
+# The seq2seq here is implemented throught the use of 2 `RecurrentNeuralNetwork`
+
+# Training parameters:
+training:
+  problem:
+    type: &p_type WikiTextLanguageModeling
+    data_folder: &data_folder ~/data/language_modeling/wikitext-2
+    dataset: &dataset wikitext-2
+    subset: train
+    sentence_length: 50
+    batch_size:  64
+
+  # optimizer parameters:
+  optimizer:
+    name: Adam
+    lr: 1.0e-3
+
+  # settings parameters
+  terminal_conditions:
+    loss_stop: 1.0e-2
+    episode_limit: 1000000
+    epoch_limit: 100
+
+# Validation parameters:
+validation:
+  partial_validation_interval: 100
+  problem:
+    type: *p_type
+    data_folder: *data_folder
+    dataset: *dataset
+    subset: valid
+    sentence_length: 50
+    batch_size:  64
+
+# Testing parameters:
+testing:
+  problem:
+    type: *p_type 
+    data_folder: *data_folder
+    dataset: *dataset
+    subset: test
+    sentence_length: 50
+    batch_size: 64
+
+pipeline:
+  name: wikitext_language_modeling_seq2seq
+
+  # Source encoding - model 1.
+  source_sentence_embedding:
+    type: SentenceEmbeddings
+    priority: 1.1
+    embeddings_size: 50
+    pretrained_embeddings: glove.6B.50d.txt
+    data_folder: *data_folder
+    source_vocabulary_files: wiki.train.tokens,wiki.valid.tokens,wiki.test.tokens
+    vocabulary_mappings_file: wiki.all.tokenized_words
+    additional_tokens: <eos>
+    export_word_mappings_to_globals: True
+    streams:
+      inputs: sources
+      outputs: embedded_sources
+
+  # Target encoding.
+  target_indexer:
+    type: SentenceIndexer
+    priority: 2.1
+    data_folder: *data_folder
+    import_word_mappings_from_globals: True
+    streams:
+      inputs: targets
+      outputs: indexed_targets
+
+  # Publish the hidden size of the seq2seq
+  global_publisher:
+    type: GlobalVariablePublisher
+    priority: 1
+    # Add input_size to globals, so classifier will use it.
+    keys: s2s_hidden_size
+    values: 300
+
+  # FF, to resize the embeddings to whatever the hidden size of te seq2seq is.
+  ff_resize_s2s_input:
+    type: FeedForwardNetwork 
+    priority: 2.5
+    s2s_hidden_size: 300
+    use_logsoftmax: False
+    dimensions: 3
+    streams:
+      inputs: embedded_sources
+      predictions: embedded_sources_resized
+    globals:
+      input_size: embeddings_size
+      prediction_size: s2s_hidden_size
+
+  # LSTM Encoder
+  lstm_encoder:
+    type: RecurrentNeuralNetwork
+    priority: 3
+    initial_state: Trainable
+    hidden_size: 300
+    num_layers: 3
+    use_logsoftmax: False
+    output_last_state: True
+    prediction_mode: Last
+    streams:
+      inputs: embedded_sources_resized
+      predictions: s2s_encoder_output
+      output_state: s2s_state_output
+    globals:
+      input_size: s2s_hidden_size
+      prediction_size: s2s_hidden_size 
+
+  # LSTM Decoder
+  lstm_decoder:
+    type: RecurrentNeuralNetwork
+    priority: 4
+    initial_state: Input
+    hidden_size: 300
+    num_layers: 3
+    use_logsoftmax: False
+    input_mode: Autoregression_First
+    autoregression_length: 50
+    prediction_mode: Dense
+    streams:
+      inputs: s2s_encoder_output
+      predictions: s2s_decoder_output
+      input_state: s2s_state_output
+    globals:
+      input_size: s2s_hidden_size
+      prediction_size: s2s_hidden_size 
+
+  # FF, to resize the from the hidden size of the seq2seq to the size of the target vector
+  ff_resize_s2s_output:
+    type: FeedForwardNetwork 
+    use_logsoftmax: True
+    dimensions: 3
+    priority: 5
+    streams:
+      inputs: s2s_decoder_output
+    globals:
+      input_size: s2s_hidden_size
+      prediction_size: vocabulary_size
+
+  # Loss
+  nllloss:
+    type: NLLLoss
+    priority: 6
+    num_targets_dims: 2
+    streams:
+      targets: indexed_targets
+      loss: loss
+
+  # Prediction decoding.
+  prediction_decoder:
+    type: SentenceIndexer
+    priority: 10
+    # Reverse mode.
+    reverse: True
+    # Use distributions as inputs.
+    use_input_distributions: True
+    data_folder: *data_folder
+    import_word_mappings_from_globals: True
+    streams:
+      inputs: predictions
+      outputs: prediction_sentences
+
+
+  # Statistics.
+  batch_size:
+    type: BatchSizeStatistics
+    priority: 100.0
+
+  #accuracy:
+  #  type: AccuracyStatistics
+  #  priority: 100.1
+  #  streams:
+  #    targets: indexed_targets
+
+  bleu:
+    type: BLEUStatistics
+    priority: 100.2
+    streams:
+      targets: indexed_targets
+
+
+  # Viewers.
+  viewer:
+    type: StreamViewer
+    priority: 100.3
+    input_streams: sources,targets,indexed_targets,prediction_sentences
+
+#: pipeline