From f148dbd23a90180fc2b37276e184322b6f4ec36b Mon Sep 17 00:00:00 2001
From: josh <jmeyer@mozilla.com>
Date: Fri, 22 Mar 2019 22:31:09 +0100
Subject: [PATCH] getting the outputs shaped correctly

---
 .compute      | 41 ++++++++++++++++++++++-------------------
 DeepSpeech.py | 21 +++++++++++++++++----
 util/flags.py |  4 ++++
 3 files changed, 43 insertions(+), 23 deletions(-)

diff --git a/.compute b/.compute
index 629e3d4f..d5854db1 100755
--- a/.compute
+++ b/.compute
@@ -2,6 +2,11 @@
 
 set -xe
 
+# check HTTP_PROXY
+if ! (( $( env | grep -iq "^http_proxy=" ) )); then
+    source /etc/profile
+fi
+
 apt-get install -y python3-venv
 python3 -m venv /tmp/venv
 source /tmp/venv/bin/activate
@@ -11,28 +16,26 @@ pip install tensorflow-gpu==1.13.0-rc2
 
 # Install ds_ctcdecoder package from TaskCluster
 pip install $(python3 util/taskcluster.py --decoder)
+pip install pandas
 
 mkdir -p ../keep/summaries
 
-data="${SHARED_DIR}/data"
-fis="${data}/LDC/fisher"
-swb="${data}/LDC/LDC97S62/swb"
-lbs="${data}/OpenSLR/LibriSpeech/librivox"
+CV="${SHARED_DIR}/data/mozilla/CommonVoice/v2.0"
 
 python -u DeepSpeech.py \
        --notest\
-  --train_welsh\
-  --dev_welsh \
-  --train_kyrgyz\
-  --dev_kyrgyz \
-  --train_batch_size 24 \
-  --dev_batch_size 48 \
-  --test_batch_size 48 \
-  --n_hidden 2048 \
-  --learning_rate 0.0001 \
-  --dropout_rate 0.2 \
-  --epoch 13 \
-  --display_step 0 \
-  --validation_step 1 \
-  --checkpoint_dir "../keep" \
-  --summary_dir "../keep/summaries"
+       --train_welsh "${CV}/cy/clips/train.csv"\
+       --dev_welsh "${CV}/cy/clips/dev.csv"\
+       --train_kyrgyz "${CV}/ky/clips/train.csv"\
+       --dev_kyrgyz "${CV}/ky/clips/dev.csv"\
+       --train_batch_size 24 \
+       --dev_batch_size 48 \
+       --test_batch_size 48 \
+       --n_hidden 2048 \
+       --learning_rate 0.0001 \
+       --dropout_rate 0.2 \
+       --epoch 13 \
+       --display_step 0 \
+       --validation_step 1 \
+       --checkpoint_dir "../keep" \
+       --summary_dir "../keep/summaries"
diff --git a/DeepSpeech.py b/DeepSpeech.py
index 86c1ec09..a4e756e4 100755
--- a/DeepSpeech.py
+++ b/DeepSpeech.py
@@ -16,7 +16,7 @@ import progressbar
 import shutil
 import tensorflow as tf
 import traceback
-
+import pandas
 from ds_ctcdecoder import ctc_beam_search_decoder, Scorer
 from six.moves import zip, range
 from tensorflow.python.tools import freeze_graph
@@ -187,11 +187,24 @@ def calculate_mean_edit_distance_and_loss(model_feeder, tower, dropout, reuse):
     # Calculate the logits of the batch using BiRNN
     logits, _ = BiRNN(batch_x, batch_seq_len, dropout, reuse)
 
+    # SPECIAL FOR LANG-ID
+    
+    batch_shape = tf.shape(batch_x)
+    batch_size = batch_shape[0]
+    n_steps = batch_shape[1]
+    output = logits
+    # permute time and batch
+    output = tf.reshape(output, [n_steps, batch_size, -1])
+    output = tf.transpose(output, [1, 0, 2])
+    output = tf.reshape(output, [batch_size, n_steps, -1])
+    # mean over time steps
+    output = tf.reduce_sum(output, axis=1) / tf.cast(batch_seq_len, tf.float32)
+    total_loss = tf.nn.sigmoid_cross_entropy_with_logits(
+        labels=tf.cast(batch_y, tf.float32), logits=output
+    )
+
     # Compute the CTC loss using TensorFlow's `ctc_loss`
     #total_loss = tf.nn.ctc_loss(labels=batch_y, inputs=logits, sequence_length=batch_seq_len)
-    total_loss = tf.nn.sigmoid_cross_entropy_with_logits(
-        labels=tf.cast(batch_y, tf.float32), logits
-    )
 
     output_probs = tf.sigmoid(output)
     prediction = tf.cast(output_probs > 0.5, tf.int32)
diff --git a/util/flags.py b/util/flags.py
index 88c1ccb4..77a05250 100644
--- a/util/flags.py
+++ b/util/flags.py
@@ -10,6 +10,10 @@ def create_flags():
     # Importer
     # ========
 
+    tf.app.flags.DEFINE_string  ('train_welsh',      '',          'comma separated list of files specifying the dataset used for training. multiple files will get merged')
+    tf.app.flags.DEFINE_string  ('dev_welsh',        '',          'comma separated list of files specifying the dataset used for validation. multiple files will get merged')
+    tf.app.flags.DEFINE_string  ('train_kyrgyz',      '',          'comma separated list of files specifying the dataset used for training. multiple files will get merged')
+    tf.app.flags.DEFINE_string  ('dev_kyrgyz',        '',          'comma separated list of files specifying the dataset used for validation. multiple files will get merged')
     tf.app.flags.DEFINE_string  ('train_files',      '',          'comma separated list of files specifying the dataset used for training. multiple files will get merged')
     tf.app.flags.DEFINE_string  ('dev_files',        '',          'comma separated list of files specifying the dataset used for validation. multiple files will get merged')
     tf.app.flags.DEFINE_string  ('test_files',       '',          'comma separated list of files specifying the dataset used for testing. multiple files will get merged')