TF 1.0 !

lstm_ctc_to_chars.py

@@ -38,7 +38,7 @@
 # inputs = tf.transpose(inputs, [0, 2, 1]) #  inputs must be a `Tensor` of shape: `[batch_size, max_time, ...]`
 inputs = tf.transpose(inputs, [2, 0, 1])  # [max_time, batch_size, features] to split:
 # Split data because rnn cell needs a list of inputs for the RNN inner loop
-inputs = tf.split(0, max_input_length, inputs)  # n_steps * (batch_size, features)
+inputs = tf.split(axis=0, num_or_size_splits=max_input_length, value=inputs)  # n_steps * (batch_size, features)
 
 num_hidden = 100  # features
 cell = tf.nn.rnn_cell.LSTMCell(num_hidden, state_is_tuple=True)
@@ -89,7 +89,7 @@
 	bias = tf.Variable(uniform_bias, name="bias_dense_%d" % i)
 	y_ = outputY = tf.matmul(output, weights, name="dense_%d" % i) + bias
 
-	cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_, y[:, i, :]), name="cost")  # prediction, target
+	cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_, labels=y[:, i, :]), name="cost")  # prediction, target
 	costs.append(cost)
 	logits.append(y_)
 
@@ -108,7 +108,7 @@
 
 ####Optimizing
 logits = y_
-logits3d = tf.pack(logits)
+logits3d = tf.stack(logits)
 seqLengths = [20] * batch_size
 cost = tf.reduce_mean(ctc.ctc_loss(logits3d, targetY, seqLengths))
 # CTCLoss op expects the reserved blank label to be the largest value! REALLY?
@@ -132,22 +132,22 @@
 try:
 	saver = tf.train.Saver(tf.global_variables())
 except:
-	saver = tf.train.Saver(tf.all_variables())
+	saver = tf.train.Saver(tf.global_variables())
 snapshot = "lstm_mfcc"
 checkpoint = tf.train.latest_checkpoint(checkpoint_dir="checkpoints")
 if checkpoint:
 	print("LOADING checkpoint " + checkpoint + "")
 	try: saver.restore(session, checkpoint)
 	except: print("incompatible checkpoint")
 try: session.run([tf.global_variables_initializer()])
-except: session.run([tf.initialize_all_variables()])  # tf <12
+except: session.run([tf.global_variables_initializer()])  # tf <12
 
 # train
 step = 0  # show first
 try: summaries = tf.summary.merge_all()
-except: summaries = tf.merge_all_summaries()  # tf<12
+except: summaries = tf.summary.merge_all()  # tf<12
 try: summary_writer = tf.summary.FileWriter("logs", session.graph)  #
-except: summary_writer = tf.train.SummaryWriter("logs", session.graph)  # tf<12
+except: summary_writer = tf.summary.FileWriter("logs", session.graph)  # tf<12
 while step < steps:
 	batch_xs, batch_ys = next(batch)
 

lstm_mfcc_ctc_to_words.py

@@ -37,7 +37,7 @@
 # inputs = tf.transpose(inputs, [0, 2, 1]) #  inputs must be a `Tensor` of shape: `[batch_size, max_time, ...]`
 inputs = tf.transpose(inputs, [2, 0, 1]) # [max_time, batch_size, features] to split:
 # Split data because rnn cell needs a list of inputs for the RNN inner loop
-inputs = tf.split(0, max_input_length, inputs)  # n_steps * (batch_size, features)
+inputs = tf.split(axis=0, num_or_size_splits=max_input_length, value=inputs)  # n_steps * (batch_size, features)
 
 num_hidden = 100 #features
 cell = tf.nn.rnn_cell.LSTMCell(num_hidden, state_is_tuple=True)
@@ -75,7 +75,7 @@
 	bias = tf.Variable(tf.random_uniform([classes], minval=-1. / width, maxval=1. / width), name="bias_dense_%d"%i)
 	y_ = outputY = tf.matmul(output, weights, name="dense_%d" % i) + bias
 
-	cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_, y[:,i,:]), name="cost")  # prediction, target
+	cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_, labels=y[:,i,:]), name="cost")  # prediction, target
 	costs.append(cost)
 	logits.append(y_)
 
@@ -94,7 +94,7 @@
 
 ####Optimizing
 logits=y_
-logits3d = tf.pack(logits)
+logits3d = tf.stack(logits)
 seqLengths=[20]*batch_size
 cost = tf.reduce_mean(ctc.ctc_loss(logits3d, targetY, seqLengths))
 # CTCLoss op expects the reserved blank label to be the largest value! REALLY?
@@ -116,23 +116,23 @@
 steps = 9999999
 session=tf.Session()
 try:saver = tf.train.Saver(tf.global_variables())
-except:saver = tf.train.Saver(tf.all_variables())
+except:saver = tf.train.Saver(tf.global_variables())
 snapshot = "lstm_mfcc"
 checkpoint = tf.train.latest_checkpoint(checkpoint_dir="checkpoints")
 if checkpoint:
 	print("LOADING " + checkpoint + " !!!")
 	try:saver.restore(session, checkpoint)
 	except: print("incompatible checkpoint")
 try: session.run([tf.global_variables_initializer()])
-except: session.run([tf.initialize_all_variables()])# tf <12
+except: session.run([tf.global_variables_initializer()])# tf <12
 
 
 #train
 step = 0  # show first
 try:summaries = tf.summary.merge_all()
-except:summaries = tf.merge_all_summaries() # tf<12
+except:summaries = tf.summary.merge_all() # tf<12
 try:summary_writer = tf.summary.FileWriter("logs", session.graph)  #
-except:summary_writer = tf.train.SummaryWriter("logs", session.graph)  # tf<12
+except:summary_writer = tf.summary.FileWriter("logs", session.graph)  # tf<12
 while step < steps:
 	batch_xs, batch_ys = next(batch)
 

lstm_mfcc_to_chars.py

@@ -34,7 +34,7 @@
 # inputs = tf.transpose(inputs, [0, 2, 1]) #  inputs must be a `Tensor` of shape: `[batch_size, max_time, ...]`
 inputs = tf.transpose(inputs, [2, 0, 1]) # [max_time, batch_size, features] to split:
 # Split data because rnn cell needs a list of inputs for the RNN inner loop
-inputs = tf.split(0, max_length, inputs)  # n_steps * (batch_size, features)
+inputs = tf.split(axis=0, num_or_size_splits=max_length, value=inputs)  # n_steps * (batch_size, features)
 
 num_hidden = 100 #features
 cell = tf.nn.rnn_cell.LSTMCell(num_hidden, state_is_tuple=True)
@@ -65,32 +65,32 @@
 # optimize
 # if use_word: y=target=tf.placeholder(tf.float32, shape=(batch_size,(None,32))) # -> seq2seq!
 y=target=tf.placeholder(tf.float32, shape=(batch_size,classes))
-cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_,y),name="cost")  # prediction, target
-tf.scalar_summary('cost', cost)
+cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_,labels=y),name="cost")  # prediction, target
+tf.summary.scalar('cost', cost)
 optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)
 prediction = y_
 # Evaluate model
 correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(target, 1))
 accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
-tf.scalar_summary('accuracy', accuracy)
+tf.summary.scalar('accuracy', accuracy)
 
 steps = 9999999
 session=tf.Session()
-saver = tf.train.Saver(tf.all_variables())
+saver = tf.train.Saver(tf.global_variables())
 snapshot = "lstm_mfcc"
 checkpoint = tf.train.latest_checkpoint(checkpoint_dir="checkpoints")
 if checkpoint:
 	print("LOADING " + checkpoint + " !!!")
 	try:saver.restore(session, checkpoint)
 	except: print("incompatible checkpoint")
 try: session.run([tf.global_variables_initializer()])
-except: session.run([tf.initialize_all_variables()])
+except: session.run([tf.global_variables_initializer()])
 
 
 #train
 step = 0  # show first
-summaries = tf.merge_all_summaries()
-summary_writer = tf.train.SummaryWriter("logs", session.graph)  #
+summaries = tf.summary.merge_all()
+summary_writer = tf.summary.FileWriter("logs", session.graph)  #
 while step < steps:
 	batch_xs, batch_ys = next(batch)
 	# tf.train.shuffle_batch_join(example_list, batch_size, capacity=min_queue_size + batch_size * 16, min_queue_size)

lstm_to_chars.py

@@ -77,7 +77,7 @@
 	#  Reshape to 2-D tensor (nTimeSteps*Size, nfeatures)
 	inputXrs = tf.reshape(inputX, [-1, nFeatures])
 	#  Split to get a list of 'n_steps' tensors of shape (_size, n_hidden)
-	inputList = tf.split(0, maxTimeSteps, inputXrs)
+	inputList = tf.split(axis=0, num_or_size_splits=maxTimeSteps, value=inputXrs)
 	targetIxs = tf.placeholder(tf.int64)
 	targetVals = tf.placeholder(tf.int32)
 	targetShape = tf.placeholder(tf.int64)
@@ -104,13 +104,13 @@
 	print("building fbH1rs ")
 	fbH1rs = [tf.reshape(t, [Size, 2, nHidden]) for t in fbH1]
 	print("building outH1 ")
-	outH1 = [tf.reduce_sum(tf.mul(t, weightsOutH1), reduction_indices=1) + biasesOutH1 for t in fbH1rs]
+	outH1 = [tf.reduce_sum(tf.multiply(t, weightsOutH1), axis=1) + biasesOutH1 for t in fbH1rs]
 	print("building logits ")
 	logits = [tf.matmul(t, weightsClasses) + biasesClasses for t in outH1]
 	print("len(outH1) %d"% len(outH1))
 	####Optimizing
 	print("building loss")
-	logits3d = tf.pack(logits)
+	logits3d = tf.stack(logits)
 	loss = tf.reduce_mean(ctc.ctc_loss(logits3d, targetY, seqLengths))
 	out = tf.identity(loss, 'ctc_loss_mean')
 	optimizer = tf.train.MomentumOptimizer(learningRate, momentum).minimize(loss)
@@ -128,13 +128,13 @@
 ####Run session
 with tf.Session(graph=graph) as session:
 	try: merged = tf.summary.merge_all()
-	except: merged = tf.merge_all_summaries()
+	except: merged = tf.summary.merge_all()
 	try:writer = tf.summary.FileWriter("/tmp/basic_new", session.graph)
-	except: writer = tf.train.SummaryWriter("/tmp/basic_new", session.graph)
+	except: writer = tf.summary.FileWriter("/tmp/basic_new", session.graph)
 	try:saver = tf.train.Saver()  # defaults to saving all variables
 	except:
 		print("tf.train.Saver() broken in tensorflow 0.12")
-		saver = tf.train.Saver(tf.all_variables())# WTF stupid API breaking
+		saver = tf.train.Saver(tf.global_variables())# WTF stupid API breaking
 	ckpt = tf.train.get_checkpoint_state('./checkpoints')
 
 	start = 0
@@ -150,7 +150,7 @@
 	else:
 		print('Initializing')
 		try: session.run(tf.global_variables_initializer())
-		except:session.run(tf.initialize_all_variables())
+		except:session.run(tf.global_variables_initializer())
 	for epoch in range(nEpochs):
 		print('Epoch', epoch + 1, '...')
 		errors = np.zeros(len(edData))

spectro_gan.py

@@ -71,7 +71,7 @@ def check_accuracy():
 draw=0
 
 # Train
-tf.initialize_all_variables().run()
+tf.global_variables_initializer().run()
 steps=30000
 e=0
 for i in range(steps):

speech2text-seq2seq.py

@@ -92,7 +92,7 @@ def res_block(tensor, size, rate, dim=num_dim):
 # CTC loss
 loss = logit.sg_ctc(target=y, seq_len=seq_len)
 tf.train.AdamOptimizer(learning_rate).minimize(loss)
-saver = tf.train.Saver(tf.all_variables())
+saver = tf.train.Saver(tf.global_variables())
 
 # train
 tf.sg_train(log_interval=30, lr=0.0001, loss=loss, ep_size=1000, max_ep=200, early_stop=False)

speech_data.py

@@ -401,8 +401,8 @@ def many_hot_to_word(word):
 def dense_to_one_hot(batch, batch_size, num_labels):
 	sparse_labels = tf.reshape(batch, [batch_size, 1])
 	indices = tf.reshape(tf.range(0, batch_size, 1), [batch_size, 1])
-	concatenated = tf.concat(1, [indices, sparse_labels])
-	concat = tf.concat(0, [[batch_size], [num_labels]])
+	concatenated = tf.concat(axis=1, values=[indices, sparse_labels])
+	concat = tf.concat(axis=0, values=[[batch_size], [num_labels]])
 	output_shape = tf.reshape(concat, [2])
 	sparse_to_dense = tf.sparse_to_dense(concatenated, output_shape, 1.0, 0.0)
 	return tf.reshape(sparse_to_dense, [batch_size, num_labels])
@@ -411,8 +411,8 @@ def dense_to_one_hot(batch, batch_size, num_labels):
 def dense_to_one_hot(batch, batch_size, num_labels):
 	sparse_labels = tf.reshape(batch, [batch_size, 1])
 	indices = tf.reshape(tf.range(0, batch_size, 1), [batch_size, 1])
-	concatenated = tf.concat(1, [indices, sparse_labels])
-	concat = tf.concat(0, [[batch_size], [num_labels]])
+	concatenated = tf.concat(axis=1, values=[indices, sparse_labels])
+	concat = tf.concat(axis=0, values=[[batch_size], [num_labels]])
 	output_shape = tf.reshape(concat, [2])
 	sparse_to_dense = tf.sparse_to_dense(concatenated, output_shape, 1.0, 0.0)
 	return tf.reshape(sparse_to_dense, [batch_size, num_labels])

speech_encoder.py

@@ -94,7 +94,7 @@ def eval(feed):
 
 
 def train_spectrogram_encoder():
-  tf.initialize_all_variables().run()
+  tf.global_variables_initializer().run()
   print("Pretrain")
   for i in range(6000-1):
     batch_xs, batch_ys = speech.train.next_batch(100)

wave_GANerate.py

@@ -79,7 +79,7 @@ def play_pcm(data):
 
 
 # Train
-tf.initialize_all_variables().run()
+tf.global_variables_initializer().run()
 steps=3000#000
 batch=speech_data.wave_batch_generator(target=speech_data.Target.digits)
 negative=[0]*batch_size # input was fake