def get_model():
return ipu.keras.Sequential(
[Embedding(256, 128),
LSTM(EMBEDDING_DIM, return_sequences=True),
LSTM(EMBEDDING_DIM, return_sequences=True),
TimeDistributed(Dense(256, activation='softmax'))],
accumulation_count = REPEAT_COUNT)
def get_model():
input_layer = Input(shape=(80), dtype=tf.int32, batch_size=minibatch_size)
x = Embedding(max_features, 128)(input_layer)
x = LSTM(128, dropout=0.2)(x)
x = Dense(16, activation='relu')(x)
x = Dense(1, activation='sigmoid')(x)
return ipu.keras.Model(input_layer, x)
def get_model():
input_layer = Input(shape=(80), dtype=tf.int32, batch_size=minibatch_size)
with ipu.keras.PipelineStage(0):
x = Embedding(max_features, 128)(input_layer)
with ipu.keras.PipelineStage(1):
x = LSTM(128, dropout=0.2)(x)
x = Dense(1, activation='sigmoid')(x)
return ipu.keras.PipelineModel(
input_layer,
x,
gradient_accumulation_count=gradient_accumulation_count)
def get_model():
input_layer = Input(shape=(80), dtype=dtypes.int32, batch_size=32)
with ipu.keras.PipelineStage(0):
x = Embedding(max_features, 128)(input_layer)
x = LSTM(128, dropout=0.2)(x)
with ipu.keras.PipelineStage(1):
a = Dense(16, activation='relu')(x)
with ipu.keras.PipelineStage(2):
b = Dense(16, activation='relu')(x)
with ipu.keras.PipelineStage(3):
x = Concatenate()([a, b])
x = Dense(1, activation='sigmoid')(x)
return ipu.keras.PipelineModel(input_layer,
x,
gradient_accumulation_count=16,
device_mapping=[0, 1, 1, 0])
def get_model():
return ipu.keras.SequentialPipelineModel(
[[Embedding(max_features, 128)],
[LSTM(128, dropout=0.2),
Dense(1, activation='sigmoid')]],
gradient_accumulation_count=gradient_accumulation_count)
def get_model():
return ipu.keras.Sequential([
Embedding(max_features, 128),
LSTM(128, dropout=0.2),
Dense(1, activation='sigmoid')
])
def get_model():
return ipu.keras.PipelinedModel(
[[Embedding(max_features, 128)],
[LSTM(128, dropout=0.2),
Dense(1, activation='sigmoid')]],
pipeline_depth=pipeline_depth)
