# In[4]:
tf.reset_default_graph()
batch_size = tf.placeholder_with_default(16, [], name='batch_size')
(input_op, seq_len,
label), input_prods = data.ops.get_even_batch_producer(paths=paths,
batch_size=batch_size)
# In[5]:
cnn_params = {
'out_dims': [256, 256, 256, 128],
'kernel_sizes': 64,
'pool_sizes': 1
}
c = cnn.model(seq_len=seq_len, input_op=input_op, **cnn_params)
#a = tf.transpose(c.output, perm=[0, 2, 1])
#a = tf.nn.top_k(a, k=8, sorted=False, name='MAX_POOL').values
#a = tf.transpose(a, perm=[0, 2, 1])
a = tf.reduce_mean(c.output, axis=1)
fc = classifier.model(input_op=a, fc_sizes=[])
logits = fc.logits
pred = fc.pred
MODEL_PATH = '/tmp/balanced/' + c.name + fc.name
MODEL_EXISTS = os.path.exists(MODEL_PATH)
if MODEL_EXISTS:
print('Model directory is not empty, removing old files')
shutil.rmtree(MODEL_PATH)
data /= data.std()
# ### Set up predictor
print('Building model graph...')
tf.reset_default_graph()
batch_size = tf.placeholder_with_default(1, [], name='batch_size')
keep_prob = tf.placeholder_with_default(1., [], name='keep_prob')
input_op = tf.placeholder(tf.float32, [1, None])
seq_len = tf.placeholder(tf.float32, [1])
cnn_params = {'out_dims': [32, 64, 64], 'kernel_sizes': 64, 'pool_sizes': 1}
c = cnn.model(seq_len=seq_len,
input_op=input_op,
keep_prob=keep_prob,
model_name='CNN_block',
**cnn_params)
RESIDUAL_POOL = 4
residual_input = c.output[..., None, :]
for i in range(1, 4):
residual_input = tf.contrib.layers.max_pool2d(
residual_input,
kernel_size=[RESIDUAL_POOL, 1],
stride=[RESIDUAL_POOL, 1])
c = cnn.model(seq_len=seq_len,
input_op=residual_input,
residual=True,