def test_is_pickleable(self, filters, strides):
model = CNNModel(filters=filters,
strides=strides,
name='cnn_model',
padding='VALID',
hidden_w_init=tf.constant_initializer(1),
hidden_nonlinearity=None)
outputs = model.build(self._input_ph).outputs
with tf.compat.v1.variable_scope('cnn_model/cnn/h0', reuse=True):
bias = tf.compat.v1.get_variable('bias')
bias.load(tf.ones_like(bias).eval())
output1 = self.sess.run(outputs,
feed_dict={self._input_ph: self.obs_input})
h = pickle.dumps(model)
with tf.compat.v1.Session(graph=tf.Graph()) as sess:
model_pickled = pickle.loads(h)
# pylint: disable=unsubscriptable-object
input_shape = self.obs_input.shape[1:] # height, width, channel
input_ph = tf.compat.v1.placeholder(tf.float32,
shape=(None, ) + input_shape,
name='input')
outputs = model_pickled.build(input_ph).outputs
output2 = sess.run(outputs, feed_dict={input_ph: self.obs_input})
assert np.array_equal(output1, output2)
def test_output_value(self, filters, in_channels, strides):
model = CNNModel(filters=filters,
strides=strides,
name='cnn_model',
padding='VALID',
hidden_w_init=tf.constant_initializer(1),
hidden_nonlinearity=None)
outputs = model.build(self._input_ph).outputs
output = self.sess.run(outputs,
feed_dict={self._input_ph: self.obs_input})
filter_sum = 1
# filter value after 3 layers of conv
for filter_iter, in_channel in zip(filters, in_channels):
filter_sum *= filter_iter[1][0] * filter_iter[1][1] * in_channel
height_size = self.input_height
width_size = self.input_width
for filter_iter, stride in zip(filters, strides):
height_size = int((height_size - filter_iter[1][0]) / stride) + 1
width_size = int((width_size - filter_iter[1][1]) / stride) + 1
flatten_shape = height_size * width_size * filters[-1][0]
# flatten
expected_output = np.full((self.batch_size, flatten_shape),
filter_sum,
dtype=np.float32)
assert np.array_equal(output, expected_output)
def test_is_pickleable(self, filter_sizes, in_channels, out_channels,
strides):
model = CNNModel(filter_dims=filter_sizes,
num_filters=out_channels,
strides=strides,
name='cnn_model',
padding='VALID',
hidden_w_init=tf.constant_initializer(1),
hidden_nonlinearity=None)
outputs = model.build(self._input_ph)
with tf.variable_scope('cnn_model/cnn/h0', reuse=True):
bias = tf.get_variable('bias')
self.sess.run(tf.assign(bias, tf.ones_like(bias)))
output1 = self.sess.run(outputs,
feed_dict={self._input_ph: self.obs_input})
h = pickle.dumps(model)
with tf.Session(graph=tf.Graph()) as sess:
model_pickled = pickle.loads(h)
input_shape = self.obs_input.shape[1:] # height, width, channel
input_ph = tf.placeholder(tf.float32,
shape=(None, ) + input_shape,
name='input')
outputs = model_pickled.build(input_ph)
output2 = sess.run(outputs, feed_dict={input_ph: self.obs_input})
assert np.array_equal(output1, output2)