def run_test():
cfg = {"rng": np.random}
input_shape = 2
output_shape = 2
instructions = \
{
"init_std": 0.05,
"layers" : [1],
"layer_functions" : ['relu'],
"layer_extras" : ['bn'],
"output_function" : 'linear',
"output_extras" : 'bn',
}
policy = FeedForward(input_shape, output_shape, None, cfg)
policy.build_model(instructions)
print("BUILT POLICY LAYERS:")
for layer in policy.model:
print(layer)
flat = np.random.randn(policy.num_params)
print("POLICY FLAT BEFORE SETTING:", policy.get_trainable_flat())
policy.set_trainable_flat(flat)
print("POLICY FLAT AFTER SETTING:", policy.get_trainable_flat())
print("FLAT SHOULD NOW BE:", flat)
def test_save_load_vbn():
cfg = {"rng": np.random}
input_shape = 8
output_shape = 8
instructions = \
{
"init_std": 0.05,
"layers": [64, 64],
"layer_functions": ['relu', 'relu'],
"layer_extras": ['bn', 'bn'],
"output_function": 'linear',
"output_extras": 'bn',
}
policy = FeedForward(input_shape, output_shape, None, cfg)
policy.build_model(instructions)
print("BUILT POLICY LAYERS:")
for layer in policy.model:
print(layer)
vbn = [np.random.randn(input_shape) for _ in range(1000)]
policy.compute_virtual_normalization(vbn)
inp = np.ones(input_shape)
out = policy.activate(inp)
print("\nOUTPUT ON ONES WITH VBN BEFORE SAVING:", out)
policy.save("data/experiments/exp_name/epochs/epoch_0/policy")
out = policy.activate(inp)
print("\nOUTPUT ON ONES WITH VBN AFTER SAVING:", out)
policy.set_trainable_flat(policy.get_trainable_flat() +
np.random.randn(policy.num_params))
out = policy.activate(inp)
print("\nJIGGLED OUTPUT ON ONES WITH VBN BEFORE LOADING", out)
policy.load("data/experiments/exp_name/epochs/epoch_0/policy")
out = policy.activate(inp)
print("OUTPUT ON ONES WITH VBN AFTER LOADING:", out)