How to remove multiple layers from pretrained ResNet50V2 Keras model

Question

I'm trying to remove multiple layers form a pre-trained Keras model ( ResNet50V2 ), but no matter what I do it's not working. I've read countless other questions on stack overflow , github issues , and forum posts related to this topic in the past month, and I still can't make it work... So I'll ask directly. What might I be doing wrong?

from ray.rllib.models.tf.tf_modelv2 import TFModelV2
from ray.rllib.utils.framework import try_import_tf
from ray.rllib.models import ModelCatalog
tf = try_import_tf()

def resnet_core(x):
    x = tf.keras.applications.resnet_v2.preprocess_input(x)
    resnet = tf.keras.applications.ResNet50V2(
        include_top=False,
        weights="imagenet",
    )
    remove_n = 130
    for i in range(remove_n):
        resnet._layers.pop()
        print(len(resnet._layers))
    s = tf.keras.models.Model(resnet.input, resnet._layers[-1].output, name='resnet-core')
    for layer in s.layers:
        print('adding layer',layer.name)
    for layer in s.layers[:]:
        layer.trainable = False
    s.build(None)

    return s(x)

class ImpalaCNN(TFModelV2):

    def __init__(self, obs_space, action_space, num_outputs, model_config, name):
        super().__init__(obs_space, action_space, num_outputs, model_config, name)

        inputs = tf.keras.layers.Input(shape=obs_space.shape, name="observations")
        x = inputs
        x = resnet_core(x)
        x = tf.keras.layers.Flatten()(x)
        x = tf.keras.layers.ReLU()(x)
        x = tf.keras.layers.Dense(units=256, activation="relu", name="hidden")(x)

        logits = tf.keras.layers.Dense(units=num_outputs, name="pi")(x)
        value = tf.keras.layers.Dense(units=1, name="vf")(x)

        self.base_model = tf.keras.Model(inputs, [logits, value])
        self.register_variables(self.base_model.variables)

    def forward(self, input_dict, state, seq_lens):
        obs = tf.cast(input_dict["obs"], tf.float32)
        logits, self._value = self.base_model(obs)
        return logits, state

    def value_function(self):
        return tf.reshape(self._value, [-1])


# Register model in ModelCatalog
ModelCatalog.register_custom_model("impala_cnn_tf", ImpalaCNN)

The error I'm getting is:

  ...
  File "/Users/manu/anaconda3/envs/procgen/lib/python3.7/site-packages/ray/rllib/evaluation/rollout_worker.py", line 376, in __init__
    self._build_policy_map(policy_dict, policy_config)
  File "/Users/manu/anaconda3/envs/procgen/lib/python3.7/site-packages/ray/rllib/evaluation/rollout_worker.py", line 859, in _build_policy_map
    policy_map[name] = cls(obs_space, act_space, merged_conf)
  File "/Users/manu/anaconda3/envs/procgen/lib/python3.7/site-packages/ray/rllib/policy/tf_policy_template.py", line 143, in __init__
    obs_include_prev_action_reward=obs_include_prev_action_reward)
  File "/Users/manu/anaconda3/envs/procgen/lib/python3.7/site-packages/ray/rllib/policy/dynamic_tf_policy.py", line 163, in __init__
    framework="tf")
  File "/Users/manu/anaconda3/envs/procgen/lib/python3.7/site-packages/ray/rllib/models/catalog.py", line 317, in get_model_v2
    registered))
ValueError: It looks like variables {<tf.Variable 'default_policy/
conv4_block4_1_conv/kernel:0' ... } 
were created as part of <impala_cnn_tf.ImpalaCNN object at 
0x19a8ccc90> but does not appear in model.variables() 
({<tf.Variable 'default_policy/pi/
kernel:0' shape=(256, 15) dtype=float32> ...}). Did you forget to call
 model.register_variables() on the variables in question?

The error seems to indicate some variables from the layers I'm trying to skip were not registered, but that's because I don't want to use them? Any ideas?

More context in case it helps:

• If I set remove_n = 0 I don't see the error (but of course the whole ResNet50V2 is being used)
• I'm a newbie with Keras and ML. This might be a very dumb question.
• The reason I'm trying to remove many layers, not just the last one, is that I want the model to fit on a small GPU.
• I'm trying to train the model using rllib for the aicrowd procgen competition (using the competition to learn Keras/RL)
• Full code here: https://github.com/maraoz/neurips2020-procgen-starter-kit/blob/master/models/impala_cnn_tf.py#L55
• Full error log here: https://pastebin.com/X0Dk7wdd

Thanks in advance!

Answer 1

Rather than popping off layers, you could try accessing the 130th layer from the last layer. Then, you can build a new model using the input of your original model and the output of this layer.

model = tf.keras.models.Model(resnet.input, resnet.layers[-130].output)

This will do essentially the same thing as what you tried but its much easier and safer since you aren't accessing any private properties of the model itself.