Reputation: 461
I'm not quite sure what my problem is here. When trying to run the model below, I get an error saying:
ValueError: Input 0 of layer "model" is incompatible with the layer: expected shape=(None, 250, 319, 3), found shape=(250, 319, 3)
The size of the image being returned from state() is (250, 319, 3). Which I have clearly defined in the get_actor() models input below. Some things I have tried:
One thing I have read, is that Keras may be expecting a batch of images, and not just a single one, so you may need to expand the dimension to (1, 250, 319, 3). I tried doing that using tf.expand_dims and I simply got another error saying it was expecting (None, 1, 250, 319, 3).
Another thing I tried was simply changing the input size to (None, 250, 319, 3) and it seemed to work until it produced another error in associaation with layer3 = layers.Flatten()(layer2) saying:
ValueError: The last dimension of the inputs to a Dense layer should be defined. Found None. Full input shape received: (None, None)
That could be a totally separate issue. Any suggestion on the best way to proceed. I'm not sure why it insists on including None in the input shape?
def state():
# can set to N number of captures per step
image = tf.convert_to_tensor(screen_capture())
img = tf.image.resize(image, [round(image.shape[0] / 2), round(image.shape[1] / 2)], preserve_aspect_ratio=True)
img_norm = tf.keras.utils.normalize(img, axis=0)
return img_norm
def get_actor():
# Initialize weights between -3e-3 and 3-e3
last_init = tf.random_uniform_initializer(minval=-0.003, maxval=0.003)
# Convolutions
inputs = layers.Input(shape=(250, 319, 3))
layer1 = layers.Conv2D(32, 3, strides=2, activation="relu")(inputs)
layer2 = layers.Conv2D(64, 3, strides=4, activation="relu")(layer1)
layer3 = layers.Flatten()(layer2)
# Fully connected layers
layer4 = layers.Dense(256, activation="relu")(layer3)
layer5 = layers.Dense(256, activation="relu")(layer4)
action = layers.Dense(3, activation="tanh", kernel_initializer=last_init)(layer5)
outputs = action
model = tf.keras.Model(inputs, outputs)
return model
actor_model = get_actor()
sampled_actions = tf.squeeze(actor_model(state()))
Upvotes: 1
Views: 2213
Reputation: 26708
Running your code with tf.expand_dims
does not seem to have any issues, so the error must be somewhere else:
import tensorflow as tf
def state():
return tf.expand_dims(tf.keras.utils.normalize(tf.random.normal((250, 319, 3))), axis=0)
def get_actor():
# Initialize weights between -3e-3 and 3-e3
last_init = tf.random_uniform_initializer(minval=-0.003, maxval=0.003)
# Convolutions
inputs = tf.keras.layers.Input(shape=(250, 319, 3))
layer1 = tf.keras.layers.Conv2D(32, 3, strides=2, activation="relu")(inputs)
layer2 = tf.keras.layers.Conv2D(64, 3, strides=4, activation="relu")(layer1)
layer3 = tf.keras.layers.Flatten()(layer2)
# Fully connected layers
layer4 = tf.keras.layers.Dense(256, activation="relu")(layer3)
layer5 = tf.keras.layers.Dense(256, activation="relu")(layer4)
action = tf.keras.layers.Dense(3, activation="tanh", kernel_initializer=last_init)(layer5)
outputs = action
model = tf.keras.Model(inputs, outputs)
return model
actor_model = get_actor()
sampled_actions = tf.squeeze(actor_model(state()))
print(sampled_actions)
tf.Tensor([ 0.00166097 0.00444322 -0.00574574], shape=(3,), dtype=float32)
Upvotes: 1