rcnn-sat

Code for feedforward and recurrent neural network models used in the paper Recurrent neural networks can explain flexible trading of speed and accuracy in biological vision.

The code has been tested with TensorFlow 1.13 and Python 3.6.

Using the code

The following code snippet shows how to build the Keras model and generate a prediction for a random image. A full example of extracting activations from a pre-trained model is given here.

import numpy as np
import tensorflow as tf
tf.enable_v2_behavior()
from rcnn_sat import preprocess_image, bl_net

# create the model with randomly initialised weights
input_layer = tf.keras.layers.Input((128, 128, 3))
model = bl_net(input_layer, classes=565, cumulative_readout=True)

# predict a random image
img = np.random.randint(0, 256, [1, 128, 128, 3], dtype=np.uint8)
model.predict(preprocess_image(img)) # softmax for each time step

Pre-trained model weights

The checkpoint files for pre-trained eco-set and imagenet models are hosted here.

Notes on pre-trained models:

• Pre-trained models expect 128 x 128 images as input.
• ImageNet models have classes=1000 and ecoset models have classes=565
• BL models were trained with cumulative_readout=False but can be tested using either option
• Model predictions correspond to the order of the categories within the files in pretrained_output_categories.