andrewschreiber

#!/bin/sh
# See video https://www.youtube.com/watch?v=7PO27i2lEOs

set -e


command_exists () {
    type "$1" &> /dev/null ;
}

andrewschreiber

Open jupyter with

$ xvfb-run -s "-screen 0 1400x900x24" jupyter notebook

In Jupyter

import matplotlib.pyplot as plt
%matplotlib inline

After each step

def show_state(env, step=0):

andrewschreiber

# Create a saliency map for each data point
for i, image in enumerate(data):
    # Forward pass on image
    # Note: the activations are saved on each layer
    output = image
    for l in range(len(network.layers)):
        output = network.layers[l].forward(output)

    # Backprop to get gradient
    label_one_hot = labels[i]

andrewschreiber

# Create a saliency map for each data point
for i, image in enumerate(data):
    # Forward pass on image
    # Note: the activations from this are saved on each layer
    output = image
    for l in range(len(network.layers)):
        output = network.layers[l].forward(output)

    # Backprop to get gradient
    label_one_hot = labels[i]

andrewschreiber

    # Backprop to get gradient
    label_one_hot = labels[i]
    dy = np.array(label_one_hot)
    for l in range(len(network.layers)-1, -1, -1):
        dout = network.layers[l].backward(dy)
        dy = dout

andrewschreiber

# Create a saliency map for each data point
for i, image in enumerate(data):
    # Run a forward pass with an image
    output = image
    for l in range(len(network.layers)):
        output = network.layers[l].forward(output)

andrewschreiber

from model.data import mnist_train_test_sets
from model.network import LeNet5
from saliency.vanilla_gradient import save_vanilla_gradient
# Get MNIST dataset, preprocessed
train_images, train_labels, test_images, test_labels = mnist_train_test_sets()

# Load net with 98% acc weights
net = LeNet5(weights_path="15epoch_weights.pkl")

# Generate saliency maps for the first 10 images

andrewschreiber

# From CLI:
# hyperdash run -n 'mymodel' python mymodel.py

import hyperdash as hd
learning_rate = hd.param('learning rate', default=0.01)  # Setup hyperparameters
# Model code here
hd.metric('loss', training_loss)  # Record a metric
# Params and metrics are pretty printed at end of experiment

andrewschreiber

Keybase proof

I hereby claim:

• I am andrewschreiber on github.
• I am andrewschreiber (https://keybase.io/andrewschreiber) on keybase.
• I have a public key whose fingerprint is B124 EC8F E431 5EC0 7D0C 6205 2824 08D7 8326 AB72

To claim this, I am signing this object:

andrewschreiber

{
  "name": "ReactiveCocoa",
  "version": "4.2.2",
  "summary": "A framework for composing and transforming streams of values.",
  "description": "ReactiveCocoa (RAC) is an Objective-C framework for Functional Reactive Programming.\nIt provides APIs for composing and transforming streams of values.",
  "homepage": "https://github.com/ReactiveCocoa/ReactiveCocoa",
  "license": {
    "type": "MIT",
    "file": "LICENSE.md"
  },