yeah it's lit

dnc/controller.py

@@ -77,6 +77,7 @@ class Controller():
         Returns: int
             controller_dim: the output dimension of the controller
         '''
+        with tf.variable_scope("dnc_scope") as scope:
             test_chi = tf.zeros([self.batch_size, self.chi_dim])
             nn_output, nn_state = self.nn_step(test_chi, state=self.get_state())
         return nn_output.get_shape().as_list()[-1]

dnc/dnc.py

@@ -10,7 +10,7 @@ from memory import Memory
 import os
 
 class DNC:
-    def __init__(self, make_controller, FLAGS):
+    def __init__(self, make_controller, FLAGS, input_steps=None):
         '''
         Builds a TensorFlow graph for the Differentiable Neural Computer. Uses TensorArrays and a while loop for efficiency
         Parameters:
@@ -40,33 +40,61 @@ class DNC:
         self.X = tf.placeholder(tf.float32, [batch_size, None, xlen], name='X')
         self.y = tf.placeholder(tf.float32, [batch_size, None, ylen], name='y')
         self.tsteps = tf.placeholder(tf.int32, name='tsteps')
+        self.input_steps = input_steps if input_steps is not None else self.tsteps
 
         self.X_tensor_array = self.unstack_time_dim(self.X)
 
         # initialize states
-        nn_state = self.controller.get_state()
+        self.nn_state = self.controller.get_state()
-        dnc_state = self.memory.zero_state()
+        self.dnc_state = self.memory.zero_state()
 
         # values for which we want a history
         self.hist_keys = ['y_hat', 'f', 'g_a', 'g_w', 'w_r', 'w_w', 'u']
-        dnc_hist = [tf.TensorArray(tf.float32, self.tsteps) for _ in range(len(self.hist_keys))]
+        dnc_hist = [tf.TensorArray(tf.float32, self.tsteps, clear_after_read=False) for _ in range(len(self.hist_keys))]
 
         # loop through time
-        with tf.variable_scope("while_loop") as scope:
+        with tf.variable_scope("dnc_scope", reuse=True) as scope:
             time = tf.constant(0, dtype=tf.int32)
 
             output = tf.while_loop(
                 cond=lambda time, *_: time < self.tsteps,
                 body=self.step,
-                loop_vars=(time, nn_state, dnc_state, dnc_hist),
+                loop_vars=(time, self.nn_state, self.dnc_state, dnc_hist),
                 )
-        (_, next_nn_state, next_dnc_state, dnc_hist) = output
+        (_, self.next_nn_state, self.next_dnc_state, dnc_hist) = output
 
         # write down the history
-        controller_dependencies = [self.controller.update_state(next_nn_state)]
+        controller_dependencies = [self.controller.update_state(self.next_nn_state)]
         with tf.control_dependencies(controller_dependencies):
             self.dnc_hist = {self.hist_keys[i]: self.stack_time_dim(v) for i, v in enumerate(dnc_hist)} # convert to dict
 
+    def step2(self, time, nn_state, dnc_state, dnc_hist):
+
+        # map from tuple to dict for readability
+        dnc_state = {self.memory.state_keys[i]: v for i, v in enumerate(dnc_state)}
+        dnc_hist = {self.hist_keys[i]: v for i, v in enumerate(dnc_hist)}
+
+        # one full pass!
+        X_t = self.X_tensor_array.read(time)
+        v, zeta, next_nn_state = self.controller.step(X_t, dnc_state['r'], nn_state)
+        next_dnc_state = self.memory.step(zeta, dnc_state)
+        y_hat = self.controller.next_y_hat(v, next_dnc_state['r'])
+
+        dnc_hist['y_hat'] = dnc_hist['y_hat'].write(time, y_hat)
+        dnc_hist['f']   = dnc_hist['f'].write(time, zeta['f'])
+        dnc_hist['g_a'] = dnc_hist['g_a'].write(time, zeta['g_a'])
+        dnc_hist['g_w'] = dnc_hist['g_w'].write(time, zeta['g_w'])
+        dnc_hist['w_r'] = dnc_hist['w_r'].write(time, next_dnc_state['w_r'])
+        dnc_hist['w_w'] = dnc_hist['w_w'].write(time, next_dnc_state['w_w'])
+        dnc_hist['u']   = dnc_hist['u'].write(time, next_dnc_state['u'])
+
+        # map from dict to tuple for tf.while_loop :/
+        next_dnc_state = [next_dnc_state[k] for k in self.memory.state_keys]
+        dnc_hist = [dnc_hist[k] for k in self.hist_keys]
+
+        time += 1
+        return time, next_nn_state, next_dnc_state, dnc_hist
+
     def step(self, time, nn_state, dnc_state, dnc_hist):
         '''
         Performs the feedforward step of the DNC in order to get the DNC output
@@ -88,8 +116,15 @@ class DNC:
         dnc_state = {self.memory.state_keys[i]: v for i, v in enumerate(dnc_state)}
         dnc_hist = {self.hist_keys[i]: v for i, v in enumerate(dnc_hist)}
 
+        def use_prev_output():
+            y_prev = tf.concat((dnc_hist['y_hat'].read(time-1), tf.zeros([self.batch_size, self.xlen - self.ylen])), axis=1)
+            return tf.reshape(y_prev, (self.batch_size, self.xlen))
+
+        def use_input_array():
+            return self.X_tensor_array.read(time)
+
         # one full pass!
-        X_t = self.X_tensor_array.read(time)
+        X_t = tf.cond(time < self.input_steps, use_input_array, use_prev_output)
         v, zeta, next_nn_state = self.controller.step(X_t, dnc_state['r'], nn_state)
         next_dnc_state = self.memory.step(zeta, dnc_state)
         y_hat = self.controller.next_y_hat(v, next_dnc_state['r'])

handwriting/.ipynb_checkpoints/Untitled-checkpoint.ipynb

@@ -0,0 +1,6 @@
+{
+ "cells": [],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 1
+}

handwriting/.ipynb_checkpoints/debug-checkpoint.ipynb

@@ -0,0 +1,6 @@
+{
+ "cells": [],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 1
+}

handwriting/logs/train_scribe.txt

@@ -0,0 +1,7 @@
+Scribe: Realistic Handriting in Tensorflow
+     by Sam Greydanus
+
+
+	loaded dataset:
+		11895 individual data points
+		5947 batches

handwriting/rnn_controller.py

@@ -0,0 +1,26 @@
+import numpy as np
+import tensorflow as tf
+from controller import Controller
+from tensorflow.contrib.rnn.python.ops.core_rnn_cell import LSTMStateTuple
+
+"""
+A 1-Layer recurrent neural network (LSTM) with 64 hidden nodes
+"""
+
+class RNNController(Controller):
+
+    def init_controller_params(self):
+        self.rnn_dim = 150
+        self.lstm_cell = tf.contrib.rnn.core_rnn_cell.BasicLSTMCell(self.rnn_dim)
+        self.state = tf.Variable(tf.zeros([self.batch_size, self.rnn_dim]), trainable=False)
+        self.output = tf.Variable(tf.zeros([self.batch_size, self.rnn_dim]), trainable=False)
+
+    def nn_step(self, X, state):
+        X = tf.convert_to_tensor(X)
+        return self.lstm_cell(X, state)
+
+    def update_state(self, update):
+        return tf.group(self.output.assign(update[0]), self.state.assign(update[1]))
+
+    def get_state(self):
+        return LSTMStateTuple(self.output, self.state)

handwriting/train.py

@@ -0,0 +1,142 @@
+import tensorflow as tf
+import numpy as np
+import sys
+sys.path.insert(0, '../dnc')
+
+from dnc import DNC
+from utils import *
+from rnn_controller import RNNController
+
+# hyperparameters 
+alphabet = "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ"
+stroke_steps = 175
+tf.app.flags.DEFINE_integer("xlen", len(alphabet) + 4, "Input dimension")
+tf.app.flags.DEFINE_integer("ylen", 3, "output dimension")
+tf.app.flags.DEFINE_integer("stroke_steps", stroke_steps, "Number of time steps for stroke data")
+tf.app.flags.DEFINE_integer("ascii_steps", stroke_steps/25, "Sequence length")
+tf.app.flags.DEFINE_integer("data_scale", 50, "How to scale stroke data")
+tf.app.flags.DEFINE_integer("batch_size", 2, "Size of batch in minibatch gradient descent")
+
+tf.app.flags.DEFINE_integer("R", 1, "Number of DNC read heads")
+tf.app.flags.DEFINE_integer("W", 16, "Word length for DNC memory")
+tf.app.flags.DEFINE_integer("N", 10, "Number of words the DNC memory can store")
+
+tf.app.flags.DEFINE_integer("train", True, "Train or sample???")
+tf.app.flags.DEFINE_integer("print_every", 100, "Print training info after this number of train steps")
+tf.app.flags.DEFINE_integer("iterations", 5000000, "Number of training iterations")
+tf.app.flags.DEFINE_float("lr", 1e-4, "Learning rate (alpha) for the model")
+tf.app.flags.DEFINE_float("momentum", .9, "RMSProp momentum")
+tf.app.flags.DEFINE_integer("save_every", 1000, "Save model after this number of train steps")
+tf.app.flags.DEFINE_string("save_path", "models/model.ckpt", "Where to save checkpoints")
+tf.app.flags.DEFINE_string("data_dir", "data/", "Where to save checkpoints")
+tf.app.flags.DEFINE_string("log_dir", "logs/", "Where to find data")
+tf.app.flags.DEFINE_string("alphabet", alphabet, "Viable characters")
+FLAGS = tf.app.flags.FLAGS
+
+
+# modify dataloader
+def next_batch(FLAGS, dl):
+    X_batch = []
+    y_batch = []
+    text = []
+    _X_batch, _y_batch, _text, _one_hots = dl.next_batch()
+    for i in range(FLAGS.batch_size):
+        ascii_part = np.concatenate((_one_hots[i], np.zeros((FLAGS.ascii_steps, 3))), axis=1)
+        X_stroke_part = np.concatenate((np.zeros((FLAGS.stroke_steps, len(FLAGS.alphabet)+1)), _X_batch[i]), axis=1)
+        
+        X = np.concatenate((ascii_part, X_stroke_part), axis=0)
+        y = np.concatenate((np.zeros((FLAGS.ascii_steps, 3)), _y_batch[i]), axis=0)
+        X_batch.append(X) ; y_batch.append(y) ; text.append(_text)
+    return [X_batch, y_batch, text]
+
+logger = Logger(FLAGS)
+dl = Dataloader(FLAGS, logger, limit = 500)
+
+
+# helper funcs
+def binary_cross_entropy(y_hat, y):
+    return tf.reduce_mean(-y*tf.log(y_hat) - (1-y)*tf.log(1-y_hat))
+
+def llprint(message):
+    sys.stdout.write(message)
+    sys.stdout.flush()
+
+
+# build graph
+sess = tf.InteractiveSession()
+
+llprint("building graph...\n")
+optimizer = tf.train.RMSPropOptimizer(FLAGS.lr, momentum=FLAGS.momentum)
+dnc = DNC(RNNController, FLAGS, input_steps=FLAGS.ascii_steps)
+
+llprint("defining loss...\n")
+y_hat, outputs = dnc.get_outputs()
+# TODO: fix this loss: l2 on [:,:,:2] and then do binary cross entropy on <EOS> tags
+loss = tf.nn.l2_loss(dnc.y - y_hat)*100./(FLAGS.batch_size*(FLAGS.ascii_steps+FLAGS.stroke_steps))
+
+llprint("computing gradients...\n")
+gradients = optimizer.compute_gradients(loss)
+for i, (grad, var) in enumerate(gradients):
+    if grad is not None:
+        gradients[i] = (tf.clip_by_value(grad, -10, 10), var)
+                    
+grad_op = optimizer.apply_gradients(gradients)
+
+llprint("init variables... \n")
+sess.run(tf.global_variables_initializer())
+llprint("ready to train...")
+
+
+# model overiew
+# tf parameter overview
+total_parameters = 0 ; print "model overview..."
+for variable in tf.trainable_variables():
+    shape = variable.get_shape()
+    variable_parameters = 1
+    for dim in shape:
+        variable_parameters *= dim.value
+    print '\tvariable "{}" has {} parameters' \
+        .format(variable.name, variable_parameters)
+    total_parameters += variable_parameters
+print "total of {} parameters".format(total_parameters)
+
+
+# load saved models
+global_step = 0
+saver = tf.train.Saver(tf.global_variables())
+load_was_success = True # yes, I'm being optimistic
+try:
+    save_dir = '/'.join(FLAGS.save_path.split('/')[:-1])
+    ckpt = tf.train.get_checkpoint_state(save_dir)
+    load_path = ckpt.model_checkpoint_path
+    saver.restore(sess, load_path)
+except:
+    print "no saved model to load."
+    load_was_success = False
+else:
+    print "loaded model: {}".format(load_path)
+    saver = tf.train.Saver(tf.global_variables())
+    global_step = int(load_path.split('-')[-1]) + 1
+
+
+# train, baby, train
+loss_history = []
+for i in xrange(global_step, FLAGS.iterations + 1):
+    llprint("\rIteration {}/{}".format(i, FLAGS.iterations))
+
+    X, y, text = next_batch(FLAGS, dl)
+    tsteps = FLAGS.ascii_steps + FLAGS.stroke_steps
+
+    fetch = [loss, grad_op]
+    feed = {dnc.X: X, dnc.y: y, dnc.tsteps: tsteps}
+
+    step_loss, _ = sess.run(fetch, feed_dict=feed)
+    loss_history.append(step_loss)
+    global_step = i
+
+    if i % 100 == 0:
+        llprint("\n\tloss: {:03.4f}\n".format(np.mean(loss_history)))
+        loss_history = []
+    if i % FLAGS.save_every == 0 and i is not 0:
+        llprint("\n\tSAVING MODEL\n")
+        saver.save(sess, FLAGS.save_path, global_step=global_step)

handwriting/utils.py

@@ -0,0 +1,203 @@
+import numpy as np
+import math
+import random
+import os
+import cPickle as pickle
+import xml.etree.ElementTree as ET
+
+from utils import *
+
+class Dataloader():
+    def __init__(self, FLAGS, logger, limit = 500):
+        self.data_dir = FLAGS.data_dir
+        self.alphabet = FLAGS.alphabet
+        self.batch_size = FLAGS.batch_size
+        self.stroke_steps = FLAGS.stroke_steps
+        self.data_scale = FLAGS.data_scale # scale data down by this factor
+        self.ascii_steps = FLAGS.ascii_steps
+        self.logger = logger
+        self.limit = limit # removes large noisy gaps in the data
+
+        data_file = os.path.join(self.data_dir, "strokes_training_data.cpkl")
+        stroke_dir = self.data_dir + "/lineStrokes"
+        ascii_dir = self.data_dir + "/ascii"
+
+        if not (os.path.exists(data_file)) :
+            self.logger.write("\tcreating training data cpkl file from raw source")
+            self.preprocess(stroke_dir, ascii_dir, data_file)
+
+        self.load_preprocessed(data_file)
+        self.reset_batch_pointer()
+
+    def preprocess(self, stroke_dir, ascii_dir, data_file):
+        # create data file from raw xml files from iam handwriting source.
+        self.logger.write("\tparsing dataset...")
+        
+        # build the list of xml files
+        filelist = []
+        # Set the directory you want to start from
+        rootDir = stroke_dir
+        for dirName, subdirList, fileList in os.walk(rootDir):
+            for fname in fileList:
+                filelist.append(dirName+"/"+fname)
+
+        # function to read each individual xml file
+        def getStrokes(filename):
+            tree = ET.parse(filename)
+            root = tree.getroot()
+
+            result = []
+
+            x_offset = 1e20
+            y_offset = 1e20
+            y_height = 0
+            for i in range(1, 4):
+                x_offset = min(x_offset, float(root[0][i].attrib['x']))
+                y_offset = min(y_offset, float(root[0][i].attrib['y']))
+                y_height = max(y_height, float(root[0][i].attrib['y']))
+            y_height -= y_offset
+            x_offset -= 100
+            y_offset -= 100
+
+            for stroke in root[1].findall('Stroke'):
+                points = []
+                for point in stroke.findall('Point'):
+                    points.append([float(point.attrib['x'])-x_offset,float(point.attrib['y'])-y_offset])
+                result.append(points)
+            return result
+        
+        # function to read each individual xml file
+        def getAscii(filename, line_number):
+            with open(filename, "r") as f:
+                s = f.read()
+            s = s[s.find("CSR"):]
+            if len(s.split("\n")) > line_number+2:
+                s = s.split("\n")[line_number+2]
+                return s
+            else:
+                return ""
+                
+        # converts a list of arrays into a 2d numpy int16 array
+        def convert_stroke_to_array(stroke):
+            n_point = 0
+            for i in range(len(stroke)):
+                n_point += len(stroke[i])
+            stroke_data = np.zeros((n_point, 3), dtype=np.int16)
+
+            prev_x = 0
+            prev_y = 0
+            counter = 0
+
+            for j in range(len(stroke)):
+                for k in range(len(stroke[j])):
+                    stroke_data[counter, 0] = int(stroke[j][k][0]) - prev_x
+                    stroke_data[counter, 1] = int(stroke[j][k][1]) - prev_y
+                    prev_x = int(stroke[j][k][0])
+                    prev_y = int(stroke[j][k][1])
+                    stroke_data[counter, 2] = 0
+                    if (k == (len(stroke[j])-1)): # end of stroke
+                        stroke_data[counter, 2] = 1
+                    counter += 1
+            return stroke_data
+
+        # build stroke database of every xml file inside iam database
+        strokes = []
+        asciis = []
+        for i in range(len(filelist)):
+            if (filelist[i][-3:] == 'xml'):
+                stroke_file = filelist[i]
+#                 print 'processing '+stroke_file
+                stroke = convert_stroke_to_array(getStrokes(stroke_file))
+                
+                ascii_file = stroke_file.replace("lineStrokes","ascii")[:-7] + ".txt"
+                line_number = stroke_file[-6:-4]
+                line_number = int(line_number) - 1
+                ascii = getAscii(ascii_file, line_number)
+                if len(ascii) > 10:
+                    strokes.append(stroke)
+                    asciis.append(ascii)
+                else:
+                    self.logger.write("\tline length was too short. line was: " + ascii)
+                
+        assert(len(strokes)==len(asciis)), "There should be a 1:1 correspondence between stroke data and ascii labels."
+        f = open(data_file,"wb")
+        pickle.dump([strokes,asciis], f, protocol=2)
+        f.close()
+        self.logger.write("\tfinished parsing dataset. saved {} lines".format(len(strokes)))
+
+
+    def load_preprocessed(self, data_file):
+        f = open(data_file,"rb")
+        [self.raw_stroke_data, self.raw_ascii_data] = pickle.load(f)
+        f.close()
+
+        # goes thru the list, and only keeps the text entries that have more than stroke_steps points
+        self.stroke_data = []
+        self.ascii_data = []
+        counter = 0
+
+        for i in range(len(self.raw_stroke_data)):
+            data = self.raw_stroke_data[i]
+            if len(data) > (self.stroke_steps+2):
+                # removes large gaps from the data
+                data = np.minimum(data, self.limit)
+                data = np.maximum(data, -self.limit)
+                data = np.array(data,dtype=np.float32)
+                data[:,0:2] /= self.data_scale
+                
+                self.stroke_data.append(data)
+                self.ascii_data.append(self.raw_ascii_data[i])
+
+        # minus 1, since we want the ydata to be a shifted version of x data
+        self.num_batches = int(len(self.stroke_data) / self.batch_size)
+        self.logger.write("\tloaded dataset:")
+        self.logger.write("\t\t{} individual data points".format(len(self.stroke_data)))
+        self.logger.write("\t\t{} batches".format(self.num_batches))
+
+    def next_batch(self):
+        # returns a randomized, stroke_steps-sized portion of the training data
+        x_batch = []
+        y_batch = []
+        ascii_list = []
+        for i in xrange(self.batch_size):
+            data = self.stroke_data[self.idx_perm[self.pointer]]
+            idx = random.randint(0, len(data)-self.stroke_steps-2)
+            x_batch.append(np.copy(data[:self.stroke_steps]))
+            y_batch.append(np.copy(data[1:self.stroke_steps+1]))
+            ascii_list.append(self.ascii_data[self.idx_perm[self.pointer]][:self.ascii_steps])
+            self.tick_batch_pointer()
+        one_hots = [to_one_hot(s, self.ascii_steps, self.alphabet) for s in ascii_list]
+        return x_batch, y_batch, ascii_list, one_hots
+
+    def tick_batch_pointer(self):
+        self.pointer += 1
+        if (self.pointer >= len(self.stroke_data)):
+            self.reset_batch_pointer()
+    def reset_batch_pointer(self):
+        self.idx_perm = np.random.permutation(len(self.stroke_data))
+        self.pointer = 0
+
+# utility function for converting input ascii characters into vectors the network can understand.
+# index position 0 means "unknown"
+def to_one_hot(s, ascii_steps, alphabet):
+    steplimit=3e3; s = s[:3e3] if len(s) > 3e3 else s # clip super-long strings
+    seq = [alphabet.find(char) + 1 for char in s]
+    if len(seq) >= ascii_steps:
+        seq = seq[:ascii_steps]
+    else:
+        seq = seq + [0]*(ascii_steps - len(seq))
+    one_hot = np.zeros((ascii_steps,len(alphabet)+1))
+    one_hot[np.arange(ascii_steps),seq] = 1
+    return one_hot
+
+# abstraction for logging
+class Logger():
+    def __init__(self, FLAGS):
+        self.logf = '{}train_scribe.txt'.format(FLAGS.log_dir) if FLAGS.train else '{}sample_scribe.txt'.format(FLAGS.log_dir)
+        with open(self.logf, 'w') as f: f.write("Scribe: Realistic Handriting in Tensorflow\n     by Sam Greydanus\n\n\n")
+
+    def write(self, s, print_it=True):
+        if print_it:
+            print s
+        with open(self.logf, 'a') as f:
+            f.write(s + "\n")

repeat-copy/rnn_models/checkpoint

@@ -1,2 +1,4 @@
-model_checkpoint_path: "model.ckpt-7000"
+model_checkpoint_path: "model.ckpt-10000"
-all_model_checkpoint_paths: "model.ckpt-7000"
+all_model_checkpoint_paths: "model.ckpt-8000"
+all_model_checkpoint_paths: "model.ckpt-9000"
+all_model_checkpoint_paths: "model.ckpt-10000"