/
run_individual_courses.py
309 lines (263 loc) · 11.7 KB
/
run_individual_courses.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
import tensorflow as tf
import common
import util
import numpy as np
import pandas
import sklearn.metrics
NUM_EPOCHS = 10000
BATCH_SIZE = 100
def mySolveLS (A, b):
return np.linalg.solve(A.T.dot(A) + np.eye(A.shape[1])*ALPHA, A.T.dot(b))
def optBilinearLS (X, y, numHidden):
yhat = np.zeros_like(y)
lastRMSE = float('inf')
TOLERANCE = 1e-3
L = np.random.random((X.shape[1], numHidden))
p = np.random.random(numHidden)
RMSE = calcRMSE(y, yhat)
while np.abs(lastRMSE - RMSE) > TOLERANCE:
# Step 1
p = mySolveLS(X.dot(L), y)
# Step 2
Lvec = mySolveLS(np.kron(p, X), y)
L = Lvec.reshape((L.shape[1], L.shape[0])).T
yhat = X.dot(L).dot(p)
lastRMSE = RMSE
RMSE = calcRMSE(y, yhat)
print RMSE
return L, p
def makeVariable (shape, stddev, wd, name, collectionNames = [""]):
var = tf.Variable(tf.random_normal(shape, stddev=stddev), name=name)
weight_decay = tf.mul(tf.nn.l2_loss(var), wd)
# Caller may wish to add to multiple collections
for collectionName in collectionNames:
tf.add_to_collection("losses{}".format(collectionName), weight_decay)
return var
def runMLR (train_x, train_y, test_x, test_y, numEpochs = NUM_EPOCHS):
print "MLR"
with tf.Graph().as_default():
session = tf.InteractiveSession()
x = tf.placeholder("float", shape=[None, train_x.shape[1]])
y_ = tf.placeholder("float", shape=[None, train_y.shape[1]])
W1 = makeVariable([train_x.shape[1],numHidden], stddev=0.5, wd=1e0)
b1 = makeVariable([train_y.shape[1]], stddev=0.5, wd=1e0)
y = tf.nn.softmax(tf.matmul(x,W1) + b1)
cross_entropy = -tf.reduce_sum(y_*tf.log(tf.clip_by_value(y,1e-10,1.0)), name='cross_entropy')
tf.add_to_collection('losses', cross_entropy)
total_loss = tf.add_n(tf.get_collection('losses'), name='total_loss')
train_step = tf.train.GradientDescentOptimizer(learning_rate=.001).minimize(total_loss)
session.run(tf.initialize_all_variables())
for i in range(numEpochs):
offset = i*BATCH_SIZE % (train_x.shape[0] - BATCH_SIZE)
train_step.run({x: train_x[offset:offset+BATCH_SIZE, :], y_: train_y[offset:offset+BATCH_SIZE, :]})
if i % 500 == 0:
util.showProgress(cross_entropy, x, y, y_, test_x, test_y)
session.close()
def runNN (train_x, train_y, test_x, test_y, numHidden, numEpochs = NUM_EPOCHS):
print "NN({})".format(numHidden)
with tf.Graph().as_default():
session = tf.InteractiveSession()
x = tf.placeholder("float", shape=[None, train_x.shape[1]])
y_ = tf.placeholder("float", shape=[None, train_y.shape[1]])
W1 = makeVariable([train_x.shape[1],numHidden], stddev=0.5, wd=1e1, name="W1")
b1 = makeVariable([numHidden], stddev=0.5, wd=1e1, name="b1")
W2 = makeVariable([numHidden,train_y.shape[1]], stddev=0.5, wd=1e0, name="W2")
#level1 = tf.nn.relu(tf.matmul(x,W1) + b1)
level1 = tf.matmul(x,W1) + b1
y = tf.nn.softmax(tf.matmul(level1,W2))
cross_entropy = -tf.reduce_mean(y_*tf.log(tf.clip_by_value(y,1e-10,1.0)), name='cross_entropy')
tf.add_to_collection('losses', cross_entropy)
total_loss = tf.add_n(tf.get_collection('losses'), name='total_loss')
train_step = tf.train.MomentumOptimizer(learning_rate=.001, momentum=0.01).minimize(total_loss)
#train_step = tf.train.AdamOptimizer(learning_rate=.001).minimize(total_loss)
session.run(tf.initialize_all_variables())
for i in range(numEpochs):
offset = i*BATCH_SIZE % (train_x.shape[0] - BATCH_SIZE)
train_step.run({x: train_x[offset:offset+BATCH_SIZE, :], y_: train_y[offset:offset+BATCH_SIZE, :]})
if i % 500 == 0:
util.showProgress(cross_entropy, x, y, y_, test_x, test_y)
session.close()
def loadDataset (filename, courseId, T, requiredCols = None):
d = pandas.io.parsers.read_csv(filename)
d.start_time = d.start_time.astype(np.datetime64)
# Only analyze rows belonging to users who participated in the courseId after T
afterIdxs = np.nonzero(d.start_time >= T)[0]
e = d.iloc[afterIdxs]
idxs = np.nonzero(e.course_id == courseId)[0]
userIds = set(e.iloc[idxs].user_id)
d = d.iloc[np.nonzero(d.user_id.isin(userIds))[0]]
# Compute labels (i.e., whether each user explored the specified course)
userIdsLabelsMap = dict(zip(e.user_id, e.explored))
# Only analyze rows occurring before T
idxs = np.nonzero(d.start_time < T)[0]
d = d.iloc[idxs]
# Assemble demographics matrix
demographics = d.drop_duplicates("user_id")
y = demographics.user_id.map(userIdsLabelsMap).as_matrix()
demographics = demographics[common.DEMOGRAPHIC_FIELDS + [ "user_id" ]]
demographics = pandas.get_dummies(demographics, columns=["continent", "LoE", "gender"])
demographics = demographics.sort("user_id")
# Assemble prior course matrix
nDaysAct = d.ndays_act
courses = pandas.get_dummies(d[['course_id', 'user_id']], columns=["course_id"])
courseCols = [ col for col in courses.columns if "course_id" in col ]
courses[courseCols] *= np.tile(np.atleast_2d(nDaysAct).T, (1, len(courseCols))) # Multiply course indicator vars by ndays_act
courses = courses.sort("user_id")
# Aggregate within each user
courses = courses.groupby(['user_id']).sum()
# Combine demographics and course matrices into x
courses = courses.reset_index()
demographics = demographics.reset_index()
x = demographics # DEMOGRAPHICS ONLY
#x = pandas.concat([ demographics, courses ], axis=1) # ALL FEATURES
x = x.drop('index', 1)
x = x.drop('user_id', 1)
if requiredCols != None:
for colName in x.columns:
if colName not in requiredCols:
x = x.drop(colName, 1)
for colName in requiredCols:
if colName not in x.columns:
x[colName] = pandas.Series(np.zeros(x.shape[0]), index=x.index)
x = x.reindex_axis(sorted(x.columns), axis=1) # Sort by column names
return x.as_matrix(), util.makeLabels(y), x.columns
def computeT (courseId):
# Find the 1st percentile, over all .csv files, of the start_time of the specified course
startTimes = []
for filenameRoot in [ "train", "test", "holdout" ]:
filename = "{}_individual.csv".format(filenameRoot)
d = pandas.io.parsers.read_csv(filename)
idxs = np.nonzero(d.course_id == courseId)[0]
d = d.iloc[idxs]
startTimes += list(d.start_time.astype(np.datetime64))
START_T = np.sort(startTimes)[int(0.01 * len(startTimes))] # Use 1st percentile als proxy for course start date
T = START_T + np.timedelta64(28, 'D') # Add 4 weeks
return T
def initializeAllData (courseId):
T = computeT(courseId)
train_x, train_y, colNames = loadDataset("train_individual.csv", courseId, T)
test_x, test_y, _ = loadDataset("test_individual.csv", courseId, T, list(colNames))
# Normalize data
normalize = True
if normalize:
mx = np.mean(train_x, axis=0)
sx = np.std(train_x, axis=0)
sx[sx == 0] = 1
train_x -= np.tile(np.atleast_2d(mx), (train_x.shape[0], 1))
train_x /= np.tile(np.atleast_2d(sx), (train_x.shape[0], 1))
# Scale testing data using parameters estimated on training set
test_x -= np.tile(np.atleast_2d(mx), (test_x.shape[0], 1))
test_x /= np.tile(np.atleast_2d(sx), (test_x.shape[0], 1))
return train_x, train_y, test_x, test_y, colNames
def runLLL_NN (all_train_x, all_train_y, all_test_x, all_test_y, numHidden, courseIds):
n = len(courseIds)
with tf.Graph().as_default():
session = tf.InteractiveSession()
# Initialize all variables
xs = []
ys_ = []
collectionNames = ["losses_{}".format(i) for i in range(n) ]
for i in range(n):
xs.append(tf.placeholder("float", shape=[None, all_train_x[i].shape[1]]))
ys_.append(tf.placeholder("float", shape=[None, all_train_y[i].shape[1]]))
W1 = makeVariable([all_train_x[0].shape[1],numHidden], stddev=0.5, wd=1e1, name="W1", collectionNames=collectionNames)
b1 = makeVariable([numHidden], stddev=0.5, wd=1e1, name="b1", collectionNames=collectionNames)
W2s = []
level1s = []
level2s = []
ys = []
for i in range(n):
level1s.append(tf.matmul(xs[i],W1) + b1)
W2s.append(makeVariable([numHidden,all_train_y[i].shape[1]], stddev=0.5, wd=1e1, name="W2_{}".format(i), collectionNames=["losses_{}".format(i)]))
level2s.append(tf.matmul(level1s[i],W2s[i]))
#level2s.append(tf.matmul(tf.nn.relu(level1s[i]), W2s[i]))
ys.append(tf.nn.softmax(level2s[i]))
# Initialize loss functions
cross_entropies = []
total_losses = []
optimizers = []
global_steps = []
for i in range(n):
cross_entropies.append(-tf.reduce_mean(ys_[i]*tf.log(tf.clip_by_value(ys[i],1e-10,1.0)), name="cross_entropy_{}".format(i)))
tf.add_to_collection("losses_{}".format(i), cross_entropies[i])
total_losses.append(tf.add_n(tf.get_collection("losses_{}".format(i)), name="total_losses_{}".format(i)))
global_steps.append(tf.Variable(0, trainable=False))
learning_rate = tf.train.exponential_decay(0.001, global_steps[i], 1000, 0.98)
optimizers.append(tf.train.MomentumOptimizer(learning_rate, momentum=0.1).minimize(total_losses[i]))
session.run(tf.initialize_all_variables())
for i in range(NUM_EPOCHS):
if i % 500 == 0:
print "..."
for j in range(n):
offset = i*BATCH_SIZE % (all_train_x[j].shape[0] - BATCH_SIZE)
optimizers[j].run({xs[j]: all_train_x[j][offset:offset+BATCH_SIZE, :], ys_[j]: all_train_y[j][offset:offset+BATCH_SIZE, :]})
if i % 500 == 0:
util.showProgress(total_losses[j], xs[j], ys[j], ys_[j], all_test_x[j], all_test_y[j])
#util.showProgress(cross_entropies[j], xs[j], ys[j], ys_[j], all_train_x[j], all_train_y[j])
session.close()
def normalize (x, mx, sx):
return (x - np.tile(np.atleast_2d(mx), (x.shape[0], 1))) / np.tile(np.atleast_2d(sx), (x.shape[0], 1))
def loadLLLData ():
if 'all_train_x' not in globals():
# Get list of all course_id's
d = pandas.io.parsers.read_csv("train_individual.csv")
global courseIds
courseIds = []
# Pick some courses
for courseId in np.unique(d.course_id):
if len(np.nonzero(d.course_id == courseId)[0]) > 10000 and\
"CS50" not in courseId and\
"2016" not in courseId:
courseIds.append(courseId)
courseIds = courseIds[0:30]
# Initialize training and testing matrices
global all_train_x, all_train_y, all_test_x, all_test_y
all_train_x = []
all_train_y = []
all_test_x = []
all_test_y = []
_, _, colNames = loadDataset("train_individual.csv", courseIds[0], computeT(courseIds[0]))
# Collect training and testing data for all courses
print "Loading data for..."
allTrainX = 0
for i, courseId in enumerate(courseIds):
print courseId
T = computeT(courseId)
train_x, train_y, _ = loadDataset("train_individual.csv", courseId, T, list(colNames))
test_x, test_y, _ = loadDataset("test_individual.csv", courseId, T, list(colNames))
# Collect all training features so we can do mean/variance normalization
if type(allTrainX) == int:
allTrainX = train_x
allTrainX = np.vstack((allTrainX, train_x))
all_train_x.append(train_x)
all_train_y.append(train_y)
all_test_x.append(test_x)
all_test_y.append(test_y)
# Normalize all data
mx = np.mean(allTrainX, axis=0)
sx = np.std(allTrainX, axis=0)
sx[sx == 0] = 1
for i in range(len(all_train_x)):
all_train_x[i] = normalize(all_train_x[i], mx, sx)
all_test_x[i] = normalize(all_test_x[i], mx, sx)
return all_train_x, all_train_y, all_test_x, all_test_y, courseIds
# Life-long learning (LLL) experiment
def runLLLExperiments ():
all_train_x, all_train_y, all_test_x, all_test_y, courseIds = loadLLLData()
for numHidden in range(3, 6):
print "numHidden = {}".format(numHidden)
#print "LLL:"
#runLLL_NN(all_train_x, all_train_y, all_test_x, all_test_y, numHidden, courseIds)
print "Regular NN:"
for j in range(len(courseIds)):
runNN(all_train_x[j], all_train_y[j], all_test_x[j], all_test_y[j], numHidden)
def runNNExperiments ():
#COURSE_ID = "HarvardX/SW12.5x/2T2014"
COURSE_ID = "HarvardX/ER22.1x/1T2014"
if 'train_x' not in globals():
train_x, train_y, test_x, test_y, colNames = initializeAllData(COURSE_ID)
for numHidden in range(2, 20, 2):
runNN(train_x, train_y, test_x, test_y, numHidden)
if __name__ == "__main__":
runLLLExperiments()
#runNNExperiments()