validSummary = tf.summary.scalar("test error", validErrorPH)
tfSummaryWriter = tf.summary.FileWriter("summary_{1}/{0}".format(
saveFname, suffix))
resumeIterN = 0
maxIterN = 30000
with tf.Session(config=tfConfig) as sess:
if resumeIterN == 0:
sess.run(tf.global_variables_initializer())
else:
tfSaver.restore(
sess,
"models_{2}/{0}_it{1}k.ckpt".format(saveFname, resumeIterN // 1000,
suffix))
print("resuming from iteration {0}...".format(resumeIterN))
tfSummaryWriter.add_graph(sess.graph)
params.baseLRST = 0.0001
# training loop
for i in range(resumeIterN, maxIterN):
currLearningRate = params.baseLRST, params.baseLR # this can be modified to be scheduled learning rates
randIdx = np.random.randint(trainN, size=[params.batchSize])
trainBatch = {
imageRawBatch: trainData["image"][randIdx],
labelBatch: trainData["label"][randIdx],
learningRate: currLearningRate
}
# run one step
_, trainBatchLoss, summary = sess.run([trainStep, loss, lossSummary],
feed_dict=trainBatch)
if (i + 1) % 10 == 0:
tfSummaryWriter.add_summary(summary, i + 1)
if (i + 1) % 100 == 0:
