def main():
Ntrain = 20000
# Input data from file
fname_train_image = "./data/train-images.idx3-ubyte"
fname_train_label = "./data/train-labels.idx1-ubyte"
fname_test_image = "./data/t10k-images.idx3-ubyte"
fname_test_label = "./data/t10k-labels.idx1-ubyte"
train_sample, valid_sample = LoadSample(fname_train_image,
fname_train_label,
validate=True)
test_sample = LoadSample(fname_test_image, fname_test_label)
print(train_sample.images.shape)
while input("q: quit, other : continue ") != 'q':
i = np.random.randint(train_sample.images.shape[0])
arr = train_sample.images[i, :]
lab = train_sample.labels[i, :]
for j in range(10):
if lab[j] == 1: print("Number = ", j)
arr = 255 * np.ones([28, 28]) - arr.reshape(28, 28)
plt.gray()
plt.imshow(arr)
plt.show()
i += 1
mnist = base.Datasets(train=train_sample,
validation=valid_sample,
test=test_sample)
model = Model("model")
model.build()
model.define_loss()
train_step = model.Get_train_step()
## save graph
saver = model.Get_Saver()
graph_location = "./logs/"
sum_accuracy = tf.summary.scalar("accuracy", model.accuracy)
sum_xentropy = tf.summary.scalar("cross entropy", model.cross_entropy)
train_writer = tf.summary.FileWriter(graph_location)
train_writer.add_graph(tf.get_default_graph())
print('Saving graph to: %s' % graph_location)
##
with tf.Session() as sess:
init = model.Get_initializer()
sess.run(init)
for i in range(Ntrain):
batch = mnist.train.next_batch(50)
if i % 100 == 0:
feed_dict = model.Get_feed_dict(batch[0], batch[1], 1)
train_accuracy = model.accuracy.eval(feed_dict=feed_dict)
tmp_sum_accuracy = sum_accuracy.eval(feed_dict=feed_dict)
tmp_sum_xentropy = sum_xentropy.eval(feed_dict=feed_dict)
train_writer.add_summary(tmp_sum_accuracy, i)
train_writer.add_summary(tmp_sum_xentropy, i)
print('step %d, training accuracy %g' % (i, train_accuracy))
feed_dict = model.Get_feed_dict(batch[0], batch[1], 0.5)
train_step.run(feed_dict=feed_dict
) # dropout 50% neurons within 1024->10 layer
feed_dict = model.Get_feed_dict(mnist.test.images, mnist.test.labels,
1)
print('test accuracy %g' % model.accuracy.eval(feed_dict=feed_dict))
savePath = saver.save(sess, "./out/Model_out")
train_writer.flush()
train_writer.close()
def main() :
fname_test_image = "./data/t10k-images.idx3-ubyte"
fname_test_label = "./data/t10k-labels.idx1-ubyte"
test_sample = LoadSample(fname_test_image,fname_test_label)
print(test_sample.images.shape)
model = Model("model")
model.build()
model.define_loss()
train_step = model.Get_train_step()
# restore
modeldir = "./out/"
with tf.Session() as sess :
model.Restore(sess,modeldir)
W_conv1 = sess.run(model.weight["conv1"])
W_conv2 = sess.run(model.weight["conv2"])
b_conv1 = sess.run(model.bias["conv1"])
b_conv2 = sess.run(model.bias["conv2"])
x = tf.cast( np.array( [128.]*28*28 ).reshape([1,-1]), tf.float32 );
x = tf.reshape(x,[-1,28,28,1])
filter_1 = tf.nn.relu( tf.nn.conv2d(x, W_conv1, strides=[1,1,1,1],padding="SAME") + b_conv1 )
pool_1 = tf.nn.max_pool( filter_1, ksize=[1,2,2,1], strides=[1,2,2,1], padding="SAME")
print("filter_1: ",filter_1.shape)
print("pool_1 : ",pool_1.shape)
print("Finish loading model\n")
# convolution 1
image_1 = np.zeros([32,5,5])
for i in range(32) :
tmp_image = np.zeros([5,5])
for im in range(100) :
tmp_pict = test_sample.images[im,:].reshape([28,28])
for ix in range(2,26) :
for iy in range(2,26) :
tmp_weight = 0
for ipx in range(-2,3) :
for ipy in range(-2,3) :
tmp_weight = tmp_weight + W_conv1[ipx,ipy,0,i] * tmp_pict[ix+ipx,iy+ipy]
if tmp_weight + b_conv1[i] <= 0: tmp_weight = 0
else : tmp_weight = tmp_weight + b_conv1[i]
tmp_image = tmp_image + tmp_pict[ix-2:ix+3,iy-2:iy+3] * tmp_weight
max_pt = np.max(tmp_image)
min_pt = np.min(tmp_image)
print("conv1 filter %d: max: %.2f min: %.2f" % (i,max_pt,min_pt) )
scale = 1
if max_pt <= 0 :
scale = -1./min_pt
else :
scale = 1./max_pt
tmp_image = tmp_image * scale * 128 + 128
tmp_image = tf.cast( tf.reshape(tmp_image,[5,5]), tf.int32 )
image_1[i,:,:] = sess.run(tmp_image)
fig = plt.imshow(image_1[i,:,:], interpolation='none', cmap=plt.cm.gray_r)
plt.colorbar()
plt.savefig("filter/filter1_"+str(i)+".png")
plt.clf()
# convolution 2
image_2 = np.zeros([64,14,14])
for im in range(100) :
tmp_pict = test_sample.images[im,:].reshape([28,28])
x = tf.reshape(tmp_pict,[-1,28,28,1])
tmp_conv1 = sess.run( tf.nn.relu( tf.nn.conv2d(x, W_conv1, strides=[1,1,1,1],padding="SAME") + b_conv1 ) )
tmp_conv1 = tmp_conv1.reshape([28,28,32])
pool_val = np.zeros([32,14,14])
pool_x = np.zeros([32,14,14])
pool_y = np.zeros([32,14,14])
# get max_pool
for ifilter in range(32) :
for ipoolx in range(14) :
for ipooly in range(14) :
tmp_pool_val = -9999
tmp_pool_x = 0
tmp_pool_y = 0
for tmp_i in range(2) :
for tmp_j in range(2) :
tmp_val = tmp_conv1[ipoolx*2+tmp_i,ipooly*2+tmp_j,ifilter]
if tmp_val > tmp_pool_val :
tmp_pool_val = tmp_val
tmp_pool_x = tmp_i
tmp_pool_y = tmp_j
pool_val[ifilter,ipoolx,ipooly] = tmp_pool_val
pool_x [ifilter,ipoolx,ipooly] = tmp_pool_x
pool_y [ifilter,ipoolx,ipooly] = tmp_pool_y
print("make pool: image ",im)
# convolution 2
for ix in range(2,12) :
for iy in range(2,12) :
tmp_weight = np.zeros([32,64])
tmp_sum = 0
for i in range(64) :
tmp_image = np.zeros([14,14])
for ifilter in range(32) :
Wx = np.sum( (W_conv2[:,:,ifilter,i] * pool_val[ifilter,ix+-2:ix+3,iy-2:iy+3]) )
tmp_sum += Wx
tmp_weight[ifilter,i] += Wx
if tmp_sum + b_conv2[i] <= 0 : continue
for ipx in range(-2,3) :
for ipy in range(-2,3) :
for ifilter in range(32) :
tmp_px = int( 4+ipx*2+pool_x[ifilter,ix+ipx,iy+ipy] )
tmp_py = int( 4+ipy*2+pool_y[ifilter,ix+ipx,iy+ipy] )
tmp_image[tmp_px:tmp_px+5,tmp_py:tmp_py+5] += image_1[ifilter,:,:].astype(float) * tmp_weight[ifilter,i]
image_2[i,:,:] += tmp_image
for i in range(64) :
tmp_image = image_2[i,:,:]
max_pt = np.max(tmp_image)
min_pt = np.min(tmp_image)
print("conv2 filter %d: max: %.2f min: %.2f" % (i,max_pt,min_pt) )
scale = 1
if max_pt < 0 :
scale = -1./min_pt
elif min_pt > 0 :
scale = 1./max_pt
else : continue
tmp_image = tmp_image * scale * 128 + 128
tmp_image = tf.cast( tmp_image, tf.int32 )
image_2[i,:,:] = sess.run(tmp_image)
fig = plt.imshow(image_2[i,:,:], interpolation='none', cmap=plt.cm.gray_r)
plt.colorbar()
plt.savefig("filter/filter2_"+str(i)+".png")
plt.clf()
