for d in range(9):
_, filter_result_d, _ = tf.split(relu_output_x, [d + 1, 1, 16 - d - 2],
-1)
filter_sum_d = -1 * tf.reduce_sum(filter_result_d) + tf.reduce_sum(
tf.square(gen_image)) * 2
gen_ops.append(
tf.train.AdamOptimizer(0.0001).minimize(filter_sum_d,
var_list=gen_image))
gen_loss_op.append(filter_sum_d)
# create saver
saver = util.TrainSaver("ckpt_zoo/mnist_cnn_bn/cnn_bn.ckpt", "main")
#%%
# train !
training_grads = util.get_grads(grad_and_var)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver.try_load(sess)
saver.remove_old_ckpt(sess, global_step=global_step)
for i in range(TRAIN_STEPS):
train_images, train_labels = mnist.train.next_batch(BATCH_SIZE)
train_images = train_images.reshape(-1, 28, 28, 1)
_, np_loss = sess.run([train_step, loss],
feed_dict={
inputs: train_images,
labels: train_labels,
is_training: True
})
plt.plot(onnx_predict_values, label='tf', linewidth=4)
plt.plot(tf_predict_values, label='onnx')
plt.legend()
plt.show()
#%%
#获得参数
with tf.Session() as sess:
if False == saver.try_load(sess):
raise Exception("not train?")
train_input_data, train_label_data = sess.run([train_input, train_label])
vars, grads = sess.run(
[util.get_vars(grad_and_var),
util.get_grads(grad_and_var)],
feed_dict={
input: train_input_data,
lable: train_label_data
})
names = util.get_names(grad_and_var)
for v, name in zip(vars, names):
plt.title("var:[%s]" % name)
plt.hist(v.flatten(), bins=50)
plt.show()
print("----------------------")
for g, name in zip(grads, names):
plt.title("grad:[%s]" % name)
plt.hist(g.flatten(), bins=50)
inf_d2_output = act(tf.matmul(inf_d3_output, d2) + db2)
inf_d1_output = tf.nn.sigmoid(tf.matmul(inf_d2_output, d1) + db1)
output = tf.identity(inf_d1_output, 'output')
with tf.name_scope("main_train"):
opt = tf.train.AdamOptimizer(0.01)
grad_and_var = opt.compute_gradients(loss)
train_step = opt.apply_gradients(grad_and_var, global_step=global_step)
# create saver
saver = util.TrainSaver("ckpt_zoo/mnist_auto_encoder/mnist_auto_encoder.ckpt",
"main")
#%%
# train !
training_grads = util.get_grads(grad_and_var)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver.try_load(sess)
saver.remove_old_ckpt(sess, global_step=global_step)
for i in range(TRAIN_STEPS):
train_images, _ = mnist.train.next_batch(BATCH_SIZE)
_, np_loss = sess.run([train_step, loss],
feed_dict={input: train_images})
if i % (TRAIN_STEPS / 10) == 0:
loss_value, step, grad_list = sess.run(
[loss, global_step, training_grads],
feed_dict={input: train_images})
print("After %d training step(s), loss: %g" % (step, loss_value))
gen_image_step = tf.train.AdamOptimizer(0.0001).minimize(filter_sum, var_list=gen_image)
gen_ops.append(gen_image_step)
gen_loss_op.append(filter_sum)
for d in range(2):
_, filter_result_d, _ = tf.split(relu_output_x, [d + 1, 1, 4 - d -2], -1)
filter_sum_d = -1 * tf.reduce_sum(filter_result_d) + tf.reduce_sum(tf.square(gen_image)) * 2
gen_ops.append(tf.train.AdamOptimizer(0.0001).minimize(filter_sum_d, var_list=gen_image))
gen_loss_op.append(filter_sum_d)
# create saver
saver = util.TrainSaver("ckpt_zoo/mnist_cnn/cnn.ckpt", "main")
#%%
# train !
training_grads = util.get_grads(grad_and_var)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver.try_load(sess)
saver.remove_old_ckpt(sess, global_step=global_step)
for i in range(TRAIN_STEPS):
train_images, train_labels = mnist.train.next_batch(BATCH_SIZE)
train_images = train_images.reshape(-1, 28, 28, 1)
_, np_loss = sess.run([train_step, loss], feed_dict={inputs: train_images, labels: train_labels})
if i % (TRAIN_STEPS/10) == 0:
loss_value, step, acc_value, grad_list = sess.run([loss, global_step, acc, training_grads], feed_dict={inputs: train_images, labels: train_labels})
print("After %d training step(s), loss: %g, acc:%f." % (step, loss_value, acc_value))
saver.save(sess, global_step=global_step)