AttributeError: 'ConvLSTMLayer' object has no attribute 'initial_state'?

[dengyueyun666]

In ConvLSTMLayer, there is a parameter named 'initial_state'.
When initial_state = None, it works.
When initial_state = previous_ConvLSTMLayer.final_state, it reports this error.

I think the source code in ConvLSTMLayer does not handle well the situation when initial_state is not None

[DEKHTIARJonathan]

Can you give a simple snippet to reproduce?

[DEKHTIARJonathan]

Please fix your code and keep an appropriate formating.
It is impossible to run your code.

• imports are missing
• vars are missing
• conv2d(...) this is not code
• etc.

If you want some help, you need to make sure that we can easily reproduce your problem

Make something simple and minimal that I can execute with python main.py

[dengyueyun666]

import tensorflow as tf
import tensorlayer as tl
from tensorlayer.layers import *

config = tf.ConfigProto(log_device_placement=True, allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)

train_x = tf.random_uniform(shape=(10, 32, 32, 3))
train_y = tf.random_uniform(shape=(10, 32, 32, 1))

ConvLSTM_last_state = None

net_in = InputLayer(train_x, name='inputs')

net = Conv2d(net_in, 16, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv1')
net = ReshapeLayer(net, shape=(10, 1, 32, 32, 16))
net = ConvLSTMLayer(net, cell_shape=(32, 32), feature_map=16, filter_size=(3, 3), cell_fn=BasicConvLSTMCell,
                    n_steps=1, initial_state=ConvLSTM_last_state, return_last=False, return_seq_2d=False)

ConvLSTM_last_state = net.final_state


net = ReshapeLayer(net, shape=(10, 1, 32, 32, 16))
net = ConvLSTMLayer(net, cell_shape=(32, 32), feature_map=16, filter_size=(3, 3), cell_fn=BasicConvLSTMCell,
                    n_steps=1, initial_state=ConvLSTM_last_state, return_last=False, return_seq_2d=False)

net = ReshapeLayer(net, shape=(10, 32, 32, -1))
net = Conv2d(net_in, 1, (1, 1), (1, 1), act=tf.nn.relu, padding='SAME', name='output')

output = net.outputs

loss = tl.cost.mean_squared_error(output, train_y, is_mean=True)
train_params = net.all_params
train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss, var_list=train_params)

tl.layers.initialize_global_variables(sess)

for epoch in range(100):
    l = sess.run(loss)
    print(l)

sess.close()

[DEKHTIARJonathan]

@zsdonghao @lgarithm @luomai I can reproduce the error. Could one of you have a look why it happens ?
It would be nice to fix it before releasing 1.8.6rc0

Thanks @dengyueyun666 for this bug report, we are investigating this ;) If you think you can fix, you are more than welcome to submit a PR

[zsdonghao]

@vyokky could you have a look on it?

[dengyueyun666]

I have submitted a PR