def forward(self, input):
# Create a placeholder for videos.
k = self.batch_size
inputs = tf.placeholder(tf.float32, [self.batch_size, self.timesteps] +
[] + [channels])
cell = ConvLSTMCell(self.shape, self.filters, self.kernel)
outputs, state = tf.nn.dynamic_rnn(cell, inputs, dtype=inputs.dtype)
# There's also a ConvGRUCell that is more memory efficient.
from cell import ConvGRUCell
cell = ConvGRUCell(shape, filters, kernel)
outputs, state = tf.nn.dynamic_rnn(cell, inputs, dtype=inputs.dtype)
# It's also possible to enter 2D input or 4D input instead of 3D.
shape = [100]
kernel = [3]
inputs = tf.placeholder(tf.float32,
[batch_size, timesteps] + shape + [channels])
cell = ConvLSTMCell(shape, filters, kernel)
outputs, state = tf.nn.bidirectional_dynamic_rnn(cell,
cell,
inputs,
dtype=inputs.dtype)
shape = [50, 50, 50]
kernel = [1, 3, 5]
inputs = tf.placeholder(tf.float32,
[batch_size, timesteps] + shape + [channels])
cell = ConvGRUCell(shape, filters, kernel)
outputs, state = tf.nn.bidirectional_dynamic_rnn(cell,
cell,
inputs,
dtype=inputs.dtype)
def t2():
# There's also a ConvGRUCell that is more memory efficient.
cell = ConvGRUCell(shape, filters, kernel)
outputs, state = tf.nn.dynamic_rnn(cell, inputs, dtype=inputs.dtype)
with tf.Session() as sess:
inp = np.random.normal(size=(batch_size, timesteps, height, width,
channels))
sess.run(tf.global_variables_initializer())
output, cell_and_hidden_state = sess.run([outputs, state],
feed_dict={inputs: inp})
print("output shape:", output.shape
) # output:[time, batch_size, height, width, width, num_filter]
print("cell_and_hidden_state:", cell_and_hidden_state)
def t4():
shape = [50, 50, 50]
kernel = [1, 3, 5]
inputs = tf.placeholder(tf.float32,
[batch_size, timesteps] + shape + [channels])
cell = ConvGRUCell(shape, filters, kernel)
outputs, state = tf.nn.bidirectional_dynamic_rnn(cell,
cell,
inputs,
dtype=inputs.dtype)
with tf.Session() as sess:
inp = np.random.normal(size=(batch_size, timesteps, height, width,
channels))
sess.run(tf.global_variables_initializer())
output, cell_and_hidden_state = sess.run([outputs, state],
feed_dict={inputs: inp})
print("output shape:", output.shape
) # output:[time, batch_size, height, width, width, num_filter]
print("cell_and_hidden_state:", cell_and_hidden_state)
inputs = tf.placeholder(tf.float32, [batch_size, timesteps] + shape + [channels])
#print(inputs)
# Make placeholder batch major again after RNN. (see https://github.com/tensorflow/tensorflow/pull/5142)
inputs = tf.transpose(inputs, (1, 0, 2, 3, 4))
#print(inputs)
# There's also a ConvGRUCell that is more memory efficient.
#from cell import ConvGRUCell
#with tf.variable_scope('cell_0'):
# def add_cell(shapes, filters, kernels, initializer, reuse=None):
# return ConvGRUCell(shapes, filters, kernels, initializer=initializer, reuse=reuse)
with tf.variable_scope('cell_0'):
cell_0 = ConvGRUCell(shape, filter, kernel, initializer=tf.truncated_normal_initializer(stddev=0.01))
#print(cell_0)
outputs, state = tf.nn.dynamic_rnn(cell_0, inputs, initial_state=cell_0.zero_state(inputs.shape[1].value, dtype=tf.float32), dtype=inputs.dtype, time_major=True)
outputs = tf.concat([outputs, outputs], 2)
#print(outputs)
#
#
with tf.variable_scope('cell_1'):
cell_1 = ConvGRUCell(shape, 3, kernel_1, initializer=tf.truncated_normal_initializer(stddev=0.01))
outputs_1, state_1 = tf.nn.dynamic_rnn(cell_1, outputs, initial_state=cell_1.zero_state(outputs.shape[1].value, dtype=tf.float32), dtype=inputs.dtype, time_major=True)
with tf.variable_scope('cell_1', reuse=True):
cell_3 = ConvGRUCell(shape, 3, kernel_1)#, reuse=True)
#print(cell_1)
outputs_3, state_3 = tf.nn.dynamic_rnn(cell_3, outputs, initial_state=state_1, dtype=inputs.dtype, time_major=True)
print(outputs_1)
kernel = [3, 3]
channels = 3
filters = 12
# Create a placeholder for videos.
inputs = tf.placeholder(tf.float32,
[batch_size, timesteps] + shape + [channels])
# Add the ConvLSTM step.
from cell import ConvLSTMCell
cell = ConvLSTMCell(shape, filters, kernel)
outputs, state = tf.nn.dynamic_rnn(cell, inputs, dtype=inputs.dtype)
# There's also a ConvGRUCell that is more memory efficient.
from cell import ConvGRUCell
cell = ConvGRUCell(shape, filters, kernel)
outputs, state = tf.nn.dynamic_rnn(cell, inputs, dtype=inputs.dtype)
# It's also possible to enter 2D input or 4D input instead of 3D.
shape = [100]
kernel = [3]
inputs = tf.placeholder(tf.float32,
[batch_size, timesteps] + shape + [channels])
cell = ConvLSTMCell(shape, filters, kernel)
outputs, state = tf.nn.bidirectional_dynamic_rnn(cell,
cell,
inputs,
dtype=inputs.dtype)
shape = [50, 50, 50]
kernel = [1, 3, 5]
height = 100 channels = 512 filters = 3 kernel = [3, 3] slice_dimension=3 # Create a placeholder for videos. inputs = tf.placeholder(tf.float32, [batch_size, timesteps, width, height, channels]) y_slice=tf.placeholder(tf.int32,[batch_size,timesteps,slice_dimension]) output_y=tf.placeholder(tf.int32,[batch_size,timesteps,width,height,channels]) # Flatten input because tf.nn.dynamic_rnn only accepts 3D input. inputs = flatten(inputs) # Add the ConvLSTM step. cell = ConvGRUCell(height, width, filters, kernel) outputs, state = tf.nn.dynamic_rnn(cell, inputs, dtype=inputs.dtype) # Reshape outputs to videos again, because tf.nn.dynamic_rnn only accepts 3D input. outputs = expand(outputs, height, width, filters) for o_batch,y_batch in zip(outputs,y_slice): for o_time,y_time in zip(o_batch,y_batch): unpack_outputs=tf.unstack(outputs) unpack_y_slice=tf.unstack(y_slice) out1=tf.slice(outputs[0][0],y_slice[0][0],[10,10,3]) out2=tf.slice(outputs[0][2],y_slice[0][2],[10,10,3]) out4=tf.slice(outputs[0][3],y_slice[0][3],[10,10,3])