Key Error on Tensorflow in Style transfer problem - tensorflow

def evaluate(encoder, decoder, in_lang, max_length=MAX_LENGTH):
if use_cuda:
in_lang = in_lang.cuda()
input_variable = Variable(in_lang)
input_variable = input_variable.unsqueeze(0)
input_length = input_variable.size(1)
encoder_hidden = encoder.initHidden()
encoder_outputs = Variable(torch.zeros(max_length, encoder.hidden_size))
encoder_outputs = encoder_outputs.cuda() if use_cuda else encoder_outputs
for ei in range(input_length):
encoder_output, encoder_hidden = encoder(input_variable[:, ei],
encoder_hidden)
encoder_outputs[ei] = encoder_output[0][0]
decoder_input = Variable(torch.LongTensor([[SOS_token]])) # SOS
decoder_input = decoder_input.cuda() if use_cuda else decoder_input
decoder_hidden = encoder_hidden
decoded_words = []
decoder_attentions = torch.zeros(max_length, max_length)
if use_attn:
for di in range(max_length):
decoder_output, decoder_hidden, decoder_attention = decoder(
decoder_input, decoder_hidden, encoder_outputs)
decoder_attentions[di] = decoder_attention.data
topv, topi = decoder_output.data.topk(1)
ni = topi[0][0]
if ni == EOS_token:
decoded_words.append('<EOS>')
break
else:
decoded_words.append(lang_dataset.output_lang.index2word[ni])
decoder_input = Variable(torch.LongTensor([[ni]]))
decoder_input = decoder_input.cuda() if use_cuda else decoder_input
else:
for di in range(max_length):
decoder_output, decoder_hidden = decoder(decoder_input,
decoder_hidden)
topv, topi = decoder_output.data.topk(1)
ni = topi[0][0]
if ni == EOS_token:
decoded_words.append('<EOS>')
break
else:
decoded_words.append(lang_dataset.output_lang.index2word[ni])
decoder_input = Variable(torch.LongTensor([[ni]]))
decoder_input = decoder_input.cuda() if use_cuda else decoder_input
if use_attn:
return decoded_words, decoder_attentions[:di + 1]
else:
return decoded_words
def evaluateRandomly(encoder, decoder, n=100):
for i in range(n):
pair_idx = random.choice(list(range(len(lang_dataset))))
pair = lang_dataset.pairs[pair_idx]
in_lang, out_lang = lang_dataset[pair_idx]
print('>', pair[0])
print('=', pair[1])
if use_attn:
output_words, attentions = evaluate(encoder, decoder, in_lang)
else:
output_words = evaluate(encoder, decoder, in_lang)
output_sentence = ' '.join(output_words)
print('<', output_sentence)
print('')
I was training a seq2seq model using a parallel corpus and in the last few steps of evaluation, I get a key error. I am not that proficient in tensorflow and would love some advice.
Thank you
I was trying to do formality style transfer on a toy txt file. We were implementing code from a github repo to see if it works, and we have been dodging errors by some previous knowledge of machine learning. But this one error got me stumped and it seems to be intrinsically about tensorflow working. We were expecting to get this
< output sentence

Related

DQN is not training well

import tensorflow as tf
import keras
import numpy as np
import gym
import random
from keras.layers import *
model = keras.models.Sequential()
model.add(Dense(12,activation = 'tanh',input_shape = (4,)))
model.add(Dense(2,activation = 'linear'))
model.compile(optimizer = 'adam',loss = 'MSE',metrics = ['accuracy'])
env = gym.make("CartPole-v1")
def preprocessing(state):
return np.reshape(state,(1,4))
replay_mem = []
replay_size = 32
reward_list = []
local_mean = list()
for episode in range(2000):
done = False
state = env.reset()
count = 0
reward_tot = 0
model.save("cartpole-dqn.h5")
while not done:
count += 1
e = (1/(episode/200+1))
#epslion greedy search
Q = model.predict(preprocessing(state))
if e>np.random.rand(1):
action = env.action_space.sample()
else:
action = np.argmax(Q)
#take_action and set reward
state_next, reward, done, info = env.step(action)
if done:
reward = - 100
#replay_mem
replay_mem.append([state,action,reward,state_next,done])
if len(replay_mem)>2048:
del replay_mem[0]
state = state_next
reward_tot += reward
state = state_next
Q_list = []
state_list = []
#set this_replay_Size
if len(replay_mem)<replay_size:
this_replay_size = len(replay_mem)
else:
this_replay_size = replay_size
#sample random batch from replay_memory
for sample in random.sample(replay_mem,this_replay_size):
state_m,action_m,reward_m,state_next_m,done_m = sample
if done:
Q[0,action] = reward_m
else:
Q_new = model.predict(preprocessing(state_next_m))
Q[0,action] = reward_m + 0.97*np.max(Q_new)
Q_list.append(Q)
state_list.append(state_m)
#convert to nupmy array and train
Q_list = np.array(Q_list)
state_list = np.array(state_list)
hist = model.fit(state_list,Q_list,epochs = 5,verbose = 0)
#print("Done :",done," Reward :",reward," Reward_total :",reward_tot)
local_mean.append(reward_tot)
reward_list.append(reward_tot)
if episode%10 == 0:
print("Episode :",episode+1," Reward_total :", reward_tot," Reward_local_mean :",np.mean(local_mean))
local_mean = list()
print("*******End Learning")
this is my full code of my model
i implement dqn algorithm
what's wrong with my code? i've train this model abbout 1000episode but there no progress
what should i change?
should i train more episode?
or is there anything wrong with my implementation?
i've been working on this project soo long help me please

Triplet-Loss using pre-trained network

I am trying to use the Triple-Loss technique to fine-tune an EfficientNet network for human Re-ID using Keras. Here is the code I am using:
This is the generator:
class SampleGen(object):
def __init__(self, file_class_mapping):
self.file_class_mapping = file_class_mapping
self.class_to_list_files = defaultdict(list)
self.list_all_files = list(file_class_mapping.keys())
self.range_all_files = list(range(len(self.list_all_files)))
for file, class_ in file_class_mapping.items():
self.class_to_list_files[class_].append(file)
self.list_classes = list(set(self.file_class_mapping.values()))
self.range_list_classes = range(len(self.list_classes))
self.class_weight = np.array([len(self.class_to_list_files[class_]) for class_ in self.list_classes])
self.class_weight = self.class_weight / np.sum(self.class_weight)
def get_sample(self):
class_idx = np.random.choice(self.range_list_classes, 1, p=self.class_weight)[0]
examples_class_idx = np.random.choice(range(len(self.class_to_list_files[self.list_classes[class_idx]])), 2)
positive_example_1, positive_example_2 = \
self.class_to_list_files[self.list_classes[class_idx]][examples_class_idx[0]], \
self.class_to_list_files[self.list_classes[class_idx]][examples_class_idx[1]]
negative_example = None
while negative_example is None or self.file_class_mapping[negative_example] == \
self.file_class_mapping[positive_example_1]:
negative_example_idx = np.random.choice(self.range_all_files, 1)[0]
negative_example = self.list_all_files[negative_example_idx]
return positive_example_1, negative_example, positive_example_2
def read_and_resize(filepath):
im = Image.open((filepath)).convert('RGB')
im = im.resize((image_size, image_size))
return np.array(im, dtype="float32")
def augment(im_array):
if np.random.uniform(0, 1) > 0.9:
im_array = np.fliplr(im_array)
return im_array
def gen(triplet_gen):
while True:
list_positive_examples_1 = []
list_negative_examples = []
list_positive_examples_2 = []
for i in range(batch_size):
positive_example_1, negative_example, positive_example_2 = triplet_gen.get_sample()
path_pos1 = join(path_train, positive_example_1)
path_neg = join(path_train, negative_example)
path_pos2 = join(path_train, positive_example_2)
positive_example_1_img = read_and_resize(path_pos1)
negative_example_img = read_and_resize(path_neg)
positive_example_2_img = read_and_resize(path_pos2)
positive_example_1_img = augment(positive_example_1_img)
negative_example_img = augment(negative_example_img)
positive_example_2_img = augment(positive_example_2_img)
list_positive_examples_1.append(positive_example_1_img)
list_negative_examples.append(negative_example_img)
list_positive_examples_2.append(positive_example_2_img)
A = preprocess_input(np.array(list_positive_examples_1))
B = preprocess_input(np.array(list_positive_examples_2))
C = preprocess_input(np.array(list_negative_examples))
label = None
yield {'anchor_input': A, 'positive_input': B, 'negative_input': C}, label
This is how I create the model:
def get_model():
base_model = efn.EfficientNetB3(weights='imagenet', include_top=False)
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
x = Dropout(0.6)(x)
x = Dense(embedding_dim)(x)
x = Lambda(lambda x: K.l2_normalize(x, axis=1), name="enc_out")(x)
embedding_model = Model(base_model.input, x, name="embedding")
input_shape = (image_size, image_size, 3)
anchor_input = Input(input_shape, name='anchor_input')
positive_input = Input(input_shape, name='positive_input')
negative_input = Input(input_shape, name='negative_input')
anchor_embedding = embedding_model(anchor_input)
positive_embedding = embedding_model(positive_input)
negative_embedding = embedding_model(negative_input)
inputs = [anchor_input, positive_input, negative_input]
outputs = [anchor_embedding, positive_embedding, negative_embedding]
triplet_model = Model(inputs, outputs)
triplet_model.add_loss(K.mean(triplet_loss(outputs)))
return embedding_model, triplet_model
And this is how I'm trying to run the training:
if __name__ == '__main__':
data = pd.read_csv(path_csv)
train, test = train_test_split(data, train_size=0.7, random_state=1337)
file_id_mapping_train = {k: v for k, v in zip(train.Image.values, train.Id.values)}
file_id_mapping_test = {k: v for k, v in zip(test.Image.values, test.Id.values)}
gen_tr = gen(SampleGen(file_id_mapping_train))
gen_te = gen(SampleGen(file_id_mapping_test))
embedding_model, triplet_model = get_model()
for i, layer in enumerate(embedding_model.layers):
print(i, layer.name, layer.trainable)
for layer in embedding_model.layers[379:]:
layer.trainable = True
for layer in embedding_model.layers[:379]:
layer.trainable = False
triplet_model.compile(loss=None, optimizer=Adam(0.0001))
history = triplet_model.fit(x=gen_tr,
validation_data=gen_te,
epochs=10,
verbose=1,
steps_per_epoch=200,
validation_steps=20,
callbacks=create_callbacks())
The csv contains two columns (Image, Id) and I am generating triplets on the go using a generator. The layer 379 is the last layer of the network so I just leave that as trainable. I let it run for some epochs and it seems like it doesn't converge, it stays around 2.30. On epochs like 20, the loss is even higher than what I've started with. Here you can see what I mean: train example Is there anything wrong with the way I think about the problem?
Thank you!

A2C is not working due to critic loss is not converging

I'm trying to do my own implementation of the Advantage Actor Critic algorithm by using tensorflow. I used the code in https://github.com/BoYanSTKO/Practical_RL-coursera/blob/master/week5_policy_based/practice_a3c.ipynb as a rough template on how I should write the algorithm.
I tried it on the simple CartPole-v0 gym environment but by implementation fails badly. The critics loss just explodes and becomes way to large while the actors loss is rather low.
I'm not sure what I'm doing wrong here. Any help? :)
I've tried separating the actor and critic from each other by having 2 different networks. This did not help either. Have also tried fine tuning some stuff like gamma and learning rate without any success.
!/usr/bin/python
import tensorflow as tf
import numpy as np
import gym
import random
from tensorboardX import SummaryWriter
class ActorCritic():
def __init__(self,state_dim,n_actions,learning_rate,gamma=0.99):
with tf.variable_scope("ActorCritic"):
self.states_ph = tf.placeholder(tf.float32,(None,state_dim),name="states")
self.action_ph = tf.placeholder(tf.int32,(None,),name="actions")
self.n_actions = n_actions
self.reward_ph = tf.placeholder(tf.float32,(None,),name="rewards")
self.next_state_values = tf.placeholder(tf.float32,(None,),name="rewards")
self.is_done_ph = tf.placeholder(tf.float32,(None,),name="rewards")
net = tf.layers.dense(self.states_ph,24,activation=tf.nn.relu)
self.logits = tf.layers.dense(net,n_actions,activation=None)
self.state_values = tf.layers.dense(net,1,activation=None)
self.action_probs = tf.nn.softmax(self.logits)
self.log_prob = tf.nn.log_softmax(self.logits)
self.entropy = -tf.reduce_sum(self.action_probs*self.log_prob,axis=-1,name="entropy")
self.logp_actions = tf.reduce_sum(self.log_prob*tf.one_hot(self.action_ph,depth=n_actions),axis=-1)
self.target_state_values = self.reward_ph + gamma*(1.0-self.is_done_ph)*self.next_state_values
self.advantage = self.target_state_values - self.state_values
self.actor_loss = -tf.reduce_mean(self.logp_actions * tf.stop_gradient(self.advantage)) - 0.01*tf.reduce_mean(self.entropy)
self.critic_loss = tf.reduce_mean(self.advantage**2.0)
self.train_opt = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(self.actor_loss+self.critic_loss)
def train(self,states,actions,rewards,is_done,nxt_state_values_batch):
sess = tf.get_default_session()
return sess.run([self.critic_loss,self.actor_loss,self.train_opt],feed_dict={
self.next_state_values:nxt_state_values_batch,
self.states_ph:states,
self.action_ph:actions,
self.reward_ph:rewards,
self.is_done_ph:is_done})
def predict_state_values(self,states):
sess = tf.get_default_session()
return sess.run(self.state_values,feed_dict={self.states_ph:states})
def sample_actions(self,states):
sess = tf.get_default_session()
action_probs = sess.run(self.action_probs,{self.states_ph:states})
return [ np.random.choice(range(self.n_actions),p=action_prob) for action_prob in action_probs ]
class EnvBatch():
def __init__(self,env_name,n_envs):
self.envs = [gym.make(env_name) for env in range(n_envs)]
self.n_actions = self.envs[0].action_space.n
self.state_dim = self.envs[0].observation_space.shape[0]
def reset(self):
return [env.reset().tolist() for env in self.envs ]
def step(self,actions):
states_batch, rewards_batch, is_done_batch = [], [], []
for action, env in zip(actions,self.envs):
s, r , d, _ = env.step(action)
if d:
s = env.reset()
states_batch.append(s)
rewards_batch.append(r)
is_done_batch.append(d)
return np.array(states_batch), np.array(rewards_batch), np.array(is_done_batch)
def evaluate_performance(env_name,agent,nr_runs=10):
env = gym.make(env_name)
rewards = []
for _ in range(nr_runs):
state = env.reset()
is_done = False
acc_reward = 0.0
while not is_done:
action = agent.sample_actions([state])
nxt_state, reward, is_done, _ = env.step(action[0])
state = nxt_state
acc_reward += reward
rewards.append(acc_reward)
return np.mean(rewards)
tf.reset_default_graph()
env = EnvBatch("CartPole-v0",10)
agent = ActorCritic(env.state_dim,env.n_actions,learning_rate=0.001)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
state_batch = env.reset()
writer = SummaryWriter()
for i in range(100000):
actions = agent.sample_actions(state_batch)
nxt_state_batch, rewards_batch, is_done_batch = env.step(actions)
nxt_state_values = agent.predict_state_values(nxt_state_batch).ravel()
critic_loss, actor_loss, _ = agent.train(state_batch,actions,rewards_batch,is_done_batch,nxt_state_values)
writer.add_scalar("actor_loss",actor_loss,i)
writer.add_scalar("critic_loss",critic_loss,i)
if i%50==0:
test_reward = evaluate_performance("CartPole-v0",agent)
writer.add_scalar("test_reward",test_reward,i)
if test_reward > 195:
print "Done!"
states_batch = nxt_state_batch
sess.close()
writer.close()

(De-)Convutional lstm autoencoder - error jumps

I'm trying to build a convolutional lstm autoencoder (that also predicts future and past) with Tensorflow, and it works to a certain degree, but the error sometimes jumps back up, so essentially, it never converges.
The model is as follows:
The encoder starts with a 64x64 frame from a 20 frame bouncing mnist video for each time step of the lstm. Every stacking layer of LSTM halfs it and increases the depth via 2x2 convolutions with a stride of 2. (so -->32x32x3 -->...--> 1x1x96)
On the other hand, the lstm performs 3x3 convolutions with a stride of 1 on its state. Both results are concatenated to form the new state. In the same way, the decoder uses transposed convolutions to go back to the original format. Then the squared error is calculated.
The error starts at around 2700 and it takes around 20 hours (geforce1060) to get down to ~1700. At which point the jumping back up (and it sometimes jumps back up to 2300 or even ridiculous values like 440300) happens often enough that I can't really get any lower. Also at that point, it can usually pinpoint where the number should be, but its too fuzzy to actually make out the digit...
I tried different learning rates and optimizers, so if anybody knows why that jumping happens, that'd make me happy :)
Here is a graph of the loss with epochs:
import tensorflow as tf
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
import os
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
#based on code by loliverhennigh (Github)
class ConvCell(tf.contrib.rnn.RNNCell):
count = 0 #exists only to remove issues with variable scope
def __init__(self, shape, num_features, transpose = False):
self.shape = shape
self.num_features = num_features
self._state_is_tuple = True
self._transpose = transpose
ConvCell.count+=1
self.count = ConvCell.count
#property
def state_size(self):
return (tf.contrib.rnn.LSTMStateTuple(self.shape[0:4],self.shape[0:4]))
#property
def output_size(self):
return tf.TensorShape(self.shape[1:4])
#here comes to the actual conv lstm implementation, if transpose = true, it performs a deconvolution on the input
def __call__(self, inputs, state, scope=None):
with tf.variable_scope(scope or type(self).__name__+str(self.count)):
c, h = state
state_shape = h.shape
input_shape = inputs.shape
#filter variables and convolutions on data coming from the same cell, a time step previous
h_filters = tf.get_variable("h_filters",[3,3,state_shape[3],self.num_features])
h_filters_gates = tf.get_variable("h_filters_gates",[3,3,state_shape[3],3])
h_partial = tf.nn.conv2d(h,h_filters,[1,1,1,1],'SAME')
h_partial_gates = tf.nn.conv2d(h,h_filters_gates,[1,1,1,1],'SAME')
c_filters = tf.get_variable("c_filters",[3,3,state_shape[3],3])
c_partial = tf.nn.conv2d(c,c_filters,[1,1,1,1],'SAME')
#filters and convolutions/deconvolutions on data coming fromthe cell input
if self._transpose:
x_filters = tf.get_variable("x_filters",[2,2,self.num_features,input_shape[3]])
x_filters_gates = tf.get_variable("x_filters_gates",[2,2,3,input_shape[3]])
x_partial = tf.nn.conv2d_transpose(inputs,x_filters,[int(state_shape[0]),int(state_shape[1]),int(state_shape[2]),self.num_features],[1,2,2,1],'VALID')
x_partial_gates = tf.nn.conv2d_transpose(inputs,x_filters_gates,[int(state_shape[0]),int(state_shape[1]),int(state_shape[2]),3],[1,2,2,1],'VALID')
else:
x_filters = tf.get_variable("x_filters",[2,2,input_shape[3],self.num_features])
x_filters_gates = tf.get_variable("x_filters_gates",[2,2,input_shape[3],3])
x_partial = tf.nn.conv2d(inputs,x_filters,[1,2,2,1],'VALID')
x_partial_gates = tf.nn.conv2d(inputs,x_filters_gates,[1,2,2,1],'VALID')
#some more lstm gate business
gate_bias = tf.get_variable("gate_bias",[1,1,1,3])
h_bias = tf.get_variable("h_bias",[1,1,1,self.num_features*2])
gates = h_partial_gates + x_partial_gates + c_partial + gate_bias
i,f,o = tf.split(gates,3,axis=3)
#concatenate the units coming from the spacial and the temporal dimension to build a unified state
concat = tf.concat([h_partial,x_partial],3) + h_bias
new_c = tf.nn.relu(concat)*tf.sigmoid(i)+c*tf.sigmoid(f)
new_h = new_c * tf.sigmoid(o)
new_state = tf.contrib.rnn.LSTMStateTuple(new_c,new_h)
return new_h, new_state #its redundant, but thats how tensorflow likes it, apparently
#global variables
LEARNING_RATE = 0.005
ITERATIONS_PER_EPOCH = 80
BATCH_SIZE = 75
TEST = False #manual switch to go from training to testing
if TEST:
BATCH_SIZE = 1
inputs = tf.placeholder(tf.float32, (20, BATCH_SIZE, 64, 64,1))
shape0 = [BATCH_SIZE,64,64,2]
shape1 = [BATCH_SIZE,32,32,6]
shape2 = [BATCH_SIZE,16,16,12]
shape3 = [BATCH_SIZE,8,8,24]
shape4 = [BATCH_SIZE,4,4,48]
shape5 = [BATCH_SIZE,2,2,96]
shape6 = [BATCH_SIZE,1,1,192]
#apparently tf.multirnncell has very specific requirements for the initial states oO
initial_state1 = (tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape1),tf.zeros(shape1)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape2),tf.zeros(shape2)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape3),tf.zeros(shape3)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape4),tf.zeros(shape4)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape5),tf.zeros(shape5)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape6),tf.zeros(shape6)))
initial_state2 = (tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape5),tf.zeros(shape5)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape4),tf.zeros(shape4)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape3),tf.zeros(shape3)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape2),tf.zeros(shape2)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape1),tf.zeros(shape1)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape0),tf.zeros(shape0)))
#encoding part of the autoencoder graph
cell1 = ConvCell(shape1,3)
cell2 = ConvCell(shape2,6)
cell3 = ConvCell(shape3,12)
cell4 = ConvCell(shape4,24)
cell5 = ConvCell(shape5,48)
cell6 = ConvCell(shape6,96)
mcell = tf.contrib.rnn.MultiRNNCell([cell1,cell2,cell3,cell4,cell5,cell6])
rnn_outputs, rnn_states = tf.nn.dynamic_rnn(mcell, inputs[0:20,:,:,:],initial_state=initial_state1,dtype=tf.float32, time_major=True)
#decoding part of the autoencoder graph, forward block and backwards block
cell9a = ConvCell(shape5,48,transpose = True)
cell10a = ConvCell(shape4,24,transpose = True)
cell11a = ConvCell(shape3,12,transpose = True)
cell12a = ConvCell(shape2,6,transpose = True)
cell13a = ConvCell(shape1,3,transpose = True)
cell14a = ConvCell(shape0,1,transpose = True)
mcella = tf.contrib.rnn.MultiRNNCell([cell9a,cell10a,cell11a,cell12a,cell13a,cell14a])
cell9b = ConvCell(shape5,48,transpose = True)
cell10b = ConvCell(shape4,24,transpose = True)
cell11b= ConvCell(shape3,12,transpose = True)
cell12b = ConvCell(shape2,6,transpose = True)
cell13b = ConvCell(shape1,3,transpose = True)
cell14b = ConvCell(shape0,1,transpose = True)
mcellb = tf.contrib.rnn.MultiRNNCell([cell9b,cell10b,cell11b,cell12b,cell13b,cell14b])
def PredictionLayer(rnn_outputs,viewPoint = 11, reverse = False):
predLength = viewPoint-2 if reverse else 20-viewPoint #vision is the input for the decoder
vision = tf.concat([rnn_outputs[viewPoint-1:viewPoint,:,:,:],tf.zeros([predLength,BATCH_SIZE,1,1,192])],0)
if reverse:
rnn_outputs2, rnn_states = tf.nn.dynamic_rnn(mcellb, vision, initial_state = initial_state2, time_major=True)
else:
rnn_outputs2, rnn_states = tf.nn.dynamic_rnn(mcella, vision, initial_state = initial_state2, time_major=True)
mean = tf.reduce_mean(rnn_outputs2,4)
if TEST:
return mean
if reverse:
return tf.reduce_sum(tf.square(mean-inputs[viewPoint-2::-1,:,:,:,0]))
else:
return tf.reduce_sum(tf.square(mean-inputs[viewPoint-1:20,:,:,:,0]))
if TEST:
mean = tf.concat([PredictionLayer(rnn_outputs,11,True)[::-1,:,:,:],createPredictionLayer(rnn_outputs,11)],0)
else: #training part of the graph
error = tf.zeros([1])
for i in range(8,15): #range size of 7 or less works, 9 or more does not, no idea why
error += PredictionLayer(rnn_outputs, i)
error += PredictionLayer(rnn_outputs, i, True)
train_fn = tf.train.RMSPropOptimizer(learning_rate=LEARNING_RATE).minimize(error)
################################################################################
## TRAINING LOOP ##
################################################################################
#code based on siemanko/tf_lstm.py (Github)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
saver = tf.train.Saver(restore_sequentially=True, allow_empty=True,)
session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
session.run(tf.global_variables_initializer())
vids = np.load("mnist_test_seq.npy") #20/10000/64/64 , moving mnist dataset from http://www.cs.toronto.edu/~nitish/unsupervised_video/
vids = vids[:,0:6000,:,:] #training set
saver.restore(session,tf.train.latest_checkpoint('./conv_lstm_multiples_v2/'))
#saver.restore(session,'.\conv_lstm_multiples\iteration-74')
for epoch in range(1000):
if TEST:
break
epoch_error = 0
#randomize batches each epoch
vids = np.swapaxes(vids,0,1)
np.random.shuffle(vids)
vids = np.swapaxes(vids,0,1)
for i in range(ITERATIONS_PER_EPOCH):
#running the graph and feeding data
err,_ = session.run([error, train_fn], {inputs: np.expand_dims(vids[:,i*BATCH_SIZE:(i+1)*BATCH_SIZE,:,:],axis=4)})
print(err)
epoch_error += err
#training error each epoch and regular saving
epoch_error /= (ITERATIONS_PER_EPOCH*BATCH_SIZE*4096*20*7)
if (epoch+1) % 5 == 0:
saver.save(session,'.\conv_lstm_multiples_v2\iteration',global_step=epoch)
print("saved")
print("Epoch %d, train error: %f" % (epoch, epoch_error))
#testing
plt.ion()
f, axarr = plt.subplots(2)
vids = np.load("mnist_test_seq.npy")
for i in range(6000,10000):
img = session.run([mean], {inputs: np.expand_dims(vids[:,i:i+1,:,:],axis=4)})
for j in range(20):
axarr[0].imshow(img[0][j,0,:,:])
axarr[1].imshow(vids[j,i,:,:])
plt.show()
plt.pause(0.1)
Usually this happens when gradients' magnitude is very high at some point and causes your network parameters to change a lot. To verify that it is indeed the case, you can produce the same plot of gradient magnitudes and see if they jump right before the loss jump. Assuming this is the case, the classic approach is to use gradient clipping (or go all the way to natural gradient).

Tensorflow: Calling externally set-up function in `tf.scan` (e.g using `tf.make_template`) results in error

I have a RNN like structure that has some building blocks (component neural networks) that are passed in by the user. Here is a minimal example:
import tensorflow as tf
tf.reset_default_graph()
def initialize(shape):
init = tf.random_normal(shape, mean=0, stddev=0.1, dtype=tf.float32)
return init
def test_rnn_with_external(input, hiddens, external_fct):
"""
A simple rnn that makes the standard update, then
feeds the new hidden state through some external
function.
"""
dim_in = input.get_shape().as_list()[-1]
btsz = input.get_shape().as_list()[1]
shape = (dim_in + hiddens, hiddens)
_init = initialize(shape)
W = tf.get_variable("rnn_w", initializer=_init)
_init = tf.zeros([hiddens])
b = tf.get_variable("rnn_b", initializer=_init)
def _step(previous, input):
concat = tf.concat(1, [input, previous])
h_t = tf.tanh(tf.add(tf.matmul(concat, W), b))
h_t = external_fct(h_t)
return h_t
h_0 = tf.zeros([btsz, hiddens])
states = tf.scan(_step,
input,
initializer=h_0,
name="states")
return states
# the external function, relying on the templating mechanism.
def ext_fct(hiddens):
"""
"""
def tmp(input):
shape = (hiddens, hiddens)
_init = initialize(shape)
W = tf.get_variable("ext_w", initializer=_init)
b = 0
return tf.add(tf.matmul(input, W), b, name="external")
return tf.make_template(name_="external_fct", func_=tmp)
# run from here on
t = 5
btsz = 4
dim = 2
hiddens = 3
x = tf.placeholder(tf.float32, shape=(t, btsz, dim))
ext = ext_fct(hiddens)
states = test_rnn_with_external(x, hiddens, external_fct=ext)
sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())
with the error ending in:
InvalidArgumentError: All inputs to node external_fct/ext_w/Assign must be from the same frame.
With Frame, I would associate an area on the stack. So I thought that maybe tf.make_template does something very wired, and thus it is not useable here. The external function can be rewritten a bit and then called more directly, like so:
import tensorflow as tf
tf.reset_default_graph()
def initialize(shape):
init = tf.random_normal(shape, mean=0, stddev=0.1, dtype=tf.float32)
return init
def test_rnn_with_external(input, hiddens, external_fct):
dim_in = input.get_shape().as_list()[-1]
btsz = input.get_shape().as_list()[1]
shape = (dim_in + hiddens, hiddens)
_init = initialize(shape)
W = tf.get_variable("rnn_w", initializer=_init)
_init = tf.zeros([hiddens])
b = tf.get_variable("rnn_b", initializer=_init)
def _step(previous, input):
"""
"""
concat = tf.concat(1, [input, previous])
h_t = tf.tanh(tf.add(tf.matmul(concat, W), b))
h_t = external_fct(h_t, hiddens)
return h_t
h_0 = tf.zeros([btsz, hiddens])
states = tf.scan(_step,
input,
initializer=h_0,
name="states")
return states
def ext_fct_new(input, hiddens):
"""
"""
shape = (hiddens, hiddens)
_init = initialize(shape)
W = tf.get_variable("ext_w_new", initializer=_init)
b = 0
return tf.add(tf.matmul(input, W), b, name="external_new")
t = 5
btsz = 4
dim = 2
hiddens = 3
x = tf.placeholder(tf.float32, shape=(t, btsz, dim))
states = test_rnn_with_external(x, hiddens, external_fct=ext_fct_new)
sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())
However, still the same error InvalidArgumentError: All inputs to node ext_w_new/Assign must be from the same frame.
Of course, moving contents of the external function into the _step part (and tf.get_variableing before) works. But then the flexibility (necessary in the original code) is gone.
What am I doing wrong? Any help/tips/pointers is greatly appreciated.
(Note: Asked this on github, too: https://github.com/tensorflow/tensorflow/issues/4478)
Using a tf.constant_initializer solves the problem. This is described here.