We Keep Coding

Keras - ValueError: Could not interpret loss function identifier

I am trying to build the autoencoder structure detailed in this IEEE article. The autoencoder uses a separable loss function where it is required that I create a custom loss function for the "cluster loss" term of the separable loss function as a function of the average output of the encoder. I create my own layer called RffConnected that calculates the cluster loss and utilizes the add_loss method. Otherwise, this RffConnected layer should act as just a normal deep layer.
Here are my relevant code snippets:
import matplotlib.pyplot as plot
from mpl_toolkits.axes_grid1 import ImageGrid
import numpy as np
import math
from matplotlib.figure import Figure
import tensorflow as tf
import keras
from keras import layers
import random
import time
from os import listdir
#loads data from a text file
def loadData(basePath, samplesPerFile, sampleRate):
real = []
imag = []
fileOrder = []
for file in listdir(basePath):
if((file != "READ_ME") and ((file != "READ_ME.txt"))):
fid = open(basePath + "\\" + file, "r")
fileOrder.append(file)
t = 0
sampleEvery = samplesPerFile / sampleRate
temp1 = []
temp2 = []
times = []
for line in fid.readlines():
times.append(t)
samples = line.split("\t")
temp1.append(float(samples[0]))
temp2.append(float(samples[1]))
t = t + sampleEvery
real.append(temp1)
imag.append(temp2)
fid.close()
real = np.array(real)
imag = np.array(imag)
return real, imag, times, fileOrder
#####################################################################################################
#Breaks up and randomizes data
def breakUpData(real, imag, times, numPartitions, basePath):
if(len(real) % numPartitions != 0):
raise ValueError("Error: The length of the dataset must be divisible by the number of partitions.")
newReal = []
newImag = []
newTimes = []
fileOrder = listdir(basePath)
dataFiles = []
interval = int(len(real[0]) / numPartitions)
for i in range(0, interval):
newTimes.append(times[i])
for i in range(0, len(real)):
tempI = []
tempQ = []
for j in range(0, len(real[0])):
tempI.append(real[i, j])
tempQ.append(imag[i, j])
if((j + 1) % interval == 0):
newReal.append(tempI)
newImag.append(tempQ)
#fileName = fileOrder[i][0: fileOrder[i].find("_") + 3]
dataFiles.append(fileOrder[i])
tempI = []
tempQ = []
#randomizes the broken up dataset and the file list
for i in range(0, len(newReal)):
r = random.randint(0, len(newReal) - 1)
tempReal = newReal[i]
tempImag = newImag[i]
newReal[i] = newReal[r]
newImag[i] = newImag[r]
newReal[r] = tempReal
newImag[r] = tempImag
tempFile = dataFiles[i]
dataFiles[i] = dataFiles[r]
dataFiles[r] = tempFile
#return np.array(newReal), np.array(newImag), newTimes, dataFiles
return newReal, newImag, newTimes, dataFiles
#####################################################################################################
#custom loss layer for the RffAe-S that calculates the clustering loss term
class RffConnected(layers.Layer):
def __init__(self, output_dim, batchSize, beta, alpha):
super(RffConnected, self).__init__()
# self.total = tf.Variable(initial_value=tf.zeros((input_dim,)), trainable=False)
#array = np.zeros(output_dim)
self.iters = 0.0
self.beta = beta
self.alpha = alpha
self.batchSize = batchSize
self.output_dim = output_dim
self.sum = tf.zeros(output_dim, tf.float64)
self.moving_average = tf.zeros(output_dim, tf.float64)
self.clusterloss = tf.zeros(output_dim, tf.float64)
self.sum = tf.cast(self.sum, tf.float32)
self.moving_average = tf.cast(self.moving_average, tf.float32)
self.clusterloss = tf.cast(self.clusterloss, tf.float32)
# self.sum = keras.Input(shape=(self.output_dim,))
# self.moving_average = keras.Input(shape=(self.output_dim,))
# self.clusterloss = keras.Input(shape=(self.output_dim,))
def build(self, input_shape):
self.kernel = self.add_weight(name = 'kernel', \
shape = (int(input_shape[-1]), self.output_dim), \
initializer = 'normal', trainable = True)
#self.kernel = tf.cast(self.kernel, tf.float64)
super(RffConnected, self).build(int(input_shape[-1]))
def call(self, inputs):
#keeps track of training epochs
self.iters = self.iters + 1
#inputs = tf.cast(inputs, tf.float64)
#where this custom layer acts as a normal layer- the loss then uses this
#calc = keras.backend.dot(inputs, self.kernel)
calc = tf.matmul(inputs, self.kernel)
#cumulative sum of deep encoded features
#self.sum = state_ops.assign(self.sum, tf.reshape(tf.math.add(self.sum, calc), tf.shape(self.sum)))
#self.sum = tf.ops.state_ops.assign(self.sum, tf.math.add(self.sum, calc))
#self.sum.assign_add(calc)
self.sum = tf.math.add(self.sum, calc)
#calculate the moving average and loss if we have already trained one batch
if(self.iters >= self.batchSize):
self.moving_average = tf.math.divide(self.sum, self.iters)
self.clusterloss = tf.math.exp(\
tf.math.multiply(-1 * self.beta, tf.math.reduce_sum(tf.math.square(tf.math.subtract(inputs, self.moving_average)))))
#self.add_loss(tf.math.multiply(self.clusterloss, self.alpha))
self.add_loss(self.clusterloss.numpy() * self.alpha)
return calc
#####################################################################################################
def customloss(y_true, y_pred):
loss = tf.square(y_true - y_pred)
print(loss)
return loss
#####################################################################################################
realTraining = np.array(real[0:2200])
realTesting = np.array(real[2200:-1])
imagTraining = np.array(imag[0:2200])
imagTesting = np.array(imag[2200:-1])
numInputs = len(realTraining[0])
i_sig = keras.Input(shape=(numInputs,))
q_sig = keras.Input(shape=(numInputs,))
iRff = tf.keras.layers.experimental.RandomFourierFeatures(numInputs, \
kernel_initializer='gaussian', scale=9.0)(i_sig)
rff1 = keras.Model(inputs=i_sig, outputs=iRff)
qRff = tf.keras.layers.experimental.RandomFourierFeatures(numInputs, \
kernel_initializer='gaussian', scale=9.0)(q_sig)
rff2 = keras.Model(inputs=q_sig, outputs=qRff)
combined = layers.Concatenate()([iRff, qRff])
combineRff = tf.keras.layers.experimental.RandomFourierFeatures(4 * numInputs, \
kernel_initializer='gaussian', scale=10.0)(combined)
preprocess = keras.Model(inputs=[iRff, qRff], outputs=combineRff)
#print(realTraining[0:5])
preprocessedTraining = preprocess.predict([realTraining, imagTraining])
preprocessedTesting = preprocess.predict([realTesting, imagTesting])
################## Entering Encoder ######################
encoderIn = keras.Input(shape=(4*numInputs,))
#connected1 = layers.Dense(100, activation="sigmoid")(encoderIn)
clusterLossLayer = RffConnected(100, 30, 1.00, 100.00)(encoderIn)
#clusterLossLayer = myRffConnected(256)(connected1)
encoder = keras.Model(inputs=encoderIn, outputs=clusterLossLayer)
################## Entering Decoder ######################
connected2 = layers.Dense(125, activation="sigmoid")(clusterLossLayer)
relu1 = layers.ReLU()(connected2)
dropout = layers.Dropout(0.2)(relu1)
reshape1 = layers.Reshape((25, 5, 1))(dropout)
bn1 = layers.BatchNormalization()(reshape1)
trans1 = layers.Conv2DTranspose(1, (4, 2))(bn1)
ups1 = layers.UpSampling2D(size=(2, 1))(trans1)
relu2 = layers.ReLU()(ups1)
bn2 = layers.BatchNormalization()(relu2)
trans2 = layers.Conv2DTranspose(1, (4, 2))(bn2)
ups2 = layers.UpSampling2D(size=(2, 1))(trans2)
relu3 = layers.ReLU()(ups2)
bn3 = layers.BatchNormalization()(relu3)
trans3 = layers.Conv2DTranspose(1, (5, 2))(bn3)
ups3 = layers.UpSampling2D(size=(2, 1))(trans3)
relu4 = layers.ReLU()(ups3)
bn4 = layers.BatchNormalization()(relu4)
trans4 = layers.Conv2DTranspose(1, (7, 1))(bn4)
reshape2 = layers.Reshape((4*numInputs, 1, 1))(trans4)
autoencoder = keras.Model(inputs=encoderIn, outputs=reshape2)
encoded_input = keras.Input(shape=(None, 100))
decoder_layer = autoencoder.layers[-1]
#autoencoder.summary()
autoencoder.compile(optimizer='adam', loss=[autoencoder.losses[-1], customloss], metrics=['accuracy', 'accuracy'])
autoencoder.fit(preprocessedTraining, preprocessedTraining, epochs=100, batch_size=20, shuffle=True, validation_data=(preprocessedTesting, preprocessedTesting))
It seems like it runs for two training epochs then it gives me an error. I end up getting this error when I run it:
ValueError: Could not interpret loss function identifier: Tensor("rff_connected_137/Const:0", shape=(100,), dtype=float32)
I've already spent a considerable amount of time debugging this thing, although if you spot any more errors I would appreciate a heads-up. Thank you in advance.

According to the documentation of the keras Keras Model Training-Loss, the 'loss' attribute can take the value of float tensor (except for the sparse loss functions returning integer arrays) with a specific shape.
If it is necessary to combine two loss functions, it would be better to perform mathematical calculations within your custom loss function to return an output of float tensor. This reference might be a help Keras CustomLoss definition.

Transformer PyTorch Error- ValueError: too many values to unpack (expected 2)

I am having issues getting my model to run. I am not sure which model to use in the translate_sentence function, I have tried model.transformer, model.encoder_de, etc. It is based off of the Transformer class and the forward() function I believe but I am getting a type error. These are the directions:
As in the forward(self, src, tgt) function of the
TransformerModel class, you need to create the appropriate
mask and encode the source sentence (just once).
You also need to create the appropriate mask and encode the
output sentence for sequential predictions. Unlike the source,
for every iteration, you need to re-encode the previous output and
pass both the source sentence and previous output into the
Transformer.
from torch.nn import Transformer
class TransformerModel(nn.Module):
def __init__(self, ntoken_in, ntoken_out, ninp, nhead, npf_dim, nlayers, src_pad_idx, trg_pad_idx, dropout=0.5):
super(TransformerModel, self).__init__()
# --------------- param -----------------
# ntoken_in: the idx of the input word after tokenization
# ntoken_out: the idx of the input word w.r.t. the tokenization
# ninp: the number of expected features in the encoder/decoder inputs
# nhead: the number of multiAttention heads
# npf_dim: the dimension of the feedforward layer
# src_pad_idx: the token for padding in source language
# trg_pad_idx: the token for padding in target language
# ----------------------------------------
self.model_type = 'Transformer'
self.pos_encoder = PositionalEncoding(ninp, dropout)
self.transformer = Transformer(d_model=ninp, nhead=nhead, num_encoder_layers=nlayers, num_decoder_layers=nlayers,
dim_feedforward=npf_dim, dropout=dropout, activation='relu')
self.encoder_en = nn.Embedding(ntoken_in, ninp) # tok_embedding for input
self.encoder_de = nn.Embedding(ntoken_out, ninp) # tok_embedding for output
self.ninp = ninp
self.decoder = nn.Linear(ninp, ntoken_out)
self.src_pad_idx = src_pad_idx
self.tgt_pad_idx = trg_pad_idx
self.init_weights()
def _generate_src_key_mask(self, src):
# for key_padding_mask in transformer
# the positions with the value of True will be ignored while the position
# with the value of False will be unchanged. We mask all padding words.
# The output dim is b*s
src_mask = (src == self.src_pad_idx)
return src_mask.T
def _generate_tgt_mask(self, tgt, sz):
# Beside key_padding_mask in transformer, the output or teacher input
# should be masked sequentially to prevent the model get any information
# from the future words it is going to predict
tgt_key_mask = tgt == self.tgt_pad_idx
# We provide FloatTensor attn_mask. It will be added to the attention weight.
mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
attn_mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0)).to(tgt.device)
return attn_mask, tgt_key_mask.T
def init_weights(self):
initrange = 0.1
self.encoder_en.weight.data.uniform_(-initrange, initrange)
self.encoder_de.weight.data.uniform_(-initrange, initrange)
self.decoder.bias.data.zero_()
self.decoder.weight.data.uniform_(-initrange, initrange)
def forward(self, src, tgt):
# src
src_key_mask = self._generate_src_key_mask(src)
src = self.encoder_en(src) * math.sqrt(self.ninp) # use a learned encoder put stoi index to a feature space s*b --> s*b*e
src = self.pos_encoder(src) # add the pos feature toward feature space
# tgt
tgt_mask, tgt_key_mask = self._generate_tgt_mask(tgt, tgt.size(0))
tgt = self.encoder_de(tgt) * math.sqrt(self.ninp)
tgt = self.pos_encoder(tgt)
output = self.transformer(src, tgt, tgt_mask=tgt_mask,
src_key_padding_mask = src_key_mask,
tgt_key_padding_mask = tgt_key_mask)
output = self.decoder(output)
return output
class PositionalEncoding(nn.Module):
# The positional encoding as described in the paper
# https://arxiv.org/pdf/1706.03762.pdf
def __init__(self, d_model, dropout=0.1, max_len=5000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0).transpose(0, 1)
self.register_buffer('pe', pe)
def forward(self, x):
x = x + self.pe[:x.size(0), :]
return self.dropout(x)
# Here we intialize our model
INPUT_DIM = len(SRC.vocab)
OUTPUT_DIM = len(TRG.vocab)
print(INPUT_DIM, OUTPUT_DIM)
HID_DIM = 256
N_LAYERS = 3
N_HEADS = 8
N_PF_DIM = 512
DROPOUT = 0.1
SRC_PAD_IDX = SRC.vocab.stoi[SRC.pad_token]
TRG_PAD_IDX = TRG.vocab.stoi[TRG.pad_token]
model =TransformerModel(ntoken_in = INPUT_DIM, ntoken_out=OUTPUT_DIM, ninp=HID_DIM,
nhead=N_HEADS, npf_dim=N_PF_DIM, nlayers=N_LAYERS,
src_pad_idx=SRC_PAD_IDX, trg_pad_idx=TRG_PAD_IDX, dropout=DROPOUT).to(device)
def count_parameters(model: nn.Module):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f'The model has {count_parameters(model):,} trainable parameters')
def initialize_weights(m):
if hasattr(m, 'weight') and m.weight.dim() > 1:
nn.init.xavier_uniform_(m.weight.data)
model.apply(initialize_weights)
----------
def translate_sentence(sentence, src_field, trg_field, model, device, max_len = 50):
model.eval()
if isinstance(sentence, str):
nlp = spacy.load('de')
tokens = [token.text.lower() for token in nlp(sentence)]
else:
tokens = [token.lower() for token in sentence]
#tokens = [src_field.init_token] + tokens + [src_field.eos_token]
src_indexes = [src_field.vocab.stoi[token] for token in tokens]
src_tensor = torch.LongTensor(src_indexes).unsqueeze(1).to(device)
with torch.no_grad():
#model.?
hidden, cell = model.encoder_en(src_tensor)
# create a list to hold the output sentence, initialized with an <sos> token
trg_indexes = [trg_field.vocab.stoi[trg_field.init_token]]
for i in range(max_len):
trg_tensor = torch.LongTensor(trg_indexes).unsqueeze(1).to(device)
with torch.no_grad():
#model.?
output, hidden, cell = model.encoder_de(trg_tensor, hidden, cell)
pred_token = output.argmax(1).item()
trg_indexes.append(pred_token)
if pred_token == trg_field.vocab.stoi[trg_field.eos_token]:
break
trg_tokens = [trg_field.vocab.itos[i] for i in trg_indexes]
return trg_tokens[1:]
----------
#getting error here
example_idx = 18
src = vars(train_data.examples[example_idx])['src']
trg = vars(train_data.examples[example_idx])['trg']
print(f'src = {src}')
print(f'trg = {trg}')
translation = translate_sentence(src, TRG, SRC, model, device)
print(f'predicted trg = {translation}')

Why does the result when restoring a saved DDPG model differ significantly from the result when saving it?

I save the trained model after a certain number of episodes with the special save() function of the DDPG class (the network is saved when the reward reaches zero), but when I restore the model again using saver.restore(), the network gives out a reward equal to approximately -1800. Why is this happening, maybe I'm doing something wrong? My network:
import tensorflow as tf
import numpy as np
import gym
epsiode_steps = 500
# learning rate for actor
lr_a = 0.001
# learning rate for critic
lr_c = 0.002
gamma = 0.9
alpha = 0.01
memory = 10000
batch_size = 32
render = True
class DDPG(object):
def __init__(self, no_of_actions, no_of_states, a_bound, ):
self.memory = np.zeros((memory, no_of_states * 2 + no_of_actions + 1), dtype=np.float32)
# initialize pointer to point to our experience buffer
self.pointer = 0
self.sess = tf.Session()
self.noise_variance = 3.0
self.no_of_actions, self.no_of_states, self.a_bound = no_of_actions, no_of_states, a_bound,
self.state = tf.placeholder(tf.float32, [None, no_of_states], 's')
self.next_state = tf.placeholder(tf.float32, [None, no_of_states], 's_')
self.reward = tf.placeholder(tf.float32, [None, 1], 'r')
with tf.variable_scope('Actor'):
self.a = self.build_actor_network(self.state, scope='eval', trainable=True)
a_ = self.build_actor_network(self.next_state, scope='target', trainable=False)
with tf.variable_scope('Critic'):
q = self.build_crtic_network(self.state, self.a, scope='eval', trainable=True)
q_ = self.build_crtic_network(self.next_state, a_, scope='target', trainable=False)
self.ae_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Actor/eval')
self.at_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Actor/target')
self.ce_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Critic/eval')
self.ct_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Critic/target')
# update target value
self.soft_replace = [
[tf.assign(at, (1 - alpha) * at + alpha * ae), tf.assign(ct, (1 - alpha) * ct + alpha * ce)]
for at, ae, ct, ce in zip(self.at_params, self.ae_params, self.ct_params, self.ce_params)]
q_target = self.reward + gamma * q_
td_error = tf.losses.mean_squared_error(labels=(self.reward + gamma * q_), predictions=q)
self.ctrain = tf.train.AdamOptimizer(lr_c).minimize(td_error, name="adam-ink", var_list=self.ce_params)
a_loss = - tf.reduce_mean(q)
# train the actor network with adam optimizer for minimizing the loss
self.atrain = tf.train.AdamOptimizer(lr_a).minimize(a_loss, var_list=self.ae_params)
tf.summary.FileWriter("logs2", self.sess.graph)
# initialize all variables
self.sess.run(tf.global_variables_initializer())
self.saver = tf.train.Saver()
self.saver.restore(self.sess, "Pendulum/nn.ckpt")
def choose_action(self, s):
a = self.sess.run(self.a, {self.state: s[np.newaxis, :]})[0]
a = np.clip(np.random.normal(a, self.noise_variance), -2, 2)
return a
def learn(self):
# soft target replacement
self.sess.run(self.soft_replace)
indices = np.random.choice(memory, size=batch_size)
batch_transition = self.memory[indices, :]
batch_states = batch_transition[:, :self.no_of_states]
batch_actions = batch_transition[:, self.no_of_states: self.no_of_states + self.no_of_actions]
batch_rewards = batch_transition[:, -self.no_of_states - 1: -self.no_of_states]
batch_next_state = batch_transition[:, -self.no_of_states:]
self.sess.run(self.atrain, {self.state: batch_states})
self.sess.run(self.ctrain, {self.state: batch_states, self.a: batch_actions, self.reward: batch_rewards,
self.next_state: batch_next_state})
# we define a function store_transition which stores all the transition information in the buffer
def store_transition(self, s, a, r, s_):
trans = np.hstack((s, a, [r], s_))
index = self.pointer % memory
self.memory[index, :] = trans
self.pointer += 1
if self.pointer > memory:
self.noise_variance *= 0.99995
self.learn()
# we define the function build_actor_network for builing our actor network and after crtic network
def build_actor_network(self, s, scope, trainable)
with tf.variable_scope(scope):
l1 = tf.layers.dense(s, 30, activation=tf.nn.tanh, name='l1', trainable=trainable)
a = tf.layers.dense(l1, self.no_of_actions, activation=tf.nn.tanh, name='a', trainable=trainable)
return tf.multiply(a, self.a_bound, name="scaled_a")
def build_crtic_network(self, s, a, scope, trainable):
with tf.variable_scope(scope):
n_l1 = 30
w1_s = tf.get_variable('w1_s', [self.no_of_states, n_l1], trainable=trainable)
w1_a = tf.get_variable('w1_a', [self.no_of_actions, n_l1], trainable=trainable)
b1 = tf.get_variable('b1', [1, n_l1], trainable=trainable)
net = tf.nn.tanh(tf.matmul(s, w1_s) + tf.matmul(a, w1_a) + b1)
q = tf.layers.dense(net, 1, trainable=trainable)
return q
def save(self):
self.saver.save(self.sess, "Pendulum/nn.ckpt")
env = gym.make("Pendulum-v0")
env = env.unwrapped
env.seed(1)
no_of_states = env.observation_space.shape[0]
no_of_actions = env.action_space.shape[0]
a_bound = env.action_space.high
ddpg = DDPG(no_of_actions, no_of_states, a_bound)
total_reward = []
no_of_episodes = 300
# for each episodes
for i in range(no_of_episodes):
# initialize the environment
s = env.reset()
# episodic reward
ep_reward = 0
for j in range(epsiode_steps):
env.render()
# select action by adding noise through OU process
a = ddpg.choose_action(s)
# peform the action and move to the next state s
s_, r, done, info = env.step(a)
# store the the transition to our experience buffer
# sample some minibatch of experience and train the network
ddpg.store_transition(s, a, r, s_)
# update current state as next state
s = s_
# add episodic rewards
ep_reward += r
if int(ep_reward) == 0 and i > 200:
ddpg.save()
print("save")
quit()
if j == epsiode_steps - 1:
total_reward.append(ep_reward)
print('Episode:', i, ' Reward: %i' % int(ep_reward))
break

Adding layers and bidirectionality to custom LSTM cell in pytorch

I use a very custom LSTM-cell inspired by http://mlexplained.com/2019/02/15/building-an-lstm-from-scratch-in-pytorch-lstms-in-depth-part-1/.
I use it to look at intermediate gating values. My question is, how would I expand this class to have an option for adding more layers and for adding bidirectionality? Should it be wrapped in a new class or added in the present one?
class Dim(IntEnum):
batch = 0
seq = 1
class simpleLSTM(nn.Module):
def __init__(self, input_sz: int, hidden_sz: int):
super().__init__()
self.input_size = input_sz
self.hidden_size = hidden_sz
# input gate
self.W_ii = Parameter(torch.Tensor(input_sz, hidden_sz))
self.W_hi = Parameter(torch.Tensor(hidden_sz, hidden_sz))
self.b_i = Parameter(torch.Tensor(hidden_sz))
# forget gate
self.W_if = Parameter(torch.Tensor(input_sz, hidden_sz))
self.W_hf = Parameter(torch.Tensor(hidden_sz, hidden_sz))
self.b_f = Parameter(torch.Tensor(hidden_sz))
# ???
self.W_ig = Parameter(torch.Tensor(input_sz, hidden_sz))
self.W_hg = Parameter(torch.Tensor(hidden_sz, hidden_sz))
self.b_g = Parameter(torch.Tensor(hidden_sz))
# output gate
self.W_io = Parameter(torch.Tensor(input_sz, hidden_sz))
self.W_ho = Parameter(torch.Tensor(hidden_sz, hidden_sz))
self.b_o = Parameter(torch.Tensor(hidden_sz))
self.init_weights()
self.out = nn.Linear(hidden_sz, len(TRG.vocab))
def init_weights(self):
for p in self.parameters():
if p.data.ndimension() >= 2:
nn.init.xavier_uniform_(p.data)
else:
nn.init.zeros_(p.data)
def forward(self, x, init_states=None ):
"""Assumes x is of shape (batch, sequence, feature)"""
seq_sz, bs, = x.size()
hidden_seq = []
prediction = []
if init_states is None:
h_t, c_t = torch.zeros(self.hidden_size).to(x.device), torch.zeros(self.hidden_size).to(x.device)
else:
h_t, c_t = init_states
for t in range(seq_sz): # iterate over the time steps
x_t = x[t, :].float()
#LOOK HERE!!!
i_t = torch.sigmoid(x_t # self.W_ii + h_t # self.W_hi + self.b_i)
f_t = torch.sigmoid(x_t # self.W_if + h_t # self.W_hf + self.b_f)
g_t = torch.tanh(x_t # self.W_ig + h_t # self.W_hg + self.b_g)
o_t = torch.sigmoid(x_t # self.W_io + h_t # self.W_ho + self.b_o)
c_t = f_t * c_t + i_t * g_t
h_t = o_t * torch.tanh(c_t)
hidden_seq.append(h_t.unsqueeze(Dim.batch))
pred_t = self.out(h_t.unsqueeze(Dim.batch))
#pred_t = F.softmax(pred_t)
prediction.append(pred_t)
hidden_seq = torch.cat(hidden_seq, dim=Dim.batch)
prediction = torch.cat(prediction, dim=Dim.batch)
# reshape from shape (sequence, batch, feature) to (batch, sequence, feature)
hidden_seq = hidden_seq.transpose(Dim.batch, Dim.seq).contiguous()
prediction = prediction.transpose(Dim.batch, Dim.seq).contiguous()
return prediction, hidden_seq, (h_t, c_t)
I call it and train using the following as an example.
lstm = simpleLSTM(1, 100)
hidden_size = lstm.hidden_size
optimizer = optim.Adam(lstm.parameters())
h_0, c_0 = (torch.zeros(hidden_size, requires_grad=True),
torch.zeros(hidden_size, requires_grad=True))
grads = []
h_t, c_t = h_0, c_0
N_EPOCHS = 10
for epoch in range(N_EPOCHS):
epoch_loss = 0
for i, batch in enumerate(train):
optimizer.zero_grad()
src, src_len = batch.src
trg = batch.trg
trg = trg.view(-1)
predict, output, hidden_states = lstm(src)
predict = predict.t().unsqueeze(1)
predict= predict.view(-1, predict.shape[-1])
loss = criterion(predict,trg)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
print(epoch_loss)

The easiest would be to create another module (say Bidirectional) and pass any cell you want to it.
Implementation itself is quite easy to do. Notice that I'm using concat operation for joining bi-directional output, you may want to specify other modes like summation etc.
Please read the comments in the code below, you may have to change it appropriately.
import torch
class Bidirectional(torch.nn.Module):
def __init__(self, cell):
super().__init__()
self.cell = cell
def __call__(self, x, init_states=None):
prediction, hidden_seq, (h_t, c_t) = self.cell(x, init_states)
backward_prediction, backward_hidden_seq, (
backward_h_t,
backward_c_t,
# Assuming sequence is first dimension, otherwise change 0 appropriately
# Reverses sequences so the LSTM cell acts on the reversed sequence
) = self.cell(torch.flip(x, (0,)), init_states)
return (
# Assuming you transpose so it has (batch, seq, features) dimensionality
torch.cat((prediction, backward_prediction), 2),
torch.cat((hidden_seq, backward_hidden_seq), 2),
# Assuming it has (batch, features) dimensionality
torch.cat((h_t, backward_ht), 1),
torch.cat((c_t, backward_ct), 1),
)
When it comes to multiple layers you could do something similiar in principle:
import torch
class Multilayer(torch.nn.Module):
def __init__(self, *cells):
super().__init__()
self.cells = torch.nn.ModuleList(cells)
def __call__(self, x, init_states=None):
inputs = x
for cell in self.cells:
prediction, hidden_seq, (h_t, c_t) = cell(inputs, init_states)
inputs = hidden_seq
return prediction, hidden_seq, (h_t, c_t)
Please note you have to pass created cell objects into Multilayer e.g.:
# For three layers of LSTM, each needs features to be set up correctly
multilayer_LSTM = Multilayer(LSTM(), LSTM(), LSTM())
You may also pass classes instead of instances into constructor and create those inside Multilayer (so hidden_size matches automatically), but those ideas should get you started.

Gradients error using TensorArray Tensorflow

i am trying to implement multidimentional lstm in tensorflow, I am using TensorArray to remember previous states, i am using a complicated way to get two neigbours state(above and from left). tf.cond want that both posible condition to exist and to have the same number of inputs. this is why i added one more cell.zero_state to the (last index +1) of the states. then i using a function to get the correct indexes to the states. when i am trying to use an optimizer in order to minimize a cost, i getting that error:
InvalidArgumentError (see above for traceback): TensorArray
MultiDimentionalLSTMCell-l1-multi-l1/state_ta_262#gradients: Could not
read from TensorArray index 809 because it has not yet been written
to.
Can someone tell how to fix it?
Ps: without optimizer it works!
class MultiDimentionalLSTMCell(tf.nn.rnn_cell.RNNCell):
"""
Note that state_is_tuple is always True.
"""
def __init__(self, num_units, forget_bias=1.0, activation=tf.nn.tanh):
self._num_units = num_units
self._forget_bias = forget_bias
self._activation = activation
#property
def state_size(self):
return tf.nn.rnn_cell.LSTMStateTuple(self._num_units, self._num_units)
#property
def output_size(self):
return self._num_units
def __call__(self, inputs, state, scope=None):
"""Long short-term memory cell (LSTM).
#param: imputs (batch,n)
#param state: the states and hidden unit of the two cells
"""
with tf.variable_scope(scope or type(self).__name__):
c1,c2,h1,h2 = state
# change bias argument to False since LN will add bias via shift
concat = tf.nn.rnn_cell._linear([inputs, h1, h2], 5 * self._num_units, False)
i, j, f1, f2, o = tf.split(1, 5, concat)
new_c = (c1 * tf.nn.sigmoid(f1 + self._forget_bias) +
c2 * tf.nn.sigmoid(f2 + self._forget_bias) + tf.nn.sigmoid(i) *
self._activation(j))
new_h = self._activation(new_c) * tf.nn.sigmoid(o)
new_state = tf.nn.rnn_cell.LSTMStateTuple(new_c, new_h)
return new_h, new_state
def multiDimentionalRNN_whileLoop(rnn_size,input_data,sh,dims=None,scopeN="layer1"):
"""Implements naive multidimentional recurent neural networks
#param rnn_size: the hidden units
#param input_data: the data to process of shape [batch,h,w,chanels]
#param sh: [heigth,width] of the windows
#param dims: dimentions to reverse the input data,eg.
dims=[False,True,True,False] => true means reverse dimention
#param scopeN : the scope
returns [batch,h/sh[0],w/sh[1],chanels*sh[0]*sh[1]] the output of the lstm
"""
with tf.variable_scope("MultiDimentionalLSTMCell-"+scopeN):
cell = MultiDimentionalLSTMCell(rnn_size)
shape = input_data.get_shape().as_list()
if shape[1]%sh[0] != 0:
offset = tf.zeros([shape[0], sh[0]-(shape[1]%sh[0]), shape[2], shape[3]])
input_data = tf.concat(1,[input_data,offset])
shape = input_data.get_shape().as_list()
if shape[2]%sh[1] != 0:
offset = tf.zeros([shape[0], shape[1], sh[1]-(shape[2]%sh[1]), shape[3]])
input_data = tf.concat(2,[input_data,offset])
shape = input_data.get_shape().as_list()
h,w = int(shape[1]/sh[0]),int(shape[2]/sh[1])
features = sh[1]*sh[0]*shape[3]
batch_size = shape[0]
x = tf.reshape(input_data, [batch_size,h,w, features])
if dims is not None:
x = tf.reverse(x, dims)
x = tf.transpose(x, [1,2,0,3])
x = tf.reshape(x, [-1, features])
x = tf.split(0, h*w, x)
sequence_length = tf.ones(shape=(batch_size,), dtype=tf.int32)*shape[0]
inputs_ta = tf.TensorArray(dtype=tf.float32, size=h*w,name='input_ta')
inputs_ta = inputs_ta.unpack(x)
states_ta = tf.TensorArray(dtype=tf.float32, size=h*w+1,name='state_ta',clear_after_read=False)
outputs_ta = tf.TensorArray(dtype=tf.float32, size=h*w,name='output_ta')
states_ta = states_ta.write(h*w, tf.nn.rnn_cell.LSTMStateTuple(tf.zeros([batch_size,rnn_size], tf.float32),
tf.zeros([batch_size,rnn_size], tf.float32)))
def getindex1(t,w):
return tf.cond(tf.less_equal(tf.constant(w),t),
lambda:t-tf.constant(w),
lambda:tf.constant(h*w))
def getindex2(t,w):
return tf.cond(tf.less(tf.constant(0),tf.mod(t,tf.constant(w))),
lambda:t-tf.constant(1),
lambda:tf.constant(h*w))
time = tf.constant(0)
def body(time, outputs_ta, states_ta):
constant_val = tf.constant(0)
stateUp = tf.cond(tf.less_equal(tf.constant(w),time),
lambda: states_ta.read(getindex1(time,w)),
lambda: states_ta.read(h*w))
stateLast = tf.cond(tf.less(constant_val,tf.mod(time,tf.constant(w))),
lambda: states_ta.read(getindex2(time,w)),
lambda: states_ta.read(h*w))
currentState = stateUp[0],stateLast[0],stateUp[1],stateLast[1]
out , state = cell(inputs_ta.read(time),currentState)
outputs_ta = outputs_ta.write(time,out)
states_ta = states_ta.write(time,state)
return time + 1, outputs_ta, states_ta
def condition(time,outputs_ta,states_ta):
return tf.less(time , tf.constant(h*w))
result , outputs_ta, states_ta = tf.while_loop(condition, body, [time,outputs_ta,states_ta])
outputs = outputs_ta.pack()
states = states_ta.pack()
y = tf.reshape(outputs, [h,w,batch_size,rnn_size])
y = tf.transpose(y, [2,0,1,3])
if dims is not None:
y = tf.reverse(y, dims)
return y
def tanAndSum(rnn_size,input_data,scope):
outs = []
for i in range(2):
for j in range(2):
dims = [False]*4
if i!=0:
dims[1] = True
if j!=0:
dims[2] = True
outputs = multiDimentionalRNN_whileLoop(rnn_size,input_data,[2,2],
dims,scope+"-multi-l{0}".format(i*2+j))
outs.append(outputs)
outs = tf.pack(outs, axis=0)
mean = tf.reduce_mean(outs, 0)
return tf.nn.tanh(mean)
graph = tf.Graph()
with graph.as_default():
input_data = tf.placeholder(tf.float32, [20,36,90,1])
#input_data = tf.ones([20,36,90,1],dtype=tf.float32)
sh = [2,2]
out1 = tanAndSum(20,input_data,'l1')
out = tanAndSum(25,out1,'l2')
cost = tf.reduce_mean(out)
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)
#out = multiDimentionalRNN_raw_rnn(2,input_data,sh,dims=[False,True,True,False],scopeN="layer1")
#cell = MultiDimentionalLSTMCell(10)
#out = cell.zero_state(2, tf.float32).c
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
ou,k,_ = session.run([out,cost,optimizer],{input_data:np.ones([20,36,90,1],dtype=np.float32)})
print(ou.shape)
print(k)

You should add parameter parallel_iterations=1 to your while loop call.
Such as:
result, outputs_ta, states_ta = tf.while_loop(
condition, body, [time,outputs_ta,states_ta], parallel_iterations=1)
This is required because inside body you perform read and write operations on the same tensor array (states_ta). And in case of parallel loop execution(parallel_iterations > 1) some thread may try to read info from tensorArray, that was not written to it by another one.
I've test your code snippet with parallel_iterations=1 on tensorflow 0.12.1 and it works as expected.