Develop Reference

RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed)

I am trying to get to grips with Pytorch and I wanted to try to reproduce this code:
https://github.com/andy-psai/MountainCar_ActorCritic/blob/master/RL%20Blog%20FINAL%20MEDIUM%20code%2002_12_19.ipynb
in Pytorch.
I am having a problem in that this error is being returned:
RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad(). Specify retain_graph=True if you need to backward through the graph a second time or if you need to access saved tensors after calling backward.
A similar question said to use zero_grad() again after the optimizer step, but this hasn't resolved the issue.
I've included the entire code below so hopefully it should be reproduceable.
Any advice would be much appreciated.
import gym
import os
import os.path as osp
import time
import math
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.distributions import Normal
env = gym.envs.make("MountainCarContinuous-v0")
# Value function
class Value(nn.Module):
def __init__(self, dim_states):
super(Value, self).__init__()
self.net = nn.Sequential(
nn.Linear(dim_states, 400),
nn.ReLU(),
nn.Linear(400,400),
nn.ReLU(),
nn.Linear(400, 1)
)
self.optimizer = optim.Adam(self.parameters(), lr = 1e-3)
self.criterion = nn.MSELoss()
def forward(self, state):
return self.net(torch.from_numpy(state).float())
def compute_return(self, output, target):
self.optimizer.zero_grad()
loss = self.criterion(output, target)
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
# Policy network
class Policy(nn.Module):
def __init__(self, dim_states, env):
super(Policy, self).__init__()
self.hidden1 = nn.Linear(dim_states, 40)
self.hidden2 = nn.Linear(40, 40)
self.mu = nn.Linear(40, 1)
self.sigma = nn.Linear(40,1)
self.env = env
self.optimizer = optim.Adam(self.parameters(), lr = 2e-5)
def forward(self, state):
state = torch.from_numpy(state).float()
x = F.relu(self.hidden1(state))
x = F.relu(self.hidden2(x))
mu = self.mu(x)
sigma = F.softmax(self.sigma(x), dim=-1)
action_dist = Normal(mu, sigma)
action_var = action_dist.rsample()
action_var = torch.clip(action_var,
self.env.action_space.low[0],
self.env.action_space.high[0])
return action_var, action_dist
def compute_return(self, action, dist, td_error):
self.optimizer.zero_grad()
loss_actor = -dist.log_prob(action)*td_error
loss_actor.backward()
self.optimizer.step()
self.optimizer.zero_grad()
# Normalise the state space
import sklearn
import sklearn.preprocessing
state_space_samples = np.array(
[env.observation_space.sample() for x in range(10000)])
scaler = sklearn.preprocessing.StandardScaler()
scaler.fit(state_space_samples)
# Normaliser
def scale_state(state):
scaled = scaler.transform([state])
return scaled
##################################
# Parameters
lr_actor = 0.00002
lr_critic = 0.001
actor = Policy(2, env)
critic = Value(2)
# Training loop params
gamma = 0.99
num_episodes = 300
episode_history = []
for episode in range(num_episodes):
# Receive initial state from E
state = env.reset()
reward_total = 0
steps = 0
done = False
while not done:
action, dist = actor(state)
# print(np.squeeze(action))
next_state, reward, done, _ = env.step(
np.array([action.item()]))
if episode % 50 == 0:
env.render()
steps += 1
reward_total += reward
# TD Target
target = reward + gamma * np.squeeze(critic(next_state), axis=0)
td_error = target - np.squeeze(critic(state), axis=0)
# Update actor
actor.compute_return(action, dist, td_error)
# Update critic
critic.compute_return(np.squeeze(critic(state), axis=0), target)
episode_history.append(reward_total)
print(f"Episode: {episode}, N Steps: {steps}, Cumulative reward {reward_total}")
if np.mean(episode_history[-100:]) > 90 and len(episode_history) > 101:
print("Solved")
print(f"Mean cumulative reward over 100 episodes {np.mean(episode_history[-100:])}")

Problem lies in this snippet. When you create target variable, there is a forward pass through critic which generates a computation graph and critic(next_state) is the leaf node of that graph making target a part of the graph (you can check this by printing target which will show you grad_fn=<AddBackward0>). Finally, when you call critic.compute_return(critic_out, target), a new computation graph is generated and passing target(which is a part of the previous computation graph) causes a Runtime error.
Solution is to call detach() on critic(next_state), this will free target variable and it will no longer be a part of the computation graph(again check by printing target).
target = reward + gamma * np.squeeze(critic(next_state).detach(), axis=0)
td_error = target - np.squeeze(critic(state), axis=0)
# Update actor
actor.compute_return(action, dist, td_error)
# Update critic
critic_out = np.squeeze(critic(state), axis=0)
print(critic_out)
critic.compute_return(critic_out, target)

RuntimeError: Found dtype Double but expected Float - Pytorch RL

I am trying to get an actor critic variant of the pendulum running, but I seem to be running into a particular problem.
RuntimeError: Found dtype Double but expected Float
I saw this had come up multiple times before so I have been through those and have attempted to change the data types of my loss (kept in comments) but it is still not working. Could anyone point out how to resolve this so that I can learn from it?
Full code below
import gym, os
import numpy as np
from itertools import count
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.distributions import Normal
from collections import namedtuple
SavedAction = namedtuple('SavedAction', ['log_prob', 'value'])
LOG_SIG_MAX = 2
LOG_SIG_MIN = -20
class ActorCritic(nn.Module):
"""
Implementing both heads of the actor critic model
"""
def __init__(self, state_space, action_space):
super(ActorCritic, self).__init__()
self.state_space = state_space
self.action_space = action_space
# HL 1
self.linear1 = nn.Linear(self.state_space, 128)
# HL 2
self.linear2 = nn.Linear(128, 256)
# Outputs
self.critic_head = nn.Linear(256, 1)
self.action_mean = nn.Linear(256, self.action_space)
self.action_std = nn.Linear(256, self.action_space)
# Saving
self.saved_actions = []
self.rewards = []
# Optimizer
self.optimizer = optim.Adam(self.parameters(), lr = 1e-3)
self.eps = np.finfo(np.float32).eps.item()
def forward(self, state):
"""
Forward pass for both actor and critic
"""
# State to Layer 1
l1_output = F.relu(self.linear1(state))
# Layer 1 to Layer 2
l2_output = F.relu(self.linear2(l1_output))
# Layer 2 to Action
mean = self.action_mean(l2_output)
std = self.action_std(l2_output)
std = torch.clamp(std, min=LOG_SIG_MIN, max = LOG_SIG_MAX)
std = std.exp()
# Layer 2 to Value
value_est = self.critic_head(l2_output)
return value_est, mean, std
def select_action(self,state):
state = torch.from_numpy(state).float().unsqueeze(0)
value_est, mean, std = self.forward(state)
value_est = value_est.reshape(-1)
# Make prob Normal dist
dist = Normal(mean, std)
action = dist.sample()
action = torch.tanh(action)
ln_prob = dist.log_prob(action)
ln_prob = ln_prob.sum()
self.saved_actions.append(SavedAction(ln_prob, value_est))
action = action.numpy()
return action[0]
def compute_returns(self, gamma): # This is the error causing code
"""
Calculate losses and do backprop
"""
R = 0
saved_actions = self.saved_actions
policy_losses = []
value_losses = []
returns = []
for r in self.rewards[::-1]:
# Discount value
R = r + gamma*R
returns.insert(0,R)
returns = torch.tensor(returns)
returns = (returns - returns.mean())/(returns.std()+self.eps)
for (log_prob, value), R in zip(saved_actions, returns):
advantage = R - value.item()
advantage = advantage.type(torch.FloatTensor)
policy_losses.append(-log_prob*advantage)
value_losses.append(F.mse_loss(value, torch.tensor([R])))
self.optimizer.zero_grad()
loss = torch.stack(policy_losses).sum() + torch.stack(value_losses).sum()
loss = loss.type(torch.FloatTensor)
loss.backward()
self.optimizer.step()
del self.rewards[:]
del self.saved_actions[:]
env = gym.make("Pendulum-v0")
state_space = env.observation_space.shape[0]
action_space = env.action_space.shape[0]
# Train Expert AC
model = ActorCritic(state_space, action_space)
train = True
if train == True:
# Main loop
window = 50
reward_history = []
for ep in count():
state = env.reset()
ep_reward = 0
for t in range(1,1000):
if ep%50 == 0:
env.render()
action = model.select_action(state)
state, reward, done, _ = env.step(action)
model.rewards.append(reward)
ep_reward += reward
if done:
break
print(reward)
model.compute_returns(0.99) # Error begins here
reward_history.append(ep_reward)
# Result information
if ep % 50 == 0:
mean = np.mean(reward_history[-window:])
print(f"Episode: {ep} Last Reward: {ep_reward} Rolling Mean: {mean}")
if np.mean(reward_history[-100:])>199:
print(f"Environment solved at episode {ep}, average run length > 200")
break
Complete error log, some elements redacted for privacy. Originally the loop actor critic and main loop were in separate files. Comments added to appropriate error causing lines.
Traceback (most recent call last):
File "pendulum.py", line 59, in <module>
model.compute_returns(0.99)
File "/home/x/Software/git/x/x/solvers/actorcritic_cont.py", line 121, in compute_returns
loss.backward()
File "/home/x/.local/lib/python3.8/site-packages/torch/_tensor.py", line 255, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
File "/home/x/.local/lib/python3.8/site-packages/torch/autograd/__init__.py", line 147, in backward
Variable._execution_engine.run_backward(
RuntimeError: Found dtype Double but expected Float

Answering here in case anyone has similar issues in the future.
The output of the reward in OpenAI Gym Pendulum-v0 is a double, so when you compute the return over the episode you need to change that to a float tensor.
I did this just by:
returns = torch.tensor(returns)
returns = (returns - returns.mean())/(returns.std()+self.eps)
returns = returns.type(torch.FloatTensor)

Why does an ANN validation accuracy oscillate?

The following training curve is generated using the same Tensorflow + Keras script written in Python:
RED line uses five features.
GREEN line uses seven features.
BLUE line uses nine features.
Can anyone tell me the probable cause of the oscillation of the GREEN line so that I can troubleshoot my script?
Source code:
import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
#os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0" # Use both gpus for training.
import sys, random
import time
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.callbacks import ModelCheckpoint
import numpy as np
from lxml import etree, objectify
# <editor-fold desc="GPU">
# resolve GPU related issues.
try:
physical_devices = tf.config.list_physical_devices('GPU')
for gpu_instance in physical_devices:
tf.config.experimental.set_memory_growth(gpu_instance, True)
except Exception as e:
pass
# END of try
# </editor-fold>
# <editor-fold desc="Lxml helper">
class LxmlHelper:
#classmethod
def objectify_xml(cls, input_path_dir):
file_dom = etree.parse(input_path_dir) # parse xml and convert it into DOM
file_xml_bin = etree.tostring(file_dom, pretty_print=False, encoding="ascii") # encode DOM into ASCII object
file_xml_text = file_xml_bin.decode() # convert binary ASCII object into ASCII text
objectified_xml = objectify.fromstring(file_xml_text) # convert text into a Doxygen object
return objectified_xml
# </editor-fold>
# <editor-fold desc="def encode(letter)">
def encode(letter: str):
if letter == 'H':
return [1.0, 0.0, 0.0]
elif letter == 'E':
return [0.0, 1.0, 0.0]
elif letter == 'C':
return [0.0, 0.0, 1.0]
elif letter == '-':
return [0.0, 0.0, 0.0]
# END of function
def encode_string_1(pattern_str: str):
# Iterate over the string
one_hot_binary_str = []
for ch in pattern_str:
try:
one_hot_binary_str = one_hot_binary_str + encode(ch)
except Exception as e:
print(pattern_str, one_hot_binary_str, ch)
# END of for loop
return one_hot_binary_str
# END of function
def encode_string_2(pattern_str: str):
# Iterate over the string
one_hot_binary_str = []
for ch in pattern_str:
temp_encoded_vect = [encode(ch)]
one_hot_binary_str = one_hot_binary_str + temp_encoded_vect
# END of for loop
return one_hot_binary_str
# END of function
# </editor-fold>
# <editor-fold desc="def load_data()">
def load_data_k(fname: str, class_index: int, feature_start_index: int, **selection):
"""Loads data for training and validation
:param fname: (``string``) - name of the file with the data
:param selection: (``kwargs``) - see below
:return: four tensorflow tensors: training input, training output, validation input and validation output
:Keyword Arguments:
* *top_n_lines* (``number``) --
take top N lines of the input and disregard the rest
* *random_n_lines* (``number``) --
take random N lines of the input and disregard the rest
* *validation_part* (``float``) --
separate N_lines * given_fraction of the input lines from the training set and use
them for validation. When the given_fraction = 1.0, then the same input set of
N_lines is used both for training and validation (this is the default)
"""
i = 0
file = open(fname)
if "top_n_lines" in selection:
lines = [next(file) for _ in range(int(selection["top_n_lines"]))]
elif "random_n_lines" in selection:
tmp_lines = file.readlines()
lines = random.sample(tmp_lines, int(selection["random_n_lines"]))
else:
lines = file.readlines()
data_x, data_y, data_z = [], [], []
for l in lines:
row = l.strip().split() # return a list of words from the line.
x = [float(ix) for ix in row[feature_start_index:]] # convert 3rd to 20th word into a vector of float numbers.
y = encode(row[class_index]) # convert the 3rd word into binary.
z = encode_string_1(row[class_index+1])
data_x.append(x) # append the vector into 'data_x'
data_y.append(y) # append the vector into 'data_y'
data_z.append(z) # append the vector into 'data_z'
# END for l in lines
num_rows = len(data_x)
given_fraction = selection.get("validation_part", 1.0)
if given_fraction > 0.9999:
valid_x, valid_y, valid_z = data_x, data_y, data_z
else:
n = int(num_rows * given_fraction)
data_x, data_y, data_z = data_x[n:], data_y[n:], data_z[n:]
valid_x, valid_y, valid_z = data_x[:n], data_y[:n], data_z[:n]
# END of if-else block
tx = tf.convert_to_tensor(data_x, np.float32)
ty = tf.convert_to_tensor(data_y, np.float32)
tz = tf.convert_to_tensor(data_z, np.float32)
vx = tf.convert_to_tensor(valid_x, np.float32)
vy = tf.convert_to_tensor(valid_y, np.float32)
vz = tf.convert_to_tensor(valid_z, np.float32)
return tx, ty, tz, vx, vy, vz
# END of the function
# </editor-fold>
# <editor-fold desc="def create_model()">
def create_model(n_hidden_1, n_hidden_2, num_classes, num_features):
# create the model
model = Sequential()
model.add(tf.keras.layers.InputLayer(input_shape=(num_features,)))
model.add(tf.keras.layers.Dense(n_hidden_1, activation='sigmoid'))
model.add(tf.keras.layers.Dense(n_hidden_2, activation='sigmoid'))
###model.add(tf.keras.layers.Dense(n_hidden_3, activation='sigmoid'))
model.add(tf.keras.layers.Dense(num_classes, activation='softmax'))
# instantiate the optimizer
opt = keras.optimizers.SGD(learning_rate=LEARNING_RATE)
# compile the model
model.compile(
optimizer=opt,
loss="categorical_crossentropy",
metrics="categorical_accuracy"
)
# return model
return model
# </editor-fold>
if __name__ == "__main__":
# <editor-fold desc="(input/output parameters)">
my_project_routine = LxmlHelper.objectify_xml("my_project_evaluate.xml")
# input data
INPUT_DATA_FILE = str(my_project_routine.input.input_data_file)
INPUT_PATH = str(my_project_routine.input.input_path)
CLASS_INDEX = int(my_project_routine.input.class_index)
FEATURE_INDEX = int(my_project_routine.input.feature_index)
# output data
OUTPUT_PATH = str(my_project_routine.output.output_path)
MODEL_FILE = str(my_project_routine.output.model_file)
TRAINING_PROGRESS_FILE = str(my_project_routine.output.training_progress_file)
# Learning parameters
LEARNING_RATE = float(my_project_routine.training_params.learning_rate)
EPOCH_SIZE = int(my_project_routine.training_params.epoch_size)
BATCH_SIZE = int(my_project_routine.training_params.batch_size)
INPUT_LINES_COUNT = int(my_project_routine.input.input_lines_count)
VALIDATION_PART = float(my_project_routine.training_params.validation_part)
SAVE_PERIOD = str(my_project_routine.output.save_period)
# NN parameters
HIDDEN_LAYER_1_NEURON_COUNT = int(my_project_routine.hidden_layers.one)
HIDDEN_LAYER_2_NEURON_COUNT = int(my_project_routine.hidden_layers.two)
###HIDDEN_LAYER_3_NEURON_COUNT = int(my_project_routine.hidden_layers.three)
CLASS_COUNT = int(my_project_routine.class_count)
FEATURES_COUNT = int(my_project_routine.features_count)
input_file_path_str = os.path.join(INPUT_PATH, INPUT_DATA_FILE)
training_progress_file_path_str = os.path.join(OUTPUT_PATH, TRAINING_PROGRESS_FILE)
model_file_path = os.path.join(OUTPUT_PATH, MODEL_FILE)
# command-line arg processing
input_file_name_str = None
if len(sys.argv) > 1:
input_file_name_str = sys.argv[1]
else:
input_file_name_str = input_file_path_str
# END of if-else
# </editor-fold>
# <editor-fold desc="(load data from file)">
# load training data from the disk
train_x, train_y, _, validate_x, validate_y, _ = \
load_data_k(
fname=input_file_name_str,
class_index=CLASS_INDEX,
feature_start_index=FEATURE_INDEX,
random_n_lines=INPUT_LINES_COUNT,
validation_part=VALIDATION_PART
)
print("training data size : ", len(train_x))
print("validation data size : ", len(validate_x))
# </editor-fold>
### STEPS_PER_EPOCH = len(train_x) // BATCH_SIZE
### VALIDATION_STEPS = len(validate_x) // BATCH_SIZE
# <editor-fold desc="(model creation)">
# load previously saved NN model
model = None
try:
model = keras.models.load_model(model_file_path)
print("Loading NN model from file.")
model.summary()
except Exception as ex:
print("No NN model found for loading.")
# END of try-except
# </editor-fold>
# <editor-fold desc="(model run)">
# # if there is no model loaded, create a new model
if model is None:
csv_logger = keras.callbacks.CSVLogger(training_progress_file_path_str)
checkpoint = ModelCheckpoint(
model_file_path,
monitor='loss',
verbose=1,
save_best_only=True,
mode='auto',
save_freq='epoch'
)
callbacks_vector = [
csv_logger,
checkpoint
]
# Set mirror strategy
#strategy = tf.distribute.MirroredStrategy(devices=["/device:GPU:0","/device:GPU:1"])
#with strategy.scope():
print("New NN model created.")
# create sequential NN model
model = create_model(
n_hidden_1=HIDDEN_LAYER_1_NEURON_COUNT,
n_hidden_2=HIDDEN_LAYER_2_NEURON_COUNT,
##n_hidden_3=HIDDEN_LAYER_3_NEURON_COUNT,
num_classes=CLASS_COUNT,
num_features=FEATURES_COUNT
)
# Train the model with the new callback
history = model.fit(
train_x, train_y,
validation_data=(validate_x, validate_y),
batch_size=BATCH_SIZE,
epochs=EPOCH_SIZE,
callbacks=[callbacks_vector],
shuffle=True,
verbose=2
)
print(history.history.keys())
# END of ... with
# END of ... if
# </editor-fold>
Plotting Script
import os
from argparse import ArgumentParser
import random
from typing import List
import matplotlib.pyplot as plt
import numpy as np
import math
import sys
import datetime
class Quad:
def __init__(self, x_vector, y_vector, color_char, label_str):
self.__x_vector = x_vector
self.__y_vector = y_vector
self.__color_char = color_char
self.__label_str = label_str
def get_x_vector(self):
return self.__x_vector
def get_y_vector(self):
return self.__y_vector
def get_color_char(self):
return self.__color_char
def get_label_str(self):
return self.__label_str
class HecaPlotClass:
def __init__(self):
self.__x_label_str: str = None
self.__y_label_str: str = None
self.__title_str: str = None
self.__trio_vector: List[Quad] = []
self.__plotter = plt
#property
def x_label_str(self):
return self.__x_label_str
#x_label_str.setter
def x_label_str(self, t):
self.__x_label_str = t
#property
def y_label_str(self):
return self.__y_label_str
#y_label_str.setter
def y_label_str(self, t):
self.__y_label_str = t
#property
def title_str(self):
return self.__title_str
#title_str.setter
def title_str(self, t):
self.__title_str = t
def add_y_axes(self, trio_obj: Quad):
self.__trio_vector.append(trio_obj)
def generate_plot(self):
for obj in self.__trio_vector:
x_vector = obj.get_x_vector()
y_vector = obj.get_y_vector()
label_str = obj.get_label_str()
# print(label_str)
# print(len(x_vector))
# print(len(y_vector))
self.__plotter.plot(
x_vector,
y_vector,
color=obj.get_color_char(),
label=label_str
)
# END of ... for loop
# Naming the x-axis, y_1_vector-axis and the whole graph
self.__plotter.xlabel(self.__x_label_str)
self.__plotter.ylabel(self.__y_label_str)
self.__plotter.title(self.__title_str)
# Adding legend, which helps us recognize the curve according to it's color
self.__plotter.legend()
# To load the display window
#self.__plotter.show()
def save_png(self, output_directory_str):
output_file_str = os.path.join(output_directory_str, self.__title_str + '.png')
self.__plotter.savefig(output_file_str)
def save_pdf(self, output_directory_str):
output_file_str = os.path.join(output_directory_str, self.__title_str + '.pdf')
self.__plotter.savefig(output_file_str)
class MainClass(object):
__colors_vector = ['red', 'green', 'blue', 'cyan', 'magenta', 'yellow', 'orange', 'lightgreen', 'crimson']
__working_dir = r"."
__file_names_vector = ["training_progress-32.txt", "training_progress-64.txt", "training_progress-128.txt"]
__input_files_vector = []
__output_directory = None
__column_no_int = 0
__split_percentage_at_tail_int = 100
__is_pdf_output = False
__is_png_output = False
# <editor-fold desc="def load_data()">
#classmethod
def __load_data(cls, fname: str, percetage_int:int, column_no_int:int):
np_array = np.loadtxt(
fname,
# usecols=range(1,11),
dtype=np.float32,
skiprows=1,
delimiter=","
)
size_vector = np_array.shape
array_len_int = size_vector[0]
rows_count_int = int(percetage_int * array_len_int / 100)
np_array = np_array[-rows_count_int:]
x = np_array[:, 0]
y = np_array[:, column_no_int]
return x, y
# END of the function
# </editor-fold>
# <editor-fold desc="(__parse_args())">
#classmethod
def __parse_args(cls):
# initialize argument parser
my_parser = ArgumentParser()
my_parser.add_argument("-c", help="column no.", type=int)
my_parser.add_argument('-i', nargs='+', help='a list of input files', required=True)
my_parser.add_argument("-o", help="output directory", type=str)
my_parser.add_argument("-n", help="percentage of data to split from tail", type=float)
my_parser.add_argument("--pdf", help="PDF output", action='store_true')
my_parser.add_argument("--png", help="PNG output", action='store_true')
# parse the argument
args = my_parser.parse_args()
cls.__input_files_vector = args.i
cls.__output_directory = args.o
cls.__split_percentage_at_tail_int = args.n
cls.__column_no_int = args.c
cls.__is_pdf_output = args.pdf
cls.__is_png_output = args.png
# </editor-fold>
#classmethod
def main(cls):
cls.__parse_args()
if cls.__input_files_vector is None:
cls.__input_files_vector = cls.__file_names_vector
if cls.__output_directory is None:
cls.__output_directory = cls.__working_dir
if cls.__split_percentage_at_tail_int is None:
cls.__split_percentage_at_tail_int = 100
if cls.__column_no_int is None:
cls.__column_no_int = 1
my_project_plot_obj = HecaPlotClass()
i = 0
for file_path_str in cls.__input_files_vector:
print(file_path_str)
x_vector, y_vector = cls.__load_data(os.path.join(cls.__working_dir, file_path_str), cls.__split_percentage_at_tail_int, cls.__column_no_int)
my_project_plot_obj.x_label_str = "Epoch"
my_project_plot_obj.y_label_str = "Accuracy"
my_project_plot_obj.title_str = "training_plot-{date:%Y-%m-%d_%H:%M:%S}".format(date=datetime.datetime.now())
my_project_plot_obj.x_axis_vector = x_vector
if i == 0:
random_int = 0
else:
random_int = i % (len(cls.__colors_vector)-1)
# END of ... if
print("random_int : ", random_int)
my_project_plot_obj.add_y_axes(Quad(x_vector, y_vector, cls.__colors_vector[random_int], file_path_str))
i = i + 1
# END of ... for loop
my_project_plot_obj.generate_plot()
my_project_plot_obj.save_png(cls.__output_directory)
my_project_plot_obj.save_pdf(cls.__output_directory)
if __name__ == "__main__":
MainClass.main()

The primary reason could be improper (non-random ~ ordered) distribution of data.
If you notice the accuracy beyond epoch 180, there is a orderly switching between the accuracy between ~0.43 (approx.) and ~0.33 (~approx.), and occasionally ~0.23 (approx.). The more important thing to notice is that the accuracy is decreasing (there's no improvement in validation accuracy) as we increase the epochs.
The accuracy can increase in such cases if you (1) reduce batch size, or (2) use a better optimizer like Adam. And check the learning rate.
These changes can help the shift and oscillation, as well.
Additionally, Running average of the accuracy can be plotted to avoid the oscillation. This is again a mitigation scheme rather than a correction scheme. But, what it does is removes the order (partition of the data) and mixes the nearby data.
Lastly, I would also reshuffle the data and normalize after each layer. See if that helps.

Generally, sharp jumps and flat lines in the accuracy usually mean that a group of examples is classified as a given class at a same time. If your dataset contains, say, 50 examples with the same combination of 7 features then they would go into the same class at the same time. This is what probably causes sharp jumps - identical or similar examples clustered together.
So for example, if you have 50 men aged 64, and a decision boundary to classify them as more prone to an illness shifts from >65 to >63, then accuracy changes rapidly as all of them change classification at the same time.
Regarding the oscillation of the curve - due to the fact above, oscillation will be amplified by small changes in learning. Your network learns based on cross entropy, which means that it minimizes the difference between target and your predictions. This means that it operates on the difference between probability and target (say, 0.3 vs class 0) instead of class and target like accuracy (so, 0 vs 0) in the same example. Cross entropy is much more smooth as it is not affected by the issue outlined above.

Keras Double DQN average reward decreases over time and is unable to converge

I am attempting to teach a Double DQN agent to run a gridworld where there is one seeker (the agent) who will try to collect all the hiders which are randomly spawned. Every step has a path_cost of -0.1 and if a hider is collected a reward of 1 is received. The DQN net receives an array with the shape (world_width,world_height,1) as the state which is a complete translation of the environment viewed from above where empty space is described as 0, seeker as 2, and hider as 3. The agent is then supposed to choose one action, either left, up, right, or down. An example configuration of the environment is shown in the image below.
gridworld
However, when training my agent the reward initially decreases in correlation to the decreasing exploration and therefore it can be assumed that when the agent follows the DQN net it will perform worse than when choosing actions randomly. Here are a few examples of the reward graphs I have received when training with different hyperparameters (y-axis is total steps where each episode is 100 steps unless it finishes).
Reward Graph
As seen the agent becomes worse at solving the environment and it is approximately when epsilon becomes equal to my min_epsilon the curve stabilizes (meaning almost no exploration or random moves).
I have tried different hyperparameters but without any apparent differences in results and would there appreciate it if someone could give me a pointer to where the problem might be.
The hyperparameters I have been mostly using is:
wandb.config.epsilon = 1.0
wandb.config.epsilon_decay = 0.99
wandb.config.batch_size = 32
wandb.config.learning_rate = 1e-3
wandb.config.gamma = 0.8
wandb.config.min_epsilon = 1e-1
wandb.config.buffersize = 10000
wandb.config.epochs = 1
wandb.config.reward_discount = 0.01
wandb.config.episodes = 1000
And here is my code:
import tensorflow as tf
from tensorflow.keras.layers import Input, Dense, Conv2D, MaxPooling2D, Flatten
from tensorflow.keras.optimizers import Adam
from collections import deque
from termcolor import colored
import wandb
from wandb.keras import WandbCallback
import numpy as np
import copy, os, random
from argparse import ArgumentParser
from plotter import plotter
from HNS import HNS
tf.keras.backend.set_floatx('float64')
wandb.init(name=name, project=project)
wandb.env.name = "HNS"
wandb.env.world_size = (8, 8)
wandb.env.state_dim = (8, 8, 1)
wandb.env.hider_count = 2
wandb.env.action_dim = 4
wandb.env.random_spawn = True
wandb.env.max_steps = 100
wandb.config.node = node
wandb.config.epsilon = 1.0
wandb.config.epsilon_decay = 0.99
wandb.config.batch_size = 32
wandb.config.learning_rate = 1e-3
wandb.config.gamma = 0.8
wandb.config.min_epsilon = 1e-1
wandb.config.buffersize = 10000
wandb.config.epochs = 1
wandb.config.reward_discount = 0.01
wandb.config.episodes = 1000
wandb.config.conv1_kernel = (8,8)
wandb.config.conv1_filters = 16
wandb.config.conv1_strides = 4
wandb.config.conv1_activation = "relu"
wandb.config.conv1_padding = "same"
wandb.config.conv2_kernel = (4,4)
wandb.config.conv2_filters = 32
wandb.config.conv2_strides = 4
wandb.config.conv2_activation = "relu"
wandb.config.conv2_padding = "same"
wandb.config.dense1_neurons = 16
wandb.config.dense1_activation = "relu"
wandb.config.loss = "mse"
parser = ArgumentParser()
parser.add_argument('--hider_count', type=int, default=wandb.env.hider_count)
parser.add_argument('--max_steps', type=int, default=wandb.env.max_steps)
parser.add_argument('--epsilon_decay', type=float, default=wandb.config.epsilon_decay)
parser.add_argument('--min_epsilon', type=float, default=wandb.config.min_epsilon)
parser.add_argument('--learning_rate', type=float, default=wandb.config.learning_rate)
parser.add_argument('--gamma', type=float, default=wandb.config.gamma)
parser.add_argument('--reward_discount', type=float, default=wandb.config.reward_discount)
parser.add_argument('--episodes', type=int, default=wandb.config.episodes)
parser.add_argument('--batch_size', type=int, default=wandb.config.batch_size)
args, unknown = parser.parse_known_args()
wandb.config.update(args, allow_val_change=True)
class ReplayBuffer:
def __init__(self):
self.buffer = deque(maxlen=wandb.config.buffersize)
def put(self, state, action, reward, next_state, done):
self.buffer.append([state, action, reward, next_state, done])
def sample(self):
sample = random.sample(self.buffer, wandb.config.batch_size)
states, actions, rewards, next_states, done = map(np.asarray, zip(*sample))
return states, actions, rewards, next_states, done
def size(self):
return len(self.buffer)
class ActionStatemodel:
def __init__(self):
self.epsilon = wandb.config.epsilon
self.model = self.create_model()
def create_model(self):
# Init model
model = tf.keras.Sequential()
# Set up layers
model.add(Conv2D(filters=wandb.config.conv1_filters, kernel_size=wandb.config.conv1_kernel, activation=wandb.config.conv1_activation,
strides=wandb.config.conv1_strides, padding=wandb.config.conv1_padding, name="conv_1", input_shape=wandb.env.state_dim))
model.add(Conv2D(filters=wandb.config.conv2_filters, kernel_size=wandb.config.conv2_kernel, activation=wandb.config.conv2_activation,
strides=wandb.config.conv2_strides, padding=wandb.config.conv2_padding, name="conv_2"))
model.add(Flatten())
model.add(Dense(units=wandb.config.dense1_neurons, activation=wandb.config.dense1_activation, name="dense_1"))
model.add(Dense(wandb.env.action_dim, name="dense_2"))
# Finalize model
model.compile(loss=wandb.config.loss, optimizer=Adam(wandb.config.learning_rate))
model.summary()
return model
# Get q-values from state
def predict(self, state):
return self.model.predict(state)
# Get action from
def get_action(self, state):
# Predict action
state = np.expand_dims(state, axis=0)
q_value = self.predict(state)
if np.random.random() < self.epsilon: return random.randint(0, wandb.env.action_dim - 1), 1
else: return np.argmax(q_value), 0
def train(self, states, targets):
history = self.model.fit(states, targets, epochs=wandb.config.epochs, callbacks=[WandbCallback()], verbose=2, use_multiprocessing=True)
return history.history["loss"][0]
class Agent:
def __init__(self, env):
self.env = env
self.predict_net = ActionStatemodel()
self.target_net = ActionStatemodel()
self.target_update()
self.buffer = ReplayBuffer()
# Copy weights from model to target_model
def target_update(self):
weights = self.predict_net.model.get_weights()
self.target_net.model.set_weights(weights)
def replay(self):
loss = 0
for _ in range(5):
states, actions, rewards, next_states, done = self.buffer.sample()
# Collect predicted actions from predict_net
predicted_q_values = self.predict_net.predict(next_states)
predicted_actions = np.argmax(predicted_q_values, axis=1)
# Get q values from target_net of above predicted actions
target_q_values = self.target_net.predict(next_states)
target_action_q_values = [np.take(target_q_values[i], predicted_actions[i]) for i in range(len(target_q_values))]
# Create targets based on q values, reward and done
targets = predicted_q_values.copy()
targets[range(wandb.config.batch_size), actions] = rewards + (1 - done) * target_action_q_values * args.gamma
loss += self.predict_net.train(states, targets)
return loss
def train(self):
# Save weights for heatmap rendering
# Main training loop
for ep in range(wandb.config.episodes):
# Initialization
done, total_reward, step, loss, exploration = False, 0, 0, 0, 0
state = self.env.reset()
while not done and step < wandb.env.max_steps:
# Predict and perform action
action, e = self.predict_net.get_action(state)
exploration += e
next_state, reward, done, _ = self.env.step(action)
self.buffer.put(state, action, reward * wandb.config.reward_discount, next_state, done)
total_reward += reward
if self.buffer.size() >= 1000 and step % 10 == 0:
loss = self.replay()
state = next_state
step += 1
self.target_update()
# Update epsilon
self.predict_net.epsilon = max(wandb.config.epsilon_decay * self.predict_net.epsilon, wandb.config.min_epsilon)
# Calculate weights change and log weights
pre_weights = self.get_weights(self.predict_net.model.layers)
tar_weights = self.get_weights(self.target_net.model.layers)
# LOG
print(colored("EP" + str(ep) + "-Reward: " + str(total_reward) + " Done: " + str(done), "green"))
wandb.log({"episode" : ep,
"buffersize" : self.buffer.size(),
"EpReward" : total_reward,
"epsilon" : self.predict_net.epsilon,
"done" : int(done),
"Exploration" : exploration / _,
"loss" : loss,
"pre_weights" : pre_weights,
"tar_weights" : tar_weights
})
# "weigthUpdate" : wandb.Image(neuron_map),
# Get weights and names for every layer of nn model
def get_weights(self, layers):
weigths = []
names = []
for layer in layers:
wb = layer.get_weights()
if wb:
weigths.append(wb[0].flatten())
names.append(layer.name)
return weigths, names
if __name__ == "__main__":
env = HNS(random_spawn=wandb.env.random_spawn, world_size=wandb.env.world_size, hider_count=wandb.env.hider_count)
agent = Agent(env=env)
agent.train()
agent.target_net.model.save(os.path.join(wandb.run.dir, "model.h5"))

How to convert this Keras code to Chainer code? (LSTM Autoencoder)

Here, I have LSTM Autoencoder written in Keras. I want to convert the code to Chainer.
import numpy as np
from keras.layers import Input, GRU
from keras.models import Model
input_feat = Input(shape=(30, 2000))
l = GRU( 100, return_sequences=True, activation="tanh", recurrent_activation="hard_sigmoid")(input_feat)
l = GRU(2000, return_sequences=True, activation="tanh", recurrent_activation="hard_sigmoid")(l)
model = Model(input_feat, l)
model.compile(optimizer="RMSprop", loss="mean_squared_error")
feat = np.load("feat.npy")
model.fit(feat, feat[:, ::-1, :], epochs=200, batch_size=250)
feat is numpy whose dimension is (269, 30, 2000). I could run above code and the result was reasonable. I had written below Chainer code.
import numpy as np
from chainer import Chain, Variable, optimizers
import chainer.functions as F
import chainer.links as L
class GRUAutoEncoder(Chain):
def __init__(self):
super().__init__()
with self.init_scope():
self.encode = L.GRU(2000, 100)
self.decode = L.GRU(100, 2000)
def __call__(self, h, mode):
if mode == "encode":
h = F.tanh(self.encode(h))
return h
if mode == "decode":
h = F.tanh(self.decode(h))
return h
def reset(self):
self.encode.reset_state()
self.decode.reset_state()
def main():
feat = np.load("feat.npy") #(269, 30, 2000)
gru_autoencoder = GRUAutoEncoder()
optimizer = optimizers.RMSprop(lr=0.01).setup(gru_autoencoder)
N = len(feat)
batch_size = 250
for epoch in range(200):
index = np.random.randint(0, N-batch_size+1)
input_splices = feat[index:index+batch_size] #(250, 30, 2000)
#Encoding
input_vector = np.zeros((30, batch_size, 2000), dtype="float32")
h = []
for i in range(frame_rate):
input_vector[i] = input_splices[:, i, :] #(250, 1, 2000)
tmp = Variable(input_vector[i])
h.append(gru_autoencoder(tmp, "encode")) #(250, 100)
#Decoding
output_vector = []
for i in range(frame_rate):
tmp = h[i]
output_vector.append(gru_autoencoder(tmp, "decode"))
x = input_vector[0]
t = output_vector[0]
for i in range(len(output_vector)):
x = F.concat((x,input_vector[i]), axis=1)
t = F.concat((t,output_vector[i]), axis=1)
loss = F.mean_squared_error(x, t)
gru_autoencoder.cleargrads()
loss.backward()
optimizer.update()
gru_autoencoder.reset()
if __name__ == "__main__":
main()
But the result of above code was not reasonable. I think the Chainer code has something wrong but I cannot find where it is.
In Keras code,
model.fit(feat, feat[:, ::-1, :])
So, I tried to reverse output_vector in Chainer code,
output_vector.reverse()
but the result was not still reasonable.

.. note: This answer is a translation of [Japanese SO].(https://ja.stackoverflow.com/questions/52162/keras%E3%81%AE%E3%82%B3%E3%83%BC%E3%83%89%E3%82%92chainer%E3%81%AB%E6%9B%B8%E3%81%8D%E6%8F%9B%E3%81%88%E3%81%9F%E3%81%84lstm-autoencoder%E3%81%AE%E5%AE%9F%E8%A3%85/52213#52213)
You should avoid using L.GRU and should use L.NStepGRU, because for L.GRU you have to write "recurrence-aware" code. In other words, you have to apply L.GRU multiple times to one timeseries, therefore "batch" must be treated with great care. L.NStepGRU (with n_layers=1) wraps the batch-processing, so it would be user-friendly.
An instance of L.StepGRU takes two input arguments: one is initial state, and the other is a list of timeserieses, which composes a batch. Conventionally, the initial state is None.
Therefore, the whole answer for your question is as follows.
### dataset.py
from chainer.dataset import DatasetMixin
import numpy as np
class MyDataset(DatasetMixin):
N_SAMPLES = 269
N_TIMESERIES = 30
N_DIMS = 2000
def __init__(self):
super().__init__()
self.data = np.random.randn(self.N_SAMPLES, self.N_TIMESERIES, self.N_DIMS) \
.astype(np.float32)
def __len__(self):
return self.N_SAMPLES
def get_example(self, i):
return self.data[i, :, :]
### model.py
import chainer
from chainer import links as L
from chainer import functions as F
from chainer.link import Chain
class MyModel(Chain):
N_IN_CHANNEL = 2000
N_HIDDEN_CHANNEL = 100
N_OUT_CHANNEL = 2000
def __init__(self):
super().__init__()
self.encoder = L.NStepGRU(n_layers=1, in_size=self.N_IN_CHANNEL, out_size=self.N_HIDDEN_CHANNEL, dropout=0)
self.decoder = L.NStepGRU(n_layers=1, in_size=self.N_HIDDEN_CHANNEL, out_size=self.N_OUT_CHANNEL, dropout=0)
def to_gpu(self, device=None):
self.encoder.to_gpu(device)
self.decoder.to_gpu(device)
def to_cpu(self):
self.encoder.to_cpu()
self.decoder.to_cpu()
#staticmethod
def flip_list(source_list):
return [F.flip(source, axis=1) for source in source_list]
def __call__(self, source_list):
"""
.. note:
This implementation makes use of "auto-encoding"
by avoiding redundant copy in GPU device.
In the typical implementation, this function should receive
both of ``source_list`` and ``target_list``.
"""
target_list = self.flip_list(source_list)
_, h_list = self.encoder(hx=None, xs=source_list)
_, predicted_list = self.decoder(hx=None, xs=h_list)
diff_list = [F.mean_squared_error(target, predicted).reshape((1,)) for target, predicted in zip(target_list, predicted_list)]
loss = F.sum(F.concat(diff_list, axis=0))
chainer.report({'loss': loss}, self)
return loss
### converter.py (referring examples/seq2seq/seq2seq.py)
from chainer.dataset import to_device
def convert(batch, device):
"""
.. note:
batch must be list(batch_size) of array
"""
if device is None:
return batch
else:
return [to_device(device, x) for x in batch]
### train.py
from chainer.iterators import SerialIterator
from chainer.optimizers import RMSprop
from chainer.training.updaters import StandardUpdater
from chainer.training.trainer import Trainer
dataset = MyDataset()
BATCH_SIZE = 32
iterator = SerialIterator(dataset, BATCH_SIZE)
model = MyModel()
optimizer = RMSprop()
optimizer.setup(model)
updater = StandardUpdater(iterator, optimizer, convert, device=0)
trainer = Trainer(updater, (100, 'iteration'))
from chainer.training.extensions import snapshot_object
trainer.extend(snapshot_object(model, "model_iter_{.updater.iteration}"), trigger=(10, 'iteration'))
from chainer.training.extensions import LogReport, PrintReport, ProgressBar
trainer.extend(LogReport(['epoch', 'iteration', 'main/loss'], (1, 'iteration')))
trainer.extend(PrintReport(['epoch', 'iteration', 'main/loss']), trigger=(1, 'iteration'))
trainer.extend(ProgressBar(update_interval=1))
trainer.run()