Related
I am trying to make predcitions on a time-series data (univariate) using Tensorflow. However, when I see the results, the predictions have a time shift (please see the figure). Anyone ideas what the problem is?
Below is the function I use to prepare the set of inputs and outputs (X and Y) as well as normalizing the data
# define a fucntion to prepare the X and Y for training the model
def df_to_X_y_group(df_as_np, window_size, num): # input is numpy array. j_start and j_end show the features, num is the number of wavelength in a group
X=[] # make an empty list
y=[]
for j in range(0,df_as_np.shape[1]-num,num): #(start, stop, step). defining a window over features of size num wihtout any overlaps
for i in range(len(df_as_np)-window_size): # there is overlap between windows in time
row=[a for a in df_as_np[i:i+window_size, j:j+num]]
X.append(row)
label=[r for r in df_as_np[i+window_size, j:j+num]]
y.append(label)
return np.array(X), np.array(y)
# convert the data to X and y format
window_size_group=50
output_group=2
num=2
X_group , y_group= df_to_X_y_group(df_as_np, window_size_group, num) # 70085 rows (number of training samples), 6 is the size of the window (number of time steps) and 5 is the number of features
X_group.shape , y_group.shape
# Split the data to train, val and test by 70,20 and 10% of the data
X_train_group , y_train_group = X_group[:int(.7 * X_group.shape[0])], y_group[:int(.7 * X_group.shape[0])]
X_val_group , y_val_group = X_group[int(.7 * X_group.shape[0]):int(.9 * X_group.shape[0])] , y_group[int(.7 * X_group.shape[0]):int(.9 * X_group.shape[0])]
X_test_group , y_test_group = X_group[int(.9 * X_group.shape[0]):] , y_group[int(.9 * X_group.shape[0]):]
# Normalize the data
X_train_mean_group = np.mean(X_train_group)
X_train_std_group = np.std(X_train_group)
def preprocess_group(X):
X_norm = (X - X_train_mean_group) / X_train_std_group
return X_norm
# Now convert the data back to before normalization
def postprocess_group(pred):
actual = (pred * X_train_std_group) + X_train_mean_group
return actual
# apply the preprocess to the inputs and outputs
X_train_norm_group=preprocess_group(X_train_group)
X_val_norm_group=preprocess_group(X_val_group)
X_test_norm_group=preprocess_group(X_test_group)
y_train_norm_group=preprocess_group(y_train_group)
y_val_norm_group=preprocess_group(y_val_group)
y_test_norm_group=preprocess_group(y_test_group)
X_train_group.shape , y_train_group.shape , X_val_group.shape , y_val_group.shape , X_test_group.shape , y_test_group .shape
Here is the model:
# Build the model
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import *
from tensorflow.keras.callbacks import ModelCheckpoint # to save the model
from tensorflow.keras.losses import MeanSquaredError
from tensorflow.keras.losses import Huber
from tensorflow.keras.metrics import RootMeanSquaredError
from tensorflow.keras.optimizers import Adam
model_group=Sequential()
model_group.add(InputLayer((X_group.shape[1],X_group.shape[2]))) # size of the input (windows (time steps) of size 6 and 1024 features )
model_group.add(LSTM(128))
model_group.add(Dense(20, 'relu'))
model_group.add(Dense(output_group))
model_group.summary()
# Train the model
cp2=ModelCheckpoint('model/group/overlapnum2' , save_best_only=True) # it saves the best model based on val error
model_group.compile(loss=MeanSquaredError(), optimizer=Adam(learning_rate=0.0001), metrics=[RootMeanSquaredError()])
model_group.fit(X_train_norm_group, y_train_norm_group, validation_data=(X_val_norm_group, y_val_norm_group), epochs=15, , callbacks=[cp2])
Below is the function to predict and plot the data:
# We make a function which predict and plot
from sklearn.metrics import mean_squared_error as mse
def plot_predictions(model, X, y, start=0, end=100):
predictions = model.predict(X[start:end])
predictions_actual=postprocess(predictions)
y_actual=postprocess(y[start:end])
plt.plot(predictions_actual, c='r', label='Predictions')
plt.plot(y_actual, label= 'Actuals')
plt.xlabel("Order of the data (TIme)")
plt.ylabel("Number of Photons")
plt.legend() # so the label shows up
return mse(y_actual, predictions_actual)
# plot
plot_predictions(model, X_test_norm, y_test_norm, start=0, end=300)
Here is the predcitions results which are shifted in time (for the true results, y(t) should be replaced with y(t+1)):
enter image description here
I have tried tuning several hyperparameters such as batch size, window size, or adding layers but none of these helped.
I am working on a multi-label classification problem. My gt labels are of shape 14 x 10 x 128, where 14 is the batch_size, 10 is the sequence_length, and 128 is the vector with values 1 if the item in sequence belongs to the object and 0 otherwise.
My output is also of same shape: 14 x 10 x 128. Since, my input sequence was of varying length I had to pad it to make it of fixed length 10. I'm trying to find the loss of the model as follows:
total_loss = 0.0
unpadded_seq_lengths = [3, 4, 5, 7, 9, 3, 2, 8, 5, 3, 5, 7, 7, ...] # true lengths of sequences
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
criterion = nn.BCEWithLogitsLoss()
for data in training_dataloader:
optimizer.zero_grad()
# shape of input 14 x 10 x 128
output = model(data)
batch_loss = 0.0
for batch_idx, sequence in enumerate(output):
# sequence shape is 10 x 128
true_seq_len = unpadded_seq_lengths[batch_idx]
# only keep unpadded gt and predicted labels since we don't want loss to be influenced by padded values
predicted_labels = sequence[:true_seq_len, :] # for example, 3 x 128
gt_labels = gt_labels_padded[batch_idx, :true_seq_len, :] # same shape as above, gt_labels_padded has shape 14 x 10 x 128
# loop through unpadded predicted and gt labels and calculate loss
for item_idx, predicted_labels_seq_item in enumerate(predicted_labels):
# predicted_labels_seq_item and gt_labels_seq_item are 1D vectors of length 128
gt_labels_seq_item = gt_labels[item_idx]
current_loss = criterion(predicted_labels_seq_item, gt_labels_seq_item)
total_loss += current_loss
batch_loss += current_loss
batch_loss.backward()
optimizer.step()
Can anybody please check to see if I'm calculating loss correctly. Thanks
Update:
Is this the correct approach for calculating accuracy metrics?
# batch size: 14
# seq length: 10
for epoch in range(10):
TP = FP = TN = FN = 0.
for x, y, mask in tr_dl:
# mask shape: (10,)
out = model(x) # out shape: (14, 10, 128)
y_pred = (torch.sigmoid(out) >= 0.5).float().type(torch.int64) # consider all predictions above 0.5 as 1, rest 0
y_pred = y_pred[mask] # y_pred shape: (14, 10, 10, 128)
y_labels = y[mask] # y_labels shape: (14, 10, 10, 128)
# do I flatten y_pred and y_labels?
y_pred = y_pred.flatten()
y_labels = y_labels.flatten()
for idx, prediction in enumerate(y_pred):
if prediction == 1 and y_labels[idx] == 1:
# calculate IOU (overlap of prediction and gt bounding box)
iou = 0.78 # assume we get this iou value for objects at idx
if iou >= 0.5:
TP += 1
else:
FP += 1
elif prediction == 1 and y_labels[idx] == 0:
FP += 1
elif prediction == 0 and y_labels[idx] == 1:
FN += 1
else:
TN += 1
EPOCH_ACC = (TP + TN) / (TP + TN + FP + FN)
It is usually recommended to stick with batch-wise operations and avoid going into single-element processing steps while in the main training loop. One way to handle this case is to make your dataset return padded inputs and labels with additionally a mask that will come useful for loss computation. In other words, to compute the loss term with sequences of varying sizes, we will use a mask instead of doing individual slices.
Dataset
The way to proceed is to make sure you build the mask in the dataset and not in the inference loop. Here I am showing a minimal implementation that you should be able to transfer to your dataset without much hassle:
class Dataset(data.Dataset):
def __init__(self):
super().__init__()
def __len__(self):
return 100
def __getitem__(self, index):
i = random.randint(5, SEQ_LEN) # for demo puporse, generate x with random length
x = torch.rand(i, EMB_SIZE)
y = torch.randint(0, N_CLASSES, (i, EMB_SIZE))
# pad data to fit in batch
pad = torch.zeros(SEQ_LEN-len(x), EMB_SIZE)
x_padded = torch.cat((pad, x))
y_padded = torch.cat((pad, y))
# construct tensor to mask loss
mask = torch.cat((torch.zeros(SEQ_LEN-len(x)), torch.ones(len(x))))
return x_padded, y_padded, mask
Essentially in the __getitem__, we not only pad the input x and target y with zero values, we also construct a simple mask containing the positions of the padded values in the currently processed element.
Notice how:
x_padded, shaped (SEQ_LEN, EMB_SIZE)
y_padded, shaped (SEQ_LEN, N_CLASSES)
mask, shaped (SEQ_LEN,)
are all three tensors which are shape invariant across the dataset, yet mask contains the padding information necessary for us to compute the loss function appropriately.
Inference
The loss you've used nn.BCEWithLogitsLoss, is the correct one since it's a multi-dimensional loss used for binary classification. In other words, you can use it here in this multi-label classification task, considering each one of the 128 logits as an individual binary prediction. Do not use nn.CrossEntropyLoss) as suggested elsewhere, since the softmax will push a single logit (i.e. class), which is the behaviour required for single-label classification tasks.
Therefore, in the training loop, we simply have to apply the mask to our loss.
for x, y, mask in dl:
y_pred = model(x)
loss = mask*bce(y_pred, y)
# backpropagation, loss postprocessing, logs, etc.
This is what you need for the first part of the question, there are already loss functions implemented in tensorflow: https://medium.com/#aadityaura_26777/the-loss-function-for-multi-label-and-multi-class-f68f95cae525. Yours is just tf.nn.weighted_cross_entropy_with_logits, but you need to set the weight.
The second part of the question is not straightforward, because there's conditioning on the IOU, generally, when you do machine learning, you should heavily depend on matrix multiplication, in your case, you probably need to pre-calculate the IOU -> 1 or 0 as a vector, then multiply with the y_pred , element-wise, this will give you the modified y_pred . After that, you can use any accuracy available function to calculate the final result.
if you can use the CROSSENTROPYLOSS instead of BCEWithLogitsLoss there is something called ignore_index. you can use it to exclude your padded sequences. the difference between the 2 losses is the activation function used (softmax vs sigmoid). but I think you can still use the CROSSENTROPYLOSSfor binary classification as well.
I am new to machine learning so as a first project I've tried to built a handwritten digit recognition neural network based on the mnist dataset and when I test it with the test images provided by the data set itself it seems to work pretty well (that's what the function test_predict is for). Now I would like to step it up and have the network recognise some actual handwritten digits that I've taken photos of.
The function partial_img_rec takes on an image containing multiple digits and it will be called by multiple_digits. I know it might seem weird that I use recursion here and I'm sure there are some more efficient ways to do this but that's not the matter. In order to test partial_img_rec I provide some photos of individual digits that are stored in the folder .\individual_test and they all look something like this:
The problem is: My neural network's prediction for every single one of my test images is "5". The probability is always about 22% no matter the actual digit displayed. I totally get why the results are not as great as those achieved with the mnist dataset's test images but I certainly didn't expect this. Do you have any idea why this is happening? Any advise is welcome.
Thank you in advance.
Here's my code (edited, now working):
# import keras and the MNIST dataset
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from keras.utils import np_utils
# numpy is necessary since keras uses numpy arrays
import numpy as np
# imports for pictures
from PIL import Image
from PIL import ImageOps
# imports for tests
import random
import os
class mnist_network():
def __init__(self):
""" load data, create and train model """
# load data
(X_train, y_train), (X_test, y_test) = mnist.load_data()
# flatten 28*28 images to a 784 vector for each image
num_pixels = X_train.shape[1] * X_train.shape[2]
X_train = X_train.reshape((X_train.shape[0], num_pixels)).astype('float32')
X_test = X_test.reshape((X_test.shape[0], num_pixels)).astype('float32')
# normalize inputs from 0-255 to 0-1
X_train = X_train / 255
X_test = X_test / 255
# one hot encode outputs
y_train = np_utils.to_categorical(y_train)
y_test = np_utils.to_categorical(y_test)
num_classes = y_test.shape[1]
# create model
self.model = Sequential()
self.model.add(Dense(num_pixels, input_dim=num_pixels, kernel_initializer='normal', activation='relu'))
self.model.add(Dense(num_classes, kernel_initializer='normal', activation='softmax'))
# Compile model
self.model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
# train the model
self.model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=10, batch_size=200, verbose=2)
self.train_img = X_train
self.train_res = y_train
self.test_img = X_test
self.test_res = y_test
def test_all(self):
""" evaluates the success rate using all the test data """
scores = self.model.evaluate(self.test_img, self.test_res, verbose=0)
print("Baseline Error: %.2f%%" % (100-scores[1]*100))
def predict_result(self, img, num_pixels = None, show=False):
""" predicts the number in a picture (vector) """
assert type(img) == np.ndarray and img.shape == (784,)
"""if show:
# show the picture!!!! some problem here
plt.imshow(img, cmap='Greys')
plt.show()"""
num_pixels = img.shape[0]
# the actual number
res_number = np.argmax(self.model.predict(img.reshape(-1,num_pixels)), axis = 1)
# the probabilities
res_probabilities = self.model.predict(img.reshape(-1,num_pixels))
return (res_number[0], res_probabilities.tolist()[0]) # we only need the first element since they only have one
def test_predict(self, amount_test = 100):
""" test some random numbers from the test part of the data set """
assert type(amount_test) == int and amount_test <= 10000
cnt_right = 0
cnt_wrong = 0
for i in range(amount_test):
ind = random.randrange(0,10000) # there are 10000 images in the test part of the data set
""" correct_res is the actual result stored in the data set
It's represented as a list of 10 elements one of which being 1, the rest 0 """
correct_list = self.test_res.tolist()
correct_list = correct_list[ind] # the correct sublist
correct_res = correct_list.index(1.0)
predicted_res = self.predict_result(self.test_img[ind])[0]
if correct_res != predicted_res:
cnt_wrong += 1
print("Error in predict ! \
index = ", ind, " predicted result = ", predicted_res, " correct result = ", correct_res)
else:
cnt_right += 1
print("The machine predicted correctly ",cnt_right," out of ",amount_test," examples. That is a success rate of ", (cnt_right/amount_test)*100,"%.")
def partial_img_rec(self, image, upper_left, lower_right, results=[]):
""" partial is a part of an image """
left_x, left_y = upper_left
right_x, right_y = lower_right
print("current test part: ", upper_left, lower_right)
print("results: ", results)
# condition to stop recursion: we've reached the full width of the picture
width, height = image.size
if right_x > width:
return results
partial = image.crop((left_x, left_y, right_x, right_y))
# rescale image to 28 *28 dimension
partial = partial.resize((28,28), Image.ANTIALIAS)
partial.show()
# transform to vector
partial = ImageOps.invert(partial)
partial = np.asarray(partial, "float32")
partial = partial / 255.
partial[partial < 0.5] = 0.
# flatten image to 28*28 = 784 vector
num_pixels = partial.shape[0] * partial.shape[1]
partial = partial.reshape(num_pixels)
step = height // 10
# is there a number in this part of the image?
res, prop = self.predict_result(partial)
print("result: ", res, ". probabilities: ", prop)
# only count this result if the network is >= 50% sure
if prop[res] >= 0.5:
results.append(res)
# step is 80% of the partial image's size (which is equivalent to the original image's height)
step = int(height * 0.8)
print("found valid result")
else:
# if there is no number found we take smaller steps
step = height // 20
print("step: ", step)
# recursive call with modified positions ( move on step variables )
return self.partial_img_rec(image, (left_x+step, left_y), (right_x+step, right_y), results=results)
def test_individual_digits(self):
""" test partial_img_rec with some individual digits (square shaped images)
saved in the folder 'individual_test' following the pattern 'number_digit.jpg' """
cnt_right, cnt_wrong = 0,0
folder_content = os.listdir(".\individual_test")
for imageName in folder_content:
# image file must be a jpg or png
assert imageName[-4:] == ".jpg" or imageName[-4:] == ".png"
correct_res = int(imageName[0])
image = Image.open(".\\individual_test\\" + imageName).convert("L")
# only square images in this test
if image.size[0] != image.size[1]:
print(imageName, " has the wrong proportions: ", image.size,". It has to be a square.")
continue
predicted_res = self.partial_img_rec(image, (0,0), (image.size[0], image.size[1]), results=[])
if predicted_res == []:
print("No prediction possible for ", imageName)
else:
predicted_res = predicted_res[0]
if predicted_res != correct_res:
print("error in partial_img-rec! Predicted ", predicted_res, ". The correct result would have been ", correct_res)
cnt_wrong += 1
else:
cnt_right += 1
print("correctly predicted ",imageName)
print(cnt_right, " out of ", cnt_right + cnt_wrong," digits were correctly recognised. The success rate is therefore ", (cnt_right / (cnt_right + cnt_wrong)) * 100," %.")
def multiple_digits(self, img):
""" takes as input an image without unnecessary whitespace surrounding the digits """
#assert type(img) == myImage
width, height = img.size
# start with the first quadratic part of the image
res_list = self.partial_img_rec(img, (0,0),(height ,height))
res_str =""
for elem in res_list:
res_str += str(elem)
return res_str
network = mnist_network()
network.test_individual_digits()
EDIT
#Geecode's answer was very helpful and the network now predicts correctly some of the pictures including the one shown above. Yet the overall success rate is lower than 50%. Do you have any ideas how to improve this?
Examples for images returning bad results:
Nothing wrong with your image in itself, your model can correctly classify it.
The issue is that you made a Floor Division on your partial:
partial = partial // 255
which always results in 0. So you always get a black image.
You have to do a "normal" division and some preparation, because your model was trained on black i.e. 0. valued pixel backgrounded negative images:
# transform to vector
partial = ImageOps.invert(partial)
partial = np.asarray(partial, "float32")
partial = partial / 255.
partial[partial < 0.5] = 0.
After then your model will classify correctly:
Out:
result: 1 . probabilities: [0.000431705528171733, 0.7594985961914062, 0.0011404436081647873, 0.00018972357793245465, 0.03162384033203125, 0.008697531186044216, 0.0014472954208031297, 0.18429973721504211, 0.006838776171207428, 0.005832481198012829]
found valid result
Note, that of course you can play on the image preparation yet, that was not the purpose of this answer.
Update:
My detailed answer regarding how to achive better performance in this task, see here.
Here's the code I'm working with (pulled from Kaggle mostly):
inputs = Input((IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS))
...
outputs = Conv2D(4, (1, 1), activation='sigmoid') (c9)
model = Model(inputs=[inputs], outputs=[outputs])
model.compile(optimizer='adam', loss='dice', metrics=[mean_iou])
results = model.fit(X_train, Y_train, validation_split=0.1, batch_size=8, epochs=30, class_weight=class_weights)
I have 4 classes that are very imbalanced. Class A equals 70%, class B = 15%, class C = 10%, and class D = 5%. However, I care most about class D. So I did the following type of calculations: D_weight = A/D = 70/5 = 14 and so on for the weight for class B and A. (if there are better methods to select these weights, then feel free)
In the last line, I'm trying to properly set class_weights and I'm doing it as so: class_weights = {0: 1.0, 1: 6, 2: 7, 3: 14}.
However, when I do this, I get the following error.
class_weight not supported for 3+ dimensional targets.
Is it possible that I add a dense layer after the last layer and just use it as a dummy layer so I can pass the class_weights and then only use the output of the last conv2d layer to do the prediction?
If this is not possible, how would I modify the loss function (I'm aware of this post, however, just passing in the weights in to the loss function won't cut it, because the loss function is called separately for each class) ? Currently, I'm using the following loss function:
def dice_coef(y_true, y_pred):
smooth = 1.
y_true_f = K.flatten(y_true)
y_pred_f = K.flatten(y_pred)
intersection = K.sum(y_true_f * y_pred_f)
return (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
def bce_dice_loss(y_true, y_pred):
return 0.5 * binary_crossentropy(y_true, y_pred) - dice_coef(y_true, y_pred)
But I don't see any way in which I can input class weights. If someone wants the full working code see this post. But remember to change the final conv2d layer's num classes to 4 instead of 1.
You can always apply the weights yourself.
The originalLossFunc below you can import from keras.losses.
The weightsList is your list with the weights ordered by class.
def weightedLoss(originalLossFunc, weightsList):
def lossFunc(true, pred):
axis = -1 #if channels last
#axis= 1 #if channels first
#argmax returns the index of the element with the greatest value
#done in the class axis, it returns the class index
classSelectors = K.argmax(true, axis=axis)
#if your loss is sparse, use only true as classSelectors
#considering weights are ordered by class, for each class
#true(1) if the class index is equal to the weight index
classSelectors = [K.equal(i, classSelectors) for i in range(len(weightsList))]
#casting boolean to float for calculations
#each tensor in the list contains 1 where ground true class is equal to its index
#if you sum all these, you will get a tensor full of ones.
classSelectors = [K.cast(x, K.floatx()) for x in classSelectors]
#for each of the selections above, multiply their respective weight
weights = [sel * w for sel,w in zip(classSelectors, weightsList)]
#sums all the selections
#result is a tensor with the respective weight for each element in predictions
weightMultiplier = weights[0]
for i in range(1, len(weights)):
weightMultiplier = weightMultiplier + weights[i]
#make sure your originalLossFunc only collapses the class axis
#you need the other axes intact to multiply the weights tensor
loss = originalLossFunc(true,pred)
loss = loss * weightMultiplier
return loss
return lossFunc
For using this in compile:
model.compile(loss= weightedLoss(keras.losses.categorical_crossentropy, weights),
optimizer=..., ...)
Changing the class balance directly on the input data
You can change the balance of the input samples too.
For instance, if you have 5 samples from class 1 and 10 samples from class 2, pass the samples for class 5 twice in the input arrays.
.
Using the sample_weight argument.
Instead of working "by class", you can also work "by sample".
Create an array of weights for each sample in your input array: len(x_train) == len(weights)
And fit passing this array to the sample_weight argument.
(If it's fit_generator, the generator will have to return the weights along with the train/true pairs: return/yield inputs, targets, weights)
So this question is about GANs.
I am trying to do a trivial example for my own proof of concept; namely, generate images of hand written digits (MNIST). While most will approach this via deep convolutional gans (dgGANs), I am just trying to achieve this via the 1D array (i.e. instead of 28x28 gray-scale pixel values, a 28*28 1d array).
This git repo features a "vanilla" gans which treats the MNIST dataset as a 1d array of 784 values. Their output values look pretty acceptable so I wanted to do something similar.
Import statements
from __future__ import print_function
import matplotlib as mpl
from matplotlib import pyplot as plt
import mxnet as mx
from mxnet import nd, gluon, autograd
from mxnet.gluon import nn, utils
import numpy as np
import os
from math import floor
from random import random
import time
from datetime import datetime
import logging
ctx = mx.gpu()
np.random.seed(3)
Hyper parameters
batch_size = 100
epochs = 100
generator_learning_rate = 0.001
discriminator_learning_rate = 0.001
beta1 = 0.5
latent_z_size = 100
Load data
mnist = mx.test_utils.get_mnist()
# convert imgs to arrays
flattened_training_data = mnist["test_data"].reshape(10000, 28*28)
define models
G = nn.Sequential()
with G.name_scope():
G.add(nn.Dense(300, activation="relu"))
G.add(nn.Dense(28 * 28, activation="tanh"))
D = nn.Sequential()
with D.name_scope():
D.add(nn.Dense(128, activation="relu"))
D.add(nn.Dense(64, activation="relu"))
D.add(nn.Dense(32, activation="relu"))
D.add(nn.Dense(2, activation="tanh"))
loss = gluon.loss.SoftmaxCrossEntropyLoss()
init stuff
G.initialize(mx.init.Normal(0.02), ctx=ctx)
D.initialize(mx.init.Normal(0.02), ctx=ctx)
trainer_G = gluon.Trainer(G.collect_params(), 'adam', {"learning_rate": generator_learning_rate, "beta1": beta1})
trainer_D = gluon.Trainer(D.collect_params(), 'adam', {"learning_rate": discriminator_learning_rate, "beta1": beta1})
metric = mx.metric.Accuracy()
dynamic plot (for juptyer notebook)
import matplotlib.pyplot as plt
import time
def dynamic_line_plt(ax, y_data, colors=['r', 'b', 'g'], labels=['Line1', 'Line2', 'Line3']):
x_data = []
y_max = 0
y_min = 0
x_min = 0
x_max = 0
for y in y_data:
x_data.append(list(range(len(y))))
if max(y) > y_max:
y_max = max(y)
if min(y) < y_min:
y_min = min(y)
if len(y) > x_max:
x_max = len(y)
ax.set_ylim(y_min, y_max)
ax.set_xlim(x_min, x_max)
if ax.lines:
for i, line in enumerate(ax.lines):
line.set_xdata(x_data[i])
line.set_ydata(y_data[i])
else:
for i in range(len(y_data)):
l = ax.plot(x_data[i], y_data[i], colors[i], label=labels[i])
ax.legend()
fig.canvas.draw()
train
stamp = datetime.now().strftime('%Y_%m_%d-%H_%M')
logging.basicConfig(level=logging.DEBUG)
# arrays to store data for plotting
loss_D = nd.array([0], ctx=ctx)
loss_G = nd.array([0], ctx=ctx)
acc_d = nd.array([0], ctx=ctx)
labels = ['Discriminator Loss', 'Generator Loss', 'Discriminator Acc.']
%matplotlib notebook
fig, ax = plt.subplots(1, 1)
ax.set_xlabel('Time')
ax.set_ylabel('Loss')
dynamic_line_plt(ax, [loss_D.asnumpy(), loss_G.asnumpy(), acc_d.asnumpy()], labels=labels)
for epoch in range(epochs):
tic = time.time()
data_iter.reset()
for i, batch in enumerate(data_iter):
####################################
# Update Disriminator: maximize log(D(x)) + log(1-D(G(z)))
####################################
# extract batch of real data
data = batch.data[0].as_in_context(ctx)
# add noise
# Produce our noisey input to the generator
latent_z = mx.nd.random_normal(0,1,shape=(batch_size, latent_z_size), ctx=ctx)
# soft and noisy labels
# real_label = mx.nd.ones((batch_size, ), ctx=ctx) * nd.random_uniform(.7, 1.2, shape=(1)).asscalar()
# fake_label = mx.nd.ones((batch_size, ), ctx=ctx) * nd.random_uniform(0, .3, shape=(1)).asscalar()
# real_label = nd.random_uniform(.7, 1.2, shape=(batch_size), ctx=ctx)
# fake_label = nd.random_uniform(0, .3, shape=(batch_size), ctx=ctx)
real_label = mx.nd.ones((batch_size, ), ctx=ctx)
fake_label = mx.nd.zeros((batch_size, ), ctx=ctx)
with autograd.record():
# train with real data
real_output = D(data)
errD_real = loss(real_output, real_label)
# train with fake data
fake = G(latent_z)
fake_output = D(fake.detach())
errD_fake = loss(fake_output, fake_label)
errD = errD_real + errD_fake
errD.backward()
trainer_D.step(batch_size)
metric.update([real_label, ], [real_output,])
metric.update([fake_label, ], [fake_output,])
####################################
# Update Generator: maximize log(D(G(z)))
####################################
with autograd.record():
output = D(fake)
errG = loss(output, real_label)
errG.backward()
trainer_G.step(batch_size)
####
# Plot Loss
####
# append new data to arrays
loss_D = nd.concat(loss_D, nd.mean(errD), dim=0)
loss_G = nd.concat(loss_G, nd.mean(errG), dim=0)
name, acc = metric.get()
acc_d = nd.concat(acc_d, nd.array([acc], ctx=ctx), dim=0)
# plot array
dynamic_line_plt(ax, [loss_D.asnumpy(), loss_G.asnumpy(), acc_d.asnumpy()], labels=labels)
name, acc = metric.get()
metric.reset()
logging.info('Binary training acc at epoch %d: %s=%f' % (epoch, name, acc))
logging.info('time: %f' % (time.time() - tic))
output
img = G(mx.nd.random_normal(0,1,shape=(100, latent_z_size), ctx=ctx))[0].reshape((28, 28))
plt.imshow(img.asnumpy(),cmap='gray')
plt.show()
Now this doesn't get nearly as good as the repo's example from above. Although fairly similar.
Thus I was wondering if you could take a look and figure out why:
the colors are inverted
why the results are sub par
I have been fiddling around with this trying a lot of various things to improve the results (I will list this in a second), but for the MNIST dataset this really shouldn't be needed.
Things I have tried (and I have also tried a host of combinations):
increasing the generator network
increasing the discriminator network
using soft labeling
using noisy labeling
batch norm after every layer in the generator
batch norm of the data
normalizing all values between -1 and 1
leaky relus in the generator
drop out layers in the generator
increased learning rate of discriminator compared to generator
decreased learning rate of i compared to generator
Please let me know if you have any ideas.
1) If you look into original dataset:
training_set = mnist["train_data"].reshape(60000, 28, 28)
plt.imshow(training_set[10,:,:], cmap='gray')
you will notice that the digits are white on a black background. So, technically speaking, your results are not inversed - they match the pattern of original images you used as a real data.
If you want to invert colors for visualization purposes, you can easily do that by changing the pallete to reversed one by adding '_r' (it works for all color palletes):
plt.imshow(img.asnumpy(), cmap='gray_r')
You also can play with ranges of colors by changing vmin and vmax parameters. They control how big the difference between colors should be. By default it is calculated automatically based on provided set.
2) "Why the results are sub par" - I think this is exactly the reason why the community started to use dcGANs. To me the results in the git repo you provided are quite noisy. Surely, they are different from what you receive, and you can achieve the same quality just by changing your activation functions from tanh to sigmoid as in the example on github:
G = nn.Sequential()
with G.name_scope():
G.add(nn.Dense(300, activation="relu"))
G.add(nn.Dense(28 * 28, activation="sigmoid"))
D = nn.Sequential()
with D.name_scope():
D.add(nn.Dense(128, activation="relu"))
D.add(nn.Dense(64, activation="relu"))
D.add(nn.Dense(32, activation="relu"))
D.add(nn.Dense(2, activation="sigmoid"))
Sigmoid never goes below zero and it works better in this scenario. Here is a sample picture I get if I train updated model for 30 epochs (the rest of the hyperparameters are same).
If you decide to explore dcGAN to get even better results, take a look here - https://mxnet.incubator.apache.org/tutorials/unsupervised_learning/gan.html It is a well explained tutorial on how to build dcGAN with Mxnet and Gluon. By using dcGAN you will get way better results than that.