I'm currently writing a script to quantise a Keras model down to 8 bits. I'm doing a fairly basic linear scaling on the weights, by assuming a normal distribution of weights and biases, and then interpolating all the values within 2 standard deviations of the mean, to the range [-128, 127].
This all works, and I run the model through inference, but my image out is crazy bad. I know there will be a small performance hit, but I'm seeing roughly 10x performance degradation.
My question is, after this scaling of the weights, do I need to do the inverse scaling operation to my output? None of the papers I've been reading seem to mention this, but I'm unsure why else my results would be so bad.
The network is for image demosaicing. It takes in a RAW image, and is meant to output an image with very low noise, and no demosaicing artefacts. My full precision model is very good, with image PSNRs of around 40-43dB, but after quantisation, I'm getting 4-8dB, and incredibly bad looking images.
Code for anyone who's bothered to read it
for i in layer_index:
count = count+1
layer = model.get_layer(index = i);
weights = layer.get_weights();
weights_act = weights[0];
bias_act = weights[1];
std = np.std(weights_act)
if (std > max_std):
max_std = std
mean = np.mean(weights_act)
mean_of_mean = mean_of_mean + mean
mean_of_mean = mean_of_mean / count
max_bound = mean_of_mean + 2*max_std
min_bound = mean_of_mean - 2*max_std
print(max_bound, min_bound)
for i in layer_index:
layer = model.get_layer(index = i);
weights = layer.get_weights();
weights_act = weights[0];
bias_act = weights[1];
weights_shape = weights_act.shape;
bias_shape = bias_act.shape;
new_weights = np.empty(weights_shape, dtype = np.int8)
print(new_weights.dtype)
new_biass = np.empty(bias_shape, dtype = np.int8)
for a in range(weights_shape[0]):
for b in range(weights_shape[1]):
for c in range(weights_shape[2]):
for d in range(weights_shape[3]):
new_weight = (((weights_act[a,b,c,d] - min_bound) * (127 - (-128)) / (max_bound - min_bound)) + (-128))
new_weights[a,b,c,d] = np.int8(new_weight)
#print(new_weights[a,b,c,d], weights_act[a,b,c,d])
for e in range(bias_shape[0]):
new_bias = (((bias_act[e] - min_bound) * (127 - (-128)) / (max_bound - min_bound)) + (-128))
new_biass[e] = np.int8(new_bias)
new_weight_layer = (new_weights, new_biass)
layer.set_weights(new_weight_layer)
You dont do what you think you are doing, I'll explain.
If you wish to take pre-trained model and quantize it you have to add scales after each operation that involves weights, lets take for example the convolution operation.
As we know convolution operation is linear in my explantion i will ignore the bias for the sake of simplicity (adding him is relatively easy), Let's assume X is our input Y is our output and W is the weights, convolution can be written as:
Y=W*X
where '*' represent the convolution operation, what you are basically doing is taking the weights and multiple them by some scalar (lets call it 'a') and shift them by some other scalar (let's call it 'b') so in your model you use W' where: W'= Wa+b
So if we return to the convolution operation we get that in your quantized network you basically do the next operation: Y' = W'*X = (Wa+b)*X
Because convolution is linear we get: Y' = a(W*X) + b*X'
Don't forget that in your network you want to receive Y not Y' at the output of the convolution therefore you must do shift + re scale to get the correct answer.
So after that explanation (which i hope was clear enough) i hope you can understand what is the problem in your network, you do this scale and shift to all of weights and you never compensate for it, I think your confusion is because your read papers that trained models in quantized mode from the beginning and didn't take pretrained model quantized it.
For you problem i think tensorflow graph transform tool might help, take a look at:
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/graph_transforms/README.md
If you wish to read more about quantizing pre trained model you can find more information in (for more academic info just go to scholar.google.com:
https://www.tensorflow.org/lite/performance/post_training_quantization
Related
Could someone please help me, in semantic segmentation tasks, should the image and mask be normalized in the batch generator class or only one of them should be normalized?
I'm using the following code to normalize image and mask:
mean_val, std_val = img.mean(), img.std()
img = (img - mean_val)/std_val
for example :
here image and corresponding masks are normalized for the prostate cancer segmentation task
while here only the masks are normalized
which one is the correct practice?
def __getitem__(self,i):
index= self.indexes[i * self.batch_size : (i + 1) * self.batch_size]
X = np.empty((self.batch_size, self.crop_dim[0], self.crop_dim[1],3)).astype(np.uint8)
Y = np.empty((self.batch_size, self.crop_dim[0], self.crop_dim[1],5)).astype(np.uint8)
for i,ID in enumerate(index):
dim= (self.crop_dim[0],self.crop_dim[1])
img=cv2.imread(self.img_list[ID],cv2.COLOR_BGR2RGB)
img = cv2.resize(img,dim)
mask=imageio.imread(self.labels[ID],as_gray=False, pilmode="RGB")
mask = cv2.resize(mask,dim)
mask= create_labels(mask)
# Augement training patches only
if self.augmentation:
sample = self.augmentation(image=img_numpy, mask=mask_numpy)
img_numpy, mask_numpy = sample['image'], sample['mask']
mean_val, std_val = img.mean(), img.std()
img = (img - mean_val)/std_val
mean_val_mask, std_val_mask = mask.mean(), mask.std()
mask= (mask - mean_val_mask)/std_val_mask
X[i,]=img_numpy
Y[i,]=mask_numpy
No!
You do not want to normalize the labels - you want to predict them directly, you do not want the target (per pixel) to change based on the global statistics of the mask.
Why normalizing?
It is common practice to normalize the inputs to a DNN model. This is motivated by the desire to control the "dynamic range" of the activations at different layers. This, in turn, helps the optimization process to converge in a more rapid and stable manner.
You can find an in-depth analysis of this normalization in this excellent paper:
He, K., Zhang, X., Ren, S. and Sun, J., Delving deep into rectifiers: Surpassing human-level performance on imagenet classification (ICCV 2015).
This rationale does not apply to the labels.
I have been trying to understand a blog on soft actor critic where we have a neural network representing a policy that outputs mean and std of gaussian distribution of action for a given state. Since direct back-propagation through stochastic node is not possible , reparamterization trick is applied as follows:
`normal = Normal(0, 1)
z = normal.sample()
action = torch.tanh(mean+ std*z.to(device))
log_prob = Normal(mean, std).log_prob(mean+ std*z.to(device)) - torch.log(1 - action.pow(2) + epsilon)
return action, log_prob, z, mean, log_std`
I want to know how the log_prob term was derived. Any help would be highly appreciated.
I have been tinkering around a lot with tensorflow in the past few days however I am quite unsure whether a function I wrote would break the backpropagation in a Neural network. I thought I'd ask here before I try to integrate this function in a NN. So the basic setup is I want to add two matricies with
op = tf.add(tfObject, tfImageBackground)
where tfImageBackground is some constant image. (i.e. an RGBA image of size 800, 800 with R = G = B = A = 0) and the tfObject is again a matrix with the same dimenstion however we get that with the function I am unsure about
def getObject(vector):
objectId = vector[0]
x = vector[1]
y = vector[2]
xEnd = baseImageSize-(x+objectSize)
yStart =baseImageSize- (y+objectSize)
padding = tf.convert_to_tensor([[x, xEnd], [yStart, y],[0,0]])
RTensor = tfObjectMatrix[objectId,:,:,0:1]
GTensor = tfObjectMatrix[objectId,:,:,1:2]
BTensor = tfObjectMatrix[objectId,:,:,2:3]
ATensor = tfObjectMatrix[objectId,:,:,3:4]
paddedR = tf.pad(tensor = RTensor,
paddings= padding,
mode='Constant',
name='padAverageRed',
constant_values=255)
...
generates padding for every channel
...
finalTensor=tf.concat([paddedR, paddedG, paddedB, paddedA], 2)
return finalTensor
The tfObjectMatrix is a list of images which never change.
I did check wether I was able to generate a tf.gradient from the op, which turned out to work. I am unsure if that is sufficient for backpropagation to work though.
Thanks for you time and effort. Any input at all would be greatly appreciated.
TensorFlow will backpropagate to everything by default. As per your code, everything will receive gradients with a training operation from an optimizer. So to answer your question, backpropagation will work.
The only thing to consider, is that you say tfObjectMatrix is a list of images that will not change. So you might not want it to receive any gradients. Therefore you might want to look into tf.stop_gradient() and maybe use it like OM = tf.stop_gradient( tfObjectMatrix ) and work with that OM in your function.
I implemented Fully-Convolution Network at TensorFlow. It use encdoder-decoder structure.
When training, I use always same image size (224x224, using random crop) and everything works nicely.
In interference phase, I want to predict one image at a time, because I want to use full-image (not croped). For example, such image have size [406,256]. And here is problem.
In Encoder-Decoder architecture I add two tesors (z = x + y). When training, sizes of both tensor matches. When predicting my single image, sizes does not match (tensor sizes: [1,47,47,64] vs [1,46,46,64]). I think it is cause by some rounding done in Conv and Pool layer.
What should I change in my architecture to works for any image size I want? Should I change rounding parameters? Or add 'cropping' of tensor?
Link to implementation of architecture:
https://gist.github.com/melgor/0e43cadf742fe3336148ab64dd63138f
(the problem occur in line 166)
I found the solution for variable input size:)
What we really need was a 'Crop-layer', that crop one tensor to match other. I found really similar layer here: http://tf-unet.readthedocs.io/en/latest/_modules/tf_unet/layers.html
(crop_and_concat).
I have just made it `crop_and_add' and it is working:
def crop_and_add(x1,x2):
x1_shape = tf.shape(x1)
x2_shape = tf.shape(x2)
# offsets for the top left corner of the crop
offsets = [0, (x1_shape[1] - x2_shape[1]) // 2, (x1_shape[2] - x2_shape[2]) // 2, 0]
size = [-1, x2_shape[1], x2_shape[2], -1]
x1_crop = tf.slice(x1, offsets, size)
return x1_crop + x2
All addition in model I replaced by above layer (so merging encoder and decoder data).
Also, the input to model need to be defined as:
image = tf.placeholder(tf.float32, shape=[1, None, None, 3], name="input_image")
So we know that we will pass single image and that image have 3 channels. but we do not know neither width nor height. And it works very nice! (40 FPS on K80 as AWS P2, size of image is 224x{}-shoter side of image have 224)
FYI, I was also trying to run ENET (2x faster than LinkNet), but in TensorFlow it is slower. I think it is because of PReLu (which is slow at TF). Also it does not support arbitraty size of image becauese of UnPool layer, which need to have predefined output size by list of integers (not placeholders). So LinkNet look better in case of Speed and Performacance in TF.
In Tensorflow, I've wrote a big model for 2 image classes problem. My question is concerned with the following code snippet:
X, y, X_val, y_val = prepare_data()
probs = calc_probs(model, session, X)
accuracy = float(np.equal(np.argmax(probs, 1), np.argmax(y, 1)).sum()) / probs.shape[0]
loss = log_loss(y, probs)
X is an np.array of shape: (25000,244,244,3). That code results in accuracy=0.5834 (towards random accuracy) and loss=2.7106. But
when I shuffle the data, by adding these 3 lines after the first line:
sample_idx = random.sample(range(0, X.shape[0]), 25000)
X = X[sample_idx]
y = y[sample_idx]
, the results become convenient: accuracy=0.9933 and loss=0.0208.
Why shuffling data can give significantly higher accuracy ? or what can be a reason for that ?
The function calc_probs is mainly a run call:
probs = session.run(model.probs, feed_dict={model.X: X})
Update:
After hours of debugging, I figured out that evaluating a single image gives different result. For example, if you run the following line of code multiple times, you get a different result each time:
session.run(model.props, feed_dict={model.X: [X[20]])
My data is normally sorted, X contains class 1 samples first then class 2. And in calc_probs function, I run using each batch of the data sequentially. So, without shuffling, each run has data of a single class.
I've also noted that with shuffling, if batch size is very small, I get the random accuracy.
There is some mathematical justification for this in the context of randomized Kaczmarz algorithm. Regular Kaczmarz algorithm is an old algorithm which can be seen as an non-shuffling SGD on a least squares problem, and there are guaranteed faster convergence rates that come out if you use randomization, follow references in http://www.cs.ubc.ca/~nickhar/W15/Lecture21Notes.pdf