I have a function dice
def dice(yPred,yTruth,thresh):
smooth = tf.constant(1.0)
threshold = tf.constant(thresh)
yPredThresh = tf.to_float(tf.greater_equal(yPred,threshold))
mul = tf.mul(yPredThresh,yTruth)
intersection = 2*tf.reduce_sum(mul) + smooth
union = tf.reduce_sum(yPredThresh) + tf.reduce_sum(yTruth) + smooth
dice = intersection/union
return dice, yPredThresh
which works. An example is given here
with tf.Session() as sess:
thresh = 0.5
print("Dice example")
yPred = tf.constant([0.1,0.9,0.7,0.3,0.1,0.1,0.9,0.9,0.1],shape=[3,3])
yTruth = tf.constant([0.0,1.0,1.0,0.0,0.0,0.0,1.0,1.0,1.0],shape=[3,3])
diceScore, yPredThresh= dice(yPred=yPred,yTruth=yTruth,thresh= thresh)
diceScore_ , yPredThresh_ , yPred_, yTruth_ = sess.run([diceScore,yPredThresh,yPred, yTruth])
print("\nScore = {0}".format(diceScore_))
>>> Score = 0.899999976158
I would like to be able to loop over the third arguement of dice, thresh. I do not know the best way to do this such that I can extract it from the graph. Something along the lines of the following...
def diceROC(yPred,yTruth,thresholds=np.linspace(0.1,0.9,20)):
thresholds = thresholds.astype(np.float32)
nThreshs = thresholds.size
diceScores = tf.zeros(shape=nThreshs)
for i in xrange(nThreshs):
score,_ = dice(yPred,yTruth,thresholds[i])
diceScores[i] = score
return diceScores
Evaluating diceScoreROC yields the error 'Tensor' object does not support item assignment as I can't loop into and slice a tf tensor apparently.
Instead of the loop, I would encourage you to use broadcasting abilities of tensorflow. If you redefine dice to:
def dice(yPred,yTruth,thresh):
smooth = tf.constant(1.0)
yPredThresh = tf.to_float(tf.greater_equal(yPred,thresh))
mul = tf.mul(yPredThresh,yTruth)
intersection = 2*tf.reduce_sum(mul, [0, 1]) + smooth
union = tf.reduce_sum(yPredThresh, [0, 1]) + tf.reduce_sum(yTruth, [0, 1]) + smooth
dice = intersection/union
return dice, yPredThresh
You will be able to pass 3-dimensional yPred and yTruth (assuming the tensors will be just repeated along the last dimension) and 1-dimensional thresh:
with tf.Session() as sess:
thresh = [0.1,0.9,20, 0.5]
print("Dice example")
yPred = tf.constant([0.1,0.9,0.7,0.3,0.1,0.1,0.9,0.9,0.1],shape=[3,3,1])
ypred_tiled = tf.tile(yPred, [1,1,4])
yTruth = tf.constant([0.0,1.0,1.0,0.0,0.0,0.0,1.0,1.0,1.0],shape=[3,3,1])
ytruth_tiled = tf.tile(yTruth, [1,1,4])
diceScore, yPredThresh= dice(yPred=ypred_tiled,yTruth=ytruth_tiled,thresh= thresh)
diceScore_ = sess.run(diceScore)
print("\nScore = {0}".format(diceScore_))
You'll get:
Score = [ 0.73333335 0.77777779 0.16666667 0.89999998]
Related
I am trying to implement a custom loss function in Tensorflow 2.4 using the Keras backend.
The loss function is a ranking loss; I found the following paper with a somewhat log-likelihood loss: Chen et al. Single-Image Depth Perception in the Wild.
Similarly, I wanted to sample some (in this case 50) points from an image to compare the relative order between ground-truth and predicted depth maps using the NYU-Depth dataset. Being a fan of Numpy, I started working with that but came to the following exception:
ValueError: No gradients provided for any variable: [...]
I have learned that this is caused by the arguments not being filled when calling the loss function but instead, a C function is compiled which is then used later. So while I know the dimensions of my tensors (4, 480, 640, 1), I cannot work with the data as wanted and have to use the keras.backend functions on top so that in the end (if I understood correctly), there is supposed to be a path between the input tensors from the TF graph and the output tensor, which has to provide a gradient.
So my question now is: Is this a feasible loss function within keras?
I have already tried a few ideas and different approaches with different variations of my original code, which was something like:
def ranking_loss_function(y_true, y_pred):
# Chen et al. loss
y_true_np = K.eval(y_true)
y_pred_np = K.eval(y_pred)
if y_true_np.shape[0] != None:
num_sample_points = 50
total_samples = num_sample_points ** 2
err_list = [0 for x in range(y_true_np.shape[0])]
for i in range(y_true_np.shape[0]):
sample_points = create_random_samples(y_true, y_pred, num_sample_points)
for x1, y1 in sample_points:
for x2, y2 in sample_points:
if y_true[i][x1][y1] > y_true[i][x2][y2]:
#image_relation_true = 1
err_list[i] += np.log(1 + np.exp(-1 * y_pred[i][x1][y1] + y_pred[i][x2][y2]))
elif y_true[i][x1][y1] < y_true[i][x2][y2]:
#image_relation_true = -1
err_list[i] += np.log(1 + np.exp(y_pred[i][x1][y1] - y_pred[i][x2][y2]))
else:
#image_relation_true = 0
err_list[i] += np.square(y_pred[i][x1][y1] - y_pred[i][x2][y2])
err_list = np.divide(err_list, total_samples)
return K.constant(err_list)
As you can probably tell, the main idea was to first create the sample points and then based on the existing relation between them in y_true/y_pred continue with the corresponding computation from the cited paper.
Can anyone help me and provide some more helpful information or tips on how to correctly implement this loss using keras.backend functions? Trying to include the ordinal relation information really confused me compared to standard regression losses.
EDIT: Just in case this causes confusion: create_random_samples() just creates 50 random sample points (x, y) coordinate pairs based on the shape[1] and shape[2] of y_true (image width and height)
EDIT(2): After finding this variation on GitHub, I have tried out a variation using only TF functions to retrieve data from the tensors and compute the output. The adjusted and probably more correct version still throws the same exception though:
def ranking_loss_function(y_true, y_pred):
#In the Wild ranking loss
y_true_np = K.eval(y_true)
y_pred_np = K.eval(y_pred)
if y_true_np.shape[0] != None:
num_sample_points = 50
total_samples = num_sample_points ** 2
bs = y_true_np.shape[0]
w = y_true_np.shape[1]
h = y_true_np.shape[2]
total_samples = total_samples * bs
num_pairs = tf.constant([total_samples], dtype=tf.float32)
output = tf.Variable(0.0)
for i in range(bs):
sample_points = create_random_samples(y_true, y_pred, num_sample_points)
for x1, y1 in sample_points:
for x2, y2 in sample_points:
y_true_sq = tf.squeeze(y_true)
y_pred_sq = tf.squeeze(y_pred)
d1_t = tf.slice(y_true_sq, [i, x1, y1], [1, 1, 1])
d2_t = tf.slice(y_true_sq, [i, x2, y2], [1, 1, 1])
d1_p = tf.slice(y_pred_sq, [i, x1, y1], [1, 1, 1])
d2_p = tf.slice(y_pred_sq, [i, x2, y2], [1, 1, 1])
d1_t_sq = tf.squeeze(d1_t)
d2_t_sq = tf.squeeze(d2_t)
d1_p_sq = tf.squeeze(d1_p)
d2_p_sq = tf.squeeze(d2_p)
if d1_t_sq > d2_t_sq:
# --> Image relation = 1
output.assign_add(tf.math.log(1 + tf.math.exp(-1 * d1_p_sq + d2_p_sq)))
elif d1_t_sq < d2_t_sq:
# --> Image relation = -1
output.assign_add(tf.math.log(1 + tf.math.exp(d1_p_sq - d2_p_sq)))
else:
output.assign_add(tf.math.square(d1_p_sq - d2_p_sq))
return output/num_pairs
EDIT(3): This is the code for create_random_samples():
(FYI: Because it was weird to get the shape from y_true in this case, I first proceeded to hard-code it here as I know it for the dataset which I am currently using.)
def create_random_samples(y_true, y_pred, num_points=50):
y_true_shape = (4, 480, 640, 1)
y_pred_shape = (4, 480, 640, 1)
if y_true_shape[0] != None:
num_samples = num_points
population = [(x, y) for x in range(y_true_shape[1]) for y in range(y_true_shape[2])]
sample_points = random.sample(population, num_samples)
return sample_points
I want to create a human pose skeleton estimation network and for this, I have a two-part network, first part generates 16 heatmaps as output(each heatmap for different joint and hence a key point can be extracted), using these 16 key points I wish to create a human skeleton and feed it to second half of my network. My problem is, how do I draw lines between the key points to create the skeleton? I couldn't find a way to do it on a tensor object using tensorflow or keras.
I know i'm a bit late but here is some code that I think does what you're after (in TFv2.3). Hopefully it will save someone time in the future!
It uses solely tensorflow ops, so you can use it in data loaders etc. The real pain here is that Tensorflow doesn't allow Eager Assignment, so you can't just update tensors by index. This works around that by creating two sparse tensors, one for the mask (where to apply the line) and another for the new_values (what value to apply at the line). The code for simply designing the line might not be applicable in your case (based on https://stackoverflow.com/a/47381058) but ported away from numpy.
import tensorflow as tf
def trapez(y, y0, w):
return tf.clip_by_value(tf.minimum(y + 1 + w/2 - y0, -y + 1 + w/2 + y0), 0, 1)
def apply_output(img, yy, xx, val):
stack = tf.stack([yy, xx], axis=1)
stack = tf.cast(stack, tf.int64)
values = tf.ones(stack.shape[0], tf.float32)
mask = tf.sparse.SparseTensor(indices=stack, values=values, dense_shape=img.shape)
mask = tf.sparse.reorder(mask)
mask = tf.sparse.to_dense(mask)
mask = tf.cast(mask, tf.float32)
new_values = tf.sparse.SparseTensor(indices=stack, values=val, dense_shape=img.shape)
new_values = tf.sparse.reorder(new_values)
new_values = tf.sparse.to_dense(new_values)
img = img * (1 - mask) + new_values * mask
img = tf.cast(tf.expand_dims(img * 255, axis=-1), tf.uint8)
return img
def weighted_line(img, r0, c0, r1, c1, w):
output = img
x = tf.range(c0, c1 + 1, dtype=tf.float32)
slope = (r1-r0) / (c1-c0)
w *= tf.sqrt(1 + tf.abs(slope)) / 2
y = x * slope + (c1*r0-c0*r1) / (c1-c0)
thickness = tf.math.ceil(w/2)
yy = (tf.reshape(tf.math.floor(y), [-1, 1]) + tf.reshape(tf.range(-thickness-1, thickness+2), [1, -1]))
xx = tf.repeat(x, yy.shape[1])
values = tf.reshape(trapez(yy, tf.reshape(y, [-1, 1]), w), [-1])
yy = tf.reshape(yy, [-1])
limits_y = tf.math.logical_and(yy >= 0, yy < img.shape[0])
limits_x = tf.math.logical_and(xx >= 0, xx < img.shape[1])
limits = tf.math.logical_and(limits_y, limits_x)
limits = tf.math.logical_and(limits, values > 0)
yy = tf.cast(yy[limits], tf.float32)
xx = tf.cast(xx[limits], tf.float32)
return yy, xx, values[limits], apply_output(output, yy, xx, values[limits])
Just for a sanity check, you can call it with the following, and display it using opencv
if __name__ == "__main__":
IMG = tf.zeros((500, 500), tf.float32)
yy, xx, vals, FINAL_IMG = weighted_line(IMG, 10, 20, 100, 200, 5)
jpeg_string = tf.io.encode_jpeg(FINAL_IMG)
tf.io.write_file("output.jpg", jpeg_string)
import cv2
img = cv2.imread("output.jpg")
cv2.imshow("Output", img)
cv2.waitKey(0)
I'm following this Repo on creating Yolo v3 model from scratch in PyTorch. The only problem is that the bounding boxes are not as tight (close to the objects) in most images I tried. I compared them to the tutorial on creating Yolo v3 model but using TensorFlow. The tensorflow model produces excellent bounding boxed that are as tight as possible to the objects.
I tried to understand how the calculations are different between the two, but I'm finding myself getting stuck with the differences between torch and tf.
I believe the code for the bounding boxes in the tf tutorial comes from here:
def yolo_layer(inputs, n_classes, anchors, img_size, data_format):
"""Creates Yolo final detection layer.
Detects boxes with respect to anchors.
Args:
inputs: Tensor input.
n_classes: Number of labels.
anchors: A list of anchor sizes.
img_size: The input size of the model.
data_format: The input format.
Returns:
Tensor output.
"""
n_anchors = len(anchors)
inputs = tf.layers.conv2d(inputs, filters=n_anchors * (5 + n_classes),
kernel_size=1, strides=1, use_bias=True,
data_format=data_format)
shape = inputs.get_shape().as_list()
grid_shape = shape[2:4] if data_format == 'channels_first' else shape[1:3]
if data_format == 'channels_first':
inputs = tf.transpose(inputs, [0, 2, 3, 1])
inputs = tf.reshape(inputs, [-1, n_anchors * grid_shape[0] * grid_shape[1],
5 + n_classes])
strides = (img_size[0] // grid_shape[0], img_size[1] // grid_shape[1])
box_centers, box_shapes, confidence, classes = \
tf.split(inputs, [2, 2, 1, n_classes], axis=-1)
x = tf.range(grid_shape[0], dtype=tf.float32)
y = tf.range(grid_shape[1], dtype=tf.float32)
x_offset, y_offset = tf.meshgrid(x, y)
x_offset = tf.reshape(x_offset, (-1, 1))
y_offset = tf.reshape(y_offset, (-1, 1))
x_y_offset = tf.concat([x_offset, y_offset], axis=-1)
x_y_offset = tf.tile(x_y_offset, [1, n_anchors])
x_y_offset = tf.reshape(x_y_offset, [1, -1, 2])
box_centers = tf.nn.sigmoid(box_centers)
box_centers = (box_centers + x_y_offset) * strides
anchors = tf.tile(anchors, [grid_shape[0] * grid_shape[1], 1])
box_shapes = tf.exp(box_shapes) * tf.to_float(anchors)
confidence = tf.nn.sigmoid(confidence)
classes = tf.nn.sigmoid(classes)
inputs = tf.concat([box_centers, box_shapes,
confidence, classes], axis=-1)
return inputs
While the code for the bounding boxes for the pytorch model comes from here, and the explanation:
def bbox_iou(box1, box2):
"""
Returns the IoU of two bounding boxes
"""
#Get the coordinates of bounding boxes
b1_x1, b1_y1, b1_x2, b1_y2 = box1[:,0], box1[:,1], box1[:,2], box1[:,3]
b2_x1, b2_y1, b2_x2, b2_y2 = box2[:,0], box2[:,1], box2[:,2], box2[:,3]
#get the corrdinates of the intersection rectangle
inter_rect_x1 = torch.max(b1_x1, b2_x1)
inter_rect_y1 = torch.max(b1_y1, b2_y1)
inter_rect_x2 = torch.min(b1_x2, b2_x2)
inter_rect_y2 = torch.min(b1_y2, b2_y2)
#Intersection area
inter_area = torch.clamp(inter_rect_x2 - inter_rect_x1 + 1, min=0) * torch.clamp(inter_rect_y2 - inter_rect_y1 + 1, min=0)
#Union Area
b1_area = (b1_x2 - b1_x1 + 1)*(b1_y2 - b1_y1 + 1)
b2_area = (b2_x2 - b2_x1 + 1)*(b2_y2 - b2_y1 + 1)
iou = inter_area / (b1_area + b2_area - inter_area)
return iou
def predict_transform(prediction, inp_dim, anchors, num_classes, CUDA = True):
batch_size = prediction.size(0)
stride = inp_dim // prediction.size(2)
grid_size = inp_dim // stride
bbox_attrs = 5 + num_classes
num_anchors = len(anchors)
prediction = prediction.view(batch_size, bbox_attrs*num_anchors, grid_size*grid_size)
prediction = prediction.transpose(1,2).contiguous()
prediction = prediction.view(batch_size, grid_size*grid_size*num_anchors, bbox_attrs)
anchors = [(a[0]/stride, a[1]/stride) for a in anchors]
#Sigmoid the centre_X, centre_Y. and object confidencce
prediction[:,:,0] = torch.sigmoid(prediction[:,:,0])
prediction[:,:,1] = torch.sigmoid(prediction[:,:,1])
prediction[:,:,4] = torch.sigmoid(prediction[:,:,4])
#Add the center offsets
grid = np.arange(grid_size)
a,b = np.meshgrid(grid, grid)
x_offset = torch.FloatTensor(a).view(-1,1)
y_offset = torch.FloatTensor(b).view(-1,1)
if CUDA:
x_offset = x_offset.cuda()
y_offset = y_offset.cuda()
x_y_offset = torch.cat((x_offset, y_offset), 1).repeat(1,num_anchors).view(-1,2).unsqueeze(0)
prediction[:,:,:2] += x_y_offset
#log space transform height and the width
anchors = torch.FloatTensor(anchors)
if CUDA:
anchors = anchors.cuda()
anchors = anchors.repeat(grid_size*grid_size, 1).unsqueeze(0)
prediction[:,:,2:4] = torch.exp(prediction[:,:,2:4])*anchors
prediction[:,:,5: 5 + num_classes] = torch.sigmoid((prediction[:,:, 5 : 5 + num_classes]))
prediction[:,:,:4] *= stride
return prediction
I'm trying to build a convolutional lstm autoencoder (that also predicts future and past) with Tensorflow, and it works to a certain degree, but the error sometimes jumps back up, so essentially, it never converges.
The model is as follows:
The encoder starts with a 64x64 frame from a 20 frame bouncing mnist video for each time step of the lstm. Every stacking layer of LSTM halfs it and increases the depth via 2x2 convolutions with a stride of 2. (so -->32x32x3 -->...--> 1x1x96)
On the other hand, the lstm performs 3x3 convolutions with a stride of 1 on its state. Both results are concatenated to form the new state. In the same way, the decoder uses transposed convolutions to go back to the original format. Then the squared error is calculated.
The error starts at around 2700 and it takes around 20 hours (geforce1060) to get down to ~1700. At which point the jumping back up (and it sometimes jumps back up to 2300 or even ridiculous values like 440300) happens often enough that I can't really get any lower. Also at that point, it can usually pinpoint where the number should be, but its too fuzzy to actually make out the digit...
I tried different learning rates and optimizers, so if anybody knows why that jumping happens, that'd make me happy :)
Here is a graph of the loss with epochs:
import tensorflow as tf
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
import os
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
#based on code by loliverhennigh (Github)
class ConvCell(tf.contrib.rnn.RNNCell):
count = 0 #exists only to remove issues with variable scope
def __init__(self, shape, num_features, transpose = False):
self.shape = shape
self.num_features = num_features
self._state_is_tuple = True
self._transpose = transpose
ConvCell.count+=1
self.count = ConvCell.count
#property
def state_size(self):
return (tf.contrib.rnn.LSTMStateTuple(self.shape[0:4],self.shape[0:4]))
#property
def output_size(self):
return tf.TensorShape(self.shape[1:4])
#here comes to the actual conv lstm implementation, if transpose = true, it performs a deconvolution on the input
def __call__(self, inputs, state, scope=None):
with tf.variable_scope(scope or type(self).__name__+str(self.count)):
c, h = state
state_shape = h.shape
input_shape = inputs.shape
#filter variables and convolutions on data coming from the same cell, a time step previous
h_filters = tf.get_variable("h_filters",[3,3,state_shape[3],self.num_features])
h_filters_gates = tf.get_variable("h_filters_gates",[3,3,state_shape[3],3])
h_partial = tf.nn.conv2d(h,h_filters,[1,1,1,1],'SAME')
h_partial_gates = tf.nn.conv2d(h,h_filters_gates,[1,1,1,1],'SAME')
c_filters = tf.get_variable("c_filters",[3,3,state_shape[3],3])
c_partial = tf.nn.conv2d(c,c_filters,[1,1,1,1],'SAME')
#filters and convolutions/deconvolutions on data coming fromthe cell input
if self._transpose:
x_filters = tf.get_variable("x_filters",[2,2,self.num_features,input_shape[3]])
x_filters_gates = tf.get_variable("x_filters_gates",[2,2,3,input_shape[3]])
x_partial = tf.nn.conv2d_transpose(inputs,x_filters,[int(state_shape[0]),int(state_shape[1]),int(state_shape[2]),self.num_features],[1,2,2,1],'VALID')
x_partial_gates = tf.nn.conv2d_transpose(inputs,x_filters_gates,[int(state_shape[0]),int(state_shape[1]),int(state_shape[2]),3],[1,2,2,1],'VALID')
else:
x_filters = tf.get_variable("x_filters",[2,2,input_shape[3],self.num_features])
x_filters_gates = tf.get_variable("x_filters_gates",[2,2,input_shape[3],3])
x_partial = tf.nn.conv2d(inputs,x_filters,[1,2,2,1],'VALID')
x_partial_gates = tf.nn.conv2d(inputs,x_filters_gates,[1,2,2,1],'VALID')
#some more lstm gate business
gate_bias = tf.get_variable("gate_bias",[1,1,1,3])
h_bias = tf.get_variable("h_bias",[1,1,1,self.num_features*2])
gates = h_partial_gates + x_partial_gates + c_partial + gate_bias
i,f,o = tf.split(gates,3,axis=3)
#concatenate the units coming from the spacial and the temporal dimension to build a unified state
concat = tf.concat([h_partial,x_partial],3) + h_bias
new_c = tf.nn.relu(concat)*tf.sigmoid(i)+c*tf.sigmoid(f)
new_h = new_c * tf.sigmoid(o)
new_state = tf.contrib.rnn.LSTMStateTuple(new_c,new_h)
return new_h, new_state #its redundant, but thats how tensorflow likes it, apparently
#global variables
LEARNING_RATE = 0.005
ITERATIONS_PER_EPOCH = 80
BATCH_SIZE = 75
TEST = False #manual switch to go from training to testing
if TEST:
BATCH_SIZE = 1
inputs = tf.placeholder(tf.float32, (20, BATCH_SIZE, 64, 64,1))
shape0 = [BATCH_SIZE,64,64,2]
shape1 = [BATCH_SIZE,32,32,6]
shape2 = [BATCH_SIZE,16,16,12]
shape3 = [BATCH_SIZE,8,8,24]
shape4 = [BATCH_SIZE,4,4,48]
shape5 = [BATCH_SIZE,2,2,96]
shape6 = [BATCH_SIZE,1,1,192]
#apparently tf.multirnncell has very specific requirements for the initial states oO
initial_state1 = (tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape1),tf.zeros(shape1)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape2),tf.zeros(shape2)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape3),tf.zeros(shape3)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape4),tf.zeros(shape4)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape5),tf.zeros(shape5)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape6),tf.zeros(shape6)))
initial_state2 = (tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape5),tf.zeros(shape5)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape4),tf.zeros(shape4)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape3),tf.zeros(shape3)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape2),tf.zeros(shape2)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape1),tf.zeros(shape1)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape0),tf.zeros(shape0)))
#encoding part of the autoencoder graph
cell1 = ConvCell(shape1,3)
cell2 = ConvCell(shape2,6)
cell3 = ConvCell(shape3,12)
cell4 = ConvCell(shape4,24)
cell5 = ConvCell(shape5,48)
cell6 = ConvCell(shape6,96)
mcell = tf.contrib.rnn.MultiRNNCell([cell1,cell2,cell3,cell4,cell5,cell6])
rnn_outputs, rnn_states = tf.nn.dynamic_rnn(mcell, inputs[0:20,:,:,:],initial_state=initial_state1,dtype=tf.float32, time_major=True)
#decoding part of the autoencoder graph, forward block and backwards block
cell9a = ConvCell(shape5,48,transpose = True)
cell10a = ConvCell(shape4,24,transpose = True)
cell11a = ConvCell(shape3,12,transpose = True)
cell12a = ConvCell(shape2,6,transpose = True)
cell13a = ConvCell(shape1,3,transpose = True)
cell14a = ConvCell(shape0,1,transpose = True)
mcella = tf.contrib.rnn.MultiRNNCell([cell9a,cell10a,cell11a,cell12a,cell13a,cell14a])
cell9b = ConvCell(shape5,48,transpose = True)
cell10b = ConvCell(shape4,24,transpose = True)
cell11b= ConvCell(shape3,12,transpose = True)
cell12b = ConvCell(shape2,6,transpose = True)
cell13b = ConvCell(shape1,3,transpose = True)
cell14b = ConvCell(shape0,1,transpose = True)
mcellb = tf.contrib.rnn.MultiRNNCell([cell9b,cell10b,cell11b,cell12b,cell13b,cell14b])
def PredictionLayer(rnn_outputs,viewPoint = 11, reverse = False):
predLength = viewPoint-2 if reverse else 20-viewPoint #vision is the input for the decoder
vision = tf.concat([rnn_outputs[viewPoint-1:viewPoint,:,:,:],tf.zeros([predLength,BATCH_SIZE,1,1,192])],0)
if reverse:
rnn_outputs2, rnn_states = tf.nn.dynamic_rnn(mcellb, vision, initial_state = initial_state2, time_major=True)
else:
rnn_outputs2, rnn_states = tf.nn.dynamic_rnn(mcella, vision, initial_state = initial_state2, time_major=True)
mean = tf.reduce_mean(rnn_outputs2,4)
if TEST:
return mean
if reverse:
return tf.reduce_sum(tf.square(mean-inputs[viewPoint-2::-1,:,:,:,0]))
else:
return tf.reduce_sum(tf.square(mean-inputs[viewPoint-1:20,:,:,:,0]))
if TEST:
mean = tf.concat([PredictionLayer(rnn_outputs,11,True)[::-1,:,:,:],createPredictionLayer(rnn_outputs,11)],0)
else: #training part of the graph
error = tf.zeros([1])
for i in range(8,15): #range size of 7 or less works, 9 or more does not, no idea why
error += PredictionLayer(rnn_outputs, i)
error += PredictionLayer(rnn_outputs, i, True)
train_fn = tf.train.RMSPropOptimizer(learning_rate=LEARNING_RATE).minimize(error)
################################################################################
## TRAINING LOOP ##
################################################################################
#code based on siemanko/tf_lstm.py (Github)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
saver = tf.train.Saver(restore_sequentially=True, allow_empty=True,)
session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
session.run(tf.global_variables_initializer())
vids = np.load("mnist_test_seq.npy") #20/10000/64/64 , moving mnist dataset from http://www.cs.toronto.edu/~nitish/unsupervised_video/
vids = vids[:,0:6000,:,:] #training set
saver.restore(session,tf.train.latest_checkpoint('./conv_lstm_multiples_v2/'))
#saver.restore(session,'.\conv_lstm_multiples\iteration-74')
for epoch in range(1000):
if TEST:
break
epoch_error = 0
#randomize batches each epoch
vids = np.swapaxes(vids,0,1)
np.random.shuffle(vids)
vids = np.swapaxes(vids,0,1)
for i in range(ITERATIONS_PER_EPOCH):
#running the graph and feeding data
err,_ = session.run([error, train_fn], {inputs: np.expand_dims(vids[:,i*BATCH_SIZE:(i+1)*BATCH_SIZE,:,:],axis=4)})
print(err)
epoch_error += err
#training error each epoch and regular saving
epoch_error /= (ITERATIONS_PER_EPOCH*BATCH_SIZE*4096*20*7)
if (epoch+1) % 5 == 0:
saver.save(session,'.\conv_lstm_multiples_v2\iteration',global_step=epoch)
print("saved")
print("Epoch %d, train error: %f" % (epoch, epoch_error))
#testing
plt.ion()
f, axarr = plt.subplots(2)
vids = np.load("mnist_test_seq.npy")
for i in range(6000,10000):
img = session.run([mean], {inputs: np.expand_dims(vids[:,i:i+1,:,:],axis=4)})
for j in range(20):
axarr[0].imshow(img[0][j,0,:,:])
axarr[1].imshow(vids[j,i,:,:])
plt.show()
plt.pause(0.1)
Usually this happens when gradients' magnitude is very high at some point and causes your network parameters to change a lot. To verify that it is indeed the case, you can produce the same plot of gradient magnitudes and see if they jump right before the loss jump. Assuming this is the case, the classic approach is to use gradient clipping (or go all the way to natural gradient).
The following snippet is from a fairly large piece of code but hopefully I can give all the information necessary:
y2 = tf.matmul(y1,ymask)
dist = tf.norm(ystar-y2,axis=0)
y1 and y2 are 128x30 and ymask is 30x30. ystar is 128x30. dist is 1x30. When ymask is the identity matrix, everything works fine. But when I set it to be all zeros, apart from a single 1 along the diagonal (so as to set all columns but one in y2 to be zero), I get nans for the gradient of dist with respect to y2, using tf.gradients(dist, [y2]). The specific value of dist is [0,0,7.9,0,...], with all the ystar-y2 values being around the range (-1,1) in the third column and zero elsewhere.
I'm pretty confused as to why a numerical issue would occur here, given there are no logs or divisions, is this underflow? Am I missing something in the maths?
For context, I'm doing this to try to train individual dimensions of y, one at a time, using the whole network.
longer version to reproduce:
import tensorflow as tf
import numpy as np
import pandas as pd
batchSize = 128
eta = 0.8
tasks = 30
imageSize = 32**2
groups = 3
tasksPerGroup = 10
trainDatapoints = 10000
w = np.zeros([imageSize, groups * tasksPerGroup])
toyIndex = 0
for toyLoop in range(groups):
m = np.ones([imageSize]) * np.random.randn(imageSize)
for taskLoop in range(tasksPerGroup):
w[:, toyIndex] = m * 0.1 * np.random.randn(1)
toyIndex += 1
xRand = np.random.normal(0, 0.5, (trainDatapoints, imageSize))
taskLabels = np.matmul(xRand, w) + np.random.normal(0,0.5,(trainDatapoints, groups * tasksPerGroup))
DF = np.concatenate((xRand, taskLabels), axis=1)
trainDF = pd.DataFrame(DF[:trainDatapoints, ])
# define graph variables
x = tf.placeholder(tf.float32, [None, imageSize])
W = tf.Variable(tf.zeros([imageSize, tasks]))
b = tf.Variable(tf.zeros([tasks]))
ystar = tf.placeholder(tf.float32, [None, tasks])
ymask = tf.placeholder(tf.float32, [tasks, tasks])
dataLength = tf.cast(tf.shape(ystar)[0],dtype=tf.float32)
y1 = tf.matmul(x, W) + b
y2 = tf.matmul(y1,ymask)
dist = tf.norm(ystar-y2,axis=0)
mse = tf.reciprocal(dataLength) * tf.reduce_mean(tf.square(dist))
grads = tf.gradients(dist, [y2])
trainStep = tf.train.GradientDescentOptimizer(eta).minimize(mse)
# build graph
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
randTask = np.random.randint(0, 9)
ymaskIn = np.zeros([tasks, tasks])
ymaskIn[randTask, randTask] = 1
batch = trainDF.sample(batchSize)
batch_xs = batch.iloc[:, :imageSize]
batch_ys = np.zeros([batchSize, tasks])
batch_ys[:, randTask] = batch.iloc[:, imageSize + randTask]
gradOut = sess.run(grads, feed_dict={x: batch_xs, ystar: batch_ys, ymask: ymaskIn})
sess.run(trainStep, feed_dict={x: batch_xs, ystar: batch_ys, ymask:ymaskIn})
Here's a very simple reproduction:
import tensorflow as tf
with tf.Graph().as_default():
y = tf.zeros(shape=[1], dtype=tf.float32)
dist = tf.norm(y,axis=0)
(grad,) = tf.gradients(dist, [y])
with tf.Session():
print(grad.eval())
Prints:
[ nan]
The issue is that tf.norm computes sum(x**2)**0.5. The gradient is x / sum(x**2) ** 0.5 (see e.g. https://math.stackexchange.com/a/84333), so when sum(x**2) is zero we're dividing by zero.
There's not much to be done in terms of a special case: the gradient as x approaches all zeros depends on which direction it's approaching from. For example if x is a single-element vector, the limit as x approaches 0 could either be 1 or -1 depending on which side of zero it's approaching from.
So in terms of solutions, you could just add a small epsilon:
import tensorflow as tf
def safe_norm(x, epsilon=1e-12, axis=None):
return tf.sqrt(tf.reduce_sum(x ** 2, axis=axis) + epsilon)
with tf.Graph().as_default():
y = tf.constant([0.])
dist = safe_norm(y,axis=0)
(grad,) = tf.gradients(dist, [y])
with tf.Session():
print(grad.eval())
Prints:
[ 0.]
Note that this is not actually the Euclidean norm. It's a good approximation as long as the input is much larger than epsilon.