def ACM(x, blockname, groups=32):
b, w, h, c = K.int_shape(x)
mu = tf.reduce_mean(x, axis=[1, 2], name=blockname + '_mu')
mu = tf.expand_dims(mu, axis=1)
mu = tf.expand_dims(mu, axis=1)
P = Conv2D(c // 2,
1,
padding='same',
groups=groups,
name=blockname + '_P1')(mu)
P = relu(P)
P = Conv2D(c, 1, padding='same', groups=groups, name=blockname + '_P2')(P)
P = sigmoid(P)
x_mu = x - mu
k = Conv2D(c, 1, padding='same', groups=groups,
name=blockname + '_K')(x_mu)
q = Conv2D(c, 1, padding='same', groups=groups,
name=blockname + '_Q')(x_mu)
k = softmax(k)
q = softmax(q)
k = x_mu * k
q = x_mu * q
k = K.sum(k, axis=[1, 2])
q = K.sum(q, axis=[1, 2])
k_q = k - q
y = x + k_q
y = y * P
return y
def get_mixture_coef(self, out_tensor):
""" Parses the output tensor to appropriate mixture density coefficients"""
# This uses eqns 18 -> 23 of http://arxiv.org/abs/1308.0850.
# Pen states:
z_pen_logits = out_tensor[:, :, 0:3]
# Process outputs into MDN parameters
M = self.hps['num_mixture']
dist_params = [
out_tensor[:, :, (3 + M * (n - 1)):(3 + M * n)]
for n in range(1, 7)
]
z_pi, z_mu1, z_mu2, z_sigma1, z_sigma2, z_corr = dist_params
# Softmax all the pi's and pen states:
z_pi = softmax(z_pi)
z_pen = softmax(z_pen_logits)
# Exponent the sigmas and also make corr between -1 and 1.
z_sigma1 = K.exp(z_sigma1)
z_sigma2 = K.exp(z_sigma2)
z_corr = tanh(z_corr)
r = [
z_pi, z_mu1, z_mu2, z_sigma1, z_sigma2, z_corr, z_pen, z_pen_logits
]
return r
def call(self, x, mask=None):
# Get input tensor static shape
_, N, C = x.get_shape().as_list()
head_dim = C // self.num_heads
x_qkv = self.qkv(x)
x_qkv = tf.reshape(x_qkv, shape=(-1, N, 3, self.num_heads, head_dim))
x_qkv = tf.transpose(x_qkv, perm=(2, 0, 3, 1, 4))
q, k, v = x_qkv[0], x_qkv[1], x_qkv[2]
# Query rescaling
q = q * self.scale
# multi-headed self-attention
k = tf.transpose(k, perm=(0, 1, 3, 2))
attn = (q @ k)
# Shift window
num_window_elements = self.window_size[0] * self.window_size[1]
relative_position_index_flat = tf.reshape(self.relative_position_index,
shape=(-1, ))
relative_position_bias = tf.gather(self.relative_position_bias_table,
relative_position_index_flat)
relative_position_bias = tf.reshape(relative_position_bias,
shape=(num_window_elements,
num_window_elements, -1))
relative_position_bias = tf.transpose(relative_position_bias,
perm=(2, 0, 1))
attn = attn + tf.expand_dims(relative_position_bias, axis=0)
if mask is not None:
nW = mask.get_shape()[0]
mask_float = tf.cast(
tf.expand_dims(tf.expand_dims(mask, axis=1), axis=0),
tf.float32)
attn = tf.reshape(
attn, shape=(-1, nW, self.num_heads, N, N)) + mask_float
attn = tf.reshape(attn, shape=(-1, self.num_heads, N, N))
attn = softmax(attn, axis=-1)
else:
attn = softmax(attn, axis=-1)
# Dropout after attention
attn = self.attn_drop(attn)
# Merge qkv vectors
x_qkv = (attn @ v)
x_qkv = tf.transpose(x_qkv, perm=(0, 2, 1, 3))
x_qkv = tf.reshape(x_qkv, shape=(-1, N, C))
# Linear projection
x_qkv = self.proj(x_qkv)
# Dropout after projection
x_qkv = self.proj_drop(x_qkv)
return x_qkv
def call(self, logits, mask=None):
if mask is not None:
# print mask
m = tf.cast(mask[0], logits.dtype)
# m = tf.cast(mask[0][0], logits.dtype)
mm = tf.expand_dims(m, -2) * tf.expand_dims(m, -1)
adder = (1.0 - mm) * self._small_negative
return logits + adder, softmax(logits + adder)
return logits, softmax(logits)
def __call__(self, inputs, training):
x, mask = inputs
g_x = iterative_call(self.inner_layers[0], x, training=training)
theta_x = iterative_call(self.inner_layers[1], x, training=training)
phi_x = iterative_call(self.inner_layers[2], x, training=training)
f_div_C = softmax(tf.linalg.matmul(theta_x, phi_x) / self.temperature, axis=2)
inverted_mask = 1. - mask
mask = tf.image.resize(mask, size=x.shape[1:3], method=tf.image.ResizeMethod.BILINEAR)
mask = tf.where(tf.greater(mask, 0.), tf.zeros_like(mask), mask)
mask = 1. - mask
inverted_mask = tf.image.resize(inverted_mask, size=(x.shape[1:3]), method=tf.image.ResizeMethod.BILINEAR)
mask *= inverted_mask
mask_expand = tf.reshape(mask, (x.shape[0], 1, -1))
mask_expand = tf.repeat(mask_expand, x.shape[1] * x.shape[2], axis=1)
if self.use_self:
mask_expand[:, range(x.shape[1] * x.shape[2]), range(x.shape[1] * x.shape[2])] = 1.0
f_div_C = mask_expand * f_div_C
if self.re_norm:
f_div_C = tf.keras.utils.normalize(f_div_C, axis=2, order=1)
y = tf.reshape(tf.linalg.matmul(f_div_C, g_x), (x.shape[0], *x.shape[1:3], self.inter_channels))
W_y = iterative_call(self.inner_layers[3], y, training=training)
W_y = iterative_call(self.inner_layers[4], W_y, training=training)
assert self.mode.casefold() == 'combine'
full_mask = tf.repeat(mask, self.inter_channels, axis=3)
z = full_mask * x + (1 - full_mask) * W_y
return z
def evaluate(adj,
x,
labels,
idx_train,
idx_val,
idx_test,
target,
retrain_iters=2,
norm_x='l1'):
classification_margins = []
class_distrs = []
for _ in range(retrain_iters):
print(f"... {_+1}/{retrain_iters}")
model = GCN(adj,
x,
labels,
device='GPU:0',
seed=123 + _,
norm_x=norm_x)
model.build()
his = model.train(idx_train, idx_val, verbose=0, epochs=100)
logit = softmax(model.predict(target)).numpy().ravel()
class_distrs.append(logit)
best_second_class_before = (
logit - np.eye(data.n_classes)[labels[target]]).argmax()
margin = logit[labels[target]] - logit[best_second_class_before]
classification_margins.append(margin)
model.close
class_distrs = np.asarray(class_distrs)
print(classification_margins)
return class_distrs
def generate(self, mask_in):
paddings = [
[0, 0],
[self.padding, self.padding],
[self.padding, self.padding],
[0, 0],
]
mask = tf.pad(mask_in, paddings)
# NOTE(brendan): this unfolding is for convolution
mask = tf.image.extract_patches(
mask,
sizes=(1, (2 * self.radius) + 1, (2 * self.radius) + 1, 1),
strides=4 * (1, ),
rates=4 * (1, ),
padding="VALID",
)
mask = tf.image.resize(mask, size=self.shape_up, method="nearest")
mask = mask[:, :-self.step + 1, :-self.step + 1, :]
# NOTE(brendan): convolve Gaussian weights with mask (smoothness inductive bias)
mask = self.weight * mask
mask = mask * softmax(self.coldness * mask, axis=-1)
mask = tf.reduce_sum(mask, axis=-1, keepdims=True)
m = round(self.margin)
if self.clamp:
mask = tf.clip_by_value(mask, clip_value_min=0, clip_value_max=1)
cropped = mask[:, m:m + self.shape[0], m:m + self.shape[1], :]
return cropped, mask
def call(self,
queries,
keys,
values,
mask=None,
training=False,
attentions=False):
"""
inputs
q: (queries_size x timesteps x input_dim)
keys: (keys memory_size x memory_timesteps x input_dim)
values: (values queries_size x timesteps x output_dim)
returns
output: context vector (queries_size x timesteps x out_dim)
attn: attn weights (queries_size x timesteps x out_dim)
"""
attn = tf.matmul(queries, keys, transpose_b=True)
scale_dim = keys.shape[2]
attn /= np.sqrt(scale_dim)
if mask is not None:
attn += mask
attn = activations.softmax(attn)
if training:
attn = self.dropout(attn)
output = tf.matmul(attn, values)
if attentions:
return output, attn
else:
return output
def build(self, z):
input_shape, _ = z
self.q_net = Dense(self.dout, input_shape=input_shape, use_bias=False,
kernel_constraint=MaskWeights(self.mask),
name="q_dense")
self.k_net = Dense(self.dout, input_shape=input_shape, use_bias=False,
kernel_constraint=MaskWeights(self.mask),
name="k_dense")
self.v_net = Dense(self.dout, input_shape=input_shape, use_bias=False,
kernel_constraint=MaskWeights(self.mask),
name="v_dense")
self.out_net = Dense(self.dout,
kernel_constraint=MaskWeights(self.mask),
name="head_combiner")
self.attention = Lambda(lambda x: activations.softmax(
tf.matmul(x[0], x[1], transpose_b=True) / (self.dk ** -0.5)
),
name='scale_dot_attention')
self.dot_product = Lambda(lambda x: tf.matmul(x[0], x[1]),
name='dot_product')
self.head_split = Reshape((input_shape[1], self.h * self.nh, self.dk))
self.head_swap = Lambda(lambda x: tf.transpose(x, [0, 2, 1, 3]))
self.head_merge = Reshape((input_shape[1], self.dout))
super(MaskedMultiHeadAttention, self).build(input_shape)
def call(self, inputs, mask=None, **kwargs):
query, value = inputs
embedding_size = query.shape[-1]
step_size = value.shape[1]
if self.attention_type == 'inner':
attention_score = tf.squeeze(tf.matmul(value, query, transpose_b=True), axis=-1)
else:
querys = tf.tile(query, [1, step_size, 1])
query_value = tf.concat([querys, value, querys-value, querys*value], axis=-1)
attention_score = self.out_kernel(self.mlp(query_value))
attention_score = tf.reshape(attention_score, (-1, step_size))
if mask is not None:
if mask[0] is not None:
raise ValueError('query should not support mask')
if mask[1] is not None:
min_value_matrix = tf.ones_like(attention_score) * (-2**31)
attention_score = tf.where(mask[1], attention_score, min_value_matrix)
attention_score = tf.divide(attention_score, tf.sqrt(embedding_size*1.0))
if self.norm:
weighted_att_score = activations.softmax(attention_score)
attention_vec = tf.squeeze(tf.matmul(tf.expand_dims(weighted_att_score, axis=1), value), axis=1)
if not self.keepdims:
out_shape = (-1, embedding_size)
else:
out_shape = (-1, 1, embedding_size)
return tf.reshape(attention_vec, out_shape)
def random_sample(self, sample_epochs=20, disable=False):
best_loss = -np.inf
best_s = None
s = tf.linalg.band_part(self.adj_changes, 0, -1) - tf.linalg.band_part(
self.adj_changes, 0, 0)
for _ in tqdm(range(sample_epochs),
desc='Random Sampling',
disable=disable):
random_matrix = tf.random.uniform(shape=(self.n_nodes,
self.n_nodes),
minval=0.,
maxval=1.)
sampled = tf.where(s > random_matrix, 1., 0.)
if tf.reduce_sum(sampled) > self.n_perturbations:
continue
with tf.device(self.device):
self.adj_changes.assign(sampled)
adj = self.get_perturbed_adj()
adj_norm = normalize_adj_tensor(adj)
logit = self.surrogate([self.tf_x, adj_norm, self.idx_attack])
logit = softmax(logit)
loss = self.compute_loss(logit)
if best_loss < loss:
best_loss = loss
best_s = sampled
return best_s.numpy()
def _read_inputs(self, inputs):
"""
Applies transformations to `inputs` to get control for this module.
Computes elements in the interface vector 𝜉
"""
def linear(name, first_dim=None, second_dim=None):
"""
Returns a linear transformation of `inputs`. If first_dim and second_dim
are provide, reshape the resulting Tensor
"""
linear = self._layers[name](inputs)
if first_dim and second_dim:
linear = tf.reshape(linear, [-1, first_dim, second_dim])
return linear
# v_t^i - The vectors to write to memory, for each write head `i`.
write_vectors = linear('write_vectors', self._num_writes, self._word_size)
# e_t^i - Amount to erase the memory by before writing, for each write head.
erase_vectors = linear('erase_vectors', self._num_writes, self._word_size)
# f_t^j - Amount that the memory at the locations read from at the previous
# time step can be declared unused, for each read head `j`.
free_gate = linear('free_gate')
# g_t^{a, i} - Interpolation between writing to unallocated memory and
# content-based lookup, for each write head `i`. Note: `a` is simply used to
# identify this gate with allocation vs writing (as defined below).
allocation_gate = linear('allocation_gate')
# g_t^{w, i} - Overall gating of write amount for each write head.
write_gate = linear('write_gate')
# 𝜋_t^j - Mixing between "backwards" and "forwards" positions (for
# each write head), and content-based lookup, for each read head.
num_read_modes = 1 + 2 * self._num_writes
read_mode = linear('read_mode', self._num_reads, num_read_modes)
read_mode = activations.softmax(read_mode)
# Parameters for the (read / write) "weights by content matching" modules.
write_keys = linear('write_keys', self._num_writes, self._word_size)
write_strengths = linear('write_strengths')
read_keys = linear('read_keys', self._num_reads, self._word_size)
read_strengths = linear('read_strengths')
result = dict(
read_keys=read_keys,
read_strengths=read_strengths,
write_keys=write_keys,
write_strengths=write_strengths,
write_vectors=write_vectors,
erase_vectors=erase_vectors,
free_gate=free_gate,
allocation_gate=allocation_gate,
write_gate=write_gate,
read_mode=read_mode,
)
return result
def getAction(self, game):
logits, _ = self.model(tf.convert_to_tensor(self.extractState(game)))
probs = activations.softmax(logits)
action = np.random.choice(self.action_size, p=probs.numpy()[0])
while action >= len(game.getMoves()):
action = np.random.choice(self.action_size, p=probs.numpy()[0])
return game.getMoves()[action]
def build_model(w,h,num_classes, dropout=.5, l2_reg=0.,conv_type='ds'): X = Input(shape=(h,w,3),name='X') #conv1 = conv_block(X,conv_type,64,3,1, name='conv1',num_conv_layers=2,l2_reg=l2_reg) #conv2 = conv_block(conv1,conv_type,128,3,1,name='conv2',l2_reg=l2_reg) conv1 = conv_block(X,conv_type,64,3,1, name='conv1',num_conv_layers=2,l2_reg=l2_reg,use_bn=True) conv2 = conv_block(conv1,conv_type,128,3,1,name='conv2',num_conv_layers=2,l2_reg=l2_reg,use_bn=True) conv3 = conv_block(conv2,conv_type,256,3,1,name='conv3',l2_reg=l2_reg,use_bn=True) conv4 = conv_block(conv3,conv_type,512,3,1,name='conv4',l2_reg=l2_reg,use_bn=True) conv5 = conv_block(conv4,conv_type,512,3,1,name='conv5',l2_reg=l2_reg,use_bn=True) fc1 = fc_block(conv5,conv_type,4096,(7,7), strides=(1,1), dropout=dropout, name='fc1',l2_reg=l2_reg,use_dropout=True,use_bn=True) fc2 = fc_block(fc1,conv_type,4096,(1,1), strides=(1,1), dropout=dropout, name='fc2',l2_reg=l2_reg,use_dropout=True,use_bn=True) score32 = score_block(fc2,conv_type,num_classes,name='score32',l2_reg=l2_reg) score16 = score_block(conv4,conv_type,num_classes,name='score16',l2_reg=l2_reg) score8 = score_block(conv3,conv_type,num_classes,name='score8',l2_reg=l2_reg) upscore32 = upsample_block(score32,num_classes,4,2,name='upscore32',l2_reg=l2_reg) upscore32c = crop(score16,name='upscore32c')(upscore32) fuse1 = Add(name='fuse1')([score16, upscore32c]) upscore16 = upsample_block(fuse1,num_classes,4,2,name='upscore16',l2_reg=l2_reg) upscore16c = crop(score8,name='upscore16c')(upscore16) fuse2 = Add(name='fuse2')([upscore16c, score8]) upscore8 = UpSampling2D((8,8),name='upscore8')(fuse2) upscore8c = crop(X,name='upscore8c')(upscore8) classifier = Lambda(lambda x: softmax(x))(upscore8c) fcn8 = Model(inputs=X, outputs=classifier, name = 'FCN8') return fcn8
def call(self, inputs):
X, A = inputs
N = K.shape(A)[-1]
# Check if the layer is operating in mixed or batch mode
mode = ops.autodetect_mode(X, A)
self.reduce_loss = mode in (modes.MIXED, modes.BATCH)
# Get normalized adjacency
if K.is_sparse(A):
I_ = tf.sparse.eye(N, dtype=A.dtype)
A_ = tf.sparse.add(A, I_)
else:
I_ = tf.eye(N, dtype=A.dtype)
A_ = A + I_
fltr = ops.normalize_A(A_)
# Node embeddings
Z = K.dot(X, self.kernel_emb)
Z = ops.modal_dot(fltr, Z)
if self.activation is not None:
Z = self.activation(Z)
# Compute cluster assignment matrix
S = K.dot(X, self.kernel_pool)
S = ops.modal_dot(fltr, S)
S = activations.softmax(S, axis=-1) # softmax applied row-wise
# Link prediction loss
S_gram = ops.modal_dot(S, S, transpose_b=True)
if mode == modes.MIXED:
A = tf.sparse.to_dense(A)[None, ...]
if K.is_sparse(A):
LP_loss = tf.sparse.add(A, -S_gram) # A/tf.norm(A) - S_gram/tf.norm(S_gram)
else:
LP_loss = A - S_gram
LP_loss = tf.norm(LP_loss, axis=(-1, -2))
if self.reduce_loss:
LP_loss = K.mean(LP_loss)
self.add_loss(LP_loss)
# Entropy loss
entr = tf.negative(
tf.reduce_sum(tf.multiply(S, K.log(S + K.epsilon())), axis=-1)
)
entr_loss = K.mean(entr, axis=-1)
if self.reduce_loss:
entr_loss = K.mean(entr_loss)
self.add_loss(entr_loss)
# Pooling
X_pooled = ops.modal_dot(S, Z, transpose_a=True)
A_pooled = ops.matmul_at_b_a(S, A)
output = [X_pooled, A_pooled]
if self.return_mask:
output.append(S)
return output
def convModel_APIHandler():
"""
This function represents a Flask API endpoint that serves a Keras image classifier.
If the input data is valid, this function will return the models top prediction for the
image, along with all of the probabilities for the classes.
"""
try:
# image is expected to be recieved in a base64 encoded string which will be decoded
# and converted to a tensor and then preprocessed
base64_image_string = request.form['image_data_buffer']
base64_decoded = base64.b64decode(base64_image_string)
image = Image.open(io.BytesIO(base64_decoded))
preprocessed_img = prepare_image(image)
raw_logits_tensor = model(preprocessed_img)
prob_distribution = softmax(raw_logits_tensor)
topPrediction = class_indices[prob_distribution.numpy().argmax()]
allProbsDict = get_all_probs(prob_distribution)
resp = jsonify(MostLikelyClass=topPrediction, allProbs=allProbsDict)
resp.status_code = 200
return resp
except:
resp = jsonify({
'message':
'There was an error processing your image by the server. Try a different image?'
})
resp.status_code = 500
return resp
def squeeze_net(shape):
inp = Input(shape=shape)
conv1 = Conv2D(filters=96,
kernel_size=(7, 7),
strides=(2, 2),
activation='relu',
padding='same')(inp)
pool1 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(conv1)
fire1 = fire_module(16, 64, 64)(pool1)
fire2 = fire_module(16, 64, 64)(fire1)
fire3 = fire_module(32, 128, 128)(fire2)
pool2 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(fire3)
fire4 = fire_module(32, 128, 128)(pool2)
fire5 = fire_module(48, 192, 192)(fire4)
fire6 = fire_module(48, 192, 192)(fire5)
fire7 = fire_module(64, 256, 256)(fire6)
pool3 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(fire7)
fire8 = fire_module(64, 256, 256)(pool3)
drop1 = Dropout(.5)(fire8)
# conv2 should have number of filters same as desired number of outputs
conv2 = Conv2D(filters=1000,
kernel_size=(1, 1),
strides=(1, 1),
activation='relu')(drop1)
pool4 = GlobalAveragePooling2D()(conv2)
# activation functions should be changed to switch to regression
out = softmax(pool4)
model = Model(inputs=inp, outputs=[out])
return model
def multiplicative_self_attention(units, n_hidden=None, n_output_features=None, activation=None):
"""
Compute multiplicative self attention for time series of vectors (with batch dimension)
the formula: score(h_i, h_j) = <W_1 h_i, W_2 h_j>, W_1 and W_2 are learnable matrices
with dimensionality [n_hidden, n_input_features]
Args:
units: tf tensor with dimensionality [batch_size, time_steps, n_input_features]
n_hidden: number of units in hidden representation of similarity measure
n_output_features: number of features in output dense layer
activation: activation at the output
Returns:
output: self attended tensor with dimensionality [batch_size, time_steps, n_output_features]
"""
n_input_features = K.int_shape(units)[2]
if n_hidden is None:
n_hidden = n_input_features
if n_output_features is None:
n_output_features = n_input_features
exp1 = Lambda(lambda x: expand_tile(x, axis=1))(units)
exp2 = Lambda(lambda x: expand_tile(x, axis=2))(units)
queries = Dense(n_hidden)(exp1)
keys = Dense(n_hidden)(exp2)
scores = Lambda(lambda x: K.sum(queries * x, axis=3, keepdims=True))(keys)
attention = Lambda(lambda x: softmax(x, axis=2))(scores)
mult = Multiply()([attention, exp1])
attended_units = Lambda(lambda x: K.sum(x, axis=2))(mult)
output = Dense(n_output_features, activation=activation)(attended_units)
return output
def additive_self_attention(units, n_hidden=None, n_output_features=None, activation=None):
"""
Compute additive self attention for time series of vectors (with batch dimension)
the formula: score(h_i, h_j) = <v, tanh(W_1 h_i + W_2 h_j)>
v is a learnable vector of n_hidden dimensionality,
W_1 and W_2 are learnable [n_hidden, n_input_features] matrices
Args:
units: tf tensor with dimensionality [batch_size, time_steps, n_input_features]
n_hidden: number of2784131 units in hidden representation of similarity measure
n_output_features: number of features in output dense layer
activation: activation at the output
Returns:
output: self attended tensor with dimensionality [batch_size, time_steps, n_output_features]
"""
n_input_features = K.int_shape(units)[2]
if n_hidden is None:
n_hidden = n_input_features
if n_output_features is None:
n_output_features = n_input_features
exp1 = Lambda(lambda x: expand_tile(x, axis=1))(units)
exp2 = Lambda(lambda x: expand_tile(x, axis=2))(units)
units_pairs = Concatenate(axis=3)([exp1, exp2])
query = Dense(n_hidden, activation="tanh")(units_pairs)
attention = Dense(1, activation=lambda x: softmax(x, axis=2))(query)
attended_units = Lambda(lambda x: K.sum(attention * x, axis=2))(exp1)
output = Dense(n_output_features, activation=activation)(attended_units)
return output
def get_default_model():
inputs = Input(shape=(None, None, 10))
x = Conv2D(128, 3, padding='same')(inputs)
x = Activation('relu')(x)
x = Conv2D(10, 1)(x)
outputs = softmax(x, axis=3)
return tf.keras.Model(inputs=inputs, outputs=outputs)
def call(self, inputs, **kwargs):
features = inputs[0]
relations = inputs[1]
sx = inputs[2]
sy = inputs[3]
z, _ = self.kenn_layer_1(features, relations, sx, sy)
return softmax(z)
def step(self, x, states):
ytm, stm = states
# repeat the hidden state to the length of the sequence
_stm = K.repeat(stm, self.timesteps)
# now multiplty the weight matrix with the repeated hidden state
_Wxstm = K.dot(_stm, self.W_a)
# calculate the attention probabilities
# this relates how much other timesteps contributed to this one.
et = K.dot(activations.tanh(_Wxstm + self._uxpb),
K.expand_dims(self.V_a))
at = K.exp(et)
at_sum = K.sum(at, axis=1)
at_sum_repeated = K.repeat(at_sum, self.timesteps)
at /= at_sum_repeated # vector of size (batchsize, timesteps, 1)
# calculate the context vector
context = K.squeeze(K.batch_dot(at, self.x_seq, axes=1), axis=1)
# ~~~> calculate new hidden state
# first calculate the "r" gate:
rt = activations.sigmoid(
K.dot(ytm, self.W_r)
+ K.dot(stm, self.U_r)
+ K.dot(context, self.C_r)
+ self.b_r)
# now calculate the "z" gate
zt = activations.sigmoid(
K.dot(ytm, self.W_z)
+ K.dot(stm, self.U_z)
+ K.dot(context, self.C_z)
+ self.b_z)
# calculate the proposal hidden state:
s_tp = activations.tanh(
K.dot(ytm, self.W_p)
+ K.dot((rt * stm), self.U_p)
+ K.dot(context, self.C_p)
+ self.b_p)
# new hidden state:
st = (1-zt)*stm + zt * s_tp
yt = activations.softmax(
K.dot(ytm, self.W_o)
+ K.dot(stm, self.U_o)
+ K.dot(context, self.C_o)
+ self.b_o)
if self.return_probabilities:
return at, [yt, st]
else:
return yt, [yt, st]
def read(self, keys, scale=None):
"""Read from memory.
Read the memory for given the keys. For each key in keys we will get one
result as `r = sum_i M[i] a[i]` where `M[i]` is the memory content
at location i and `a[i]` is the attention weight for key at location i.
`a` is calculated as softmax of a scaled similarity between key and
each memory content: `a[i] = exp(scale*sim[i])/(sum_i scale*sim[i])`
Args:
keys (Tensor): shape[-1] is dim.
For single key read, the shape is (batch_size, dim).
For multiple key read, the shape is (batch_szie, k, dim), where
k is the number of keys.
scale (None|float|Tensor): shape is () or keys.shape[:-1]. The
cosine similarities are multiplied with `scale` before softmax
is applied. If None, use the scale provided at constructor.
Returns:
resutl Tensor: shape is same as keys. result[..., i] is the read
result for the corresponding key.
"""
if not self._built:
self.build(keys.shape[0])
assert 2 <= len(keys.shape) <= 3
assert keys.shape[0] == self._batch_size
assert keys.shape[-1] == self.dim
if scale is None:
scale = self._scale
else:
if isinstance(scale, (int, float)):
pass
else: # assuming it's Tensor
scale = expand_dims_as(scale, keys)
sim = layers.dot([keys, self._memory],
axes=-1,
normalize=self._normalize)
sim = sim * scale
attention = activations.softmax(sim)
result = layers.dot([attention, self._memory], axes=(-1, 1))
if len(sim.shape) > 2: # multiple read keys
usage = tf.reduce_sum(attention,
axis=tf.range(1,
len(sim.shape) - 1))
else:
usage = attention
if self._snapshot_only:
self._usage.assign_add(usage)
else:
self._usage = self._usage + usage
return result
def squeeze_net_complex_skip(shape):
inp = Input(shape=shape)
conv1 = Conv2D(filters=96,
kernel_size=(7, 7),
strides=(2, 2),
activation='relu',
padding='same')(inp)
pool1 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(conv1)
skip_conv1 = Conv2D(filters=128,
kernel_size=(1, 1),
strides=(1, 1),
activation='relu')(pool1)
fire1 = fire_module(16, 64, 64)(pool1)
complex_add1 = Add()([fire1, skip_conv1])
fire2 = fire_module(16, 64, 64)(complex_add1)
add1 = Add()([complex_add1, fire2])
fire3 = fire_module(32, 128, 128)(add1)
skip_conv2 = Conv2D(filters=256,
kernel_size=(1, 1),
strides=(1, 1),
activation='relu')(add1)
complex_add2 = Add()([skip_conv2, fire3])
pool2 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(complex_add2)
fire4 = fire_module(32, 128, 128)(pool2)
add2 = Add()([pool2, fire4])
skip_conv3 = Conv2D(filters=384,
kernel_size=(1, 1),
strides=(1, 1),
activation='relu')(add2)
fire5 = fire_module(48, 192, 192)(add2)
complex_add3 = Add()([skip_conv3, fire5])
fire6 = fire_module(48, 192, 192)(complex_add3)
add3 = Add()([complex_add3, fire6])
fire7 = fire_module(64, 256, 256)(add3)
skip_conv4 = Conv2D(filters=512,
kernel_size=(1, 1),
strides=(1, 1),
activation='relu')(add3)
complex_add4 = Add()([skip_conv4, fire7])
pool3 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(complex_add4)
fire8 = fire_module(64, 256, 256)(pool3)
add4 = Add()([pool3, fire8])
drop1 = Dropout(.5)(add4)
# conv2 should have number of filters same as desired number of outputs
conv2 = Conv2D(filters=1000,
kernel_size=(1, 1),
strides=(1, 1),
activation='relu')(drop1)
pool4 = GlobalAveragePooling2D()(conv2)
# activation functions should be changed to switch to regression
out = softmax(pool4)
model = Model(inputs=inp, outputs=[out])
return model
def call(self, hidden, timesteps):
hidden_transformed = self.transform_hidden(hidden)
hidden_repeated = K.repeat(hidden_transformed, timesteps)
input_seq_transformed = self._input_seq_shaped
alignment_score = self.calculate_alignment(hidden_repeated,
input_seq_transformed)
score_vector = softmax(alignment_score, 1)
context_vector = K.sum(score_vector * self.input_seq, 1)
return context_vector
def update_surrogate(self, trainable_variables, idx):
with tf.GradientTape() as tape:
adj = self.get_perturbed_adj()
adj_norm = normalize_adj_tensor(adj)
logit = self.surrogate([self.tf_x, adj_norm, idx])
logit = softmax(logit)
loss = self.compute_loss(logit)
gradients = tape.gradient(loss, trainable_variables)
self.optimizer.apply_gradients(zip(gradients, trainable_variables))
def head_no_fc(x):
'''
No incluye fully connected, logra un
+2.5~ en val_place_acc
Sin los FC, evitamos un poco mas el overfitting
'''
x = block_no_activation(x, k=1, n_c=259, s=1, padding='same')
x = GlobalAveragePooling2D()(x)
x = Reshape((7, 37, 1))(x)
return Lambda(lambda x: softmax(x, axis=-2))(x)
def call(self, encoder_outputs, decoder_outputs, mask=None):
w1_e = self.W1(encoder_outputs)
w2_d = self.W2(decoder_outputs)
tanh_output = activations.tanh(w1_e + w2_d)
v_dot_tanh = self.V(tanh_output)
if mask is not None:
v_dot_tanh += (mask * -1e-9)
attention_weights = activations.softmax(v_dot_tanh, axis=1)
att_shape = tf.shape(attention_weights)
return tf.reshape(attention_weights, (att_shape[0], att_shape[1]))
def predict_step_on_batch(self,
x,
out_weight=None,
return_logits=True,
device="CPU"):
with tf.device(device):
out = self(x, training=False)
out = gather(out, out_weight)
if not return_logits:
out = softmax(out)
return out
def compute_gradients(self, idx):
with tf.GradientTape() as tape:
tape.watch(self.adj_changes)
adj = self.get_perturbed_adj()
adj_norm = normalize_adj_tensor(adj)
logit = self.surrogate([self.tf_x, adj_norm, idx])
logit = softmax(logit)
loss = self.compute_loss(logit)
gradients = tape.gradient(loss, self.adj_changes)
return gradients
