def lay_conv2D(
input,
name='conv2d',
kernels=(3, 5, 7), # layer kernels
filters=(36, 12,
6), # int divisible by len(kernels) or tuple of len(kernels)
dilation=1,
activation=None,
useBias=True,
gatedLU=False, # Gated Linear Unit architecture
initializer=None,
seed=12321,
verbLev=0):
if initializer is None: initializer = my_initializer(seed)
with tf.variable_scope(name):
variables = []
subOutList = []
if type(kernels) is not tuple: kernels = (kernels, )
if verbLev > 0:
print(' > %s: kernels %s, filetrs %s, dilation %s' %
(name, kernels, filters, dilation))
for k in range(len(kernels)):
with tf.variable_scope('kernel_%d' % k):
subKernel = kernels[k]
if type(filters) is not tuple:
subFilters = filters / len(kernels)
else:
subFilters = filters[k]
if gatedLU: subFilters *= 2
convLay = tf.layers.Conv2D(filters=subFilters,
kernel_size=subKernel,
dilation_rate=dilation,
activation=None,
use_bias=useBias,
kernel_initializer=initializer,
padding='valid',
data_format='channels_last')
subOutput = convLay(input)
for var in convLay.variables:
variables.append(var)
if verbLev > 1:
print(' >> subConv: filters %s, kernel %s' %
(subFilters, subKernel))
subOutList.append(subOutput)
output = tf.concat(subOutList, axis=-1)
if gatedLU:
s1, s2 = tf.split(output, num_or_size_splits=2, axis=-1)
output = s1 * tf.sigmoid(s2)
else:
if activation: output = activation(output)
variables = flatten_LOTens(variables)
return output, variables
def lay_conv1D(
input,
name='conv1D',
kernels=(3, 5, 7), # layer kernels
filters=(36, 12,
6), # int divisible by len(kernels) or tuple of len(kernels)
dilation=1,
activation=None,
use_bias=True,
gated_LU=False, # Gated Linear Unit architecture
initializer=None,
padding='valid', # 'same' adds padding, 'valid' does not
seed=12321,
verb=0):
if initializer is None: initializer = my_initializer(seed)
with tf.variable_scope(name):
sub_out_list = []
if type(kernels) is not tuple: kernels = (kernels, )
if verb > 1:
print(' > %s: kernels %s, filters %s, dilation %s' %
(name, kernels, filters, dilation))
for k in range(len(kernels)):
with tf.variable_scope('kernel_%d' % k):
sub_kernel = kernels[k]
if type(filters) is not tuple:
sub_filters = filters // len(kernels)
else:
sub_filters = filters[k]
if gated_LU: sub_filters *= 2
conv_lay = tf.layers.Conv1D(filters=sub_filters,
kernel_size=sub_kernel,
dilation_rate=dilation,
activation=None,
use_bias=use_bias,
kernel_initializer=initializer,
padding=padding,
data_format='channels_last')
sub_output = conv_lay(input)
if verb > 1:
print(' >> sub_conv: filters %s, kernel %s' %
(sub_filters, sub_kernel))
sub_out_list.append(sub_output)
output = tf.concat(sub_out_list, axis=-1)
if gated_LU:
s1, s2 = tf.split(output, num_or_size_splits=2, axis=-1)
output = s1 * tf.sigmoid(s2)
elif activation:
output = activation(output)
return output
def decN(
input,
dictW,
predN=1, # N samples for every feature
name='decN',
hLays=None, # tuple or list of ints
hActiv=tf.nn.relu,
initializer=None,
seed=12321,
verbLev=0):
if verbLev > 0: print('\nBuilding decoderN ...')
if verbLev > 1: print('decoder input:', input)
if initializer is None: initializer = my_initializer(seed)
with tf.variable_scope(name):
# hidden layers
if hLays:
for nLay in range(len(hLays)):
laySize = hLays[nLay]
input = lay_dense(input=input,
units=laySize,
activation=hActiv,
use_bias=True,
initializer=initializer,
seed=seed,
name='decoderN_Hlay_%s' % nLay)
# projection to predN x dictW
logits = lay_dense(input=input,
units=predN * dictW,
activation=None,
use_bias=True,
initializer=initializer,
seed=seed,
name='decoderNProjection')
if verbLev > 1:
print(' > projection to logits (%dx dictW):' % predN, logits)
if predN > 1:
logits = tf.reshape(logits, [tf.shape(logits)[0], -1, dictW])
if verbLev > 1:
print(' > reshaped logits (B,%dxS,dictW):' % predN, logits)
predictions = tf.argmax(logits, axis=-1, output_type=tf.int32)
if verbLev > 1: print(' > predictions:', predictions)
return logits, predictions
def lay_dense(
input,
units: int, # layer width
name='dense',
reuse=False,
activation=None,
use_bias=True,
initializer=None,
seed=12321):
if initializer is None: initializer = my_initializer(seed)
dense_lay = tf.layers.Dense(units=units,
activation=activation,
use_bias=use_bias,
kernel_initializer=initializer,
name=name,
_reuse=reuse)
output = dense_lay(input)
return output
def cards_enc(
train_flag, # train flag (bool tensor)
c_ids, # seven cards (ids tensor)
tat_case: bool = False, # task attention transformer architecture
emb_width: int = 24, # cards embedding width
t_drop: float = 0,
f_drop: float = 0,
in_proj: int = None,
n_layers: int = 8,
dense_mul: int = 4, # transformer dense multiplication
activation=tf.nn.relu,
dropout: float = 0, # transformer dropout
seed=12321,
verb=0):
if verb > 0: print('\nBuilding card encoder...')
with tf.variable_scope('cards_enc'):
zsL = []
hist_summ = []
c_emb = tf.get_variable( # cards embeddings
name='c_emb',
shape=[53, emb_width], # one card for 'no_card'
dtype=tf.float32,
initializer=my_initializer(seed=seed))
hist_summ += [tf.summary.histogram('1.c_emb', c_emb, family='c_emb')]
c_emb_look = tf.nn.embedding_lookup(params=c_emb, ids=c_ids)
if verb > 1: print(' > 1.c_emb_look:', c_emb_look)
myc_emb = tf.get_variable( # my cards embeddings
name='myc_emb',
shape=[2, c_emb.shape[-1]],
dtype=tf.float32,
initializer=my_initializer(seed=seed))
myc_emb_look = tf.nn.embedding_lookup(params=myc_emb,
ids=[0, 0, 1, 1, 1, 1, 1])
if verb > 1: print(' > myc_emb_look:', myc_emb_look)
input = c_emb_look + myc_emb_look
if t_drop or f_drop:
input = tf_drop(input=input,
time_drop=t_drop,
feat_drop=f_drop,
train_flag=train_flag,
seed=seed)
# input projection (without activation)
if in_proj:
input = lay_dense(input=input,
units=in_proj,
name='c_proj',
reuse=tf.AUTO_REUSE,
use_bias=False,
seed=seed)
if verb > 1: print(' > input projected:', input)
elif verb > 1: print(' > input:', input)
enc_out = enc_TNS(in_seq=input,
name='TAT' if tat_case else 'TNS',
seq_out=not tat_case,
add_PE=False,
n_blocks=n_layers,
n_heads=1,
dense_mul=dense_mul,
activation=activation,
max_seq_len=7,
dropout=dropout,
dropout_att=0,
drop_flag=train_flag,
seed=seed,
n_hist=3,
verb=verb)
output = enc_out['output']
zsL += enc_out['zeroes']
hist_summ += enc_out['hist_summ']
if not tat_case:
output = tf.unstack(output, axis=-2)
output = tf.concat(output, axis=-1)
if verb > 1: print(' > encT reshaped output:', output)
elif verb > 1: print(' > encT output:', output)
enc_vars = tf.global_variables(scope=tf.get_variable_scope().name)
return {
'output': output,
'enc_vars': enc_vars,
'hist_summ': hist_summ,
'zeroes': zsL
}
def enc_DRT(
input,
name='enc_DRT',
shared_lays: bool = False, # shared variables in enc_layers
n_layers=12,
lay_width: int = None, # for None matches input width
dns_scale=6, # scale(*) of first dense
activation=tf.nn.relu, # gelu is really worth a try
dropout=0.0, # dropout after two denses
training_flag=None, # training flag tensor (for dropout)
initializer=None,
seed=12321,
n_hist=4, # number of histogram layers (for TB)
verb=0):
lay_width_matched = ''
if lay_width is None:
lay_width = input.shape.as_list()[-1]
lay_width_matched = '(lay_width taken form input width)'
if verb > 0:
drp = 0.0 if not dropout else dropout
print(
f'\nBuilding DRTencoder ({n_layers}x{lay_width} drop:{drp:.2f}) {lay_width_matched}...'
)
if initializer is None: initializer = my_initializer(seed)
hist_summ = []
hist_layers = list_of_layers(n_layers, n_select=n_hist)
if verb > 1: print(' > histogram layers of DRTencoder:', hist_layers)
zsL = [] # zeroes list
with tf.variable_scope(name):
# input projection
iW = input.shape[-1]
if iW != lay_width:
input = lay_dense(input=input,
units=lay_width,
use_bias=False,
initializer=initializer,
seed=seed)
if verb > 0:
print('projected input to layWidth(%d) since it differs(%d)' %
(lay_width, iW))
input = tf.keras.layers.LayerNormalization(axis=-1)(
input) # input layer_norm
output = input # for 0 layers case
for nL in range(n_layers):
lay_name = f'DRLay_{nL}' if not shared_lays else 'DRLay_shared'
lay_out = lay_DRT(input=output,
name=lay_name,
hist_name=name,
dns_scale=dns_scale,
activation=activation,
dropout=dropout,
training_flag=training_flag,
initializer=initializer,
seed=seed)
output = lay_out['output']
if nL in hist_layers: hist_summ.append(lay_out['hist_summ'])
zsL += lay_out['zeroes']
return {'output': output, 'hist_summ': hist_summ, 'zeroes': zsL}
def lay_DRT(
input,
name='lay_DRT', # scope name, be careful when stacked since auto_reuse
hist_name=None, # family name of histogram
dns_scale=4,
activation=tf.nn.relu, # gelu is really worth a try
dropout=None, # dropout (after two denses)
training_flag=None, # training flag tensor (for dropout)
initializer=None,
seed=12321):
if not hist_name: hist_name = name
lay_width = input.shape[-1]
if initializer is None: initializer = my_initializer(seed)
hist_summ = []
with tf.variable_scope(name_or_scope=name, reuse=tf.AUTO_REUSE):
hist_summ.append(
tf.summary.histogram('a_denseSin', input, family=hist_name))
# dense (scale up)
output = lay_dense(input=input,
units=int(lay_width * dns_scale),
activation=None,
use_bias=True,
initializer=initializer,
seed=seed,
name='denseS')
hist_summ.append(
tf.summary.histogram('b_denseSout', output, family=hist_name))
# activation
output = activation(output)
zsL = [zeroes(output)] # zeroes list
hist_summ.append(
tf.summary.histogram('c_activation', output, family=hist_name))
# dense (scale down) no activ
output = lay_dense(input=output,
units=lay_width,
name='DRTdenseNA',
use_bias=True,
initializer=initializer,
seed=seed)
hist_summ.append(
tf.summary.histogram('d_denseNAout', output, family=hist_name))
# layer dropout
if dropout:
output = tf.layers.dropout(inputs=output,
rate=dropout,
training=training_flag,
seed=seed)
# residual
output = lay_res(input, output)
hist_summ.append(
tf.summary.histogram('e_residual', output, family=hist_name))
# layer_norm
output = tf.keras.layers.LayerNormalization(axis=-1)(output)
hist_summ.append(
tf.summary.histogram('f_LAYout', output, family=hist_name))
return {'output': output, 'hist_summ': hist_summ, 'zeroes': zsL}
def enc_CNN(
input: tf.Tensor,
history: tf.
Tensor = None, # optional history(state) tensor with shape [bsz, n_layers ,kernel-1, n_filters], >> masked cnn
name='enc_CNN',
# layer params
shared_lays: bool = False, # shared variables in enc_layers
n_layers: int = 12, # num of layers
kernel: int = 3, # layer kernel
n_filters: int = 128, # num of filters
activation=tf.nn.
relu, # global enc activation func, gelu is really worth a try
lay_drop: float or None = 0.0,
ldrt_scale: int or
None = 0, # DRT @enc_lay - scale(*) of first dense, for None or 0 DRT @lay won't be build
ldrt_drop: float or None = 0.0, # DRT @enc_lay - dropout
# other
training_flag: tf.Tensor or bool = None, # dropout training flag tensor
initializer=None,
seed: int = 12321,
n_hist: int = 4, # number of histogram layers
verb=0):
if verb > 0:
print(
f'\n *** enc_CNN *** Building {name} ({n_layers}x{n_filters})...')
if initializer is None: initializer = my_initializer(seed)
# manage history
history_lays = None
if history is not None:
history_lays = tf.unstack(history, axis=-3)
if verb > 1:
print(
f' > state_lays len {len(history_lays)} of: {history_lays[0]}')
hist_summ = []
hist_layers = list_of_layers(n_layers, n_select=n_hist)
if verb > 1: print(f' > histogram layers of cnn encoder: {hist_layers}')
with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
input_lays = [
] # here we will store inputs of the following layers to extract the state (history)
zsL = [] # zeroes
# input projection - to match n_filters and input width
if verb > 1: print(f' > encoder input: {input}')
if input.shape[-1] != n_filters:
input = lay_dense(input=input,
units=n_filters,
name='enc_input_projection',
initializer=initializer)
if verb > 1: print(f' > encoder projected input: {input}')
output = input # for 0 layers case
sub_output = input # first input
for depth in range(n_layers):
lay_name = f'enc_CNN_lay_{depth}' if not shared_lays else 'enc_CNN_lay_shared'
if verb > 1: print(f'<< layer {lay_name}:')
lay_input = tf.concat([history_lays[depth], sub_output],
axis=-2) if history_lays else sub_output
if verb > 1:
print(f' > sub_output (previous): {sub_output}')
print(f' > lay_input (eventually padded): {lay_input}')
input_lays.append(lay_input)
hist_lay = depth in hist_layers
with tf.variable_scope(lay_name):
if hist_lay:
hist_summ.append(
tf.summary.histogram('a_lay_in',
lay_input,
family=name))
# LN
lay_input = tf.keras.layers.LayerNormalization(
axis=-1)(lay_input)
if hist_lay:
hist_summ.append(
tf.summary.histogram('b_LN', lay_input, family=name))
# conv no activation
output = lay_conv1D(
input=lay_input,
name='conv1D',
kernels=kernel,
filters=n_filters,
activation=None,
initializer=initializer,
padding='same' if history is None else 'valid',
seed=seed,
verb=0)
if hist_lay:
hist_summ.append(
tf.summary.histogram('c_cnn', output, family=name))
# activation
if activation:
output = activation(output)
zsL += [zeroes(output)] # catch zeroes
if hist_lay:
hist_summ.append(
tf.summary.histogram('d_activation',
output,
family=name))
# dropout
if lay_drop:
output = tf.layers.dropout(inputs=output,
rate=lay_drop,
training=training_flag,
seed=seed)
if hist_lay:
hist_summ.append(
tf.summary.histogram('e_drop', output,
family=name))
# RES, here we take sub_output, since lay_input may be padded by history
output += sub_output
if hist_lay:
hist_summ.append(
tf.summary.histogram('f_residual', output,
family=name))
if verb > 1: print(f' > output (layer): {output}')
if ldrt_scale:
lay_out = lay_DRT(input=output,
name=lay_name + '_lay_DRT',
hist_name=name,
dns_scale=ldrt_scale,
activation=activation,
dropout=ldrt_drop,
training_flag=training_flag,
initializer=initializer,
seed=seed)
output = lay_out['output']
zsL += lay_out['zeroes']
if hist_lay: hist_summ.append(lay_out['hist_summ'])
sub_output = output
output = tf.keras.layers.LayerNormalization(axis=-1)(output) # final LN
# prepare fin_state
fin_state = None
if history is not None:
state = tf.stack(input_lays, axis=-3)
if verb > 1: print(f' > state (stacked): {state}')
fin_state = tf.split(state,
num_or_size_splits=[-1, kernel - 1],
axis=-2)[1]
if verb > 1: print(f' > fin_state (split): {fin_state}')
if verb > 1: print(f' > {name} output: {output}')
return {
'output': output,
'state': fin_state, # history for next
'hist_summ': hist_summ,
'zeroes': zsL
}
def cnn_DMG(
name :str,
train_ce :bool= True, # train cards encoder
c_embW :int= 12, # card emb width >> makes network width (x7)
n_lay= 12, # number of CNNR layers >> makes network deep ( >> context length)
width= None, # representation width (number of filters), for None uses cards_encoded_width
activation= tf.nn.relu,
opt_class= partial(tf.compat.v1.train.AdamOptimizer, beta1=0.7, beta2=0.7),
iLR= 3e-5,
warm_up= 100, # num of steps has to be small (since we do rare updates)
avt_SVal= 0.04,
avt_window= 20,
do_clip= True,
verb= 0,
**kwargs):
if verb>0: print(f'\nBuilding {name} cnn_DMG (graph)...')
with tf.variable_scope(name):
n_hands = tf.get_variable( # number of hands while learning
name= 'n_hands',
shape= [],
trainable= False,
initializer= tf.constant_initializer(0),
dtype= tf.int32)
cards_PH = tf.placeholder( # 7 cards placeholder
name= 'cards_PH',
dtype= tf.int32,
shape= [None, None, 7]) # [bsz,seq,7cards]
train_PH = tf.placeholder( # train placeholder
name= 'train_PH',
dtype= tf.bool,
shape= [])
ce_out = cards_enc(
train_flag= train_PH,
c_ids= cards_PH,
emb_width= c_embW)
cards_encoded = ce_out['output']
enc_vars = ce_out['enc_vars']
enc_zsL = ce_out['zeroes']
if verb>1: print(' ### num of enc_vars (%d) %s'%(len(enc_vars),short_scin(num_var_floats(enc_vars))))
if verb>1: print(' > cards encoded:', cards_encoded)
switch_PH = tf.placeholder( # switch placeholder
name= 'switch_PH',
dtype= tf.int32, # 0 for move, 1 for cards
shape= [None, None, 1]) # [bsz,seq,1]
event_PH = tf.placeholder( # event id placeholder
name= 'event_PH',
dtype= tf.int32,
shape= [None, None]) # [bsz,seq]
n_events = 1 + N_TABLE_PLAYERS + len(TBL_MOV)*(N_TABLE_PLAYERS-1)
event_emb = tf.get_variable( # event type embeddings
name= 'event_emb',
shape= [n_events, cards_encoded.shape[-1]],
dtype= tf.float32,
initializer= my_initializer())
event_in = tf.nn.embedding_lookup(params=event_emb, ids=event_PH)
if verb>1: print(' > event_in:', event_in)
# tried with tf.where and switching inputs, but speed was the same...
switch = tf.cast(switch_PH, dtype=tf.float32)
input = switch*cards_encoded + (1-switch)*event_in
if verb>1: print(' > input (merged):', input)
# projection without activation and bias
if width:
input = lay_dense(
input= input,
units= width,
use_bias= False)
if verb>1: print(' > projected input (projected):', input)
else: width = cards_encoded.shape[-1]
# layer_norm
sub_output = tf.contrib.layers.layer_norm(
inputs= input,
begin_norm_axis= -1,
begin_params_axis= -1)
state_shape = [n_lay, 2, width]
single_zero_state = tf.zeros(shape=state_shape) # [n_lay,2,width]
state_PH = tf.placeholder(
name= 'state_PH',
dtype= tf.float32,
shape= [None] + state_shape) # [bsz,n_lay,2,width]
cnn_enc_out = enc_CNN(
input= sub_output,
history= state_PH,
n_layers= n_lay,
n_filters= width,
activation= activation,
n_hist= 0)
out = cnn_enc_out['output']
fin_state = cnn_enc_out['state']
cnn_zsL = cnn_enc_out['zeroes']
if verb > 1:
print(' > out:', out)
print(' > fin_state (split):', fin_state)
# projection to logits
logits = lay_dense(
input= out,
units= len(TBL_MOV),
use_bias= False)
if verb>1: print(' > logits:', logits)
probs = tf.nn.softmax(logits)
cnn_vars = tf.trainable_variables(scope=tf.get_variable_scope().name) + [n_hands]
cnn_vars = [var for var in cnn_vars if var not in enc_vars]
if verb>1: print(' ### num of cnn_vars (%d) %s'%(len(cnn_vars),short_scin(num_var_floats(cnn_vars))))
move_PH = tf.placeholder( # move made (label)
name= 'move_PH',
dtype= tf.int32,
shape= [None, None]) # [bsz,seq]
rew_PH = tf.placeholder( # reward for move made
name= 'rew_PH',
dtype= tf.float32,
shape= [None, None]) # [bsz,seq]
# this loss is auto averaged with reduction parameter
# loss = tf.losses.SparseCategoricalCrossentropy(from_logits=True)
# loss = loss(y_true=move, y_pred=logits, sample_weight=rew)
loss = tf.losses.sparse_softmax_cross_entropy(
labels= move_PH,
logits= logits,
weights= rew_PH)
train_vars = [] + cnn_vars
if train_ce: train_vars += enc_vars
return{
'name': name,
'cards_PH': cards_PH,
'train_PH': train_PH,
'switch_PH': switch_PH,
'event_PH': event_PH,
'move_PH': move_PH,
'rew_PH': rew_PH,
'state_PH': state_PH,
'single_zero_state': single_zero_state,
'probs': probs,
'fin_state': fin_state,
'enc_zeroes': tf.concat(enc_zsL, axis=-1),
'cnn_zeroes': tf.concat(cnn_zsL, axis=-1),
'loss': loss,
'n_hands': n_hands,
'enc_vars': enc_vars,
'cnn_vars': cnn_vars,
'train_vars': train_vars}