def skew(inputs, scope="skew"):
with tf.name_scope(scope):
batch, height, width, channel = get_shape(inputs)
new_width = width + height - 1
skewed_rows = []# inputs = tf.zeros([batch, width * 2 - 1 , height, channel])
#rows = tf.unpack(tf.transpose(inputs, [1, 0, 3, 2])) # [height, batch, channel, width]
rows = tf.split(1, height, inputs) # [batch, 1, width, channel]
for i, row in enumerate(rows):
transposed_row = tf.transpose(tf.squeeze(row, [1]), [0, 2, 1])
reshaped_row = tf.reshape(transposed_row, [-1, width]) # [batch * channel, width]
padded_row = tf.pad(reshaped_row, ((0, 0), (i, height - 1 - i)))
unsqueezed_row = tf.reshape(padded_row, [-1, channel, new_width]) # [batch, channel, width*2-1]
new_row = tf.transpose(unsqueezed_row, [0, 2, 1]) # [batch, width*2-1, channel]
assert get_shape(new_row) == [batch, new_width, channel], "wrong shape of skewed row"
skewed_rows.append(new_row)
skewed_inputs = tf.pack(skewed_rows, axis=1, name="skewed_inputs")
desired_shape = [None, height, new_width, channel]
skewed_shape = get_shape(skewed_inputs)
assert skewed_shape == desired_shape, "wrong shape of skewed input. Actual {}; Expected {}".format(skewed_shape, desired_shape)
return skewed_inputs
def feature_net(self, image, rnn, a, r, state_in, scope="feature"):
shape = get_shape(image)
image = tf.cast(image, tf.float32) / 255.0
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
image = tf.reshape(image, [-1] + shape[-3:])
filter = [16, 32, 32]
kernel = [(3, 3), (3, 3), (5, 3)]
stride = [(1, 2), (1, 2), (2, 1)]
for i in range(len(filter)):
image = tf.layers.conv2d(
image,
filters=filter[i],
kernel_size=kernel[i][0],
strides=stride[i][0],
padding="valid",
activation=None,
name="conv_%d" % i)
image = tf.layers.max_pooling2d(
image,
pool_size=kernel[i][1],
strides=stride[i][1],
padding="valid",
name="maxpool_%d" % i)
image = self.resblock(
image, "res0_%d" % i)
# image = self.resblock(
# image, "res1_%d" % i)
image = tf.nn.relu(image)
new_shape = get_shape(image)
feature = tf.reshape(
image, [shape[0], shape[1], new_shape[1] * new_shape[2] * new_shape[3]])
a_onehot = tf.one_hot(
a, depth=self.act_space, dtype=tf.float32)
feature = tf.layers.dense(feature, 256, tf.nn.relu, name="feature")
feature = tf.concat([feature, a_onehot, r[:, :, None]], axis=-1)
if self.use_hrnn:
initial_state = tf.split(state_in, [1, -1], axis=-1)
feature, count_out, state_out = rnn(
feature, initial_state=initial_state)
state_out = tf.concat([count_out, state_out], axis=-1)
elif self.use_rmc:
initial_state = [state_in]
feature, state_out = rnn(
feature, initial_state=initial_state)
elif self.use_amc:
initial_state = tf.split(state_in, [1, 64, 64, -1], axis=-1)
feature, count_out, ns_out, pos_out, state_out = rnn(
feature, initial_state=initial_state)
state_out = tf.concat([count_out, ns_out, pos_out, state_out], axis=-1)
else:
c_in, h_in = tf.split(state_in, 2, axis=-1)
feature, c_out, h_out = rnn(
feature, initial_state=[c_in, h_in])
state_out = tf.concat([c_out, h_out], axis=-1)
return feature, state_out
def _last_relevant(seq, length):
batch_size = get_shape(seq)[0]
max_length = get_shape(seq)[1]
input_size = get_shape(seq)[2]
index = tf.range(0, batch_size) * max_length + (length - 1)
flat = tf.reshape(seq, [-1, input_size])
return tf.gather(flat, index)
def control_net(self, feature, scope="pixel_control"):
shape = get_shape(feature)
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
feature = tf.reshape(feature, [-1, shape[-1]])
feature = tf.layers.dense(feature,
7 * 7 * 32,
tf.nn.relu,
name="feature")
image = tf.reshape(feature, [-1, 7, 7, 32])
image = tf.nn.conv2d_transpose(
image,
filter=tf.get_variable(name="deconv", shape=[9, 9, 32,
32]),
output_shape=[get_shape(feature)[0], 21, 21, 32],
strides=2,
padding="VALID")
image = tf.nn.relu(image)
image = tf.nn.conv2d_transpose(
image,
filter=tf.get_variable(name="control",
shape=[4, 4, self.act_space, 32]),
output_shape=[
get_shape(feature)[0], 21, 21, self.act_space
],
strides=1,
padding="SAME")
image = tf.reshape(image,
shape=[shape[0], shape[1]] +
get_shape(image)[-3:])
return image
def sent_level_attention(self):
with tf.variable_scope('sent-level') as scope:
sent_inputs = tf.reshape(self.word_outputs, [-1, self.max_sent_length, 2 * self.cell_dim])
# sentence encoder
cell_fw = rnn.GRUCell(self.cell_dim, name='cell_fw')
cell_bw = rnn.GRUCell(self.cell_dim, name='cell_bw')
init_state_fw = tf.tile(tf.get_variable('init_state_fw',
shape=[1, self.cell_dim],
initializer=tf.constant_initializer(0)),
multiples=[get_shape(sent_inputs)[0], 1])
init_state_bw = tf.tile(tf.get_variable('init_state_bw',
shape=[1, self.cell_dim],
initializer=tf.constant_initializer(0)),
multiples=[get_shape(sent_inputs)[0], 1])
rnn_outputs, _ = bidirectional_rnn(cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=sent_inputs,
input_lengths=self.sent_lengths,
initial_state_fw=init_state_fw,
initial_state_bw=init_state_bw,
scope=scope)
sent_outputs, sent_att_weights = attention(inputs=rnn_outputs,
att_dim=self.att_dim,
sequence_lengths=self.sent_lengths)
self.sent_outputs = tf.layers.dropout(sent_outputs, self.dropout_rate, training=self.is_training)
def make_demo(sess, dphs, enqueue_op):
burn_in = FLAGS.burn_in
seqlen = FLAGS.seqlen + burn_in
n_step = FLAGS.n_step
gamma = FLAGS.gamma
if FLAGS.use_all_demos:
names = glob.glob("/opt/tiger/test_ppo/Demos/*.demo")
else:
names = glob.glob("/opt/tiger/test_ppo/Demos/*_0.demo")
fd = OrderedDict()
for name in names:
dicseg = QueueReader.read(name)[0]
dicseg["n_step_r"] = get_n_step_rewards(dicseg["r"], n_step, gamma)
while len(dicseg["s"]) > burn_in:
next_seg = dict()
next_seg["s"] = padding(dicseg["s"][:seqlen], seqlen,
dicseg["s"][0], np.float32)
next_seg["prev_a"] = padding(dicseg["a"][:seqlen], seqlen,
dicseg["a"][0], np.int32)
next_seg["a"] = padding(dicseg["a"][1:seqlen + 1], seqlen,
dicseg["a"][0], np.int32)
next_seg["r"] = padding(dicseg["a"][:seqlen], seqlen,
dicseg["r"][0], np.float32)
next_seg["n_step_r"] = padding(dicseg["n_step_r"][:seqlen], seqlen,
dicseg["n_step_r"][0], np.float32)
next_seg["state_in"] = np.zeros(get_shape(dphs["state_in"])[1:])
next_seg["slots"] = padding([1] * len(dicseg["s"][:-1][:seqlen]),
seqlen, 0, np.int32)
next_seg["bootstrap_s"] = padding(
dicseg["s"][seqlen:seqlen + n_step], n_step, dicseg["s"][0],
np.float32)
next_seg["bootstrap_prev_a"] = padding(
dicseg["a"][seqlen:seqlen + n_step], n_step, dicseg["a"][0],
np.int32)
next_seg["bootstrap_slots"] = padding(
[1] * len(dicseg["s"][:-1][seqlen:seqlen + n_step]), n_step, 0,
np.int32)
next_seg["a_logits"] = np.zeros(get_shape(dphs["a_logits"])[1:])
next_seg["v_cur"] = np.zeros(get_shape(dphs["v_cur"])[1:])
next_seg["v_tar"] = np.zeros(get_shape(dphs["v_tar"])[1:])
next_seg["adv"] = np.zeros(get_shape(dphs["adv"])[1:])
dicseg = {k: v[burn_in:] for k, v in dicseg.items()}
for key in dphs:
if key == "priority":
fd[dphs[key]] = [-1000000]
else:
fd[dphs[key]] = [next_seg[key]]
sess.run(enqueue_op, feed_dict=fd)
def coex(s, s1, a, act_size, layers=2, activation=tf.nn.leaky_relu, scope="coex"):
# s1 = tf.stop_gradient(s1)
s_shape = get_shape(s)
s1_shape = get_shape(s1)
assert len(s_shape) > 3
assert len(s1_shape) > 3
s_size = s_shape[-1]
s1_size = s1_shape[-1]
assert s_size == s1_size
feature_size = s_size
s = tf.reshape(
s, shape=s_shape[:-3] + [s_shape[-3] * s_shape[-2], s_size])
s1 = tf.reshape(
s1, shape=s1_shape[:-3] + [s1_shape[-3] * s1_shape[-2], s1_size])
with tf.variable_scope(scope):
with tf.variable_scope("attentive_dynamic_model"):
e_logits = tf.concat([s1 - s, s], axis=-1)
for i in range(layers - 1):
e_logits = tf.layers.dense(
e_logits,
feature_size,
activation=activation,
name="e_mlp_%d" % i)
e_logits = tf.layers.dense(
e_logits,
act_size,
activation=None,
name="e_logits")
alpha_logits = s1
for i in range(layers - 1):
alpha_logits = tf.layers.dense(
alpha_logits,
feature_size,
activation=activation,
name="alpha_mlp_%d" % i)
alpha_logits = tf.layers.dense(
alpha_logits,
act_size,
activation=None,
name="alpha_logits")
alpha = tf.nn.softmax(alpha_logits, axis=-2)
a_logits = tf.reduce_sum(e_logits * alpha, axis=-2)
i_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=a, logits=a_logits)
tf.summary.scalar("i_loss", tf.reduce_mean(i_loss))
return i_loss
def _init_word_encoder(self):
'''
Build Word Encoder part as in the paper
:return:
'''
with tf.variable_scope('word-encoder') as scope:
# collapses num docs,num of sentences and creates (number sentences, number words,embedding)
# treats each sentece independent of docs, sentence location
word_inputs = tf.reshape(self.embedded_inputs,
[-1, self.max_word_length, self.emb_size])
# containing the length of each sentence
word_lengths = tf.reshape(self.word_lengths, [-1])
# define forward and backword GRU cells
cell_fw = rnn.GRUCell(self.cell_dim, name='cell_fw')
cell_bw = rnn.GRUCell(self.cell_dim, name='cell_bw')
# initialize state of forward GRU cell as 0's, for each sentence in batch
init_state_fw = tf.tile(tf.get_variable(
'init_state_fw',
shape=[1, self.cell_dim],
initializer=tf.constant_initializer(0)),
multiples=[get_shape(word_inputs)[0], 1])
# same but for backward GRU cell
init_state_bw = tf.tile(tf.get_variable(
'init_state_bw',
shape=[1, self.cell_dim],
initializer=tf.constant_initializer(0)),
multiples=[get_shape(word_inputs)[0], 1])
# bidirectional_rnn returns outputs, state; why do we keep the output and not hidden state???
rnn_outputs, _ = bidirectional_rnn(cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=word_inputs,
input_lengths=word_lengths,
initial_state_fw=init_state_fw,
initial_state_bw=init_state_bw,
scope=scope)
# rnn_outputs.shape = [number sentences, number words, 2*self.cell_dim]
# word_outputs sentence vectors, word_att_weights alpha
# output dim for word_outputs (num sentences,1,2* hidden state cell dim); sentence vectors as in paper
word_outputs, word_att_weights = attention(
inputs=rnn_outputs,
att_dim=self.att_dim,
sequence_lengths=word_lengths)
# apply dropout, only activate during training
self.word_outputs = tf.layers.dropout(word_outputs,
self.dropout_rate,
training=self.is_training)
def get_channel_indices(args, nrms=5.):
# Initial moments
if args.chanran:
aux1, aux2 = args.chanran
ind = range(aux1, aux2 + 1)
elif args.chans:
ichans = args.chans
nrows, ncols = ut.get_shape(args, len(ichans))
args.logger.info('Number of channels = %i', len(ichans))
return ichans, nrows, ncols
else:
raise ValueError('No channels were selected')
# Get rms
rms = quick_rms(args.cube.unmasked_data[ind, :, :].value) * args.cube.unit
args.logger.info('Preliminary rms: %s', rms)
mask = args.cube >= nrms * rms
mask = np.squeeze(mask.include())
nvalid = np.sum(mask, axis=(-1, -2))
# Final channel limits
for i, nval in zip(ind, nvalid[ind]):
if nval == 0:
args.logger.info('Channel %i below threshold', i)
aux1 = i + 1
else:
break
for i, nval in zip(ind[::-1], nvalid[ind][::-1]):
if nval == 0:
args.logger.info('Channel %i below threshold', i)
aux2 = i - 1
else:
break
args.logger.info('Final channel range: %i, %i, %i', aux1, aux2,
args.every[0])
# Final results
ichans = range(aux1, aux2 + 1, args.every[0])
nrows, ncols = ut.get_shape(args, len(ichans), minimize=True)
args.logger.info('Number of channels = %i', len(ichans))
args.logger.info('Figure size = %i', len(ichans))
while nrows * ncols < len(ichans):
i0 = ichans[0]
i1 = ichans[-1]
if nvalid[i0] <= nvalid[i1]:
args.logger.info('Dropping channel %i to match shape', i0)
ichans = ichans[1:]
else:
args.logger.info('Dropping channel %i to match shape', i1)
ichans = ichans[:-1]
return ichans, nrows, ncols
def KL_from_gaussians(p_mus, p_sigmas, q_mus, q_sigmas):
k = get_shape(p_mus)[-1]
assert k == get_shape(p_sigmas)[-1]
assert k == get_shape(q_mus)[-1]
assert k == get_shape(q_sigmas)[-1]
trace_term = tf.reduce_sum(p_sigmas / q_sigmas, axis=-1)
quadratic_term = tf.reduce_sum((q_mus - p_mus)**2 / q_sigmas, axis=-1)
k_term = tf.cast(k, tf.float32)
log_det_term = tf.reduce_sum(tf.math.log(q_sigmas) - tf.math.log(p_sigmas),
axis=-1)
kl = 0.5 * (trace_term + quadratic_term - k_term + log_det_term)
return kl
def _init_sent_encoder(self):
'''
Build Sentence Encoder part as in the paper
:return:
'''
with tf.variable_scope('sent-encoder') as scope:
# input shape: (number docs, max sentence per document, 2*cell_dim)
sent_inputs = tf.reshape(
self.word_outputs,
[-1, self.max_sent_length, 2 * self.cell_dim])
# sentence encoder
cell_fw = rnn.GRUCell(self.cell_dim, name='cell_fw')
cell_bw = rnn.GRUCell(self.cell_dim, name='cell_bw')
# for each document get the hidden state array
init_state_fw = tf.tile(tf.get_variable(
'init_state_fw',
shape=[1, self.cell_dim],
initializer=tf.constant_initializer(0)),
multiples=[get_shape(sent_inputs)[0], 1])
init_state_bw = tf.tile(tf.get_variable(
'init_state_bw',
shape=[1, self.cell_dim],
initializer=tf.constant_initializer(0)),
multiples=[get_shape(sent_inputs)[0], 1])
rnn_outputs, _ = bidirectional_rnn(cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=sent_inputs,
input_lengths=self.sent_lengths,
initial_state_fw=init_state_fw,
initial_state_bw=init_state_bw,
scope=scope)
# rnn_outputs.shape = [num docs, number sentences, 2*self.cell_dim]
# Returns document vectors
# output dim for word_outputs (num docs,1,2* hidden state cell dim); sentence vectors as in paper
sent_outputs, sent_att_weights = attention(
inputs=rnn_outputs,
att_dim=self.att_dim,
sequence_lengths=self.sent_lengths)
#dropout
self.sent_outputs = tf.layers.dropout(sent_outputs,
self.dropout_rate,
training=self.is_training)
def diagonal_bilstm(inputs, hidden_dims, use_residual=False, scope='diagonal_bilstm'):
with tf.variable_scope(scope):
def reverse(inputs):
return tf.reverse(inputs, [False, False, True, False])
output_state_fw = diagonal_lstm(inputs, hidden_dims, scope='output_state_fw')
output_state_bw = reverse(diagonal_lstm(reverse(inputs), hidden_dims, scope='output_state_bw'))
if use_residual:
#conv2d(input, num_outputs, kernel_height, kernel_width, mask_type='A', scope='conv2d'):
residual_state_fw = conv2d(output_state_fw, hidden_dims * 2, 1, 1, "B", scope="residual_fw")
output_state_fw = residual_state_fw + inputs
residual_state_bw = conv2d(output_state_bw, hidden_dims * 2, 1, 1, "B", scope="residual_bw")
output_state_bw = residual_state_bw + inputs
batch, height, width, channel = get_shape(output_state_bw)
output_state_bw_except_last = tf.slice(output_state_bw, [0, 0, 0, 0], [-1, height-1, -1, -1])
output_state_bw_only_last = tf.slice(output_state_bw, [0, height-1, 0, 0], [-1, 1, -1, -1])
dummy_zeros = tf.zeros_like(output_state_bw_only_last)
output_state_bw_with_last_zeros = tf.concat(1, [output_state_bw_except_last, dummy_zeros])
return output_state_fw + output_state_bw_with_last_zeros
def categorical(logits):
shape = get_shape(logits)
if len(shape) > 2:
logits = tf.reshape(logits, shape=[-1, shape[-1]])
samples = tf.random.categorical(logits, 1)
samples = tf.reshape(samples, shape=shape[:-1] + [1])
return samples
def load_act_model(load_file, model_scope, env, nenvs=1, num_actions=5):
print('Loading from...', load_file)
ob_shape = utils.get_shape(env.observation_space)
ac_space = env.action_space
sess = tf.get_default_session()
act = CnnPolicy(sess,
ob_shape,
ac_space,
nenvs,
1,
model_scope,
reuse=False)
with tf.variable_scope(model_scope):
params = tf.trainable_variables(model_scope)
loaded_params = joblib.load(Config.MODEL_DIR + load_file)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
sess.run(restores)
return act
def max_over_samples(inputs, context: ModularContext):
""" Maximise the value of inputs over samples."""
if context.mode == ModularMode.SAMPLES_EVALUATION:
shape = get_shape(inputs)
inputs = tf.reshape(inputs, [context.sample_size, -1] + shape[1:])
return tf.reduce_max(inputs, axis=0)
return inputs
def conv2d(input, num_outputs, kernel_height, kernel_width, mask_type='A', scope='conv2d'):
with tf.variable_scope(scope):
batch_size, image_height, image_width, num_channels = get_shape(input)
center_height = kernel_height // 2
center_width = kernel_width // 2
# initialize kernel weights
weights_shape = [kernel_height, kernel_width, num_channels, num_outputs]
weights = tf.get_variable("weights", weights_shape, tf.float32, WEIGHT_INITIALIZER, None)
# pre-convolution mask
mask_shape = (kernel_height, kernel_width, num_channels, num_outputs)
mask = np.ones(mask_shape, dtype=np.float32)
mask[center_height, center_width+1:, :, :] = 0.0
mask[center_height+1:, :, :, :] = 0.0
# in type A, we do not allow a connection to the current focus of the kernel
# which is its center pixel
if mask_type.lower() == 'a':
mask[center_height, center_width, :, :] = 0.0
# apply the mask
weights *= tf.constant(mask, dtype=tf.float32)
# store the weights variable
tf.add_to_collection('conv2d_weights_mask_%s' % mask_type, weights)
stride_shape = [1, 1, 1, 1]
outputs = tf.nn.conv2d(input, weights, stride_shape, padding='SAME', name='conv2d_outputs')
tf.add_to_collection('conv2d_outputs', outputs)
return outputs
def __init__(self,
*,
env,
model,
opponent_model1,
opponent_model2=None,
nsteps,
gamma,
lam):
self.env = env
self.model = MultiModel(model, opponent_model1, opponent_model2)
nenv = env.num_envs
input_shape = utils.get_shape(env.observation_space)
self.primary_obs = np.zeros((nenv, ) + input_shape,
dtype=model.train_model.X.dtype.name)
self.opponent_obs1 = np.zeros((nenv, ) + input_shape,
dtype=model.train_model.X.dtype.name)
self.opponent_obs2 = None
if Config.NUM_SNAKES == 3:
self.opponent_obs2 = np.zeros((nenv, ) + input_shape,
dtype=model.train_model.X.dtype.name)
multi_agent_obs = env.reset()
self.use_multi_agent_obs(multi_agent_obs)
self.gamma = gamma
self.lam = lam
self.nsteps = nsteps
self.states = model.initial_state
self.dones = [False for _ in range(nenv)]
def _decode_lstm(self, x, h, context):
with tf.compat.v1.variable_scope('decode_lstm'):
w_h = tf.compat.v1.get_variable(
'w_h', [self.C, self.W], initializer=self.weight_initializer)
b_h = tf.compat.v1.get_variable('b_h', [self.W],
initializer=self.const_initializer)
w_out = tf.compat.v1.get_variable(
'w_out', [self.W, self.V], initializer=self.weight_initializer)
b_out = tf.compat.v1.get_variable(
'b_out', [self.V], initializer=self.const_initializer)
h = tf.layers.dropout(h,
self.dropout_rate,
training=self.is_training)
h_logits = tf.matmul(h, w_h) + b_h
w_ctx2out = tf.compat.v1.get_variable(
'w_ctx2out', [get_shape(context)[1], self.W],
initializer=self.weight_initializer)
h_logits += tf.matmul(context, w_ctx2out)
h_logits += x
h_logits = tf.nn.tanh(h_logits)
h_logits = tf.layers.dropout(h_logits,
self.dropout_rate,
training=self.is_training)
out_logits = tf.matmul(h_logits, w_out) + b_out
return out_logits
def plot_data(args):
# Keyword arguments for tile plotter
opts = {}
opts['nrows'], opts['ncols'] = utils.get_shape(args,
len(args.data),
default_cols=1)
assert opts['nrows'] * opts['ncols'] >= len(args.data)
args.logger.info('Rows, columns = %i, %i', opts['nrows'], opts['ncols'])
# Setup tile plotter
args.logger.debug('Initializing figure')
fig = plt.NPlotter(config=args.config[0], section=args.section[0], **opts)
# Iterate over data
xunits = []
yunits = []
for i, (loc, data) in enumerate(zip(fig.axes, args.data)):
label = fig.get_value('axlabel', None, loc, sep=',')
label = utils.get_axis_label(args, i, label)
overplots = utils.get_overplots(args, i)
ax, xunit, yunit = utils.splot(loc,
fig,
data,
args.logger,
overplots=overplots,
cols=args.columns,
errorcol=args.errcols[0])
xunits += [xunit]
yunits += [yunit]
fig.auto_config(legend=args.legend, units=(xunits, yunits))
return fig
def get_feature_size_per_file(self, f_name):
"""
Return the dimensionality of the features in a given file.
Typically, this will be the number of bins in a T-F representation
"""
shape = utils.get_shape(os.path.join(f_name.replace('.data',
'.shape')))
return shape[1]
def benchmark_onnxruntime(path_to_model,
repeat=1000,
number=1,
warmup=100,
quantize=False):
"""
Parameters
----------
path_to_model: str or onnx.ModelProto
Path to an onnx model.
repeat: int
Repetition of experiment. Default: 1000
number: int
Number of forward passes in each experiment. Default: 1
warmup: int
Number of disregarded experiments. Default: 100
quantize: bool
Dynamically quantize the model with default parameters.
Returns
-------
info: dict
Information about the size and min, max, mean, std of the time
of the experiments.
"""
assert repeat >= 2 * warmup
if quantize:
import onnx
from onnx import version_converter
from onnxruntime.quantization import quantize_dynamic
orig_model = onnx.load(path_to_model)
if orig_model.opset_import[0].version < 11:
converted_model = version_converter.convert_version(orig_model, 11)
path_to_model = '/tmp/model_conv.onnx'
with open(path_to_model, 'wb') as f:
f.write(converted_model.SerializeToString())
del orig_model, converted_model
path_to_quant_model = "/tmp/model_quant.onnx"
model = quantize_dynamic(path_to_model, path_to_quant_model)
size = os.path.getsize(path_to_quant_model)
sess = ort.InferenceSession(path_to_quant_model)
else:
size = os.path.getsize(path_to_model)
sess = ort.InferenceSession(path_to_model)
inputs = {
x.name: np.random.randn(*get_shape(x)).astype(get_type(x))
for x in sess.get_inputs()
}
def _benchmark():
output = sess.run(None, inputs)
res = dict(size=size, input_size=[tuple(x.shape) for x in inputs.values()])
res.update(benchmark_speed(_benchmark, repeat, number, warmup))
return res
def icm(s, s1, a, act_size, layers=2, activation=tf.nn.relu, scope="icm"):
"""Curiosity-driven Exploration by Self-supervised Prediction"""
# s1 = tf.stop_gradient(s1)
s_size = get_shape(s)[-1]
s1_size = get_shape(s1)[-1]
assert s_size == s1_size
feature_size = s_size
a_onehot = tf.one_hot(a, act_size, dtype=tf.float32)
with tf.variable_scope(scope):
with tf.variable_scope("forward_model"):
s1_hat = tf.concat([s, a_onehot], axis=-1)
for i in range(layers - 1):
s1_hat = tf.layers.dense(s1_hat,
feature_size,
activation=activation,
name="layer_%d" % i)
s1_hat = tf.layers.dense(s1_hat,
feature_size,
activation=None,
name="predict_target")
f_loss = 0.5 * tf.reduce_sum(tf.square(s + s1_hat - s1), axis=-1)
with tf.variable_scope("inverse_model"):
a_logits_hat = tf.concat([s, s1 - s], axis=-1)
for i in range(layers - 1):
a_logits_hat = tf.layers.dense(a_logits_hat,
feature_size,
activation=activation,
name="layers_%d" % i)
a_logits_hat = tf.layers.dense(a_logits_hat,
act_size,
activation=None,
name="predict_act_logits")
i_loss = tf.nn.softmax_cross_entropy_with_logits(
labels=a_onehot, logits=a_logits_hat)
tf.summary.scalar("f_loss", tf.reduce_mean(f_loss))
tf.summary.scalar("i_loss", tf.reduce_mean(i_loss))
return icmLoss(f_loss, i_loss)
def get_num_instances_per_file(self, f_name):
"""
Return the number of context_windows, patches, or instances generated out of a given file
"""
shape = utils.get_shape(os.path.join(f_name.replace('.data',
'.shape')))
file_frames = float(shape[0])
return np.maximum(
1, int(np.ceil((file_frames - self.patch_len) / self.patch_hop)))
def unskew(skewed_outputs, width=0, scope="unskew"):
with tf.name_scope(scope):
batch, height, skewed_width, channel = get_shape(skewed_outputs)
#rows = tf.unpack(tf.transpose(skewed_outputs, [1, 0, 2, 3,])) # [height, batch, width, channel]
rows = tf.split(1, height, skewed_outputs) # [batch, 1, width, channel]
width = width if width else height
unskewed_rows = []
# iterate through the rows
for i, row in enumerate(rows):
sliced_row = tf.slice(row, [0, 0, i, 0], [-1, -1, width, -1])
unskewed_rows.append(sliced_row)
unskewed_output = tf.concat(1, unskewed_rows, name="unskewed_output")
desired_shape = [None, height, width, channel]
output_shape = get_shape(unskewed_output)
assert output_shape == desired_shape, "wrong shape of unskewed output. Actual {}; Expected {}".format(output_shape, desired_shape)
return unskewed_output
def __init__(self, act_space, lstm, scope="agent", **kwargs):
self.act_space = act_space
self.scope = scope
self.s = kwargs.get("s")
self.prev_a = kwargs.get("prev_a")
self.state_in = kwargs.get("state_in")
self.slots = tf.cast(kwargs.get("slots"), tf.float32)
feature, self.state_out = self.feature_net(self.s, lstm, self.prev_a,
self.state_in)
self.qf = self.q_fn(feature, self.slots, self.scope + "_current")
self.current_act = tf.argmax(self.qf, axis=-1)
# base_probs = tf.ones(
# get_shape(self.prev_a) + [act_space]
# ) * epsilon / tf.cast(act_space, tf.float32)
# argmax_a = tf.argmax(self.qf, axis=-1)
# argmax_probs = tf.one_hot(
# argmax_a, depth=act_space, dtype=tf.float32
# ) * (1.0 - epsilon)
# self.current_act_probs = base_probs + argmax_probs
#
# self.current_act = tf.squeeze(
# categorical(tf.math.log(self.current_act_probs)), axis=-1)
self.bootstrap_s = kwargs.get("bootstrap_s")
if self.bootstrap_s is not None:
self.bootstrap_prev_a = kwargs.get("bootstrap_prev_a")
self.bootstrap_slots = tf.cast(kwargs.get("bootstrap_slots"),
tf.float32)
self.a = kwargs.get("a")
self.r = kwargs.get("r")
self.qa = tf.reduce_sum(
tf.one_hot(self.a, depth=self.act_space, dtype=tf.float32) *
self.qf,
axis=-1)
bootstrap_feature, _ = self.feature_net(self.bootstrap_s, lstm,
self.bootstrap_prev_a,
self.state_out)
n_step = get_shape(bootstrap_feature)[1]
feature1 = tf.concat([feature[:, n_step:, :], bootstrap_feature],
axis=1)
slots1 = tf.concat([self.slots[:, n_step:], self.bootstrap_slots],
axis=1)
self.q1f1 = self.q_fn(feature1, slots1, self.scope + "_target")
self.q1f = self.q_fn(feature1, slots1, self.scope + "_current")
self.qa1 = doubleQ(self.q1f1, self.q1f)
def categorical(tensor, num):
shape = get_shape(tensor)
if len(shape) == 2:
return tf.random.categorical(tensor, num)
elif len(shape) == 3:
new = tf.reshape(tensor, [-1, shape[-1]])
sample = tf.random.categorical(new, num)
return tf.reshape(sample, [shape[0], shape[1], num])
else:
raise ValueError(tensor.name + "should have dim 2 or 3")
def pad_left(self, x, size, dim, shift_right=False):
if self.state_old is None:
shape = get_shape(x)
shape[dim] = size
x_pad = torch.zeros(shape, dtype=x.dtype, device=x.device)
else:
x_pad = self.state_old[0]
#add left part of x
x_padded = torch.cat([x_pad, x], dim)
#get right part for padding on next iter
x_splited = torch.split(x_padded,
[get_shape(x_padded)[dim] - size, size],
dim=dim)
self.update(x_splited[-1:])
return x_splited[0] if shift_right else x_padded
def conv1d(input, num_outputs, kernel_size, scope='conv1d'):
with tf.variable_scope(scope):
batch_size, image_height, image_width, num_channels = get_shape(input)
kernel_height, kernel_width = kernel_size, 1
# initialize kernel weights
weights_shape = [kernel_height, kernel_width, num_channels, num_outputs]
weights = tf.get_variable("weights", weights_shape, tf.float32, WEIGHT_INITIALIZER, None)
stride_shape = [1, 1, 1, 1]
outputs = tf.nn.conv2d(input, weights, stride_shape, padding='SAME', name='conv1d_outputs')
return outputs
def encode(self, x):
# [B,T]->[B,1,1,T]
x = x.unsqueeze(1).unsqueeze(2)
# [B,1,1,T]->[B,2F,1,T]
X = torch.nn.functional.conv2d(x,
self._encdec_matrix,
stride=self.hop_length)
# [B,2F,1,T]->[B,2,F,T]
B, F2, _, T = get_shape(X)
return X.reshape(B, 2, F2 // 2, T)
def reconstruct_net(self, feature, scope="reconstruct"):
shape = get_shape(feature)
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
feature = tf.reshape(feature, [-1, shape[-1]])
feature = tf.layers.dense(feature,
800,
tf.nn.relu,
name="feature")
image = tf.reshape(feature, [-1, 5, 5, 32])
filter = [16, 32, 32]
size = [(84, 82), (40, 38), (18, 7)]
kernel = [(3, 3), (3, 3), (5, 3)]
stride = [(1, 2), (1, 2), (2, 1)]
for i in range(len(filter) - 1, -1, -1):
image = self.resblock(image, "res0_%d" % i)
image = tf.image.resize_nearest_neighbor(
image, [size[i][1], size[i][1]])
output_channels = filter[i - 1] if i > 0 else 1
input_channels = filter[i]
image = tf.nn.conv2d_transpose(
image,
filter=tf.get_variable(name="deconv_%d" % i,
shape=[
kernel[i][0], kernel[i][0],
output_channels,
input_channels
]),
output_shape=[
get_shape(feature)[0], size[i][0], size[i][0],
output_channels
],
strides=stride[i][0],
padding="VALID")
image = tf.reshape(image,
shape=[shape[0], shape[1]] +
get_shape(image)[-3:])
return image
