def scorer_head_model(self, features, training, reuse=False):
with tf.variable_scope("scorer_head", reuse=reuse): # define variable scope
dense1 = layers.dense_layer(features, units=1024, use_bias=True)
relu1 = layers.relu_layer(dense1)
dense2 = layers.dense_layer(relu1, units=512, use_bias=True)
relu2 = layers.relu_layer(dense2)
dense3 = layers.dense_layer(relu2, units=256, use_bias=True)
relu3 = layers.relu_layer(dense3)
dense4 = layers.dense_layer(relu3, units=1, use_bias=True)
return dense4
def generator_model(self, noise, feats, training, reuse=False): # construct the graph of the generator
a = time.time()
with tf.variable_scope("generator",
reuse=reuse): # define variable scope to easily retrieve vars of the generator
gen_inp = tf.concat([noise, feats], -1)
with tf.name_scope("preprocess_inp"):
dense1 = layers.dense_layer(gen_inp, units=4 * 4 * 1024, use_bias=False)
bn1 = layers.batch_norm_layer_mcgan(dense1, training, 0.8)
relu1 = layers.relu_layer(bn1)
reshaped = tf.reshape(relu1, shape=[-1, 1024, 4, 4]) # shape=(batch_size, 1024, 4, 4)
deconv1 = layers.deconv_block_mcgan(reshaped, training, momentum=0.8, out_channels=512, filter_size=(4, 4),
strides=(2, 2), padding="same",
use_bias=True) # shape=(batch_size, 512, 8, 8)
deconv2 = layers.deconv_block_mcgan(deconv1, training, momentum=0.8, out_channels=256, filter_size=(4, 4), strides=(2, 2),
padding="same", use_bias=True) # shape=(batch_size, 256, 16, 16)
deconv3 = layers.deconv_block_mcgan(deconv2, training, momentum=0.8, out_channels=128, filter_size=(4, 4), strides=(2, 2),
padding="same", use_bias=True) # shape=(batch_size, 128, 32, 32)
deconv4 = layers.deconv_layer(deconv3, out_channels=1, filter_size=(4, 4), strides=(2, 2), padding="same",
use_bias=True) # shape=(batch_size, 1, 64, 64)
gen_out = layers.tanh_layer(deconv4)
print("Built Generator model in {} s".format(time.time() - a))
list_ops = {"dense1": dense1, "bn1":bn1, "relu1": relu1, "reshaped": reshaped, "deconv1": deconv1, "deconv2": deconv2,
"deconv3": deconv3, "deconv4": deconv4,
"gen_out": gen_out} # list of operations, can be used to run the graph up to a certain op
# i,e get the subgraph
return gen_out, list_ops
def __call__(self, inp, training):
a = time.time()
with tf.variable_scope("StackedSRM"): # define variable scope
inter = inp # intermediate input
outputs = []
for i in range(self.nb_stacks):
with tf.name_scope("stack"):
conv = layers.conv_layer(inter,
out_channels=64,
filter_size=(4, 4),
strides=(2, 2),
padding=(1, 1),
pad_values=0,
use_bias=False)
relu = layers.relu_layer(conv)
res_module = layers.residual_module_srm(relu,
training,
out_channels=64,
nb_blocks=6,
pad_values=0,
use_bias=False)
h, w = tf.shape(res_module)[2], tf.shape(res_module)[3]
up_sample1 = layers.resize_layer(
res_module,
new_size=[2 * h, 2 * w],
resize_method=tf.image.ResizeMethod.NEAREST_NEIGHBOR
) # nearest neighbor up sampling
conv1 = layers.conv_layer(up_sample1,
out_channels=128,
filter_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
pad_values=0,
use_bias=False)
h, w = tf.shape(conv1)[2], tf.shape(conv1)[3]
up_sample2 = layers.resize_layer(
conv1,
new_size=[2 * h, 2 * w],
resize_method=tf.image.ResizeMethod.NEAREST_NEIGHBOR
) # nearest neighbor up sampling
conv2 = layers.conv_layer(up_sample2,
out_channels=1,
filter_size=(3, 3),
strides=(1, 1),
padding=(1, 1),
pad_values=0,
use_bias=False)
inter = (
layers.tanh_layer(conv2) + 1
) / 2.0 # apply tanh and renormalize so that the output is in the range [0, 1] to prepare it to be inputted to the next stack
outputs.append(inter)
print("SRM Model built in {} s".format(time.time() - a))
return outputs
def discriminator(input_vectors, sequence_length, is_real=True, **opts):
"""
Args:
input_vectors: output of the RNN either from real or generated data
sequence_length: TensorFlow queue or placeholder tensor for number of word ids for each sentence
is_real: if True, RNN outputs when feeding in actual data, if False feeds in generated data
"""
tf.logging.info(" Setting up discriminator")
rnn_final_state = (
input_vectors >>
lay.dense_layer(hidden_dims=opts["embedding_dim"]) >>
lay.recurrent_layer(sequence_length=sequence_length, hidden_dims=opts["rnn_hidden_dim"],
return_final_state=True)
)
prediction_logits = (
rnn_final_state >>
lay.dense_layer(hidden_dims=opts["output_hidden_dim"]) >>
lay.relu_layer() >>
lay.dropout_layer(opts["output_dropout_keep_prob"]) >>
lay.dense_layer(hidden_dims=opts["output_hidden_dim"]) >>
lay.relu_layer() >>
lay.dropout_layer(opts["output_dropout_keep_prob"]) >>
lay.dense_layer(hidden_dims=1)
)
if is_real:
target = tf.ones_like(prediction_logits)
else:
target = tf.zeros_like(prediction_logits)
# TODO: add accuracy
loss = (
prediction_logits >>
lay.sigmoid_cross_entropy_layer(target=target)
)
# TODO: return logits in case for WGAN and l2 GANs
return DiscriminatorTuple(rnn_final_state=rnn_final_state, prediction_logits=prediction_logits, loss=loss)
def discriminator(input_vectors, sequence_length, is_real=True, **opts):
"""
Args:
input_vectors: output of the RNN either from real or generated data
sequence_length: TensorFlow queue or placeholder tensor for number of word ids for each sentence
is_real: if True, RNN outputs when feeding in actual data, if False feeds in generated data
"""
tf.logging.info(" Setting up discriminator")
rnn_final_state = (
input_vectors >>
lay.dense_layer(hidden_dims=opts["embedding_dim"]) >>
lay.recurrent_layer(sequence_length=sequence_length, hidden_dims=opts["rnn_hidden_dim"],
return_final_state=True)
)
prediction_logits = (
rnn_final_state >>
lay.dense_layer(hidden_dims=opts["output_hidden_dim"]) >>
lay.relu_layer() >>
lay.dropout_layer(opts["output_dropout_keep_prob"]) >>
lay.dense_layer(hidden_dims=opts["output_hidden_dim"]) >>
lay.relu_layer() >>
lay.dropout_layer(opts["output_dropout_keep_prob"]) >>
lay.dense_layer(hidden_dims=1)
)
if is_real:
target = tf.ones_like(prediction_logits)
else:
target = tf.zeros_like(prediction_logits)
# TODO: add accuracy
loss = (
prediction_logits >>
lay.sigmoid_cross_entropy_layer(target=target)
)
# TODO: return logits in case for WGAN and l2 GANs
return DiscriminatorTuple(rnn_final_state=rnn_final_state, prediction_logits=prediction_logits, loss=loss)
def discriminator_model(self, inp, feats, training, reuse=False, resize=False, minibatch=False): # construct the graph of the discriminator
a = time.time()
with tf.variable_scope("discriminator",
reuse=reuse): # define variable scope to easily retrieve vars of the discriminator
if resize:
inp = layers.max_pool_layer(inp, pool_size=(2, 2), strides=(2, 2), padding=(12, 12))
inp = layers.max_pool_layer(inp, pool_size=(2, 2), strides=(2, 2))
inp = layers.max_pool_layer(inp, pool_size=(2, 2), strides=(2, 2))
inp = layers.max_pool_layer(inp, pool_size=(2, 2), strides=(2, 2))
conv1 = layers.conv_block_mcgan(inp, training, momentum=0.8, out_channels=128, filter_size=(4, 4), strides=(2, 2),
padding="same", use_bias=True, batch_norm=False,
alpha=0.3) # shape=(batch_size, 128, 32, 32)
conv2 = layers.conv_block_mcgan(conv1, training, momentum=0.8, out_channels=256, filter_size=(4, 4), strides=(2, 2),
padding="same", use_bias=True,
alpha=0.3) # shape=(batch_size, 256, 16, 16)
conv3 = layers.conv_block_mcgan(conv2, training, momentum=0.8, out_channels=512, filter_size=(4, 4), strides=(2, 2),
padding="same", use_bias=True, alpha=0.3) # shape=(batch_size, 512, 8, 8)
conv4 = layers.conv_block_mcgan(conv3, training, momentum=0.8, out_channels=1024, filter_size=(4, 4), strides=(2, 2),
padding="same", use_bias=True, alpha=0.3) # shape=(batch_size, 1024, 4, 4)
flat = tf.reshape(conv4, [-1, 1024 * 4 * 4])
if(minibatch):
minibatched = layers.minibatch(flat, num_kernels=5, kernel_dim=3)
dense1 = layers.dense_layer(minibatched, 128, use_bias=True)
else:
dense1 = layers.dense_layer(flat, 128, use_bias=True)
drop1 = layers.dropout_layer(dense1, training, dropout_rate=0.3)
LRU1 = layers.leaky_relu_layer(drop1, alpha=0.3)
dense2 = layers.dense_layer(feats, units=3, use_bias=True)
relu2 = layers.relu_layer(dense2)
bn2 = layers.batch_norm_layer_mcgan(relu2, training, 0.8)
dense3 = layers.dense_layer(bn2, units=1)
merged = tf.concat([LRU1, dense3], axis=-1)
logits = layers.dense_layer(merged, units=2, use_bias=True)
out = logits
print("Built Discriminator model in {} s".format(time.time() - a))
list_ops = {"inp": inp, "conv1": conv1, "conv2": conv2, "conv3": conv3, "conv4": conv4, "flat": flat,
"dense1":dense1, "drop1":drop1, "LRU1":LRU1, "dense2":dense2, "relu2":relu2, "bn2":bn2,
"dense3":dense3, "logits": logits, "out": out}
return out, list_ops
def get_model(self):
params = {
'batch_size': 20,
'learning_rate': 0.01,
'epochs': 20,
'num_classes': 10,
'dropout_rate': 0.1
}
Model = model.model()
Model.add(layers.convolution_layer(field=8, padding=0, stride=1, depth=1, filters=5))
Model.add(layers.relu_layer())
Model.add(layers.flatten_layer())
Model.add(layers.dropout_layer(r = params['dropout_rate']))
Model.add(layers.linear_layer(13*13*5, 10))
Model.set_loss(layers.softmax_cross_entropy())
Model.set_hyper_params(params)
return Model
def __init__(self,
input_size,
hidden_size,
output_size,
weight_init_std=0.01):
self.params = {}
self.params['W1'] = weight_init_std * np.random.randn(
input_size, hidden_size)
self.params['b1'] = np.zeros(hidden_size)
self.params['W2'] = weight_init_std * np.random.randn(
hidden_size, output_size)
self.params['b2'] = np.zeros(output_size)
#レイヤーの作成
self.layers = OrderedDict()
self.layers["affine1"] = lay.affin_layer(
self.params['W1'], self.params['b1']) #インスタンス生成時の引数はコンストラクタへの引数
self.layers["relu1"] = lay.relu_layer()
self.layers["affine2"] = lay.affin_layer(self.params['W2'],
self.params['b2'])
self.lastlayer = lay.softmax_loss_layer()
def run():
model_name = 'alexnet'
directory_caffe = './caffemodel'
directory_theano = './theanomodel'
url_prototxt = 'https://raw.githubusercontent.com/BVLC/caffe/master/models/bvlc_alexnet/deploy.prototxt'
url_caffemodel = 'http://dl.caffe.berkeleyvision.org/bvlc_alexnet.caffemodel'
filename_prototxt = '%s/%s.prototxt' % (directory_caffe, model_name)
filename_caffemodel = '%s/%s.caffemodel' % (directory_caffe, model_name)
filename_theanomodel = '%s/%s.model' % (directory_theano, model_name)
# download caffemodel
print 'downloading caffemodel'
if not os.path.exists(directory_caffe):
os.mkdir(directory_caffe)
if not os.path.exists(filename_prototxt):
p = subprocess.Popen(('wget', url_prototxt, '-O', filename_prototxt))
p.wait()
if not os.path.exists(filename_caffemodel):
p = subprocess.Popen((
'wget',
url_caffemodel,
'-O',
filename_caffemodel,
))
p.wait()
# load caffe model
print 'loading caffe model'
model_caffe = caffe.Net(filename_prototxt, filename_caffemodel, True)
conv1_W = theano.shared(model_caffe.params['conv1'][0].data[:, :, ::-1, ::-1])
conv2_W = theano.shared(model_caffe.params['conv2'][0].data[:, :, ::-1, ::-1])
conv3_W = theano.shared(model_caffe.params['conv3'][0].data[:, :, ::-1, ::-1])
conv4_W = theano.shared(model_caffe.params['conv4'][0].data[:, :, ::-1, ::-1])
conv5_W = theano.shared(model_caffe.params['conv5'][0].data[:, :, ::-1, ::-1])
conv1_b = theano.shared(model_caffe.params['conv1'][1].data.squeeze())
conv2_b = theano.shared(model_caffe.params['conv2'][1].data.squeeze())
conv3_b = theano.shared(model_caffe.params['conv3'][1].data.squeeze())
conv4_b = theano.shared(model_caffe.params['conv4'][1].data.squeeze())
conv5_b = theano.shared(model_caffe.params['conv5'][1].data.squeeze())
fc6_W = theano.shared(model_caffe.params['fc6'][0].data.squeeze())
fc7_W = theano.shared(model_caffe.params['fc7'][0].data.squeeze())
fc8_W = theano.shared(model_caffe.params['fc8'][0].data.squeeze())
fc6_b = theano.shared(model_caffe.params['fc6'][1].data.squeeze())
fc7_b = theano.shared(model_caffe.params['fc7'][1].data.squeeze())
fc8_b = theano.shared(model_caffe.params['fc8'][1].data.squeeze())
# make theano model
print 'building theano model'
model_theano = collections.OrderedDict()
model_theano['data'] = T.tensor4()
model_theano['conv1'] = layers.convolution_layer(model_theano['data'], conv1_W, conv1_b, subsample=(4, 4))
model_theano['relu1'] = layers.relu_layer(model_theano['conv1'])
model_theano['norm1'] = layers.lrn_layer(model_theano['relu1'])
model_theano['pool1'] = layers.pooling_layer(model_theano['norm1'])
model_theano['conv2'] = layers.convolution_layer(model_theano['pool1'], conv2_W, conv2_b, border='same', group=2)
model_theano['relu2'] = layers.relu_layer(model_theano['conv2'])
model_theano['norm2'] = layers.lrn_layer(model_theano['relu2'])
model_theano['pool2'] = layers.pooling_layer(model_theano['norm2'])
model_theano['conv3'] = layers.convolution_layer(model_theano['pool2'], conv3_W, conv3_b, border='same')
model_theano['relu3'] = layers.relu_layer(model_theano['conv3'])
model_theano['conv4'] = layers.convolution_layer(model_theano['relu3'], conv4_W, conv4_b, border='same', group=2)
model_theano['relu4'] = layers.relu_layer(model_theano['conv4'])
model_theano['conv5'] = layers.convolution_layer(model_theano['relu4'], conv5_W, conv5_b, border='same', group=2)
model_theano['relu5'] = layers.relu_layer(model_theano['conv5'])
model_theano['pool5'] = layers.pooling_layer(model_theano['relu5'])
model_theano['fc6'] = layers.inner_product_layer(model_theano['pool5'], fc6_W, fc6_b)
model_theano['relu6'] = layers.relu_layer(model_theano['fc6'])
model_theano['fc7'] = layers.inner_product_layer(model_theano['relu6'], fc7_W, fc7_b)
model_theano['relu7'] = layers.relu_layer(model_theano['fc7'])
model_theano['fc8'] = layers.inner_product_layer(model_theano['relu7'], fc8_W, fc8_b)
model_theano['prob'] = layers.softmax_layer(model_theano['fc8'])
# check
print 'checking model'
data = np.random.randn(*model_caffe.blobs['data'].data.shape)
data = data.astype(np.float32) * 10
model_caffe.blobs['data'].data[:] = data
model_caffe.forward()
theano_output = theano.function(
[model_theano['data']],
model_theano['prob'],
)(data)
error = (
theano_output.squeeze() -
model_caffe.blobs['prob'].data.squeeze()
).max()
assert error < 1e-6
# save
print 'saving'
if not os.path.exists(directory_theano):
os.mkdir(directory_theano)
sys.setrecursionlimit(100000)
pickle.dump(
model_theano,
open(filename_theanomodel, 'wb'),
protocol=pickle.HIGHEST_PROTOCOL,
)
print 'done'
def run():
model_name = 'alexnet'
directory_caffe = './caffemodel'
directory_theano = './theanomodel'
url_prototxt = 'https://raw.githubusercontent.com/BVLC/caffe/master/models/bvlc_alexnet/deploy.prototxt'
url_caffemodel = 'http://dl.caffe.berkeleyvision.org/bvlc_alexnet.caffemodel'
filename_prototxt = '%s/%s.prototxt' % (directory_caffe, model_name)
filename_caffemodel = '%s/%s.caffemodel' % (directory_caffe, model_name)
filename_theanomodel = '%s/%s.model' % (directory_theano, model_name)
# download caffemodel
print 'downloading caffemodel'
if not os.path.exists(directory_caffe):
os.mkdir(directory_caffe)
if not os.path.exists(filename_prototxt):
p = subprocess.Popen(('wget', url_prototxt, '-O', filename_prototxt))
p.wait()
if not os.path.exists(filename_caffemodel):
p = subprocess.Popen((
'wget',
url_caffemodel,
'-O',
filename_caffemodel,
))
p.wait()
# load caffe model
print 'loading caffe model'
model_caffe = caffe.Net(filename_prototxt, filename_caffemodel, True)
conv1_W = theano.shared(
model_caffe.params['conv1'][0].data[:, :, ::-1, ::-1])
conv2_W = theano.shared(
model_caffe.params['conv2'][0].data[:, :, ::-1, ::-1])
conv3_W = theano.shared(
model_caffe.params['conv3'][0].data[:, :, ::-1, ::-1])
conv4_W = theano.shared(
model_caffe.params['conv4'][0].data[:, :, ::-1, ::-1])
conv5_W = theano.shared(
model_caffe.params['conv5'][0].data[:, :, ::-1, ::-1])
conv1_b = theano.shared(model_caffe.params['conv1'][1].data.squeeze())
conv2_b = theano.shared(model_caffe.params['conv2'][1].data.squeeze())
conv3_b = theano.shared(model_caffe.params['conv3'][1].data.squeeze())
conv4_b = theano.shared(model_caffe.params['conv4'][1].data.squeeze())
conv5_b = theano.shared(model_caffe.params['conv5'][1].data.squeeze())
fc6_W = theano.shared(model_caffe.params['fc6'][0].data.squeeze())
fc7_W = theano.shared(model_caffe.params['fc7'][0].data.squeeze())
fc8_W = theano.shared(model_caffe.params['fc8'][0].data.squeeze())
fc6_b = theano.shared(model_caffe.params['fc6'][1].data.squeeze())
fc7_b = theano.shared(model_caffe.params['fc7'][1].data.squeeze())
fc8_b = theano.shared(model_caffe.params['fc8'][1].data.squeeze())
# make theano model
print 'building theano model'
model_theano = collections.OrderedDict()
model_theano['data'] = T.tensor4()
model_theano['conv1'] = layers.convolution_layer(model_theano['data'],
conv1_W,
conv1_b,
subsample=(4, 4))
model_theano['relu1'] = layers.relu_layer(model_theano['conv1'])
model_theano['norm1'] = layers.lrn_layer(model_theano['relu1'])
model_theano['pool1'] = layers.pooling_layer(model_theano['norm1'])
model_theano['conv2'] = layers.convolution_layer(model_theano['pool1'],
conv2_W,
conv2_b,
border='same',
group=2)
model_theano['relu2'] = layers.relu_layer(model_theano['conv2'])
model_theano['norm2'] = layers.lrn_layer(model_theano['relu2'])
model_theano['pool2'] = layers.pooling_layer(model_theano['norm2'])
model_theano['conv3'] = layers.convolution_layer(model_theano['pool2'],
conv3_W,
conv3_b,
border='same')
model_theano['relu3'] = layers.relu_layer(model_theano['conv3'])
model_theano['conv4'] = layers.convolution_layer(model_theano['relu3'],
conv4_W,
conv4_b,
border='same',
group=2)
model_theano['relu4'] = layers.relu_layer(model_theano['conv4'])
model_theano['conv5'] = layers.convolution_layer(model_theano['relu4'],
conv5_W,
conv5_b,
border='same',
group=2)
model_theano['relu5'] = layers.relu_layer(model_theano['conv5'])
model_theano['pool5'] = layers.pooling_layer(model_theano['relu5'])
model_theano['fc6'] = layers.inner_product_layer(model_theano['pool5'],
fc6_W, fc6_b)
model_theano['relu6'] = layers.relu_layer(model_theano['fc6'])
model_theano['fc7'] = layers.inner_product_layer(model_theano['relu6'],
fc7_W, fc7_b)
model_theano['relu7'] = layers.relu_layer(model_theano['fc7'])
model_theano['fc8'] = layers.inner_product_layer(model_theano['relu7'],
fc8_W, fc8_b)
model_theano['prob'] = layers.softmax_layer(model_theano['fc8'])
# check
print 'checking model'
data = np.random.randn(*model_caffe.blobs['data'].data.shape)
data = data.astype(np.float32) * 10
model_caffe.blobs['data'].data[:] = data
model_caffe.forward()
theano_output = theano.function(
[model_theano['data']],
model_theano['prob'],
)(data)
error = (theano_output.squeeze() -
model_caffe.blobs['prob'].data.squeeze()).max()
assert error < 1e-6
# save
print 'saving'
if not os.path.exists(directory_theano):
os.mkdir(directory_theano)
sys.setrecursionlimit(100000)
pickle.dump(
model_theano,
open(filename_theanomodel, 'wb'),
protocol=pickle.HIGHEST_PROTOCOL,
)
print 'done'
def run():
model_name = "vggnet"
directory_caffe = "./caffemodel"
directory_theano = "./theanomodel"
url_prototxt = "https://gist.githubusercontent.com/ksimonyan/3785162f95cd2d5fee77/raw/f02f8769e64494bcd3d7e97d5d747ac275825721/VGG_ILSVRC_19_layers_deploy.prototxt"
url_caffemodel = "http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_19_layers.caffemodel"
filename_prototxt = "%s/%s.prototxt" % (directory_caffe, model_name)
filename_caffemodel = "%s/%s.caffemodel" % (directory_caffe, model_name)
filename_theanomodel = "%s/%s.model" % (directory_theano, model_name)
# download caffemodel
print "downloading caffemodel"
if not os.path.exists(directory_caffe):
os.mkdir(directory_caffe)
if not os.path.exists(filename_prototxt):
p = subprocess.Popen(("wget", url_prototxt, "-O", filename_prototxt))
p.wait()
if not os.path.exists(filename_caffemodel):
p = subprocess.Popen(("wget", url_caffemodel, "-O", filename_caffemodel))
p.wait()
# load caffe model
model_caffe = caffe.Net(filename_prototxt, filename_caffemodel, True)
conv1_1_W = theano.shared(model_caffe.params["conv1_1"][0].data[:, :, ::-1, ::-1])
conv1_2_W = theano.shared(model_caffe.params["conv1_2"][0].data[:, :, ::-1, ::-1])
conv2_1_W = theano.shared(model_caffe.params["conv2_1"][0].data[:, :, ::-1, ::-1])
conv2_2_W = theano.shared(model_caffe.params["conv2_2"][0].data[:, :, ::-1, ::-1])
conv3_1_W = theano.shared(model_caffe.params["conv3_1"][0].data[:, :, ::-1, ::-1])
conv3_2_W = theano.shared(model_caffe.params["conv3_2"][0].data[:, :, ::-1, ::-1])
conv3_3_W = theano.shared(model_caffe.params["conv3_3"][0].data[:, :, ::-1, ::-1])
conv3_4_W = theano.shared(model_caffe.params["conv3_4"][0].data[:, :, ::-1, ::-1])
conv4_1_W = theano.shared(model_caffe.params["conv4_1"][0].data[:, :, ::-1, ::-1])
conv4_2_W = theano.shared(model_caffe.params["conv4_2"][0].data[:, :, ::-1, ::-1])
conv4_3_W = theano.shared(model_caffe.params["conv4_3"][0].data[:, :, ::-1, ::-1])
conv4_4_W = theano.shared(model_caffe.params["conv4_4"][0].data[:, :, ::-1, ::-1])
conv5_1_W = theano.shared(model_caffe.params["conv5_1"][0].data[:, :, ::-1, ::-1])
conv5_2_W = theano.shared(model_caffe.params["conv5_2"][0].data[:, :, ::-1, ::-1])
conv5_3_W = theano.shared(model_caffe.params["conv5_3"][0].data[:, :, ::-1, ::-1])
conv5_4_W = theano.shared(model_caffe.params["conv5_4"][0].data[:, :, ::-1, ::-1])
conv1_1_b = theano.shared(model_caffe.params["conv1_1"][1].data.squeeze())
conv1_2_b = theano.shared(model_caffe.params["conv1_2"][1].data.squeeze())
conv2_1_b = theano.shared(model_caffe.params["conv2_1"][1].data.squeeze())
conv2_2_b = theano.shared(model_caffe.params["conv2_2"][1].data.squeeze())
conv3_1_b = theano.shared(model_caffe.params["conv3_1"][1].data.squeeze())
conv3_2_b = theano.shared(model_caffe.params["conv3_2"][1].data.squeeze())
conv3_3_b = theano.shared(model_caffe.params["conv3_3"][1].data.squeeze())
conv3_4_b = theano.shared(model_caffe.params["conv3_4"][1].data.squeeze())
conv4_1_b = theano.shared(model_caffe.params["conv4_1"][1].data.squeeze())
conv4_2_b = theano.shared(model_caffe.params["conv4_2"][1].data.squeeze())
conv4_3_b = theano.shared(model_caffe.params["conv4_3"][1].data.squeeze())
conv4_4_b = theano.shared(model_caffe.params["conv4_4"][1].data.squeeze())
conv5_1_b = theano.shared(model_caffe.params["conv5_1"][1].data.squeeze())
conv5_2_b = theano.shared(model_caffe.params["conv5_2"][1].data.squeeze())
conv5_3_b = theano.shared(model_caffe.params["conv5_3"][1].data.squeeze())
conv5_4_b = theano.shared(model_caffe.params["conv5_4"][1].data.squeeze())
fc6_W = theano.shared(model_caffe.params["fc6"][0].data.squeeze())
fc7_W = theano.shared(model_caffe.params["fc7"][0].data.squeeze())
fc8_W = theano.shared(model_caffe.params["fc8"][0].data.squeeze())
fc6_b = theano.shared(model_caffe.params["fc6"][1].data.squeeze())
fc7_b = theano.shared(model_caffe.params["fc7"][1].data.squeeze())
fc8_b = theano.shared(model_caffe.params["fc8"][1].data.squeeze())
# make theano model
model_theano = collections.OrderedDict()
model_theano["data"] = T.tensor4()
model_theano["conv1_1"] = layers.convolution_layer(model_theano["data"], conv1_1_W, conv1_1_b, border="same")
model_theano["relu1_1"] = layers.relu_layer(model_theano["conv1_1"])
model_theano["conv1_2"] = layers.convolution_layer(model_theano["relu1_1"], conv1_2_W, conv1_2_b, border="same")
model_theano["relu1_2"] = layers.relu_layer(model_theano["conv1_2"])
model_theano["pool1"] = layers.pooling_layer(model_theano["relu1_2"], size=(2, 2), stride=(2, 2))
model_theano["conv2_1"] = layers.convolution_layer(model_theano["pool1"], conv2_1_W, conv2_1_b, border="same")
model_theano["relu2_1"] = layers.relu_layer(model_theano["conv2_1"])
model_theano["conv2_2"] = layers.convolution_layer(model_theano["relu2_1"], conv2_2_W, conv2_2_b, border="same")
model_theano["relu2_2"] = layers.relu_layer(model_theano["conv2_2"])
model_theano["pool2"] = layers.pooling_layer(model_theano["relu2_2"], size=(2, 2), stride=(2, 2))
model_theano["conv3_1"] = layers.convolution_layer(model_theano["pool2"], conv3_1_W, conv3_1_b, border="same")
model_theano["relu3_1"] = layers.relu_layer(model_theano["conv3_1"])
model_theano["conv3_2"] = layers.convolution_layer(model_theano["relu3_1"], conv3_2_W, conv3_2_b, border="same")
model_theano["relu3_2"] = layers.relu_layer(model_theano["conv3_2"])
model_theano["conv3_3"] = layers.convolution_layer(model_theano["relu3_2"], conv3_3_W, conv3_3_b, border="same")
model_theano["relu3_3"] = layers.relu_layer(model_theano["conv3_3"])
model_theano["conv3_4"] = layers.convolution_layer(model_theano["relu3_3"], conv3_4_W, conv3_4_b, border="same")
model_theano["relu3_4"] = layers.relu_layer(model_theano["conv3_4"])
model_theano["pool3"] = layers.pooling_layer(model_theano["relu3_4"], size=(2, 2), stride=(2, 2))
model_theano["conv4_1"] = layers.convolution_layer(model_theano["pool3"], conv4_1_W, conv4_1_b, border="same")
model_theano["relu4_1"] = layers.relu_layer(model_theano["conv4_1"])
model_theano["conv4_2"] = layers.convolution_layer(model_theano["relu4_1"], conv4_2_W, conv4_2_b, border="same")
model_theano["relu4_2"] = layers.relu_layer(model_theano["conv4_2"])
model_theano["conv4_3"] = layers.convolution_layer(model_theano["relu4_2"], conv4_3_W, conv4_3_b, border="same")
model_theano["relu4_3"] = layers.relu_layer(model_theano["conv4_3"])
model_theano["conv4_4"] = layers.convolution_layer(model_theano["relu4_3"], conv4_4_W, conv4_4_b, border="same")
model_theano["relu4_4"] = layers.relu_layer(model_theano["conv4_4"])
model_theano["pool4"] = layers.pooling_layer(model_theano["relu4_4"], size=(2, 2), stride=(2, 2))
model_theano["conv5_1"] = layers.convolution_layer(model_theano["pool4"], conv5_1_W, conv5_1_b, border="same")
model_theano["relu5_1"] = layers.relu_layer(model_theano["conv5_1"])
model_theano["conv5_2"] = layers.convolution_layer(model_theano["relu5_1"], conv5_2_W, conv5_2_b, border="same")
model_theano["relu5_2"] = layers.relu_layer(model_theano["conv5_2"])
model_theano["conv5_3"] = layers.convolution_layer(model_theano["relu5_2"], conv5_3_W, conv5_3_b, border="same")
model_theano["relu5_3"] = layers.relu_layer(model_theano["conv5_3"])
model_theano["conv5_4"] = layers.convolution_layer(model_theano["relu5_3"], conv5_4_W, conv5_4_b, border="same")
model_theano["relu5_4"] = layers.relu_layer(model_theano["conv5_4"])
model_theano["pool5"] = layers.pooling_layer(model_theano["relu5_4"], size=(2, 2), stride=(2, 2))
model_theano["fc6"] = layers.inner_product_layer(model_theano["pool5"], fc6_W, fc6_b)
model_theano["relu6"] = layers.relu_layer(model_theano["fc6"])
model_theano["fc7"] = layers.inner_product_layer(model_theano["relu6"], fc7_W, fc7_b)
model_theano["relu7"] = layers.relu_layer(model_theano["fc7"])
model_theano["fc8"] = layers.inner_product_layer(model_theano["relu7"], fc8_W, fc8_b)
model_theano["prob"] = layers.softmax_layer(model_theano["fc8"])
# check
data = np.random.randn(*model_caffe.blobs["data"].data.shape).astype(np.float32) * 10
model_caffe.blobs["data"].data[:] = data
model_caffe.forward()
theano_output = theano.function([model_theano["data"]], model_theano["prob"])(data)
error = (theano_output.squeeze() - model_caffe.blobs["prob"].data.squeeze()).max()
assert error < 1e-6
# save
print "saving"
if not os.path.exists(directory_theano):
os.mkdir(directory_theano)
sys.setrecursionlimit(100000)
pickle.dump(model_theano, open(filename_theanomodel, "wb"), protocol=pickle.HIGHEST_PROTOCOL)
print "done"
def build_encoder(x):
# The encoder uses the deep residual network.
outputs = []
pooling = [1, 2, 2, 1]
shape = x.get_shape().as_list()
bs = shape[0]
seq = shape[1]
temp_shape = [bs * seq] + shape[2:]
x = tf.reshape(x, temp_shape)
# print x.get_shape().as_list()
# layer 0
with tf.variable_scope("encoder_layer0", reuse=tf.AUTO_REUSE):
conv0_0 = layers.conv_layer(name="conv0_0",
x=x,
filter_shape=layers.create_variable(
"filter0_0", shape=[7, 7, 3, 96]))
conv0_0 = layers.batch_normalization(conv0_0, "conv0_0_bn")
conv0_0 = layers.relu_layer(conv0_0)
conv0_1 = layers.conv_layer(name="conv0_1",
x=conv0_0,
filter_shape=layers.create_variable(
"filter0_1", shape=[3, 3, 96, 96]))
conv0_1 = layers.batch_normalization(conv0_1, "conv0_1_bn")
conv0_1 = layers.relu_layer(conv0_1)
shortcut0 = layers.conv_layer(name="shortcut",
x=x,
filter_shape=layers.create_variable(
"filter0_2", shape=[1, 1, 3, 96]))
shortcut0 = layers.batch_normalization(shortcut0, "shortcut0_bn")
shortcut0 = layers.relu_layer(shortcut0)
layer0 = layers.pooling_layer("pooling", conv0_1 + shortcut0, pooling)
outputs.append(layer0) # [bs * size, 64, 64, 96]
# layer 1
with tf.variable_scope("encoder_layer1", reuse=tf.AUTO_REUSE):
conv1_0 = layers.conv_layer(name="conv1_0",
x=layer0,
filter_shape=layers.create_variable(
"filter1_0", shape=[3, 3, 96, 128]))
conv1_0 = layers.batch_normalization(conv1_0, "conv1_0_bn")
conv1_0 = layers.relu_layer(conv1_0)
conv1_1 = layers.conv_layer(name="conv1_1",
x=conv1_0,
filter_shape=layers.create_variable(
"filter1_1", shape=[3, 3, 128, 128]))
conv1_1 = layers.batch_normalization(conv1_1, "conv1_1_bn")
conv1_1 = layers.relu_layer(conv1_1)
shortcut1 = layers.conv_layer(name="shortcut",
x=layer0,
filter_shape=layers.create_variable(
"filter1_2", shape=[1, 1, 96, 128]))
shortcut1 = layers.batch_normalization(shortcut1, "shortcut1_bn")
shortcut1 = layers.relu_layer(shortcut1)
layer1 = layers.pooling_layer("pooling", conv1_1 + shortcut1, pooling)
outputs.append(layer1) # [bs * size, 32, 32, 128]
# layer 2
with tf.variable_scope("encoder_layer2", reuse=tf.AUTO_REUSE):
conv2_0 = layers.conv_layer(name="conv2_0",
x=layer1,
filter_shape=layers.create_variable(
"filter2_0", shape=[3, 3, 128, 256]))
conv2_0 = layers.batch_normalization(conv2_0, "conv2_0_bn")
conv2_0 = layers.relu_layer(conv2_0)
conv2_1 = layers.conv_layer(name="conv2_1",
x=conv2_0,
filter_shape=layers.create_variable(
"filter2_1", shape=[3, 3, 256, 256]))
conv2_1 = layers.batch_normalization(conv2_1, "conv2_1_bn")
conv2_1 = layers.relu_layer(conv2_1)
shortcut2 = layers.conv_layer(name="shortcut",
x=layer1,
filter_shape=layers.create_variable(
"filter2_2", shape=[1, 1, 128, 256]))
shortcut2 = layers.batch_normalization(shortcut2, "shortcut2_bn")
shortcut2 = layers.relu_layer(shortcut2)
layer2 = layers.pooling_layer("pooling", conv2_1 + shortcut2, pooling)
outputs.append(layer2) # [bs * size, 16, 16, 256]
# layer 3
with tf.variable_scope("encoder_layer3", reuse=tf.AUTO_REUSE):
conv3_0 = layers.conv_layer(name="conv3_0",
x=layer2,
filter_shape=layers.create_variable(
"filter3_0", shape=[3, 3, 256, 256]))
conv3_0 = layers.batch_normalization(conv3_0, "conv3_0_bn")
conv3_0 = layers.relu_layer(conv3_0)
conv3_1 = layers.conv_layer(name="conv3_1",
x=conv3_0,
filter_shape=layers.create_variable(
"filter3_1", shape=[3, 3, 256, 256]))
conv3_1 = layers.batch_normalization(conv3_1, "conv3_1_bn")
conv3_1 = layers.relu_layer(conv3_1)
layer3 = layers.pooling_layer("pooling", conv3_1, pooling)
outputs.append(layer3) # [bs * size, 8, 8, 256]
# layer 4
with tf.variable_scope("encoder_layer4", reuse=tf.AUTO_REUSE):
conv4_0 = layers.conv_layer(name="conv4_0",
x=layer3,
filter_shape=layers.create_variable(
"filter4_0", shape=[3, 3, 256, 256]))
conv4_0 = layers.batch_normalization(conv4_0, "conv4_0_bn")
conv4_0 = layers.relu_layer(conv4_0)
conv4_1 = layers.conv_layer(name="conv4_1",
x=conv4_0,
filter_shape=layers.create_variable(
"filter4_1", shape=[3, 3, 256, 256]))
conv4_1 = layers.batch_normalization(conv4_1, "conv4_1_bn")
conv4_1 = layers.relu_layer(conv4_1)
shortcut4 = layers.conv_layer(name="shortcut",
x=layer3,
filter_shape=layers.create_variable(
"filter4_2", shape=[1, 1, 256, 256]))
shortcut4 = layers.batch_normalization(shortcut4, "shortcut4_bn")
shortcut4 = layers.relu_layer(shortcut4)
layer4 = layers.pooling_layer("pooling", conv4_1 + shortcut4, pooling)
outputs.append(layer4) # [bs * size, 4, 4, 256]
# layer 5
with tf.variable_scope("encoder_layer5", reuse=tf.AUTO_REUSE):
conv5_0 = layers.conv_layer(name="conv5_0",
x=layer4,
filter_shape=layers.create_variable(
"filter5_0", shape=[3, 3, 256, 256]))
conv5_0 = layers.batch_normalization(conv5_0, "conv5_0_bn")
conv5_0 = layers.relu_layer(conv5_0)
conv5_1 = layers.conv_layer(name="conv5_1",
x=conv5_0,
filter_shape=layers.create_variable(
"filter5_1", shape=[3, 3, 256, 256]))
conv5_1 = layers.batch_normalization(conv5_1, "conv5_1_bn")
conv5_1 = layers.relu_layer(conv5_1)
shortcut5 = layers.conv_layer(name="shortcut",
x=layer4,
filter_shape=layers.create_variable(
"filter5_2", shape=[1, 1, 256, 256]))
shortcut5 = layers.batch_normalization(shortcut5, "shortcut5_bn")
shortcut5 = layers.relu_layer(shortcut5)
layer5 = layers.pooling_layer("pooling", conv5_1 + shortcut5, pooling)
outputs.append(layer5) # [bs * size, 2, 2, 256]
final_shape = [bs, seq, fc_layer_size[0]]
# Flatten layer and fully connected layer
flatten = layers.flatten_layer(layer5)
outputs.append(flatten)
with tf.variable_scope("fc_layer", reuse=tf.AUTO_REUSE):
layer_fc = layers.fully_connected_layer(flatten, fc_layer_size[0],
"fclayer_w", "fclayer_b")
# layer_fc = layers.batch_normalization(layer_fc, "fc_bn")
layer_fc = layers.relu_layer(layer_fc)
outputs.append(layer_fc) # [bs * size, 1024]
# [bs, size, 1024]
return tf.reshape(outputs[-1], final_shape)
