def testShapeInferenceTranspose(self):
model = model_helper.ModelHelper(name="test_model")
workspace.FeedBlob(
"tensor",
np.random.rand(4, 2, 3, 3, 5).astype(np.float32)
)
# Testing with axes undefined
brew.transpose(
model,
["tensor"],
"transpose",
)
self.InferTensorRunAndCompare(model)
# Testing with axes defined
brew.transpose(
model,
["tensor"],
"transpose",
axes=np.random.permutation(5)
)
return self.InferTensorRunAndCompare(model)
def testShapeInferenceTranspose(self):
model = model_helper.ModelHelper(name="test_model")
workspace.FeedBlob(
"tensor",
np.random.rand(4, 2, 3, 3, 5).astype(np.float32)
)
# Testing with axes undefined
brew.transpose(
model,
["tensor"],
"transpose",
)
self.InferTensorRunAndCompare(model)
# Testing with axes defined
brew.transpose(
model,
["tensor"],
"transpose",
axes=np.random.permutation(5)
)
return self.InferTensorRunAndCompare(model)
def _calc_attention_logits_from_sum_match(
model,
decoder_hidden_encoder_outputs_sum,
encoder_output_dim,
scope,
):
# [encoder_length, batch_size, encoder_output_dim]
decoder_hidden_encoder_outputs_sum = model.net.Tanh(
decoder_hidden_encoder_outputs_sum,
decoder_hidden_encoder_outputs_sum,
)
# [encoder_length, batch_size, 1]
attention_logits = brew.fc(
model,
decoder_hidden_encoder_outputs_sum,
s(scope, 'attention_logits'),
dim_in=encoder_output_dim,
dim_out=1,
axis=2,
freeze_bias=True,
)
# [batch_size, encoder_length, 1]
attention_logits_transposed = brew.transpose(
model,
attention_logits,
s(scope, 'attention_logits_transposed'),
axes=[1, 0, 2],
)
return attention_logits_transposed
def _calc_attention_logits_from_sum_match(
model,
decoder_hidden_encoder_outputs_sum,
encoder_output_dim,
scope,
):
# [encoder_length, batch_size, encoder_output_dim]
decoder_hidden_encoder_outputs_sum = model.net.Tanh(
decoder_hidden_encoder_outputs_sum,
decoder_hidden_encoder_outputs_sum,
)
# [encoder_length, batch_size, 1]
attention_logits = brew.fc(
model,
decoder_hidden_encoder_outputs_sum,
s(scope, 'attention_logits'),
dim_in=encoder_output_dim,
dim_out=1,
axis=2,
freeze_bias=True,
)
# [batch_size, encoder_length, 1]
attention_logits_transposed = brew.transpose(
model,
attention_logits,
s(scope, 'attention_logits_transposed'),
axes=[1, 0, 2],
)
return attention_logits_transposed
def prepare_input(self, model, input_blob):
if self.encoder_outputs_transposed is None:
self.encoder_outputs_transposed = brew.transpose(
model,
self.encoder_outputs,
self.scope('encoder_outputs_transposed'),
axes=[1, 2, 0],
)
if self.weighted_encoder_outputs is None:
self.weighted_encoder_outputs = brew.fc(
model,
self.encoder_outputs,
self.scope('weighted_encoder_outputs'),
dim_in=self.encoder_output_dim,
dim_out=self.encoder_output_dim,
axis=2,
)
return self.decoder_cell.prepare_input(model, input_blob)
def prepare_input(self, model, input_blob):
if self.encoder_outputs_transposed is None:
self.encoder_outputs_transposed = brew.transpose(
model,
self.encoder_outputs,
self.scope('encoder_outputs_transposed'),
axes=[1, 2, 0],
)
if self.weighted_encoder_outputs is None:
self.weighted_encoder_outputs = brew.fc(
model,
self.encoder_outputs,
self.scope('weighted_encoder_outputs'),
dim_in=self.encoder_output_dim,
dim_out=self.encoder_output_dim,
axis=2,
)
return self.decoder_cell.prepare_input(model, input_blob)
def _calc_attention_logits_from_sum_match(
model,
decoder_hidden_encoder_outputs_sum,
encoder_output_dim,
scope,
):
# [encoder_length, batch_size, encoder_output_dim]
decoder_hidden_encoder_outputs_sum = model.net.Tanh(
decoder_hidden_encoder_outputs_sum,
decoder_hidden_encoder_outputs_sum,
)
attention_v = model.param_init_net.XavierFill(
[],
s(scope, 'attention_v'),
shape=[1, encoder_output_dim],
)
model.params.append(attention_v)
attention_zeros = model.param_init_net.ConstantFill(
[],
s(scope, 'attention_zeros'),
value=0.0,
shape=[1],
)
# [encoder_length, batch_size, 1]
attention_logits = model.net.FC(
[decoder_hidden_encoder_outputs_sum, attention_v, attention_zeros],
[s(scope, 'attention_logits')],
axis=2,
)
# [batch_size, encoder_length, 1]
attention_logits_transposed = brew.transpose(
model,
attention_logits,
s(scope, 'attention_logits_transposed'),
axes=[1, 0, 2],
)
return attention_logits_transposed
def Transpose(self, *args, **kwargs):
return brew.transpose(self, *args, use_cudnn=self.use_cudnn, **kwargs)
def build_crf_net(self, input_blob, initial_state, transitions):
"""
Adds the crf_net recurrent operator to the model.
model: model_helper.ModelHelper object new operators would be added
to
input_blob: the input sequence in a format T x N x D
where T is sequence size, N - batch size and D - input dimension
##Only supports batch-size 1##
seq_lengths: blob containing sequence lengths (unused)
"""
scope = "crf_net"
def s(name):
""
# We have to manually scope due to our internal/external blob
# relationships.
return "{}/{}".format(str(scope), str(name))
step_model = model_helper.ModelHelper(name="crf_step", param_model=self.model)
input_t, cell_t_prev, _ = step_model.net.AddExternalInputs(
core.ScopedBlobReference("input_t"),
core.ScopedBlobReference("cell_t_prev"),
transitions,
)
zero_segment_id = step_model.param_init_net.ConstantFill(
[],
[s("zero_segment_id")],
value=0,
shape=[self.num_classes_padded],
dtype=core.DataType.INT32,
)
# A hack to bypass model cloning for test
step_model.param_init_net.AddExternalOutput(zero_segment_id)
""" the CRF step """
# Do tile
prev_transpose = brew.transpose(
step_model, cell_t_prev, [s("prev_transpose")], axes=(0, 2, 1)
)
prev_tiled = step_model.net.Tile(
prev_transpose, [s("prev_tiled")], tiles=self.num_classes_padded, axis=2
)
input_t_tiled = step_model.net.Tile(
input_t, [s("input_t_tiled")], tiles=self.num_classes_padded, axis=1
)
input_with_prev = step_model.net.Add(
[prev_tiled, input_t_tiled], [s("input_with_prev")]
)
all_with_transitions = step_model.net.Add(
[input_with_prev, transitions],
[s("prev_with_transitions")],
broadcast=1,
use_grad_hack=1,
)
all_with_transitions_reshaped, _ = step_model.net.Reshape(
all_with_transitions,
[s("all_with_transitions_reshaped"), s("all_with_transitions_orig")],
shape=(self.num_classes_padded, self.num_classes_padded),
)
cell_t = step_model.net.SortedSegmentRangeLogSumExp(
[all_with_transitions_reshaped, zero_segment_id], [s("cell_t")]
)
step_model.net.AddExternalOutputs(cell_t)
""" recurrent network """
cell_input_blob = initial_state
out_all, out_last = recurrent.recurrent_net(
net=self.model.net,
cell_net=step_model.net,
inputs=[(input_t, input_blob)],
initial_cell_inputs=[(cell_t_prev, cell_input_blob)],
links={cell_t_prev: cell_t},
scope=scope,
outputs_with_grads=(1,),
)
return out_last
def apply_soft_coverage_attention(
model,
encoder_output_dim,
encoder_outputs_transposed,
weighted_encoder_outputs,
decoder_hidden_state_t,
decoder_hidden_state_dim,
scope,
encoder_lengths,
coverage_t_prev,
coverage_weights,
):
weighted_decoder_hidden_state = _apply_fc_weight_for_sum_match(
model=model,
input=decoder_hidden_state_t,
dim_in=decoder_hidden_state_dim,
dim_out=encoder_output_dim,
scope=scope,
name='weighted_decoder_hidden_state',
)
# [encoder_length, batch_size, encoder_output_dim]
decoder_hidden_encoder_outputs_sum_tmp = model.net.Add(
[weighted_encoder_outputs, weighted_decoder_hidden_state],
s(scope, 'decoder_hidden_encoder_outputs_sum_tmp'),
broadcast=1,
)
# [batch_size, encoder_length]
coverage_t_prev_2d = model.net.Squeeze(
coverage_t_prev,
s(scope, 'coverage_t_prev_2d'),
dims=[0],
)
# [encoder_length, batch_size]
coverage_t_prev_transposed = brew.transpose(
model,
coverage_t_prev_2d,
s(scope, 'coverage_t_prev_transposed'),
)
# [encoder_length, batch_size, encoder_output_dim]
scaled_coverage_weights = model.net.Mul(
[coverage_weights, coverage_t_prev_transposed],
s(scope, 'scaled_coverage_weights'),
broadcast=1,
axis=0,
)
# [encoder_length, batch_size, encoder_output_dim]
decoder_hidden_encoder_outputs_sum = model.net.Add(
[decoder_hidden_encoder_outputs_sum_tmp, scaled_coverage_weights],
s(scope, 'decoder_hidden_encoder_outputs_sum'),
)
# [batch_size, encoder_length, 1]
attention_logits_transposed = _calc_attention_logits_from_sum_match(
model=model,
decoder_hidden_encoder_outputs_sum=decoder_hidden_encoder_outputs_sum,
encoder_output_dim=encoder_output_dim,
scope=scope,
)
# [batch_size, encoder_length, 1]
attention_weights_3d = _calc_attention_weights(
model=model,
attention_logits_transposed=attention_logits_transposed,
scope=scope,
encoder_lengths=encoder_lengths,
)
# [batch_size, encoder_output_dim, 1]
attention_weighted_encoder_context = _calc_weighted_context(
model=model,
encoder_outputs_transposed=encoder_outputs_transposed,
encoder_output_dim=encoder_output_dim,
attention_weights_3d=attention_weights_3d,
scope=scope,
)
# [batch_size, encoder_length]
attention_weights_2d = model.net.Squeeze(
attention_weights_3d,
s(scope, 'attention_weights_2d'),
dims=[2],
)
coverage_t = model.net.Add(
[coverage_t_prev, attention_weights_2d],
s(scope, 'coverage_t'),
broadcast=1,
)
return (
attention_weighted_encoder_context,
attention_weights_3d,
[decoder_hidden_encoder_outputs_sum],
coverage_t,
)
def Transpose(self, *args, **kwargs):
return brew.transpose(self, *args, use_cudnn=self.use_cudnn, **kwargs)
def create_caffe2_model(model,
input_shape,
use_cudnn=True,
init_params=False,
keras_channel_last=True):
arg_scope = {'order': 'NCHW', 'use_cudnn': use_cudnn}
caffe2_model = model_helper.ModelHelper(name='model',
init_params=init_params,
arg_scope=arg_scope)
num_conv_layers = 0
layer_num = 0
layer_sizes = {}
prev_layer_name = ''
for layer in model.layers:
inb_node = layer._inbound_nodes[0]
num_input_layers = len(inb_node.inbound_layers)
input_name_list = []
for ii in range(0, num_input_layers):
inp_layer = inb_node.inbound_layers[ii]
input_name_list.append(inp_layer.name)
prev_layer_name = inp_layer.name
if isinstance(inp_layer, keras.layers.Flatten):
pass
#pinb_node = inp_layer._inbound_nodes[0]
#prev_layer_name = pinb_node.inbound_layers[0].name
name = layer.name
config = layer.get_config()
inputShape = layer.input_shape
outputShape = layer.output_shape
if isinstance(layer, keras.engine.input_layer.InputLayer):
input_sizes = (input_shape[2], input_shape[3])
layer_sizes[name] = input_sizes
else:
if (input_name_list[0] not in layer_sizes):
raise ValueError("Can't find layer size for ",
input_name_list[0])
else:
input_sizes = layer_sizes[input_name_list[0]]
layer_dim = len(outputShape)
if (layer_dim == 4):
if (keras_channel_last):
out_sizes = (outputShape[1], outputShape[2])
else:
out_sizes = (outputShape[2], outputShape[3])
elif (layer_dim == 2):
out_sizes = (0, 0) #flattened
else:
raise ValueError(
'Unsupported layer dimension : {0}'.format(layer_dim))
if isinstance(layer, keras.layers.Flatten):
tmp_prev = prev_layer_name
if (keras_channel_last):
tmp_prev = prev_layer_name + '_transpose' #nb, img_h, img_w, chan <-- nb, chan, img_h, img_w
c2_layer = brew.transpose(caffe2_model,
prev_layer_name,
tmp_prev,
axes=(0, 2, 3, 1))
c2_layer = caffe2_model.net.Flatten(tmp_prev, name)
#print('FLatten previous layer ', prev_layer_name, ' current layer ', name , 'inputshape ', inputShape)
layer_sizes[name] = out_sizes
elif isinstance(layer, keras.layers.Dropout):
#print('name is ', name, ' prev_layer_name ', prev_layer_name)
c2_layer = caffe2_model.net.Dropout(prev_layer_name,
name,
is_test=True
#ratio=config['rate']
)
#same size
layer_sizes[name] = input_sizes
elif (isinstance(layer, keras.layers.convolutional.Conv2D)):
dim_in = inputShape[-1]
dim_out = outputShape[-1]
kernel = config['kernel_size'][0]
stride = config['strides'][0]
if (config['padding'] == 'same'):
pad_sizes = get_padding_sizes(input_sizes,
config['kernel_size'],
config['strides'])
elif (config['padding'] == 'valid'):
pad_sizes = ((0, 0), (0, 0))
else:
raise ValueError('unsupported padding')
#print('pad sizes ', pad_sizes)
layer_sizes[name] = out_sizes
c2_layer = brew.conv(caffe2_model,
prev_layer_name,
name,
dim_in=dim_in,
dim_out=dim_out,
kernel=kernel,
stride=stride,
pad_l=pad_sizes[0][0],
pad_r=pad_sizes[0][1],
pad_t=pad_sizes[1][0],
pad_b=pad_sizes[1][1])
if config['activation'] == 'linear':
pass
elif config['activation'] == 'relu':
c2_layer = brew.relu(caffe2_model, name, name)
elif config['activation'] == 'softmax':
#c2_layer = brew.softmax(caffe2_model, name, name)
c2_layer = brew.softmax(caffe2_model, name, 'softmax')
else:
raise ValueError(
'The only supported activation for conv layer is relu')
elif isinstance(layer, keras.layers.MaxPooling2D):
kernel = config['pool_size'][0]
stride = config['strides'][0]
pad_size = ((0, 0), (0, 0))
layer_sizes[name] = out_sizes
c2_layer = brew.max_pool(caffe2_model,
prev_layer_name,
name,
kernel=kernel,
stride=stride)
elif isinstance(layer, keras.layers.AveragePooling2D):
kernel = config['pool_size'][0]
stride = config['strides'][0]
pad_size = ((0, 0), (0, 0))
layer_sizes[name] = out_sizes
c2_layer = brew.average_pool(caffe2_model,
prev_layer_name,
name,
kernel=kernel,
stride=stride)
elif isinstance(layer, keras.layers.BatchNormalization):
dim_in = inputShape[-1]
epsilon = config['epsilon']
momentum = config['momentum']
c2_layer = brew.spatial_bn(caffe2_model,
prev_layer_name,
name,
dim_in=dim_in,
epsilon=epsilon,
momentum=momentum,
is_test=True)
#same size
layer_sizes[name] = input_sizes
elif (isinstance(layer, keras.layers.core.Dense)):
dim_in = inputShape[-1]
dim_out = outputShape[-1]
#print('input shape for dense is ', inputShape)
if (len(inputShape) == 2): #flattened input
c2_layer = brew.fc(caffe2_model,
prev_layer_name,
name,
dim_in=dim_in,
dim_out=dim_out)
else: #fully convolutional input
c2_layer = brew.conv(caffe2_model,
prev_layer_name,
name,
dim_in=dim_in,
dim_out=dim_out,
kernel=1,
stride=1)
activation = config['activation']
if activation == 'relu':
c2_layer = brew.relu(caffe2_model, name, name)
elif activation == 'softmax':
c2_layer = brew.softmax(caffe2_model, name, 'softmax')
elif activation == 'linear':
pass #
else:
raise ValueError(
'The only supported activations for fc layer are relu and softmax'
)
#same size
layer_sizes[name] = input_sizes
elif (isinstance(layer, keras.layers.advanced_activations.LeakyReLU)):
dim_in = inputShape[-1]
c2_layer = caffe2_model.net.LeakyRelu(prev_layer_name,
name,
alpha=config['alpha'])
#same size
layer_sizes[name] = input_sizes
elif (isinstance(layer, keras.layers.merge.Add)):
c2_layer = brew.sum(caffe2_model,
[input_name_list[0], input_name_list[1]], name)
#same size
layer_sizes[name] = input_sizes
layer_num = layer_num + 1
if (layer_num == len(model.layers)):
caffe2_model.net.AddExternalOutput(c2_layer)
return caffe2_model
def build_crf_net(self, input_blob, initial_state, transitions):
'''
Adds the crf_net recurrent operator to the model.
model: model_helper.ModelHelper object new operators would be added
to
input_blob: the input sequence in a format T x N x D
where T is sequence size, N - batch size and D - input dimention
##Only supports batch-size 1##
seq_lengths: blob containing sequence lengths (unused)
'''
scope = 'crf_net'
def s(name):
''
# We have to manually scope due to our internal/external blob
# relationships.
return "{}/{}".format(str(scope), str(name))
step_model = model_helper.ModelHelper(name='crf_step',
param_model=self.model)
input_t, cell_t_prev, _ = (
step_model.net.AddExternalInputs(
core.ScopedBlobReference('input_t'),
core.ScopedBlobReference('cell_t_prev'),
transitions
)
)
zero_segment_id = step_model.param_init_net.ConstantFill(
[],
[s('zero_segment_id')],
value=0,
shape=[self.num_classes_padded],
dtype=core.DataType.INT32,
)
# A hack to bypass model cloning for test
step_model.param_init_net.AddExternalOutput(zero_segment_id)
""" the CRF step """
# Do tile
prev_transpose = brew.transpose(
step_model,
cell_t_prev,
[s('prev_transpose')],
axes=(0, 2, 1),
)
prev_tiled = step_model.net.Tile(
prev_transpose,
[s('prev_tiled')],
tiles=self.num_classes_padded,
axis=2,
)
input_t_tiled = step_model.net.Tile(
input_t,
[s('input_t_tiled')],
tiles=self.num_classes_padded,
axis=1,
)
input_with_prev = step_model.net.Add(
[prev_tiled, input_t_tiled],
[s('input_with_prev')]
)
all_with_transitions = step_model.net.Add(
[input_with_prev, transitions],
[s('prev_with_transitions')],
broadcast=1,
use_grad_hack=1,
)
all_with_transitions_reshaped, _ = step_model.net.Reshape(
all_with_transitions,
[s('all_with_transitions_reshaped'), s('all_with_transitions_orig')],
shape=(self.num_classes_padded, self.num_classes_padded)
)
cell_t = step_model.net.SortedSegmentRangeLogSumExp(
[all_with_transitions_reshaped, zero_segment_id],
[s('cell_t')],
)
step_model.net.AddExternalOutputs(cell_t)
""" recurrent network """
cell_input_blob = initial_state
out_all, out_last = recurrent.recurrent_net(
net=self.model.net,
cell_net=step_model.net,
inputs=[(input_t, input_blob)],
initial_cell_inputs=[
(cell_t_prev, cell_input_blob),
],
links={
cell_t_prev: cell_t,
},
scope=scope,
outputs_with_grads=(1,)
)
return out_last
