def transform_prediction_net(model_path_prefix, quantization):
tensorflow_model_path = model_path_prefix + "/pred_net.pb"
caffe_model_path_prefix = "asr_convolution_transformer_prediction_net"
caffe_model_name = "asr_convolution_transformer_prediction_net"
asr_caffe = Tensorflow2CaffeConvolutionTransformer(
tensorflow_model_path,
caffe_model_path_prefix,
caffe_model_name,
check=False,
calc=True,
quantization=quantization)
data = {}
data["label"] = np.array([[1]])
asr_caffe.generate_prediction_net(data)
def transform_joint_net(model_path_prefix, data_path_prefix, quantization):
tensorflow_model_path = model_path_prefix + "/joint_net.pb"
caffe_model_path_prefix = "asr_convolution_transformer_joint_net"
caffe_model_name = "asr_convolution_transformer_joint_net"
asr_caffe = Tensorflow2CaffeConvolutionTransformer(
tensorflow_model_path,
caffe_model_path_prefix,
caffe_model_name,
check=False,
calc=True,
quantization=quantization)
shapes = {}
shapes["encoder"] = [1, 512]
shapes["prediction_net"] = [1, 512]
data = {}
data["encoder"] = np.load(data_path_prefix + "/encoder.npy")
data["prediction_net"] = np.load(data_path_prefix + "/pred_net.npy")
asr_caffe.print_weight_map()
asr_caffe.generate_joint_net(shapes, data)
def transform_encoder(model_path_prefix, data_path_prefix, quantization,
block_id_start, block_id_end):
tensorflow_model_path = model_path_prefix + "/encoder.pb"
caffe_model_path_prefix = "asr_convolution_transformer_encoder"
caffe_model_name = "asr_convolution_transformer_encoder"
first_frame = True
asr_caffe = Tensorflow2CaffeConvolutionTransformer(
tensorflow_model_path,
caffe_model_path_prefix,
caffe_model_name,
nchwc8=True,
first_frame=first_frame,
check=False,
calc=True,
quantization=quantization)
data = {}
if (first_frame):
data["sounds"] = np.load(data_path_prefix + "/sound0_frame0.npy")
else:
data["sounds"] = np.load(data_path_prefix + "/sound0_frame1.npy")
asr_caffe.generate_encoder(data, block_id_start, block_id_end)
def transform_reference(model_path_prefix, data_path_prefix, quantization,
FFN_decomposition):
tensorflow_model_path = model_path_prefix + "/encoder.pb"
caffe_model_path_prefix = "asr_convolution_transformer_reference"
caffe_model_name = "asr_convolution_transformer_reference"
first_frame = True
asr_caffe = Tensorflow2CaffeConvolutionTransformer(
base_params,
tensorflow_model_path,
caffe_model_path_prefix,
caffe_model_name,
nchwc8=True,
first_frame=first_frame,
check=False,
calc=True,
quantization=quantization,
FFN_decomposition=FFN_decomposition)
data = {}
data["reference"] = np.array([[
42, 39, 42, 39, 41, 38, 14, 28, 6, 36, 25, 23, 32, 15, 8, 36, 33, 39,
37, 19, 33, 8, 27, 27, 23, 5
]])
out = asr_caffe.generate_reference(data)
np.save("reference.npy", asr_caffe.get_tensor(out))
def transform_encoder(model_path_prefix, data_path_prefix, quantization,
FFN_decomposition):
tensorflow_model_path = model_path_prefix + "/encoder.pb"
caffe_model_path_prefix = "asr_convolution_transformer_encoder"
caffe_model_name = "asr_convolution_transformer_encoder"
first_frame = True
asr_caffe = Tensorflow2CaffeConvolutionTransformer(
base_params,
tensorflow_model_path,
caffe_model_path_prefix,
caffe_model_name,
nchwc8=True,
first_frame=first_frame,
check=False,
calc=True,
quantization=quantization,
FFN_decomposition=FFN_decomposition)
data = {}
data["reference"] = np.load("reference.npy")
if (first_frame):
data["sounds"] = np.load(data_path_prefix + "/frame0.npy")
else:
data["sounds"] = np.load(data_path_prefix + "/frame1.npy")
asr_caffe.generate_encoder(data)