added weight-exchange library

11 files changed, 827 insertions, 0 deletions

diff --git a/dnn/torch/weight-exchange/README.md b/dnn/torch/weight-exchange/README.md
new file mode 100644
index 00000000..f4818b5b
--- /dev/null
+++ b/dnn/torch/weight-exchange/README.md
@@ -0,0 +1,21 @@
+# weight-exchange
+
+
+
+## Weight Exchange
+Repo wor exchanging weights betweeen torch an tensorflow.keras modules, using an intermediate numpy format.
+
+Routines for loading/dumping torch weights are located in exchange/torch and can be loaded with
+```
+import exchange.torch
+```
+and routines for loading/dumping tensorflow weights are located in exchange/tf and can be loaded with
+```
+import exchange.tf
+```
+
+Note that `exchange.torch` requires torch to be installed and `exchange.tf` requires tensorflow. To avoid the necessity of installing both torch and tensorflow in the working environment, none of these submodules is imported when calling `import exchange`. Similarly, the requirements listed in `requirements.txt` do include neither Tensorflow or Pytorch.
+
+
+## C export
+The module `exchange.c_export` contains routines to export weights to C files. On the long run it will be possible to call all `dump_...` functions with either a path string or a `CWriter` instance based on which the export format is chosen. This is currently only implemented for `torch.nn.GRU`, `torch.nn.Linear` and `torch.nn.Conv1d`.
\ No newline at end of file
diff --git a/dnn/torch/weight-exchange/requirements.txt b/dnn/torch/weight-exchange/requirements.txt
new file mode 100644
index 00000000..296d6545
--- /dev/null
+++ b/dnn/torch/weight-exchange/requirements.txt
@@ -0,0 +1 @@
+numpy
\ No newline at end of file
diff --git a/dnn/torch/weight-exchange/setup.py b/dnn/torch/weight-exchange/setup.py
new file mode 100644
index 00000000..bf298627
--- /dev/null
+++ b/dnn/torch/weight-exchange/setup.py
@@ -0,0 +1,19 @@
+#!/usr/bin/env/python
+import os
+from setuptools import setup
+
+lib_folder = os.path.dirname(os.path.realpath(__file__))
+
+with open(os.path.join(lib_folder, 'requirements.txt'), 'r') as f:
+    install_requires = list(f.read().splitlines())
+
+print(install_requires)
+
+setup(name='wexchange',
+      version='1.4',
+      author='Jan Buethe',
+      author_email='jbuethe@amazon.de',
+      description='Weight-exchange library between Pytorch and Tensorflow',
+      packages=['wexchange', 'wexchange.tf', 'wexchange.torch', 'wexchange.c_export'],
+      install_requires=install_requires
+      )
diff --git a/dnn/torch/weight-exchange/wexchange/__init__.py b/dnn/torch/weight-exchange/wexchange/__init__.py
new file mode 100644
index 00000000..330d17e6
--- /dev/null
+++ b/dnn/torch/weight-exchange/wexchange/__init__.py
@@ -0,0 +1 @@
+from . import c_export
\ No newline at end of file
diff --git a/dnn/torch/weight-exchange/wexchange/c_export/__init__.py b/dnn/torch/weight-exchange/wexchange/c_export/__init__.py
new file mode 100644
index 00000000..6fb5ea79
--- /dev/null
+++ b/dnn/torch/weight-exchange/wexchange/c_export/__init__.py
@@ -0,0 +1,2 @@
+from .c_writer import CWriter
+from .common import print_gru_layer, print_dense_layer, print_conv1d_layer, print_vector
\ No newline at end of file
diff --git a/dnn/torch/weight-exchange/wexchange/c_export/c_writer.py b/dnn/torch/weight-exchange/wexchange/c_export/c_writer.py
new file mode 100644
index 00000000..624ab665
--- /dev/null
+++ b/dnn/torch/weight-exchange/wexchange/c_export/c_writer.py
@@ -0,0 +1,143 @@
+import os
+from collections import OrderedDict
+
+class CWriter:
+    def __init__(self,
+                 filename_without_extension,
+                 message=None,
+                 header_only=False,
+                 enable_binary_blob=False,
+                 create_state_struct=False,
+                 model_struct_name="Model",
+                 nnet_header="nnet.h"):
+        """
+        Writer class for creating souce and header files for weight exports to C
+
+        Parameters:
+        -----------
+
+        filename_without_extension: str
+            filename from which .c and .h files are created
+
+        message: str, optional
+            if given and not None, this message will be printed as comment in the header file
+
+        header_only: bool, optional
+            if True, only a header file is created; defaults to False
+
+        enable_binary_blob: bool, optional
+            if True, export is done in binary blob format and a model type is created; defaults to False
+
+        create_state_struct: bool, optional
+            if True, a state struct type is created in the header file; if False, state sizes are defined as macros; defaults to False
+
+        model_struct_name: str, optional
+            name used for the model struct type; only relevant when enable_binary_blob is True; defaults to "Model"
+
+        nnet_header: str, optional
+            name of header nnet header file; defaults to nnet.h
+
+        """
+
+
+        self.header_only = header_only
+        self.enable_binary_blob = enable_binary_blob
+        self.create_state_struct = create_state_struct
+        self.model_struct_name = model_struct_name
+
+        # for binary blob format, format is key=<layer name>, value=(<layer type>, <init call>)
+        self.layer_dict = OrderedDict()
+
+        # for binary blob format, format is key=<layer name>, value=<layer type>
+        self.weight_arrays = set()
+
+        # form model struct, format is key=<layer name>, value=<number of elements>
+        self.state_dict = OrderedDict()
+
+        self.header = open(filename_without_extension + ".h", "w")
+        header_name = os.path.basename(filename_without_extension) + '.h'
+
+        if message is not None:
+            self.header.write(f"/* {message} */\n\n")
+
+        self.header_guard = os.path.basename(filename_without_extension).upper() + "_H"
+        self.header.write(
+f'''
+#ifndef {self.header_guard}
+#define {self.header_guard}
+
+#include "{nnet_header}"
+
+'''
+        )
+
+        if not self.header_only:
+            self.source = open(filename_without_extension + ".c", "w")
+            if message is not None:
+                self.source.write(f"/* {message} */\n\n")
+
+            self.source.write(
+f"""
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+""")
+            self.source.write(f'#include "{header_name}"\n\n')
+
+
+    def _finalize_header(self):
+
+        # create model type
+        if self.enable_binary_blob:
+            self.header.write(f"\nstruct {self.model_struct_name} {{")
+            for name, data in self.layer_dict.items():
+                layer_type = data[0]
+                self.header.write(f"\n    {layer_type} {name};")
+            self.header.write(f"\n}};\n")
+
+            init_prototype = f"int init_{self.model_struct_name.lower()}({self.model_struct_name} *model, const WeightArray *arrays)"
+            self.header.write(f"\n{init_prototype};\n")
+
+        self.header.write(f"\n#endif /* {self.header_guard} */\n")
+
+    def _finalize_source(self):
+
+        if self.enable_binary_blob:
+            # create weight array
+            self.source.write("\n#ifndef USE_WEIGHTS_FILE\n")
+            self.source.write(f"const WeightArray {self.model_struct_name.lower()}_arrays[] = {{\n")
+            for name in self.weight_arrays:
+                self.source.write(f"#ifdef WEIGHTS_{name}_DEFINED\n")
+                self.source.write(f'    {{"{name}",  WEIGHTS_{name}_TYPE, sizeof({name}), {name}}},\n')
+                self.source.write(f"#endif\n")
+            self.source.write("    {NULL, 0, 0, NULL}\n")
+            self.source.write("};\n")
+
+            self.source.write("#endif /* USE_WEIGHTS_FILE */\n")
+
+            # create init function definition
+            init_prototype = f"int init_{self.model_struct_name.lower()}({self.model_struct_name} *model, const WeightArray *arrays)"
+            self.source.write("\n#ifndef DUMP_BINARY_WEIGHTS\n")
+            self.source.write(f"{init_prototype} {{\n")
+            for name, data in self.layer_dict.items():
+                self.source.write(f"    if ({data[1]}) return 1;\n")
+            self.source.write("    return 0;\n")
+            self.source.write("}\n")
+            self.source.write("#endif /* DUMP_BINARY_WEIGHTS */\n")
+
+
+    def close(self):
+
+        if not self.header_only:
+            self._finalize_source()
+            self.source.close()
+
+        self._finalize_header()
+        self.header.close()
+
+    def __del__(self):
+        try:
+            self.close()
+        except:
+            pass
\ No newline at end of file
diff --git a/dnn/torch/weight-exchange/wexchange/c_export/common.py b/dnn/torch/weight-exchange/wexchange/c_export/common.py
new file mode 100644
index 00000000..f469f848
--- /dev/null
+++ b/dnn/torch/weight-exchange/wexchange/c_export/common.py
@@ -0,0 +1,315 @@
+'''Copyright (c) 2017-2018 Mozilla
+   Copyright (c) 2022 Amazon
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
+   CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+'''
+
+import numpy as np
+
+from .c_writer import CWriter
+
+def print_vector(writer, vector, name, dtype='float', dotp=False, static=True):
+
+    f = writer.source
+    binary_blob = writer.enable_binary_blob
+
+    if binary_blob:
+        f.write(
+f'''
+#ifndef USE_WEIGHTS_FILE
+#define WEIGHTS_{name}_DEFINED
+#define WEIGHTS_{name}_TYPE WEIGHT_TYPE_{"qweight" if dotp else "float"}
+'''
+        )
+        writer.weight_arrays.add(name)
+
+    if dotp:
+        vector = vector.reshape((vector.shape[0]//4, 4, vector.shape[1]//8, 8))
+        vector = vector.transpose((2, 0, 3, 1))
+
+    v = np.reshape(vector, (-1))
+
+    if static:
+        f.write('static ')
+
+    f.write(f'const {dtype} {name}[{len(v)}] = {{\n    ')
+
+    for i in range(0, len(v)):
+
+        f.write(f'{v[i]}')
+
+        if (i!=len(v)-1):
+            f.write(',')
+        else:
+            break
+
+        if (i%8==7):
+            f.write("\n    ")
+        else:
+            f.write(" ")
+
+    f.write('\n};\n\n')
+    if binary_blob:
+        f.write(
+f'''
+#endif /* USE_WEIGHTS_FILE */
+'''
+        )
+
+    return vector
+
+
+
+def print_sparse_vector(writer, A, name, have_diag=True):
+    f = writer.source
+    N = A.shape[0]
+    M = A.shape[1]
+    W = np.zeros((0,), dtype='int')
+    W0 = np.zeros((0,))
+    if have_diag:
+        diag = np.concatenate([np.diag(A[:,:N]), np.diag(A[:,N:2*N]), np.diag(A[:,2*N:])])
+        A[:,:N] = A[:,:N] - np.diag(np.diag(A[:,:N]))
+        A[:,N:2*N] = A[:,N:2*N] - np.diag(np.diag(A[:,N:2*N]))
+        A[:,2*N:] = A[:,2*N:] - np.diag(np.diag(A[:,2*N:]))
+        print_vector(writer, diag, name + '_diag')
+    AQ = np.minimum(127, np.maximum(-128, np.round(A*128))).astype('int')
+    idx = np.zeros((0,), dtype='int')
+    for i in range(M//8):
+        pos = idx.shape[0]
+        idx = np.append(idx, -1)
+        nb_nonzero = 0
+        for j in range(N//4):
+            block = A[j*4:(j+1)*4, i*8:(i+1)*8]
+            qblock = AQ[j*4:(j+1)*4, i*8:(i+1)*8]
+            if np.sum(np.abs(block)) > 1e-10:
+                nb_nonzero = nb_nonzero + 1
+                idx = np.append(idx, j*4)
+                vblock = qblock.transpose((1,0)).reshape((-1,))
+                W0 = np.concatenate([W0, block.reshape((-1,))])
+                W = np.concatenate([W, vblock])
+        idx[pos] = nb_nonzero
+    f.write('#ifdef DOT_PROD\n')
+    print_vector(writer, W, name, dtype='qweight')
+    f.write('#else /*DOT_PROD*/\n')
+    print_vector(writer, W0, name, dtype='qweight')
+    f.write('#endif /*DOT_PROD*/\n')
+
+    print_vector(writer, idx, name + '_idx', dtype='int')
+    return AQ
+
+def _check_activation(activation):
+    if not activation in {"TANH", "SIGMOID", "LINEAR", "SWISH", "RELU", "SOFTMAX"}:
+        raise ValueError(f"error: unknown activation {activation}")
+
+def print_dense_layer(writer : CWriter,
+                      name : str,
+                      weight : np.ndarray,
+                      bias : np.ndarray,
+                      activation: str,
+                      format : str = 'torch'):
+
+    _check_activation(activation)
+
+    if format == 'torch':
+        weight = weight.transpose()
+
+    print_vector(writer, weight, name + "_weights")
+    print_vector(writer, bias, name + "_bias")
+
+    writer.header.write(f"\n#define {name.upper()}_OUT_SIZE {weight.shape[1]}\n")
+
+    if writer.enable_binary_blob:
+        init_call = f'dense_init(&model->{name}, arrays, "{name}_bias", "{name}_weights", {weight.shape[0]}, {weight.shape[1]}, ACTIVATION_{activation})'
+        writer.layer_dict[name] = ('DenseLayer', init_call)
+    else:
+        writer.source.write(
+f"""
+
+const DenseLayer {name} = {{
+   {name}_bias,
+   {name}_weights,
+   {weight.shape[0]},
+   {weight.shape[1]},
+   ACTIVATION_{activation}
+}};
+
+"""
+        )
+
+        writer.header.write(f"\nextern const DenseLayer {name};\n\n")
+
+
+
+
+
+def print_conv1d_layer(writer : CWriter,
+                       name : str,
+                       weight : np.ndarray,
+                       bias : np.ndarray,
+                       activation: str,
+                       format : str = 'torch'):
+
+    _check_activation(activation)
+
+    if format == "torch":
+        # convert to channels last
+        weight = np.transpose(weight, (2, 1, 0))
+
+    print_vector(writer, weight, name + "_weights")
+    print_vector(writer, bias, name + "_bias")
+
+    writer.header.write(f"\n#define {name.upper()}_OUT_SIZE {weight.shape[2]}\n")
+    writer.header.write(f"\n#define {name.upper()}_STATE_SIZE ({weight.shape[1]} * ({weight.shape[0] - 1}))\n")
+    writer.header.write(f"\n#define {name.upper()}_DELAY {(weight.shape[0] - 1) // 2}\n") # CAVE: delay is not a property of the conv layer
+
+    if writer.enable_binary_blob:
+        init_call = f'conv1d_init(&model->{name}, arrays, "{name}_bias", "{name}_weights", {weight.shape[1]}, {weight.shape[0]}, {weight.shape[2]}, ACTIVATION_{activation})'
+        writer.layer_dict[name] = ('Conv1DLayer', init_call)
+    else:
+
+        writer.source.write(
+f"""
+
+const Conv1DLayer {name} = {{
+   {name}_bias,
+   {name}_weights,
+   {weight.shape[1]},
+   {weight.shape[0]},
+   {weight.shape[2]},
+   ACTIVATION_{activation}
+}};
+
+"""
+        )
+
+        writer.header.write(f"\nextern const Conv1DLayer {name};\n\n")
+
+    return weight.shape[0] * weight.shape[1]
+
+
+def print_gru_layer(writer : CWriter,
+                    name : str,
+                    weight : np.ndarray,
+                    recurrent_weight : np.ndarray,
+                    bias : np.ndarray,
+                    recurrent_bias : np.ndarray,
+                    activation: str,
+                    format : str = 'torch',
+                    dotp : bool = False,
+                    input_sparse : bool = False,
+                    reset_after : int = 0
+                    ):
+
+    _check_activation(activation)
+
+    if format == "torch":
+        # transpose weight matrices and change gate order from rzn to zrn
+
+        N = weight.shape[0] // 3
+        for x in [weight, recurrent_weight, bias, recurrent_bias]:
+            tmp = x[0:N].copy()
+            x[0:N] = x[N:2*N]
+            x[N:2*N] = tmp
+
+        weight = weight.transpose()
+        recurrent_weight = recurrent_weight.transpose()
+
+
+    # input weights
+    if input_sparse:
+        qweight = print_sparse_vector(writer, weight, name + '_weights', have_diag=False)
+    else:
+        qweight = np.clip(np.round(128. * weight).astype('int'), -128, 127)
+
+        if dotp:
+            writer.source.write("#ifdef DOT_PROD\n")
+            print_vector(writer, qweight, name + '_weights', dtype='qweight', dotp=True)
+            writer.source.write("#else /*DOT_PROD*/\n")
+
+        print_vector(writer, weight, name + '_weights')
+
+        if dotp:
+             writer.source.write("#endif /*DOT_PROD*/\n")
+
+
+    # recurrent weights
+    recurrent_qweight = np.clip(np.round(128. * recurrent_weight).astype('int'), -128, 127)
+
+    if dotp:
+        writer.source.write("#ifdef DOT_PROD\n")
+        print_vector(writer, recurrent_qweight, name + '_recurrent_weights', dtype='qweight', dotp=True)
+        writer.source.write("#else /*DOT_PROD*/\n")
+
+    print_vector(writer, recurrent_weight, name + '_recurrent_weights')
+
+    if dotp:
+        writer.source.write("#endif /*DOT_PROD*/\n")
+
+
+    # corrected bias for unsigned int matrix multiplication
+    subias              = bias - np.sum(qweight / 128., axis=0)
+    recurrent_subias    = recurrent_bias - np.sum(recurrent_qweight / 128., axis=0)
+
+    print_vector(writer, np.concatenate((bias, recurrent_bias)), name + "_bias")
+    print_vector(writer, np.concatenate((subias, recurrent_subias)), name + "_subias")
+
+
+    # wrapping it up
+    writer.header.write(f"\n#define {name.upper()}_OUT_SIZE {N}\n")
+    writer.header.write(f"\n#define {name.upper()}_STATE_SIZE {N}\n")
+
+    if writer.enable_binary_blob:
+        if input_sparse:
+            init_call = f'gru_init(&model->{name}, arrays, "{name}_bias", "{name}_subias", "{name}_weights", "{name + "_weights_idx"}", "{name}_recurrent_weights", {weight.shape[0]}, {weight.shape[1] // 3}, ACTIVATION_{activation}, {reset_after})'
+        else:
+            init_call = f'gru_init(&model->{name}, arrays, "{name}_bias", "{name}_subias", "{name}_weights", NULL, "{name}_recurrent_weights", {weight.shape[0]}, {weight.shape[1] // 3}, ACTIVATION_{activation}, {reset_after})'
+
+        writer.layer_dict[name] = ('GRULayer', init_call)
+
+    else:
+
+        writer.source.write(
+f"""
+
+const GRULayer {name} = {{
+   {name}_bias,
+   {name}_subias,
+   {name}_weights,
+   {name + "_weights_idx" if input_sparse else "NULL"},
+   {name}_recurrent_weights,
+   {weight.shape[0]},
+   {weight.shape[1] // 3},
+   ACTIVATION_{activation},
+   {reset_after}
+}};
+
+"""
+        )
+
+        writer.header.write(f"\nextern const GRULayer {name};\n")
+
+
+    return N
+
+
diff --git a/dnn/torch/weight-exchange/wexchange/tf/__init__.py b/dnn/torch/weight-exchange/wexchange/tf/__init__.py
new file mode 100644
index 00000000..02678048
--- /dev/null
+++ b/dnn/torch/weight-exchange/wexchange/tf/__init__.py
@@ -0,0 +1,5 @@
+from .tf import dump_tf_conv1d_weights, load_tf_conv1d_weights
+from .tf import dump_tf_dense_weights, load_tf_dense_weights
+from .tf import dump_tf_embedding_weights, load_tf_embedding_weights
+from .tf import dump_tf_gru_weights, load_tf_gru_weights
+from .tf import dump_tf_weights, load_tf_weights
\ No newline at end of file
diff --git a/dnn/torch/weight-exchange/wexchange/tf/tf.py b/dnn/torch/weight-exchange/wexchange/tf/tf.py
new file mode 100644
index 00000000..d6481eb7
--- /dev/null
+++ b/dnn/torch/weight-exchange/wexchange/tf/tf.py
@@ -0,0 +1,169 @@
+import os
+
+import tensorflow as tf
+import numpy as np
+
+from wexchange.c_export import CWriter, print_gru_layer, print_dense_layer, print_conv1d_layer
+
+def dump_tf_gru_weights(where, gru, name=None, input_sparse=False, dotp=False):
+
+
+    assert gru.activation == tf.keras.activations.tanh
+    assert gru.recurrent_activation == tf.keras.activations.sigmoid
+    assert gru.reset_after == True
+
+    w_ih = gru.weights[0].numpy().transpose().copy()
+    w_hh = gru.weights[1].numpy().transpose().copy()
+    b_ih = gru.weights[2].numpy()[0].copy()
+    b_hh = gru.weights[2].numpy()[1].copy()
+
+    if isinstance(where, CWriter):
+        return print_gru_layer(where, name, w_ih, w_hh, b_ih, b_hh, 'TANH', format='tf', reset_after=1, input_sparse=input_sparse, dotp=dotp)
+    else:
+        os.makedirs(where, exist_ok=True)
+
+        # zrn => rzn
+        N = w_ih.shape[0] // 3
+        for x in [w_ih, w_hh, b_ih, b_hh]:
+            tmp = x[0:N].copy()
+            x[0:N] = x[N:2*N]
+            x[N:2*N] = tmp
+
+        np.save(os.path.join(where, 'weight_ih_rzn.npy'), w_ih)
+        np.save(os.path.join(where, 'weight_hh_rzn.npy'), w_hh)
+        np.save(os.path.join(where, 'bias_ih_rzn.npy'), b_ih)
+        np.save(os.path.join(where, 'bias_hh_rzn.npy'), b_hh)
+
+
+def load_tf_gru_weights(path, gru):
+
+    assert gru.activation == tf.keras.activations.tanh
+    assert gru.recurrent_activation == tf.keras.activations.sigmoid
+    assert gru.reset_after == True
+
+    w_ih = np.load(os.path.join(path, 'weight_ih_rzn.npy'))
+    w_hh = np.load(os.path.join(path, 'weight_hh_rzn.npy'))
+    b_ih = np.load(os.path.join(path, 'bias_ih_rzn.npy'))
+    b_hh = np.load(os.path.join(path, 'bias_hh_rzn.npy'))
+
+    # rzn => zrn
+    N = w_ih.shape[0] // 3
+    for x in [w_ih, w_hh, b_ih, b_hh]:
+        tmp = x[0:N].copy()
+        x[0:N] = x[N:2*N]
+        x[N:2*N] = tmp
+
+    gru.weights[0].assign(tf.convert_to_tensor(w_ih.transpose()))
+    gru.weights[1].assign(tf.convert_to_tensor(w_hh.transpose()))
+    gru.weights[2].assign(tf.convert_to_tensor(np.vstack((b_ih, b_hh))))
+
+
+def dump_tf_dense_weights(where, dense, name=None):
+
+    w = dense.weights[0].numpy()
+    if dense.bias is None:
+        b = np.zeros(dense.units, dtype=w.dtype)
+    else:
+        b = dense.bias.numpy()
+
+
+
+    if isinstance(where, CWriter):
+        try:
+            activation = dense.activation.__name__.upper()
+        except:
+            activation = "LINEAR"
+
+        return print_dense_layer(where, name, w, b, activation, format='tf')
+
+    else:
+        os.makedirs(where, exist_ok=True)
+
+        np.save(os.path.join(where, 'weight.npy'), w.transpose())
+        np.save(os.path.join(where, 'bias.npy'), b)
+
+
+def load_tf_dense_weights(path, dense):
+
+    w = np.load(os.path.join(path, 'weight.npy')).transpose()
+    b = np.load(os.path.join(path, 'bias.npy'))
+
+    dense.weights[0].assign(tf.convert_to_tensor(w))
+    if dense.bias is not None:
+        dense.weights[1].assign(tf.convert_to_tensor(b))
+
+
+def dump_tf_conv1d_weights(where, conv, name=None):
+
+    assert conv.data_format == 'channels_last'
+
+    w = conv.weights[0].numpy().copy()
+    if conv.bias is None:
+        b = np.zeros(conv.filters, dtype=w.dtype)
+    else:
+        b = conv.bias.numpy()
+
+    if isinstance(where, CWriter):
+        try:
+            activation = conv.activation.__name__.upper()
+        except:
+            activation = "LINEAR"
+
+        return print_conv1d_layer(where, name, w, b, activation, format='tf')
+    else:
+        os.makedirs(where, exist_ok=True)
+
+        w = np.transpose(w, (2, 1, 0))
+        np.save(os.path.join(where, 'weight_oik.npy'), w)
+        np.save(os.path.join(where, 'bias.npy'), b)
+
+
+def load_tf_conv1d_weights(path, conv):
+
+    w = np.load(os.path.join(path, 'weight_oik.npy'))
+    b = np.load(os.path.join(path, 'bias.npy'))
+
+    w = np.transpose(w, (2, 1, 0))
+
+    conv.weights[0].assign(tf.convert_to_tensor(w))
+    if conv.bias is not None:
+        conv.weights[1].assign(tf.convert_to_tensor(b))
+
+
+def dump_tf_embedding_weights(path, emb):
+    os.makedirs(path, exist_ok=True)
+
+    w = emb.weights[0].numpy()
+    np.save(os.path.join(path, 'weight.npy'), w)
+
+
+
+def load_tf_embedding_weights(path, emb):
+
+    w = np.load(os.path.join(path, 'weight.npy'))
+    emb.weights[0].assign(tf.convert_to_tensor(w))
+
+
+def dump_tf_weights(path, module):
+    if isinstance(module, tf.keras.layers.Dense):
+        dump_tf_dense_weights(path, module)
+    elif isinstance(module, tf.keras.layers.GRU):
+        dump_tf_gru_weights(path, module)
+    elif isinstance(module, tf.keras.layers.Conv1D):
+        dump_tf_conv1d_weights(path, module)
+    elif isinstance(module, tf.keras.layers.Embedding):
+        dump_tf_embedding_weights(path, module)
+    else:
+        raise ValueError(f'dump_tf_weights: layer of type {type(module)} not supported')
+
+def load_tf_weights(path, module):
+    if isinstance(module, tf.keras.layers.Dense):
+        load_tf_dense_weights(path, module)
+    elif isinstance(module, tf.keras.layers.GRU):
+        load_tf_gru_weights(path, module)
+    elif isinstance(module, tf.keras.layers.Conv1D):
+        load_tf_conv1d_weights(path, module)
+    elif isinstance(module, tf.keras.layers.Embedding):
+        load_tf_embedding_weights(path, module)
+    else:
+        raise ValueError(f'dump_tf_weights: layer of type {type(module)} not supported')
\ No newline at end of file
diff --git a/dnn/torch/weight-exchange/wexchange/torch/__init__.py b/dnn/torch/weight-exchange/wexchange/torch/__init__.py
new file mode 100644
index 00000000..61b57665
--- /dev/null
+++ b/dnn/torch/weight-exchange/wexchange/torch/__init__.py
@@ -0,0 +1,5 @@
+from .torch import dump_torch_conv1d_weights, load_torch_conv1d_weights
+from .torch import dump_torch_dense_weights, load_torch_dense_weights
+from .torch import dump_torch_gru_weights, load_torch_gru_weights
+from .torch import dump_torch_embedding_weights, load_torch_embedding_weights
+from .torch import dump_torch_weights, load_torch_weights
diff --git a/dnn/torch/weight-exchange/wexchange/torch/torch.py b/dnn/torch/weight-exchange/wexchange/torch/torch.py
new file mode 100644
index 00000000..92a20d3d
--- /dev/null
+++ b/dnn/torch/weight-exchange/wexchange/torch/torch.py
@@ -0,0 +1,146 @@
+import os
+
+import torch
+import numpy as np
+
+from wexchange.c_export import CWriter, print_gru_layer, print_dense_layer, print_conv1d_layer
+
+def dump_torch_gru_weights(where, gru, name=None, input_sparse=False, dotp=False):
+
+    assert gru.num_layers == 1
+    assert gru.bidirectional == False
+
+    w_ih = gru.weight_ih_l0.detach().cpu().numpy()
+    w_hh = gru.weight_hh_l0.detach().cpu().numpy()
+    b_ih = gru.bias_ih_l0.detach().cpu().numpy()
+    b_hh = gru.bias_hh_l0.detach().cpu().numpy()
+
+    if isinstance(where, CWriter):
+        return print_gru_layer(where, name, w_ih, w_hh, b_ih, b_hh, 'TANH', format='torch', reset_after=1, input_sparse=input_sparse, dotp=dotp)
+    else:
+        os.makedirs(where, exist_ok=True)
+
+        np.save(os.path.join(where, 'weight_ih_rzn.npy'), w_ih)
+        np.save(os.path.join(where, 'weight_hh_rzn.npy'), w_hh)
+        np.save(os.path.join(where, 'bias_ih_rzn.npy'), b_ih)
+        np.save(os.path.join(where, 'bias_hh_rzn.npy'), b_hh)
+
+
+
+def load_torch_gru_weights(where, gru):
+
+    assert gru.num_layers == 1
+    assert gru.bidirectional == False
+
+    w_ih = np.load(os.path.join(where, 'weight_ih_rzn.npy'))
+    w_hh = np.load(os.path.join(where, 'weight_hh_rzn.npy'))
+    b_ih = np.load(os.path.join(where, 'bias_ih_rzn.npy'))
+    b_hh = np.load(os.path.join(where, 'bias_hh_rzn.npy'))
+
+    with torch.no_grad():
+        gru.weight_ih_l0.set_(torch.from_numpy(w_ih))
+        gru.weight_hh_l0.set_(torch.from_numpy(w_hh))
+        gru.bias_ih_l0.set_(torch.from_numpy(b_ih))
+        gru.bias_hh_l0.set_(torch.from_numpy(b_hh))
+
+
+def dump_torch_dense_weights(where, dense, name=None, activation="LINEAR"):
+
+    w = dense.weight.detach().cpu().numpy()
+    if dense.bias is None:
+        b = np.zeros(dense.out_features, dtype=w.dtype)
+    else:
+        b = dense.bias.detach().cpu().numpy()
+
+    if isinstance(where, CWriter):
+        return print_dense_layer(where, name, w, b, activation, format='torch')
+
+    else:
+        os.makedirs(where, exist_ok=True)
+
+        np.save(os.path.join(where, 'weight.npy'), w)
+        np.save(os.path.join(where, 'bias.npy'), b)
+
+
+def load_torch_dense_weights(where, dense):
+
+    w = np.load(os.path.join(where, 'weight.npy'))
+    b = np.load(os.path.join(where, 'bias.npy'))
+
+    with torch.no_grad():
+        dense.weight.set_(torch.from_numpy(w))
+        if dense.bias is not None:
+            dense.bias.set_(torch.from_numpy(b))
+
+
+def dump_torch_conv1d_weights(where, conv, name=None, activation="LINEAR"):
+
+    w = conv.weight.detach().cpu().numpy()
+    if conv.bias is None:
+        b = np.zeros(conv.out_channels, dtype=w.dtype)
+    else:
+        b = conv.bias.detach().cpu().numpy()
+
+    if isinstance(where, CWriter):
+
+        return print_conv1d_layer(where, name, w, b, activation, format='torch')
+    else:
+        os.makedirs(where, exist_ok=True)
+
+        np.save(os.path.join(where, 'weight_oik.npy'), w)
+
+        np.save(os.path.join(where, 'bias.npy'), b)
+
+
+def load_torch_conv1d_weights(where, conv):
+
+    with torch.no_grad():
+        w = np.load(os.path.join(where, 'weight_oik.npy'))
+        conv.weight.set_(torch.from_numpy(w))
+        if type(conv.bias) != type(None):
+            b = np.load(os.path.join(where, 'bias.npy'))
+            if conv.bias is not None:
+                conv.bias.set_(torch.from_numpy(b))
+
+
+def dump_torch_embedding_weights(where, emb):
+    os.makedirs(where, exist_ok=True)
+
+    w = emb.weight.detach().cpu().numpy()
+    np.save(os.path.join(where, 'weight.npy'), w)
+
+
+def load_torch_embedding_weights(where, emb):
+
+    w = np.load(os.path.join(where, 'weight.npy'))
+
+    with torch.no_grad():
+        emb.weight.set_(torch.from_numpy(w))
+
+def dump_torch_weights(where, module, name=None, activation="LINEAR", verbose=False, **kwargs):
+    """ generic function for dumping weights of some torch.nn.Module """
+    if verbose and name is not None:
+        print(f"printing layer {name} of type {type(module)}...")
+    if isinstance(module, torch.nn.Linear):
+        return dump_torch_dense_weights(where, module, name, activation, **kwargs)
+    elif isinstance(module, torch.nn.GRU):
+        return dump_torch_gru_weights(where, module, name, **kwargs)
+    elif isinstance(module, torch.nn.Conv1d):
+        return dump_torch_conv1d_weights(where, module, name, **kwargs)
+    elif isinstance(module, torch.nn.Embedding):
+        return dump_torch_embedding_weights(where, module)
+    else:
+        raise ValueError(f'dump_tf_weights: layer of type {type(module)} not supported')
+
+def load_torch_weights(where, module):
+    """ generic function for loading weights of some torch.nn.Module """
+    if isinstance(module, torch.nn.Linear):
+        load_torch_dense_weights(where, module)
+    elif isinstance(module, torch.nn.GRU):
+        load_torch_gru_weights(where, module)
+    elif isinstance(module, torch.nn.Conv1d):
+        load_torch_conv1d_weights(where, module)
+    elif isinstance(module, torch.nn.Embedding):
+        load_torch_embedding_weights(where, module)
+    else:
+        raise ValueError(f'dump_tf_weights: layer of type {type(module)} not supported')
\ No newline at end of file