Fixed ordering of values to keras convention

kayaleitner · Mar 4, 2020 · 6f74b3c · 6f74b3c
1 parent ecfc73c
commit 6f74b3c
Show file tree

Hide file tree

Showing 17 changed files with 16,358 additions and 16,294 deletions.
diff --git a/net/README.md b/net/README.md
@@ -0,0 +1,30 @@
+# Neural Network Training
+
+The training can be done with either Pytorch or Keras, but Pytorch is recommended because of the unreliable
+API of Tensorflow (the foundation of Keras)
+
+Make sure you install all requirements:
+
+````shell script
+pip install -r requirements.txt
+````
+
+To train the network simply run in a terminal:
+
+```shell script
+
+python train_torch.py
+python quantize.py
+
+```
+
+or use the Jupyter notebook to train the network.
+
+## Quantisiation
+
+For quantisation fixed point quantisation has been used.
+
+| Network              | Accuracy |
+|----------------------|---------:|
+| Network:             | 0.9832   |
+| Quantised Network:   | 0.9349   |
diff --git a/net/extract_net_parameters.py b/net/extract_net_parameters.py
@@ -1,28 +1,20 @@
 import numpy as np
 import os
-import torch
-import keras
 import shutil
 
 # The line below is needed so that torch.load() works as expected
-from train_torch import LeNetV2, ConvBN, LinearRelu, Flatten
 
 # Chose if you want to use the training parameters from Keras or Torch
+from train_keras import load_keras
+from train_torch import save_torch_model_weights, load_torch
+
 EXTRACT_FROM_KERAS = True
 EXTRACT_FROM_TROCH = True
 
 EXPORT_DIR = 'np'
 
 SCRIPT_PATH = os.path.dirname(os.path.abspath(__file__))
 
-KERAS_SAVE_DIR = 'keras'
-KERAS_CONFIG_FILE = os.path.join(KERAS_SAVE_DIR, 'model_config.json')
-KERAS_WEIGHTS_FILE = os.path.join(KERAS_SAVE_DIR, 'weights.h5')
-
-TORCH_SAVE_DIR = os.path.join(SCRIPT_PATH, 'torch')
-TORCH_SAVE_FILE = os.path.join(TORCH_SAVE_DIR, 'LeNet.pth')
-TORCH_STATES_SAVE_FILE = os.path.join(TORCH_SAVE_DIR, 'LeNetStates.pth')
-
 
 def extract(cleanup=True, extract_keras=EXTRACT_FROM_KERAS, extract_torch=EXTRACT_FROM_TROCH):
     # Clean up and create dirs
@@ -45,107 +37,5 @@ def extract(cleanup=True, extract_keras=EXTRACT_FROM_KERAS, extract_torch=EXTRAC
         save_torch_model_weights(tmodel)
 
 
-def save_torch_model_weights(tmodel):
-    tmodel.eval()  # Put in eval mode
-    for key, weights in tmodel.state_dict().items():
-        weights = weights.numpy()
-        vals = weights.flatten(order='C')
-        np.savetxt(os.path.join(EXPORT_DIR, 't_{}.txt'.format(key)), vals,
-                   header=str(weights.shape))
-
-
-def load_torch(filepath=TORCH_SAVE_FILE):
-    from train_torch import get_lenet_model
-
-    # Create a model instance, make sure that save data and source code is in sync
-    # model = get_lenet_model()
-    # model.load_state_dict(torch.load(TORCH_STATES_SAVE_FILE))
-
-    # ToDo: Better way would be the line below but it doesnt work for me :(
-    model = torch.load(filepath)
-
-    return model
-
-
-def load_keras() -> keras.Model:
-    if not os.path.exists(KERAS_CONFIG_FILE) or not os.path.exists(KERAS_WEIGHTS_FILE):
-        raise RuntimeError("There is no trained model data! (or the model might have the wrong filename?)")
-
-    # Reload the model from the 2 files we saved
-    with open(KERAS_CONFIG_FILE) as json_file:
-        json_config = json_file.read()
-
-    model = keras.models.model_from_json(json_config)
-    model.load_weights(KERAS_WEIGHTS_FILE)
-    return model
-
-
-def _read_np_tensor(weight_file):
-    with open(weight_file) as f:
-        init_line = f.readline()
-
-    # assume the first line is a comment
-    assert init_line[0] == '#'
-    p1 = init_line.find('(')
-    p2 = init_line.find(')')
-    dims = [int(ds) for ds in init_line[p1 + 1:p2].split(sep=',') if len(ds) > 0]
-
-    return np.loadtxt(weight_file).reshape(dims)
-
-
-def reorder(x):
-    co, ci, h, w, = x.shape
-    x_ = np.zeros(shape=(h, w, ci, co), dtype=x.dtype)
-
-    for hx in range(h):
-        for wx in range(w):
-            for cix in range(ci):
-                for cox in range(co):
-                    x_[hx, wx, cix, cox] = x[cox, cix, hx, wx]
-    return x_
-
-
-def read_np_torch(ordering='BCHW', target_dtype=None):
-    d = {
-        'cn1.b': _read_np_tensor('np/t_0.0.bias.txt'),
-        'cn1.k': _read_np_tensor('np/t_0.0.weight.txt'),
-        'cn2.b': _read_np_tensor('np/t_3.0.bias.txt'),
-        'cn2.k': _read_np_tensor('np/t_3.0.weight.txt'),
-        'fc1.b': _read_np_tensor('np/t_7.0.bias.txt'),
-        'fc1.w': _read_np_tensor('np/t_7.0.weight.txt'),
-        'fc2.b': _read_np_tensor('np/t_9.0.bias.txt'),
-        'fc2.w': _read_np_tensor('np/t_9.0.weight.txt'),
-    }
-
-    if ordering is 'BCHW':
-        pass
-    elif ordering is 'BHWC':
-        k1 = np.moveaxis(d['cn1.k'], [0, 1], [3, 2])
-        k2 = np.moveaxis(d['cn2.k'], [0, 1], [3, 2])
-
-        # k1 = reorder(d['cn1.k'])
-        # k2 = reorder(d['cn2.k'])
-
-        assert k1[1, 2, 0, 4] == d['cn1.k'][4, 0, 1, 2]
-        assert k2[1, 2, 3, 4] == d['cn2.k'][4, 3, 1, 2]
-
-        d['cn1.k'] = k1
-        d['cn2.k'] = k2
-        d['fc1.w'] = np.moveaxis(d['fc1.w'], 0, 1)
-        d['fc2.w'] = np.moveaxis(d['fc2.w'], 0, 1)
-
-
-
-    else:
-        raise NotImplementedError('Expected ordering to be "BCHW" or "BHWC" but is: {}'.format(ordering))
-
-    if target_dtype is not None:
-        d_old = d
-        d = {}
-        for key, values in d_old.items():
-            d[key] = values.astype(target_dtype)
-    return d
-
-
 if __name__ == '__main__':
     extract()
diff --git a/net/final_weights/cn1.b.npy b/net/final_weights/cn1.b.npy
diff --git a/net/final_weights/cn1.k.npy b/net/final_weights/cn1.k.npy
diff --git a/net/final_weights/cn2.b.npy b/net/final_weights/cn2.b.npy
diff --git a/net/final_weights/cn2.k.npy b/net/final_weights/cn2.k.npy
diff --git a/net/final_weights/fc1.b.npy b/net/final_weights/fc1.b.npy
diff --git a/net/final_weights/fc1.w.npy b/net/final_weights/fc1.w.npy