I want to transfer pretrained weights from an old project on github :
The original Keras model code is:
def get_keras_autoencoder(self, input_size=256, nb_filter=96, k_size=5):
input_img = Input(shape=(1, input_size, input_size))
conv1 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv1')(input_img)
maxp1 = MaxPooling2D((2, 2), border_mode='same', name='maxp1')(conv1)
conv2 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv2')(maxp1)
maxp2 = MaxPooling2D((2, 2), border_mode='same', name='maxp2')(conv2)
conv3 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv3')(maxp2)
encoder = MaxPooling2D((2, 2), border_mode='same', name='encoder')(conv3)
conv4 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv4')(encoder)
upsa1 = UpSampling2D((2, 2), name='upsa1')(conv4)
conv4 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv5')(upsa1)
upsa2 = UpSampling2D((2, 2), name='upsa2')(conv4)
conv5 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv6')(upsa2)
upsa3 = UpSampling2D((2, 2), name='upsa3')(conv5)
decoder = Convolution2D(1, k_size, k_size, activation='sigmoid', border_mode='same')(upsa3)
autoencoder = Model(input_img, decoder)
return autoencoder
The ported PyTorch code I've done is:
class SAEModel(nn.Module):
def __init__(self):
super(SAEModel, self).__init__()
# encoder
self.conv1 = nn.Conv2d( 1, 96, kernel_size=5, stride=1, padding=2)
self.maxp1 = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
self.conv2 = nn.Conv2d(96, 96, kernel_size=5, stride=1, padding=2)
self.maxp2 = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
self.conv3 = nn.Conv2d(96, 96, kernel_size=5, stride=1, padding=2)
self.encod = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
# decoder
self.conv4 = nn.Conv2d(96, 96, kernel_size=5, stride=1, padding=2)
self.upsa1 = nn.Upsample(scale_factor=2)
self.conv5 = nn.Conv2d(96, 96, kernel_size=5, stride=1, padding=2)
self.upsa2 = nn.Upsample(scale_factor=2)
self.conv6 = nn.Conv2d(96, 96, kernel_size=5, stride=1, padding=2)
self.upsa3 = nn.Upsample(scale_factor=2)
self.decod = nn.Conv2d(96, 1, kernel_size=5, stride=1, padding=2)
def forward(self, x):
# encoder activations
x = F.relu(self.conv1(x))
x = self.maxp1(x)
x = F.relu(self.conv2(x))
x = self.maxp2(x)
x = F.relu(self.conv3(x))
x = self.encod(x)
# decoder activations
x = F.relu(self.conv4(x))
x = self.upsa1(x)
x = F.relu(self.conv5(x))
x = self.upsa2(x)
x = F.relu(self.conv6(x))
x = self.upsa3(x)
x = F.sigmoid(self.decod(x))
return x
The summary dump of the Keras model:
Layer (type) Output Shape Param # Connected to
====================================================================================================
input_1 (InputLayer) (None, 1, 256, 256) 0
____________________________________________________________________________________________________
conv1 (Convolution2D) (None, 1, 256, 96) 2496 input_1[0][0]
____________________________________________________________________________________________________
maxp1 (MaxPooling2D) (None, 1, 128, 96) 0 conv1[0][0]
____________________________________________________________________________________________________
conv2 (Convolution2D) (None, 1, 128, 96) 230496 maxp1[0][0]
____________________________________________________________________________________________________
maxp2 (MaxPooling2D) (None, 1, 64, 96) 0 conv2[0][0]
____________________________________________________________________________________________________
conv3 (Convolution2D) (None, 1, 64, 96) 230496 maxp2[0][0]
____________________________________________________________________________________________________
encoder (MaxPooling2D) (None, 1, 32, 96) 0 conv3[0][0]
____________________________________________________________________________________________________
conv4 (Convolution2D) (None, 1, 32, 96) 230496 encoder[0][0]
____________________________________________________________________________________________________
upsa1 (UpSampling2D) (None, 2, 64, 96) 0 conv4[0][0]
____________________________________________________________________________________________________
conv5 (Convolution2D) (None, 2, 64, 96) 230496 upsa1[0][0]
____________________________________________________________________________________________________
upsa2 (UpSampling2D) (None, 4, 128, 96) 0 conv5[0][0]
____________________________________________________________________________________________________
conv6 (Convolution2D) (None, 4, 128, 96) 230496 upsa2[0][0]
____________________________________________________________________________________________________
upsa3 (UpSampling2D) (None, 8, 256, 96) 0 conv6[0][0]
____________________________________________________________________________________________________
convolution2d_1 (Convolution2D) (None, 8, 256, 1) 2401 upsa3[0][0]
====================================================================================================
Total params: 1,157,377
Trainable params: 1,157,377
Non-trainable params: 0
And the one for the ported PyTorch mode :
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 96, 256, 256] 2,496
MaxPool2d-2 [-1, 96, 128, 128] 0
Conv2d-3 [-1, 96, 128, 128] 230,496
MaxPool2d-4 [-1, 96, 64, 64] 0
Conv2d-5 [-1, 96, 64, 64] 230,496
MaxPool2d-6 [-1, 96, 32, 32] 0
Conv2d-7 [-1, 96, 32, 32] 230,496
Upsample-8 [-1, 96, 64, 64] 0
Conv2d-9 [-1, 96, 64, 64] 230,496
Upsample-10 [-1, 96, 128, 128] 0
Conv2d-11 [-1, 96, 128, 128] 230,496
Upsample-12 [-1, 96, 256, 256] 0
Conv2d-13 [-1, 1, 256, 256] 2,401
================================================================
Total params: 1,157,377
Trainable params: 1,157,377
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.25
Forward/backward pass size (MB): 158.00
Params size (MB): 4.42
Estimated Total Size (MB): 162.67
And finally the code to convert the weights :
import os
import argparse
from pathlib import Path
from keras.layers import Input, Convolution2D, MaxPooling2D, UpSampling2D
from keras.models import Model
from keras.models import load_model
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchsummary import summary
class Musica:
def decompose_path(self, path):
return Path(path).parent, Path(path).stem, Path(path).suffix
def convert_keras_weights(self, weights_path):
autoencoder = self.get_keras_autoencoder()
autoencoder.load_weights(weights_path)
directory, filename, extension = self.decompose_path(weights_path)
autoencoder.save(os.path.join(directory, filename + '_model_included' + extension))
def convert_keras_to_torch(self, full_model_path, print_models: bool = False, print_weights: bool = False):
keras_model = load_model(full_model_path)
if print_models:
print(keras_model.summary())
weights = keras_model.get_weights()
torch_model = SAEModel()
if print_models:
summary(torch_model.cuda(), (1, 256, 256))
if print_weights:
print(f'Keras Weight {weights[0].shape} and Pytorch Weight {torch_model.conv1.weight.shape}')
print(f'Keras Weight {weights[1].shape} and Pytorch Weight {torch_model.conv1.bias.shape}')
print(f'Keras Weight {weights[2].shape} and Pytorch Weight {torch_model.conv2.weight.shape}')
print(f'Keras Weight {weights[3].shape} and Pytorch Weight {torch_model.conv2.bias.shape}')
print(f'Keras Weight {weights[4].shape} and Pytorch Weight {torch_model.conv3.weight.shape}')
print(f'Keras Weight {weights[5].shape} and Pytorch Weight {torch_model.conv3.bias.shape}')
print(f'Keras Weight {weights[6].shape} and Pytorch Weight {torch_model.conv4.weight.shape}')
print(f'Keras Weight {weights[7].shape} and Pytorch Weight {torch_model.conv4.bias.shape}')
print(f'Keras Weight {weights[8].shape} and Pytorch Weight {torch_model.conv5.weight.shape}')
print(f'Keras Weight {weights[9].shape} and Pytorch Weight {torch_model.conv5.bias.shape}')
print(f'Keras Weight {weights[10].shape} and Pytorch Weight {torch_model.conv6.weight.shape}')
print(f'Keras Weight {weights[11].shape} and Pytorch Weight {torch_model.conv6.bias.shape}')
print(f'Keras Weight {weights[12].shape} and Pytorch Weight {torch_model.decod.weight.shape}')
print(f'Keras Weight {weights[13].shape} and Pytorch Weight {torch_model.decod.bias.shape}')
# load keras weights into torch model
torch_model.conv1.weight.data = torch.from_numpy(weights[0])
torch_model.conv1.bias.data = torch.from_numpy(weights[1])
torch_model.conv2.weight.data = torch.from_numpy(weights[2])
torch_model.conv2.bias.data = torch.from_numpy(weights[3])
torch_model.conv3.weight.data = torch.from_numpy(weights[4])
torch_model.conv3.bias.data = torch.from_numpy(weights[5])
torch_model.conv4.weight.data = torch.from_numpy(weights[6])
torch_model.conv4.bias.data = torch.from_numpy(weights[7])
torch_model.conv5.weight.data = torch.from_numpy(weights[8])
torch_model.conv5.bias.data = torch.from_numpy(weights[9])
torch_model.conv6.weight.data = torch.from_numpy(weights[10])
torch_model.conv6.bias.data = torch.from_numpy(weights[11])
torch_model.decod.weight.data = torch.from_numpy(weights[12])
torch_model.decod.bias.data = torch.from_numpy(weights[13])
directory, filename, extension = self.decompose_path(full_model_path)
export_path = os.path.join(directory, filename + '_torch' + '.pth')
torch.save(torch_model.state_dict(), export_path)
def get_keras_autoencoder(self, input_size=256, nb_filter=96, k_size=5):
input_img = Input(shape=(1, input_size, input_size))
conv1 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv1')(input_img)
maxp1 = MaxPooling2D((2, 2), border_mode='same', name='maxp1')(conv1)
conv2 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv2')(maxp1)
maxp2 = MaxPooling2D((2, 2), border_mode='same', name='maxp2')(conv2)
conv3 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv3')(maxp2)
encoder = MaxPooling2D((2, 2), border_mode='same', name='encoder')(conv3)
conv4 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv4')(encoder)
upsa1 = UpSampling2D((2, 2), name='upsa1')(conv4)
conv4 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv5')(upsa1)
upsa2 = UpSampling2D((2, 2), name='upsa2')(conv4)
conv5 = Convolution2D(nb_filter, k_size, k_size, activation='relu', border_mode='same', name='conv6')(upsa2)
upsa3 = UpSampling2D((2, 2), name='upsa3')(conv5)
decoder = Convolution2D(1, k_size, k_size, activation='sigmoid', border_mode='same')(upsa3)
autoencoder = Model(input_img, decoder)
return autoencoder
The weights transpose dump:
Keras Weight (96, 1, 5, 5) and Pytorch Weight torch.Size([96, 1, 5, 5])
Keras Weight (96,) and Pytorch Weight torch.Size([96])
Keras Weight (96, 96, 5, 5) and Pytorch Weight torch.Size([96, 96, 5, 5])
Keras Weight (96,) and Pytorch Weight torch.Size([96])
Keras Weight (96, 96, 5, 5) and Pytorch Weight torch.Size([96, 96, 5, 5])
Keras Weight (96,) and Pytorch Weight torch.Size([96])
Keras Weight (96, 96, 5, 5) and Pytorch Weight torch.Size([96, 96, 5, 5])
Keras Weight (96,) and Pytorch Weight torch.Size([96])
Keras Weight (96, 96, 5, 5) and Pytorch Weight torch.Size([96, 96, 5, 5])
Keras Weight (96,) and Pytorch Weight torch.Size([96])
Keras Weight (96, 96, 5, 5) and Pytorch Weight torch.Size([96, 96, 5, 5])
Keras Weight (96,) and Pytorch Weight torch.Size([96])
Keras Weight (1, 96, 5, 5) and Pytorch Weight torch.Size([1, 96, 5, 5])
Keras Weight (1,) and Pytorch Weight torch.Size([1])
All work well and I run the pytorch model using patched image (into 256x256 blocks). Unfortunatly, the model fail to do the job (sheet music staff-line remover) and got a slightly padded output (few pixels to right/bottom). I suspect a problem with the padding='same' that introduce some padding divergence.
Here is the input image :
Here is the same image binarized :
Here is the output image :
I precise that the code of the patching to process the image work well and well tested before, so the horizontal and vertical lines are not a patching problem but the result of the bad processing of the patch inside the model.