waittim
commented
4 years ago I tried to convert the Darknet model to be a Keras model. I think the code from repo keras-yolo3 didn't consider the group in Conv2d because it bases on the old version of Keras which doesn't support the group operation. Maybe you can try this code, which is based on the latest TF version, to convert Conv2d layer.
def conv(self, block):
'''在读取darknet的yolov3.weights文件时，顺序是
1 - bias；
2 - 如果有bn，则接着读取三个scale，mean，var
3 - 读取权重
'''
# Darknet serializes convolutional weights as:
# [bias/beta, [gamma, mean, variance], conv_weights]
self.count[0] += 1
# read conv block
filters = int(block['filters'])
size = int(block['size'])
stride = int(block['stride'])
pad = int(block['pad'])
activation = block['activation']
groups = int(block['groups']) if 'groups' in block else 1
padding = 'same' if pad == 1 and stride == 1 else 'valid'
batch_normalize = 'batch_normalize' in block
prev_layer_shape = K.int_shape(self.prev_layer)
weights_shape = (size, size, int(prev_layer_shape[-1]/groups), filters)
darknet_w_shape = (filters, weights_shape[2], size, size)
weights_size = np.product(weights_shape)
print('+',self.count[0],'conv2d',
'bn' if batch_normalize else ' ',
activation,
weights_shape)
# 读取滤波器个偏置
conv_bias = self.weight_loader.parser_buffer(
shape=(filters,),
dtype='float32',
buffer_size=filters*4)
# 如果有bn，则接着读取滤波器个scale，mean，var
if batch_normalize:
bn_weight_list = self.bn(filters, conv_bias)
# 读取权重
conv_weights = self.weight_loader.parser_buffer(
shape=darknet_w_shape,
dtype='float32',
buffer_size=weights_size*4)
# DarkNet conv_weights are serialized Caffe-style:
# (out_dim, in_dim, height, width)
# We would like to set these to Tensorflow order:
# (height, width, in_dim, out_dim)
conv_weights = np.transpose(conv_weights, [2, 3, 1, 0])
conv_weights = [conv_weights] if batch_normalize else \
[conv_weights, conv_bias]
act_fn = None
if activation == 'leaky':
pass
elif activation != 'linear':
raise
if stride > 1:
self.prev_layer = ZeroPadding2D(((1,0),(1,0)))(self.prev_layer)
conv_layer = (Conv2D(
filters, (size, size),
strides=(stride, stride),
kernel_regularizer=l2(self.weight_decay),
use_bias=not batch_normalize,
weights=conv_weights,
activation=act_fn,
groups=groups,
padding=padding))(self.prev_layer)
if batch_normalize:
conv_layer = BatchNormalization(weights=bn_weight_list)(conv_layer)
self.prev_layer = conv_layer
if activation == 'linear':
self.all_layers.append(self.prev_layer)
elif activation == 'leaky':
act_layer = LeakyReLU(alpha=0.1)(self.prev_layer)
self.prev_layer = act_layer
self.all_layers.append(act_layer)It dealt with buffer is too small for requested array. However, when I was training the keras model, I met some other errors. So I abandoned this approach. If you can find a usable approach. Please let me know.
Thank you!
When I convert the weights file to other model files, it always report error: buffer is too small for requested array. I check the .cfg file,then I found there some 'groups' vars in [convolutional] layer, maybe it reads more weights than GroupConv2D needs and make the weights not enough to use, but I don't know how to split groups in yolo-fastest(filters=groups?). Is the error caused by this problem, or others, and how to solve this error?