Hi, I am having some trouble running the third cell of the notebook. Error is:

ValueErrorTraceback (most recent call last)
<ipython-input-7-f7a1b7dd0c14> in <module>()
      7 
      8 #create model with default config ( == no skip_layer and 1000 units in the last layer)
----> 9 model = AlexNet(x, keep_prob, 1000, [])
     10 
     11 #define activation of last layer as score

/mnt/ilcompf6d0/user/txiao/DockerFiles/finetune_alexnet_with_tensorflow/alexnet.py in __init__(self, x, keep_prob, num_classes, skip_layer, weights_path)
     39 
     40     # Call the create function to build the computational graph of AlexNet
---> 41     self.create()
     42 
     43   def create(self):

/mnt/ilcompf6d0/user/txiao/DockerFiles/finetune_alexnet_with_tensorflow/alexnet.py in create(self)
     44 
     45     # 1st Layer: Conv (w ReLu) -> Pool -> Lrn
---> 46     conv1 = conv(self.X, 11, 11, 96, 4, 4, padding = 'VALID', name = 'conv1')
     47     pool1 = max_pool(conv1, 3, 3, 2, 2, padding = 'VALID', name = 'pool1')
     48     norm1 = lrn(pool1, 2, 2e-05, 0.75, name = 'norm1')

/mnt/ilcompf6d0/user/txiao/DockerFiles/finetune_alexnet_with_tensorflow/alexnet.py in conv(x, filter_height, filter_width, num_filters, stride_y, stride_x, name, padding, groups)
    131   with tf.variable_scope(name) as scope:
    132     # Create tf variables for the weights and biases of the conv layer
--> 133     weights = tf.get_variable('weights', shape = [filter_height, filter_width, input_channels/groups, num_filters])
    134     biases = tf.get_variable('biases', shape = [num_filters])
    135 

/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.pyc in get_variable(name, shape, dtype, initializer, regularizer, trainable, collections, caching_device, partitioner, validate_shape, use_resource, custom_getter)
   1047       collections=collections, caching_device=caching_device,
   1048       partitioner=partitioner, validate_shape=validate_shape,
-> 1049       use_resource=use_resource, custom_getter=custom_getter)
   1050 get_variable_or_local_docstring = (
   1051     """%s

/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.pyc in get_variable(self, var_store, name, shape, dtype, initializer, regularizer, trainable, collections, caching_device, partitioner, validate_shape, use_resource, custom_getter)
    946           collections=collections, caching_device=caching_device,
    947           partitioner=partitioner, validate_shape=validate_shape,
--> 948           use_resource=use_resource, custom_getter=custom_getter)
    949 
    950   def _get_partitioned_variable(self,

/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.pyc in get_variable(self, name, shape, dtype, initializer, regularizer, reuse, trainable, collections, caching_device, partitioner, validate_shape, use_resource, custom_getter)
    354           reuse=reuse, trainable=trainable, collections=collections,
    355           caching_device=caching_device, partitioner=partitioner,
--> 356           validate_shape=validate_shape, use_resource=use_resource)
    357 
    358   def _get_partitioned_variable(

/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.pyc in _true_getter(name, shape, dtype, initializer, regularizer, reuse, trainable, collections, caching_device, partitioner, validate_shape, use_resource)
    339           trainable=trainable, collections=collections,
    340           caching_device=caching_device, validate_shape=validate_shape,
--> 341           use_resource=use_resource)
    342 
    343     if custom_getter is not None:

/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variable_scope.pyc in _get_single_variable(self, name, shape, dtype, initializer, regularizer, partition_info, reuse, trainable, collections, caching_device, validate_shape, use_resource)
    651                          " Did you mean to set reuse=True in VarScope? "
    652                          "Originally defined at:\n\n%s" % (
--> 653                              name, "".join(traceback.format_list(tb))))
    654       found_var = self._vars[name]
    655       if not shape.is_compatible_with(found_var.get_shape()):

ValueError: Variable conv1/weights already exists, disallowed. Did you mean to set reuse=True in VarScope? Originally defined at:

  File "alexnet.py", line 133, in conv
    weights = tf.get_variable('weights', shape = [filter_height, filter_width, input_channels/groups, num_filters])
  File "alexnet.py", line 46, in create
    conv1 = conv(self.X, 11, 11, 96, 4, 4, padding = 'VALID', name = 'conv1')
  File "alexnet.py", line 41, in __init__
    self.create()

This error was in both the May 22 commit (TF 1.0) and the latest commit, TF 1.12rc0. I am on TF 1.1.0.

Cheers!

I have successfully trained alexnet for dogs vs cats dataset but when I run tensorboard I can only visualize this page:

The rest of the pages are inactive:

I have tensorflow 1.3.0, and python 3.6.3 and I am using windows 10. Do you have any idea what can cause the problem?

I run fintune.py,but I get some error.Could someone plz tell me what's the problem and how to modify the code to solve it? Thank you

the error is flowing: tensorflow.python.framework.errors_impl.InvalidArgumentError: Nan in summary histogram for: fc8/biases_0

hi, thanks for your sharing. I am using your code to finetune alexnet to classify images into blur or clear. When I run finetune.py with my 'train.txt' and 'val.txt', after dozens of train_batches, I get this error :-1: tensorflow.python.framework.errors_impl.InvalidArgumentError: Input shape axis 0 must equal 4, got shape [5] [[Node: unstack_1 = UnpackT=DT_INT32, axis=0, num=4]] [[Node: IteratorGetNext = IteratorGetNextoutput_shapes=[[?,227,227,3], [?,2]], output_types=[DT_FLOAT, DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"]]

I use python3.4, TF1.4.0

Hey, so I finished fine tuning my model, but I haven't really used tensorflow that much before. I understand how to load in the images I want to test on, but how do I use the checkpoint file to classify them?

Also thoughts on parameters for this dataset; 330 training images of Benign and Malignant, along with a validation and testing set of 40 images. I changed learning rate to 0.005 as well as batch size to 40, any other thoughts? I'm still training the last two layers, I'm not sure whether this is the right move. Thanks!!

Hello,

I tried to use the same code as posted for the AlexNet network and then used the validation for the same. I have tried the same code as in the post and am getting a very bad accuracy for some reason.

('Class name:', 'zebra', ' and Probability ', 0.6496317) ('Class name:', 'sea lion', ' and Probability ', 0.34226722) ('Class name:', 'llama', ' and Probability ', 0.36589968)

I used python and not Jupyter notebook for the validation so I printed the accuracy in the command line. I'm not sure why I have a different value. Can anyone give me some on this?

hi!It's very kind of you for sharing the repository.I used it to train the Dogs vs.Cats datasets and it worked.But when I used the ILSVRC2012 dataset to fine-tune the structure, I found some errors:

Traceback (most recent call last): File "/home/chengy/anaconda3/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1322, in _do_call return fn(*args) File "/home/chengy/anaconda3/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1307, in _run_fn options, feed_dict, fetch_list, target_list, run_metadata) File "/home/chengy/anaconda3/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1409, in _call_tf_sessionrun run_metadata) tensorflow.python.framework.errors_impl.InvalidArgumentError: Nan in summary histogram for: fc8/weights_0 [[Node: fc8/weights_0 = HistogramSummary[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"](fc8/weights_0/tag, fc8/weights/read/_41)]]

I've tried to replace the loss function to the following: loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=score, labels=tf.clip_by_value(y,1e-10,1.0))) and changed the learning rate to a very low number as 1e-4.But these operations did not make it better. I just use 20 categories of the whole 1 000 categories of Imagenet to fine-tune the model.Does anyone have met the same problem or have solutions to fix it?Thanks in advance.

Using the new ImageDataGenerator (June 15 update), I get lower accuracy than the old ImageDataGenerator (using cv2).

I have changed the new ImageDataGenerator very slightly (tf.image.decode_png => tf.image.decode_jpeg, VGG_MEAN => IMAGENET_MEAN), but still get significant difference in accuracy. To benchmark the two datagenerators, I am loading in the BVLC AlexNet weights and validating on the ImageNet validation set without shuffle.

With the old cv2 ImageDataGenerator, I get 55.72% Top-1 Accuracy and 79.08% Top-5 Accuracy. With the new TF ImageDataGenerator, I get 48.58% Top-1 Accuracy and 73.21% Top-5 Accuracy.

I suspect this has to do with how TF loads in images as opposed to CV2. Here is what the two different image processing steps look like:

Old CV2 ImageDataGenerator:

def next_batch(self, batch_size):
        """
        This function gets the next n ( = batch_size) images from the path list
        and labels and loads the images into them into memory 
        """
        # Get next batch of image (path) and labels
        paths = self.images[self.pointer:self.pointer + batch_size]
        labels = self.labels[self.pointer:self.pointer + batch_size]
        
        #update pointer
        self.pointer += batch_size
        
        # Read images
        images = np.ndarray([batch_size, self.scale_size[0], self.scale_size[1], 3])
        for i in range(len(paths)):
            img = cv2.imread(paths[i])
            
            #flip image at random if flag is selected
            if self.horizontal_flip and np.random.random() < 0.5:
                img = cv2.flip(img, 1)
            
            #rescale image
            img = cv2.resize(img, (self.scale_size[0], self.scale_size[1]))
            img = img.astype(np.float32)
            
            #subtract mean, which is np.array([104., 117., 124.])
            img -= self.mean
                                                                 
            images[i] = img

        # Expand labels to one hot encoding
        one_hot_labels = np.zeros((batch_size, self.n_classes))
        for i in range(len(labels)):
            one_hot_labels[i][labels[i]] = 1

        #return array of images and labels
        return images, one_hot_labels

New TF ImageDataGenerator:

def _parse_function_inference(self, filename, label):
        """Input parser for samples of the validation/test set."""
        # convert label number into one-hot-encoding
        one_hot = tf.one_hot(label, self.num_classes)

        # load and preprocess the image
        img_string = tf.read_file(filename)
        img_decoded = tf.image.decode_jpeg(img_string, channels=3)
        img_resized = tf.image.resize_images(img_decoded, [227, 227])
        
        IMAGENET_MEAN = tf.constant([104., 117., 124.], dtype=tf.float32)
        img_float = tf.to_float(img_resized)
        # RGB -> BGR
        img_bgr = img_float[:, :, ::-1]
        img_centered = tf.subtract(img_bgr, IMAGENET_MEAN)

        return img_centered, one_hot

I believe I am using the CV2 Datagenerator correctly as I can successfully finetune and call train_generator.next_batch(batch_size) within the training and validation loops with no issue. I built off of the finetune.py file you provided for the new TF Datagenerator (I use cpu device cpu:0 and then run the datagenerator init ops), and then get batches within the training/validation loops with img_batch, label_batch = sess.run(next_batch).

Any advice? Thanks a lot for providing and maintaining this proejct. Besides for this issue, it was super easy to work with and modify!

Hi, I tried reproducing the example keeping it as simple as possible. I have train.txt as: images/cat1.png 0 images/cat2.png 0 images/cat3.png 0 images/dog1.png 1 images/dog2.png 1 images/dog3.png 1 And test.txt as: images/cat4.png 0 images/dog4.png 1 when i run the finetune.py i get this error: Traceback (most recent call last): File "finetune.py", line 163, in test_acc /= test_count ZeroDivisionError: float division by zero Tried debugging the error and found that val_batches_per_epoch is 0 and the inner body isn't executing.

Thanks for your contribution! when i reuse your code, and i want to train fc6, fc7,fc8 , but the result: Cross entropy = nan if only train fc7,fc8 , it works well. also i have test earlier layers ,, like conv1,conv2.., it still appear nan do have have any suggestion to help me fix this?

I get the following error from datagenerator.py: error: (-215) ssize.width > 0 && ssize.height > 0 in function resize

finetune_alexnet_with_tensorflow-master/datagenerator.py", line 150, in next_batch img = cv2.resize(img, (self.scale_size[0], self.scale_size[0]))

I debugged code as follows:

			img = cv2.imread(paths[i])
			print(img);
			if img is None:
				print (paths[i] + " : fail to read")
			else:
				print("Image is read");
			exit();

Error suggests that image is not read while in a separate program i verified that image is read:

import numpy as np
import cv2

img = cv2.imread('images/ld2.png')
cv2.imwrite('messigray.png',img) 

Followed this solution but couldn't work.

While I test the scrip of fintune.py on TensorFlow 1.5, the system throws the errorn as follows. I list the information inlucuding ValueError, possible questionable code and traceback as follows. Appreciate your help in advance.

1. General Message

ValueError: Dimension size must be evenly divisible by 2 but is 1 Number of ways to split should evenly divide the split dimension for 'split_1' (op: 'Split') with input shapes: [], [5,5,2,1] and with computed input

2. Code

It targets to the line of code as follows.

alexnet.py

weight_groups = tf.split(value=weights, num_or_size_splits=group, axis=3) 

3. Detailed Error Message:

$ python finetune.py --conv=4 --dropout_rate=0.03

Traceback (most recent call last): File "/home/nano/.virtualenvs/win/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py", line 1607, in _create_c_op c_op = c_api.TF_FinishOperation(op_desc) tensorflow.python.framework.errors_impl.InvalidArgumentError: Dimension size must be evenly divisible by 2 but is 1 Number of ways to split should evenly divide the split dimension for 'split_1' (op: 'Split') with input shapes: [], [5,5,2,1] and with computed input tensors: input[0] = <3>.

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File "finetune.py", line 97, in model = AlexNet(x, keep_prob, num_classes, train_layers) File "/home/nano/Documents/finetune_alexnet_with_tf/alexnet.py", line 56, in init self.create() File "/home/nano/Documents/finetune_alexnet_with_tf/alexnet.py", line 69, in create conv2 = conv(pool1, 5, 5, 256, 1, 1, name='conv2', group=2) File "/home/nano/Documents/finetune_alexnet_with_tf/alexnet.py", line 159, in conv weight_groups = tf.split(value=weights, num_or_size_splits=group, axis=3) File "/home/nano/.virtualenvs/win/lib/python3.6/site-packages/tensorflow_core/python/ops/array_ops.py", line 1684, in split axis=axis, num_split=num_or_size_splits, value=value, name=name) File "/home/nano/.virtualenvs/win/lib/python3.6/site-packages/tensorflow_core/python/ops/gen_array_ops.py", line 9898, in split "Split", split_dim=axis, value=value, num_split=num_split, name=name) File "/home/nano/.virtualenvs/win/lib/python3.6/site-packages/tensorflow_core/python/framework/op_def_library.py", line 794, in _apply_op_helper op_def=op_def) File "/home/nano/.virtualenvs/win/lib/python3.6/site-packages/tensorflow_core/python/util/deprecation.py", line 507, in new_func return func(*args, **kwargs) File "/home/nano/.virtualenvs/win/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py", line 3357, in create_op attrs, op_def, compute_device) File "/home/nano/.virtualenvs/win/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py", line 3426, in _create_op_internal op_def=op_def) File "/home/nano/.virtualenvs/win/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py", line 1770, in init control_input_ops) File "/home/nano/.virtualenvs/win/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py", line 1610, in _create_c_op raise ValueError(str(e)) ValueError: Dimension size must be evenly divisible by 2 but is 1 Number of ways to split should evenly divide the split dimension for 'split_1' (op: 'Split') with input shapes: [], [5,5,2,1] and with computed input

I run the script of finetune.py in both TensorFlow 1.5 and TensorFlow 2.1. After solving many issues, I found out the shape incompatibility issue in the script of alexnet.py. Please help fix issue at your convenience. Appreciate your help in advance.

I get to know that there is the scope conflict between the shape and the conv argument. tf.variable_scope() usually defines global variables in the with context. It influences other related variables. For instance, shape= [filter_height, filter_width, input_channels//groups, num_filters], it denotes [11,11,3,96] in the Conv1; in contrast, Conv1 includes the arguments: 11, 11, 96, 4, 4.

1. Error Message

ValueError: The initial value's shape (()) is not compatible with the explicitly supplied shape argument ([11, 11, 3, 96]).

2. Attempted Changes

I tried to make the following changes.

1). Change the order of either the shape or the Conv1 arguments, for instance, shape=[11,11,96,3] or conv(self.X, 11, 11, 4, 4, 96, name='conv1', padding='VALID')

2). Change the name of shape

kernel_shape= [filter_height, filter_width, input_channels//groups, num_filters]

3). Delete the shape and keep the argument.

[filter_height, filter_width, input_channels//groups, num_filters]

However, the following error varieties have still been persisted.

ValueError: The initial value's shape (()) is not compatible with the explicitly supplied shape argument ([11, 11, 3, 4]).

ValueError: Shapes must be equal rank, but are 4 and 0 for 'conv1/Variable/Assign' (op: 'Assign') with input shapes: [11,11,3,4], [].

ValueError: Shapes must be equal rank, but are 4 and 0 for 'conv1/Variable/Assign' (op: 'Assign') with input shapes: [11,11,3,96], [].

It is definitely the critical issue of "shape" in the second snippet. But I have not yet figured a way to solve the issues.

3. Snippets

1st snippet.

class AlexNet(object):
    .........
    def create(self):
        """Create the network graph."""
        # 1st Layer: Conv (w ReLu) -> Lrn -> Pool
        conv1 = conv(self.X, 11, 11, 4, 4, 96, name='conv1', padding='VALID')
        norm1 = lrn(conv1, 2, 2e-05, 0.75, name='norm1')
        pool1 = max_pool(norm1, 3, 3, 2, 2, name='pool1', padding='VALID')

2nd snippet:

def conv(x, filter_height, filter_width, stride_y, stride_x, num_filters, name,
    padding='SAME', groups=1):
    .........
    with tf.compat.v1.variable_scope(name) as scope:
        weights = tf.Variable('weights', shape=[filter_height, 
                                                                      filter_width, 
                                                                      input_channels//groups,
                                                                      num_filters])
        biases = tf.Variable('biases', shape=[num_filters])

4. Detailed error message

$ python finetune.py

Traceback (most recent call last): File "finetune.py", line 91, in model = AlexNet(x, keep_prob, num_classes, train_layers) File "/home/mike/Documents/finetune_alexnet_with_tf/alexnet.py", line 56, in init self.create() File "/home/mike/Documents/finetune_alexnet_with_tf/alexnet.py", line 61, in create conv1 = conv(self.X, 11, 11, 96, 4, 4, padding='VALID', name='conv1') File "/home/mike/Documents/finetune_alexnet_with_tf/alexnet.py", line 147, in conv num_filters]) File "/home/mike/miniconda3/lib/python3.7/site-packages/tensorflow_core/python/ops/variables.py", line 260, in call return cls._variable_v2_call(*args, **kwargs) File "/home/mike/miniconda3/lib/python3.7/site-packages/tensorflow_core/python/ops/variables.py", line 254, in _variable_v2_call shape=shape) File "/home/mike/miniconda3/lib/python3.7/site-packages/tensorflow_core/python/ops/variables.py", line 235, in previous_getter = lambda **kws: default_variable_creator_v2(None, **kws) File "/home/mike/miniconda3/lib/python3.7/site-packages/tensorflow_core/python/ops/variable_scope.py", line 2645, in default_variable_creator_v2 shape=shape) File "/home/mike/miniconda3/lib/python3.7/site-packages/tensorflow_core/python/ops/variables.py", line 262, in call return super(VariableMetaclass, cls).call(*args, **kwargs) File "/home/mike/miniconda3/lib/python3.7/site-packages/tensorflow_core/python/ops/resource_variable_ops.py", line 1411, in init distribute_strategy=distribute_strategy) File "/home/mike/miniconda3/lib/python3.7/site-packages/tensorflow_core/python/ops/resource_variable_ops.py", line 1549, in _init_from_args (initial_value.shape, shape)) ValueError: The initial value's shape (()) is not compatible with the explicitly supplied shape argument ([11, 11, 3, 96]).

2019-05-28 10:45:45.833285 Validation Accuracy = 0.2188 2019-05-28 10:45:45.833390 Saving checkpoint of model... 2019-05-28 10:45:47.135384 Model checkpoint saved at checkpoints/model_epoch1.ckpt 2019-05-28 10:45:47.464044 Start validation 2019-05-28 10:45:50.295000 Validation Accuracy = 0.2188 2019-05-28 10:45:50.295116 Saving checkpoint of model... 2019-05-28 10:45:51.710815 Model checkpoint saved at checkpoints/model_epoch1.ckpt 2019-05-28 10:45:52.035434 Start validation 2019-05-28 10:45:54.884074 Validation Accuracy = 0.2188 2019-05-28 10:45:54.884180 Saving checkpoint of model... 2019-05-28 10:45:56.231970 Model checkpoint saved at checkpoints/model_epoch1.ckpt 2019-05-28 10:45:56.560833 Start validation 2019-05-28 10:45:59.407273 Validation Accuracy = 0.2188 2019-05-28 10:45:59.407380 Saving checkpoint of model... 2019-05-28 10:46:01.239952 Model checkpoint saved at checkpoints/model_epoch1.ckpt 2019-05-28 10:46:01.566239 Start validation 2019-05-28 10:46:04.391418 Validation Accuracy = 0.2188 2019-05-28 10:46:04.391538 Saving checkpoint of model... 2019-05-28 10:46:05.695854 Model checkpoint saved at checkpoints/model_epoch1.ckpt 2019-05-28 10:46:06.340697 Start validation 2019-05-28 10:46:09.281800 Validation Accuracy = 0.2188 2019-05-28 10:46:09.281909 Saving checkpoint of model... 2019-05-28 10:46:10.555316 Model checkpoint saved at checkpoints/model_epoch1.ckpt 2019-05-28 10:46:10.904200 Start validation 2019-05-28 10:46:13.926193 Validation Accuracy = 0.2188 2019-05-28 10:46:13.926291 Saving checkpoint of model...

As you see,accuracy is 0.2188 and not change.What can I do to fit this condition

Hello, I use python 3.6 + tensorflow 1.9.0 to do a classification task by finetuning Alexnet. In kratzert's srouce code, the range of data is [0,255] and the image is substracted by meanfile[103,116,123]. However, in my case, I perform better accrucy when the input images divide by 255 after substracted mean(distribution is [-0.5,0.5]). If the input images did not divide by 255 (distrubution is [-128, 128]), the results is bad and more easily to suffer Nan problem in summary operation when the learning rate is higher then 0.001. The question is should I scale the image ?