tensorboard-x refactor

Author: lanpa

demo.py

@@ -3,7 +3,10 @@
 import numpy as np
 import torchvision.models as models
 from torchvision import datasets
-from tensorboard import SummaryWriter
+from tensorboardX import SummaryWriter
+#import skimage
+#from skimage import data, io
+
 resnet18 = models.resnet18(False)
 writer = SummaryWriter()
 sample_rate = 44100
@@ -12,20 +15,22 @@
 for n_iter in range(100):
     s1 = torch.rand(1) # value to keep
     s2 = torch.rand(1)
-    writer.add_scalar('data/scalar1', s1[0], n_iter) #data grouping by `slash`
-    writer.add_scalar('data/scalar2', s2[0], n_iter)
+    writer.add_scalar('data/scalar1', s1[0], n_iter) # data grouping by `slash`
+    writer.add_scalar('data/scalar2', s2, n_iter) # passing Tensor is OK!
     x = torch.rand(32, 3, 64, 64) # output from network
     if n_iter%10==0:
         x = vutils.make_grid(x, normalize=True, scale_each=True)   
-        writer.add_image('Image', x, n_iter)
+        writer.add_image('Image', x, n_iter) # Tensor
+        #writer.add_image('astronaut', skimage.data.astronaut(), n_iter) # numpy
+        #writer.add_image('imread', skimage.io.imread('screenshots/audio.png'), n_iter) # numpy
         x = torch.zeros(sample_rate*2)
         for i in range(x.size(0)):
             x[i] = np.cos(freqs[n_iter//10]*np.pi*float(i)/float(sample_rate)) # sound amplitude should in [-1, 1]
         writer.add_audio('myAudio', x, n_iter)
         writer.add_text('Text', 'text logged at step:'+str(n_iter), n_iter)
         writer.add_text('another Text', 'another text logged at step:'+str(n_iter), n_iter)
         for name, param in resnet18.named_parameters():
-            writer.add_histogram(name, param.clone().cpu().data.numpy(), n_iter)
+            writer.add_histogram(name, param, n_iter)
 
 dataset = datasets.MNIST('mnist', train=False, download=True)
 images = dataset.test_data[:100].float()

demo_embedding.py

@@ -3,7 +3,7 @@
 import torch.nn.functional as F
 import os
 from torch.autograd.variable import Variable
-from tensorboard import SummaryWriter
+from tensorboardX import SummaryWriter
 from torch.utils.data import TensorDataset, DataLoader
 
 #EMBEDDING VISUALIZATION FOR A TWO-CLASSES PROBLEM

demo_graph.py

@@ -4,7 +4,7 @@
 import numpy as np
 import torch.nn.functional as F
 import torchvision.models as models
-from tensorboard import SummaryWriter
+from tensorboardX import SummaryWriter
 
 class Mnist(nn.Module):
     def __init__(self):

setup.py

@@ -18,9 +18,9 @@
 ]
 
 setup(
-    name='tensorboard-pytorch',
-    version='0.7.1',
-    description='Log TensorBoard events with pytorch',
+    name='tensorboardX',
+    version='0.7.5',
+    description='TensorBoardX lets you watch Tensors Flow without Tensorflow',
     long_description= history,
     author='Tzu-Wei Huang',
     author_email='huang.dexter@gmail.com',

tensorboardX/summary.py

@@ -84,9 +84,7 @@ def scalar(name, scalar, collections=None):
       ValueError: If tensor has the wrong shape or type.
     """
     name = _clean_tag(name)
-    if not isinstance(scalar, float):
-        # try conversion, if failed then need handle by user.
-        scalar = float(scalar)
+    scalar = float(scalar)
     return Summary(value=[Summary.Value(tag=name, simple_value=scalar)])
 
 
@@ -182,8 +180,6 @@ def make_image(tensor):
                          encoded_image_string=image_string)
 
 def audio(tag, tensor, sample_rate=44100):
-  tensor = tensor.squeeze()
-  assert tensor.dim()==1, 'input tensor should be 1 dimensional.'
   tensor_list = [int(32767.0*x) for x in tensor]
   import io
   import wave
@@ -212,69 +208,3 @@ def text(tag, text):
   tensor = TensorProto(dtype='DT_STRING', string_val=[text.encode(encoding='utf_8')])
   return Summary(value=[Summary.Value(node_name=tag, metadata=smd, tensor=tensor)])
 
-
-'''
-def merge(inputs, collections=None, name=None):
-  # pylint: disable=line-too-long
-  """Merges summaries.
-  This op creates a
-  [`Summary`](https://www.tensorflow.org/code/tensorflow/core/framework/summary.proto)
-  protocol buffer that contains the union of all the values in the input
-  summaries.
-  When the Op is run, it reports an `InvalidArgument` error if multiple values
-  in the summaries to merge use the same tag.
-  Args:
-    inputs: A list of `string` `Tensor` objects containing serialized `Summary`
-      protocol buffers.
-    collections: Optional list of graph collections keys. The new summary op is
-      added to these collections. Defaults to `[GraphKeys.SUMMARIES]`.
-    name: A name for the operation (optional).
-  Returns:
-    A scalar `Tensor` of type `string`. The serialized `Summary` protocol
-    buffer resulting from the merging.
-  """
-  # pylint: enable=line-too-long
-  name = _clean_tag(name)
-  with _ops.name_scope(name, 'Merge', inputs):
-    # pylint: disable=protected-access
-    val = _gen_logging_ops._merge_summary(inputs=inputs, name=name)
-    _collect(val, collections, [])
-  return val
-
-
-def merge_all(key=_ops.GraphKeys.SUMMARIES):
-  """Merges all summaries collected in the default graph.
-  Args:
-    key: `GraphKey` used to collect the summaries.  Defaults to
-      `GraphKeys.SUMMARIES`.
-  Returns:
-    If no summaries were collected, returns None.  Otherwise returns a scalar
-    `Tensor` of type `string` containing the serialized `Summary` protocol
-    buffer resulting from the merging.
-  """
-  summary_ops = _ops.get_collection(key)
-  if not summary_ops:
-    return None
-  else:
-    return merge(summary_ops)
-
-
-def get_summary_description(node_def):
-  """Given a TensorSummary node_def, retrieve its SummaryDescription.
-  When a Summary op is instantiated, a SummaryDescription of associated
-  metadata is stored in its NodeDef. This method retrieves the description.
-  Args:
-    node_def: the node_def_pb2.NodeDef of a TensorSummary op
-  Returns:
-    a summary_pb2.SummaryDescription
-  Raises:
-    ValueError: if the node is not a summary op.
-  """
-
-  if node_def.op != 'TensorSummary':
-    raise ValueError("Can't get_summary_description on %s" % node_def.op)
-  description_str = _compat.as_str_any(node_def.attr['description'].s)
-  summary_description = SummaryDescription()
-  _json_format.Parse(description_str, summary_description)
-  return summary_description
-'''

tensorboardX/writer.py

@@ -28,7 +28,7 @@
 from .summary import scalar, histogram, image, audio, text
 from .graph import graph
 from .embedding import make_mat, make_sprite, make_tsv, append_pbtxt
-
+from .x2num import makenp
 
 class SummaryToEventTransformer(object):
     """Abstractly implements the SummaryWriter API.
@@ -248,6 +248,8 @@ def add_scalar(self, tag, scalar_value, global_step=None):
             global_step (int): Global step value to record
 
         """
+        scalar_value = makenp(scalar_value)
+        assert(scalar_value.squeeze().ndim==0), 'input of add_scalar should be 0D'
         self.file_writer.add_summary(scalar(tag, scalar_value), global_step)
 
     def add_histogram(self, tag, values, global_step=None, bins='tensorflow'):
@@ -262,6 +264,7 @@ def add_histogram(self, tag, values, global_step=None, bins='tensorflow'):
         """
         if bins=='tensorflow':
             bins = self.default_bins
+        values = makenp(values)
         self.file_writer.add_summary(histogram(tag, values, bins), global_step)
 
     def add_image(self, tag, img_tensor, global_step=None):
@@ -274,6 +277,7 @@ def add_image(self, tag, img_tensor, global_step=None):
         Shape:
             img_tensor: :math:`(3, H, W)`. Use ``torchvision.utils.make_grid()`` to prepare it is a good idea.
         """
+        img_tensor = makenp(img_tensor, 'IMG')
         self.file_writer.add_summary(image(tag, img_tensor), global_step)
     def add_audio(self, tag, snd_tensor, global_step=None):
         """Add audio data to summary.
@@ -286,6 +290,10 @@ def add_audio(self, tag, snd_tensor, global_step=None):
         Shape:
             snd_tensor: :math:`(1, L)`. The values should between [-1, 1]. The sample rate is currently fixed at 44100 KHz.
         """
+        snd_tensor = makenp(snd_tensor)
+        snd_tensor = snd_tensor.squeeze()
+        assert(snd_tensor.ndim==1), 'input tensor should be 1 dimensional.'
+        
         self.file_writer.add_summary(audio(tag, snd_tensor), global_step)
     def add_text(self, tag, text_string, global_step=None):
         """Add text data to summary.

tensorboardX/x2num.py

@@ -1,22 +1,44 @@
-def makenp(x, **kwargs):
+# DO NOT alter/distruct/free input object !
+
+import numpy as np
+
+def makenp(x, modality=None):
     # if already numpy, return
-    pass
+    if isinstance(x, np.ndarray):
+        if modality == 'IMG' and x.dtype == np.uint8:
+            return x.astype(np.float32)/255.0
+        return x
+    if isinstance(x, float) or isinstance(x, int):
+        return np.array([x])
+    if 'torch' in str(type(x)):
+        return pytorch_np(x, modality)
 
-def pytorch_np():
-    pass
+def pytorch_np(x, modality):
+    import torch
+    if isinstance(x, torch.autograd.variable.Variable):
+        x = x.data
+    x = x.cpu()
+    assert isinstance(x, torch.Tensor), 'invalid input type'
+    if modality == 'IMG':
+        assert x.dim()<4 and x.dim()>1, 'input tensor should be 3D for color image or 2D for gray image.'
+        if x.dim()==2:
+            x = x.unsqueeze(0)
+        x = x.permute(1,2,0) #CHW to HWC
+    return x.numpy()
 
-def torch_np():
-    pass
 
-def theano_np():
+def theano_np(x):
+    import theano
     pass
 
-def caffe2_np():
+def caffe2_np(x):
     pass
 
-def mxnet_np():
+def mxnet_np(x):
     pass
 
-def chainer_np():
+def chainer_np(x):
+    import chainer
+    #x = chainer.cuda.to_cpu(x.data)
     pass
 

tests/test_pytorch_np.py

@@ -0,0 +1,24 @@
+from tensorboardX import x2num
+import torch
+import numpy as np
+shapes = [(3, 10, 10), (1, ), (1, 2, 3, 4, 5)]
+
+def test_pytorch_np():
+
+    for shape in shapes:
+        # regular tensor
+        assert(isinstance(x2num.makenp(torch.Tensor(*shape)), np.ndarray))
+
+        # CUDA tensor
+        assert(isinstance(x2num.makenp(torch.Tensor(*shape).cuda()), np.ndarray))
+
+        # regular variable
+        assert(isinstance(x2num.makenp(torch.autograd.variable.Variable(torch.Tensor((*shape)))), np.ndarray))
+
+        # CUDA variable
+        assert(isinstance(x2num.makenp(torch.autograd.variable.Variable(torch.Tensor((*shape))).cuda()), np.ndarray))
+
+    # python primitive type
+    assert(isinstance(x2num.makenp(0), np.ndarray))
+    assert(isinstance(x2num.makenp(0.1), np.ndarray))
+