update backend

Author: Laicheng0830

docs/modules/cost.rst

@@ -11,7 +11,7 @@ we can. So we encourage you to use TensorFlow's function, , see `TensorFlow API
 
 .. autosummary::
 
-   cross_entropy
+   softmax_cross_entropy_with_logits
    sigmoid_cross_entropy
    binary_cross_entropy
    mean_squared_error
@@ -28,12 +28,11 @@ we can. So we encourage you to use TensorFlow's function, , see `TensorFlow API
    maxnorm_regularizer
    maxnorm_o_regularizer
    maxnorm_i_regularizer
-   huber_loss
 
 
 Softmax cross entropy
 ----------------------
-.. autofunction:: cross_entropy
+.. autofunction:: softmax_cross_entropy_with_logits
 
 Sigmoid cross entropy
 ----------------------
@@ -94,7 +93,3 @@ Special
 .. autofunction:: lo_regularizer
 .. autofunction:: maxnorm_o_regularizer
 .. autofunction:: maxnorm_i_regularizer
-
-Huber Loss
-^^^^^^^^^^
-.. autofunction:: huber_loss

docs/user/get_start_model.rst

@@ -120,7 +120,7 @@ Switching train/test modes
   ... # testing code here
   
   # method 2: Using packaged training modules
-  model = tl.models.Model(network=MLP, loss_fn=tl.cost.cross_entropy, optimizer=optimizer)
+  model = tl.models.Model(network=MLP, loss_fn=tl.cost.softmax_cross_entropy_with_logits, optimizer=optimizer)
   model.train(n_epoch=n_epoch, train_dataset=train_ds)
 
 Reuse weights
@@ -192,13 +192,13 @@ Save weights only
   MLP.save_weights('./model_weights.npz') # by default, file will be in hdf5 format
   MLP.load_weights('./model_weights.npz')
 
-Save model architecture and weights (optional)
+Save model weights (optional)
 -----------------------------------------------
 
 .. code-block:: python
 
   # When using packaged training modules. Saving and loading the model can be done as follows
-  model = tl.models.Model(network=MLP, loss_fn=tl.cost.cross_entropy, optimizer=optimizer)
+  model = tl.models.Model(network=MLP, loss_fn=tl.cost.softmax_cross_entropy_with_logits, optimizer=optimizer)
   model.train(n_epoch=n_epoch, train_dataset=train_ds)
   model.save_weights('./model.npz', format='npz_dict')
   model.load_weights('./model.npz', format='npz_dict')

tensorlayer/backend/ops/load_backend.py

@@ -7,6 +7,7 @@
 
 BACKEND = 'tensorflow'
 # BACKEND = 'mindspore'
+# BACKEND = 'paddle'
 
 # Check for backend.json files
 tl_backend_dir = os.path.expanduser('~')

tensorlayer/backend/ops/paddle_nn.py

@@ -431,16 +431,22 @@ class Conv2D(object):
 
     def __init__(self, strides, padding, data_format='NHWC', dilations=None, out_channel=None, k_size=None):
         self.data_format, self.padding = preprocess_2d_format(data_format, padding)
-        self.ksize = k_size[0]
         if self.data_format is 'NHWC':
-            self.dg_stride = strides[1]
-            self.dg_dilation = dilations[1]
+            self._stride = (strides[1], strides[2])
+            self._dilation = (dilations[1], dilations[2])
         elif self.data_format is 'NCHW':
-            self.dg_stride = strides[2]
-            self.dg_dilation = dilations[2]
+            self._stride = (strides[2], strides[3])
+            self._dilation = (dilations[2], dilations[3])
+
 
     def __call__(self, inputs, filters):
-        raise NotImplementedError
+        outputs = F.conv2d(x=inputs,
+                           weight=filters,
+                           stride=self._stride,
+                           dilation=self._dilation,
+                           padding=self.padding,
+                           data_format=self.data_format)
+        return outputs
 
 
 def conv2d(input, filters, strides, padding, data_format='NCHW', dilations=None):
@@ -468,7 +474,20 @@ def conv2d(input, filters, strides, padding, data_format='NCHW', dilations=None)
     -------
         A Tensor. Has the same type as input.
     """
-    raise NotImplementedError
+    data_format, padding = preprocess_2d_format(data_format, padding)
+    if data_format is 'NHWC':
+        _stride = (strides[1], strides[2])
+        _dilation = (dilations[1], dilations[2])
+    elif data_format is 'NCHW':
+        _stride = (strides[2], strides[3])
+        _dilation = (dilations[2], dilations[3])
+    outputs = F.conv2d(x=input,
+                       weight=filters,
+                       stride=_stride,
+                       dilation=_dilation,
+                       padding=padding,
+                       data_format=data_format)
+    return outputs
 
 
 class Conv3D(object):
@@ -577,10 +596,18 @@ class MaxPool(object):
     def __init__(self, ksize, strides, padding, data_format=None):
         self.data_format, self.padding = preprocess_2d_format(data_format, padding)
         self.ksize = ksize
-        self.strides = strides
+        if self.data_format is 'NHWC':
+            self._stride = (strides[1], strides[2])
+        elif self.data_format is 'NCHW':
+            self._stride = (strides[2], strides[3])
 
     def __call__(self, inputs):
-        raise NotImplementedError
+        outputs = F.max_pool2d(x=inputs,
+                               kernel_size=self.ksize,
+                               stride=self._stride,
+                               padding=self.padding,
+                               data_format=self.data_format)
+        return outputs
 
 
 def max_pool(input, ksize, strides, padding, data_format=None):
@@ -951,7 +978,7 @@ def __init__(self):
         pass
 
     def __call__(self, *args, **kwargs):
-        raise NotImplementedError
+        pd.nn.BatchNorm2D
 
 
 class GroupConv2D(object):

tensorlayer/initializers/__init__.py

@@ -5,7 +5,7 @@
 #     'Initializer', 'Zeros', 'Ones', 'Constant', 'RandomUniform', 'RandomNormal', 'TruncatedNormal',
 #     'deconv2d_bilinear_upsampling_initializer', 'He_Normal'
 # ]
-
+from .load_initializers_backend import Initializer
 from .load_initializers_backend import Zeros
 from .load_initializers_backend import Ones
 from .load_initializers_backend import Constant

tensorlayer/layers/__init__.py

@@ -18,7 +18,7 @@
 from .padding import *
 from .pooling import *
 from .quantize import *
-# from .recurrent import *
+from .recurrent import *
 from .scale import *
 from .shape import *
 from .spatial_transformer import *

tensorlayer/layers/convolution/simplified_conv.py

@@ -188,7 +188,7 @@ class Conv2d(Module):
     --------
     With TensorLayer
 
-    >>> net = tl.layers.Input([8, 3, 400, 400], name='input')
+    >>> net = tl.layers.Input([8, 400, 400, 3], name='input')
     >>> conv2d = tl.layers.Conv2d(n_filter=32, filter_size=(3, 3), strides=(2, 2), b_init=None, in_channels=3, name='conv2d_1')
     >>> print(conv2d)
     >>> tensor = tl.layers.Conv2d(n_filter=32, filter_size=(3, 3), strides=(2, 2), act=tl.ReLU, name='conv2d_2')(net)
@@ -630,7 +630,7 @@ class DeConv2d(Module):
     --------
     With TensorLayer
 
-    >>> net = tl.layers.Input([8, 3, 400, 400], name='input')
+    >>> net = tl.layers.Input([8, 400, 400, 3], name='input')
     >>> conv2d_transpose = tl.layers.DeConv2d(n_filter=32, filter_size=(3, 3), strides=(2, 2), b_init=None, in_channels=3, name='conv2d_transpose_1')
     >>> print(conv2d_transpose)
     >>> tensor = tl.layers.DeConv2d(n_filter=32, filter_size=(3, 3), strides=(2, 2), act=tl.ReLU, name='conv2d_transpose_2')(net)

tensorlayer/layers/core/core_paddle.py

@@ -191,6 +191,18 @@ def __call__(self, *inputs, **kwargs):
             return outputs
 
     def _get_weights(self, var_name, shape, init=None, trainable=True, transposed=None):
+        # TODO 2D mindspore weights shape : [out_channel, in_channel, kernel_h, kernel_w]
+        # TODO 2D mindspore transposed shape [in_channel, out_channel, kernel_h, kernel_w]
+        if len(shape) == 3:
+            shape = shape[::-1]
+        if len(shape) == 4:
+            if not transposed and self.data_format == 'NHWC':
+                shape = (shape[3], shape[0], shape[1], shape[2])
+            else:
+                shape = (shape[3], shape[2], shape[0], shape[1])
+        if len(shape) == 5:
+            shape = (shape[4], shape[3], shape[0], shape[1], shape[2])
+
         if var_name in ["filters", "weights"]:
             w_tmp = self.create_parameter(shape=shape, attr=init, is_bias=False)
         elif var_name in ["biases"]: