Adds PyTorch NN (except Pooling and Conv) (keras-team#111)

* Add NN for pytorch backend

* Add PyTorch nn functions

* Add PyTorch nn functions

* Adds PyTorch Backend

* Adds PyTorch Backend

* Adds PyTorch nn backend

* Add PyTorch backend funcitons

* Add Torch NN Backend

* Add Torch NN Backend

* Add Torch NN Backend

keras_core/backend/torch/nn.py

@@ -0,0 +1,240 @@
+import torch
+import torch.nn.functional as tnn
+
+from keras_core.backend.config import epsilon
+
+
+def relu(x):
+    return tnn.relu(x)
+
+
+def relu6(x):
+    return tnn.relu6(x)
+
+
+def sigmoid(x):
+    return tnn.sigmoid(x)
+
+
+def tanh(x):
+    return tnn.tanh(x)
+
+
+def softplus(x):
+    return tnn.softplus(x)
+
+
+def softsign(x):
+    return tnn.soft_sign(x)
+
+
+def silu(x, beta=1.0):
+    return x * sigmoid(beta * x)
+
+
+def swish(x):
+    return silu(x, beta=1)
+
+
+def log_sigmoid(x):
+    return tnn.logsigmoid(x)
+
+
+def leaky_relu(x, negative_slope=0.2):
+    return tnn.leaky_relu(x, negative_slope=negative_slope)
+
+
+def hard_sigmoid(x):
+    return tnn.hardsigmoid(x)
+
+
+def elu(x):
+    return tnn.elu(x)
+
+
+def selu(x):
+    return tnn.selu(x)
+
+
+def gelu(x, approximate=True):
+    return tnn.gelu(x, approximate)
+
+
+def softmax(x, axis=None):
+    return tnn.softmax(x, dim=axis)
+
+
+def log_softmax(x, axis=-1):
+    return tnn.log_softmax(x, dim=axis)
+
+
+def max_pool(
+    inputs,
+    pool_size,
+    strides=None,
+    padding="valid",
+    data_format="channels_last",
+):
+    raise NotImplementedError(
+        "`max_pool` not yet implemented for PyTorch Backend"
+    )
+
+
+def average_pool(
+    inputs,
+    pool_size,
+    strides,
+    padding,
+    data_format="channels_last",
+):
+    raise NotImplementedError(
+        "`average_pool` not yet implemented for PyTorch Backend"
+    )
+
+
+def conv(
+    inputs,
+    kernel,
+    strides=1,
+    padding="valid",
+    data_format="channels_last",
+    dilation_rate=1,
+):
+    raise NotImplementedError("`conv` not yet implemented for PyTorch Backend")
+
+
+def depthwise_conv(
+    inputs,
+    kernel,
+    strides=1,
+    padding="valid",
+    data_format="channels_last",
+    dilation_rate=1,
+):
+    raise NotImplementedError(
+        "`depthwise_conv` not yet implemented for PyTorch Backend"
+    )
+
+
+def separable_conv(
+    inputs,
+    depthwise_kernel,
+    pointwise_kernel,
+    strides=1,
+    padding="valid",
+    data_format="channels_last",
+    dilation_rate=1,
+):
+    depthwise_conv_output = depthwise_conv(
+        inputs,
+        depthwise_kernel,
+        strides,
+        padding,
+        data_format,
+        dilation_rate,
+    )
+    return conv(
+        depthwise_conv_output,
+        pointwise_kernel,
+        strides=1,
+        padding="valid",
+        data_format=data_format,
+        dilation_rate=dilation_rate,
+    )
+
+
+def conv_transpose(
+    inputs,
+    kernel,
+    strides=1,
+    padding="valid",
+    output_padding=None,
+    data_format="channels_last",
+    dilation_rate=1,
+):
+    raise NotImplementedError(
+        "`conv_transpose` not yet implemented for PyTorch backend"
+    )
+
+
+def one_hot(x, num_classes, axis=-1):
+    if axis != -1 or axis != x.shape[-1]:
+        raise ValueError(
+            "`one_hot` is only implemented for last axis for PyTorch backend. "
+            f"`axis` arg value {axis} should be -1 or last axis of the input "
+            f"tensor with shape {x.shape}."
+        )
+    return tnn.one_hot(x, num_classes)
+
+
+def categorical_crossentropy(target, output, from_logits=False, axis=-1):
+    target = torch.as_tensor(target)
+    output = torch.as_tensor(output)
+
+    if target.shape != output.shape:
+        raise ValueError(
+            "Arguments `target` and `output` must have the same shape. "
+            "Received: "
+            f"target.shape={target.shape}, output.shape={output.shape}"
+        )
+    if len(target.shape) < 1:
+        raise ValueError(
+            "Arguments `target` and `output` must be at least rank 1. "
+            "Received: "
+            f"target.shape={target.shape}, output.shape={output.shape}"
+        )
+
+    if from_logits:
+        log_prob = tnn.log_softmax(output, dim=axis)
+    else:
+        output = output / torch.sum(output, dim=axis, keepdim=True)
+        output = torch.clip(output, epsilon(), 1.0 - epsilon())
+        log_prob = torch.log(output)
+    return -torch.sum(target * log_prob, dim=axis)
+
+
+def sparse_categorical_crossentropy(target, output, from_logits=False, axis=-1):
+    target = torch.as_tensor(target, dtype=torch.long)
+    output = torch.as_tensor(output)
+
+    if len(target.shape) == len(output.shape) and target.shape[-1] == 1:
+        target = torch.squeeze(target, dim=-1)
+
+    if len(output.shape) < 1:
+        raise ValueError(
+            "Argument `output` must be at least rank 1. "
+            "Received: "
+            f"output.shape={output.shape}"
+        )
+    if target.shape != output.shape[:-1]:
+        raise ValueError(
+            "Arguments `target` and `output` must have the same shape "
+            "up until the last dimension: "
+            f"target.shape={target.shape}, output.shape={output.shape}"
+        )
+    if from_logits:
+        log_prob = tnn.log_softmax(output, dim=axis)
+    else:
+        output = output / torch.sum(output, dim=axis, keepdim=True)
+        output = torch.clip(output, epsilon(), 1.0 - epsilon())
+        log_prob = torch.log(output)
+    target = one_hot(target, output.shape[axis], axis=axis)
+    return -torch.sum(target * log_prob, dim=axis)
+
+
+def binary_crossentropy(target, output, from_logits=False):
+    # TODO: `torch.as_tensor` has device arg. Need to think how to pass it.
+    target = torch.as_tensor(target)
+    output = torch.as_tensor(output)
+
+    if target.shape != output.shape:
+        raise ValueError(
+            "Arguments `target` and `output` must have the same shape. "
+            "Received: "
+            f"target.shape={target.shape}, output.shape={output.shape}"
+        )
+
+    if from_logits:
+        return tnn.binary_cross_entropy_with_logits(output, target)
+    else:
+        return tnn.binary_cross_entropy(output, target)

requirements.txt

@@ -11,5 +11,4 @@ pytest
 pandas
 absl-py
 requests
-h5py
-torch
+h5py