Fix loss equations
This commit is contained in:
parent
87ec9a1b0f
commit
5fb6482ded
@ -1679,7 +1679,7 @@
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"<img src=\"images/layer1.png\" alt=\"Activations of early layers of a CNN\" width=\"300\" caption=\"Activations of early layers of a CNN\" id=\"img_layer1\">"
|
||||
"<img src=\"images/layer1.png\" alt=\"Activations of early layers of a CNN\" width=\"300\" caption=\"Activations of early layers of a CNN (curtesy of Matthew D. Zeiler and Rob Fergus)\" id=\"img_layer1\">"
|
||||
]
|
||||
},
|
||||
{
|
||||
@ -1693,7 +1693,7 @@
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"<img src=\"images/layer2.png\" alt=\"Activations of early layers of a CNN\" width=\"800\" caption=\"Activations of early layers of a CNN\" id=\"img_layer2\">"
|
||||
"<img src=\"images/layer2.png\" alt=\"Activations of early layers of a CNN\" width=\"800\" caption=\"Activations of early layers of a CNN (curtesy of Matthew D. Zeiler and Rob Fergus)\" id=\"img_layer2\">"
|
||||
]
|
||||
},
|
||||
{
|
||||
@ -1709,7 +1709,7 @@
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"<img src=\"images/chapter2_layer3.PNG\" alt=\"Activations of medium layers of a CNN\" width=\"800\" caption=\"Activations of medium layers of a CNN\" id=\"img_layer3\">"
|
||||
"<img src=\"images/chapter2_layer3.PNG\" alt=\"Activations of medium layers of a CNN\" width=\"800\" caption=\"Activations of medium layers of a CNN (curtesy of Matthew D. Zeiler and Rob Fergus)\" id=\"img_layer3\">"
|
||||
]
|
||||
},
|
||||
{
|
||||
@ -1723,7 +1723,7 @@
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"<img src=\"images/chapter2_layer4and5.PNG\" alt=\"Activations of end layers of a CNN\" width=\"800\" caption=\"Activations of end layers of a CNN\" id=\"img_layer4\">"
|
||||
"<img src=\"images/chapter2_layer4and5.PNG\" alt=\"Activations of end layers of a CNN\" width=\"800\" caption=\"Activations of end layers of a CNN (curtesy of Matthew D. Zeiler and Rob Fergus)\" id=\"img_layer4\">"
|
||||
]
|
||||
},
|
||||
{
|
||||
|
@ -476,7 +476,8 @@
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(Path('train/002844.jpg'), ['train'])"
|
||||
"(Path('/home/sgugger/.fastai/data/pascal_2007/train/008663.jpg'),\n",
|
||||
" ['car', 'person'])"
|
||||
]
|
||||
},
|
||||
"execution_count": null,
|
||||
@ -737,8 +738,8 @@
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"tensor([-1.0028, 0.3400, -0.5906, 0.7806, 3.1160, -0.1994, 1.3180, 1.6361, -1.7553, 0.2217, 2.8052, 1.3229, 0.9369, -1.4760, -0.3204, -2.3116, -3.8615, -1.5931, 0.0745, -3.6006],\n",
|
||||
" device='cuda:5', grad_fn=<SelectBackward>)"
|
||||
"tensor([ 2.0258, -1.3543, 1.4640, 1.7754, -1.2820, -5.8053, 3.6130, 0.7193, -4.3683, -2.5001, -2.8373, -1.8037, 2.0122, 0.6189, 1.9729, 0.8999, -2.6769, -0.3829, 1.2212, 1.6073],\n",
|
||||
" device='cuda:0', grad_fn=<SelectBackward>)"
|
||||
]
|
||||
},
|
||||
"execution_count": null,
|
||||
@ -772,7 +773,25 @@
|
||||
"source": [
|
||||
"def binary_cross_entropy(inputs, targets):\n",
|
||||
" inputs = inputs.sigmoid()\n",
|
||||
" return torch.where(targets==1, 1-inputs, inputs).log().mean()"
|
||||
" return torch.where(targets==1, inputs, 1-inputs).log().mean()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"binary_cross_entropy(activs, y)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"F.binary_cross_entropy_with_logits(activs, y)"
|
||||
]
|
||||
},
|
||||
{
|
||||
|
@ -313,7 +313,8 @@
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(Path('train/002844.jpg'), ['train'])"
|
||||
"(Path('/home/sgugger/.fastai/data/pascal_2007/train/008663.jpg'),\n",
|
||||
" ['car', 'person'])"
|
||||
]
|
||||
},
|
||||
"execution_count": null,
|
||||
@ -488,8 +489,8 @@
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"tensor([-1.0028, 0.3400, -0.5906, 0.7806, 3.1160, -0.1994, 1.3180, 1.6361, -1.7553, 0.2217, 2.8052, 1.3229, 0.9369, -1.4760, -0.3204, -2.3116, -3.8615, -1.5931, 0.0745, -3.6006],\n",
|
||||
" device='cuda:5', grad_fn=<SelectBackward>)"
|
||||
"tensor([ 2.0258, -1.3543, 1.4640, 1.7754, -1.2820, -5.8053, 3.6130, 0.7193, -4.3683, -2.5001, -2.8373, -1.8037, 2.0122, 0.6189, 1.9729, 0.8999, -2.6769, -0.3829, 1.2212, 1.6073],\n",
|
||||
" device='cuda:0', grad_fn=<SelectBackward>)"
|
||||
]
|
||||
},
|
||||
"execution_count": null,
|
||||
@ -509,7 +510,25 @@
|
||||
"source": [
|
||||
"def binary_cross_entropy(inputs, targets):\n",
|
||||
" inputs = inputs.sigmoid()\n",
|
||||
" return torch.where(targets==1, 1-inputs, inputs).log().mean()"
|
||||
" return torch.where(targets==1, inputs, 1-inputs).log().mean()"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"binary_cross_entropy(activs, y)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"F.binary_cross_entropy_with_logits(activs, y)"
|
||||
]
|
||||
},
|
||||
{
|
||||
|
Loading…
Reference in New Issue
Block a user