DO NOT MERGE

WIP optim module

Author: Ludwig Schubert

captum/optim/README.md

@@ -0,0 +1,13 @@
+# Captum "optim" module
+
+This is a WIP PR to integrate existing feature visualization code from the authors of `tensorflow/lucid` into captum.
+It is also an opportunity to review which parts of such interpretability tools still feel rough to implement in a system like PyTorch, and to make suggetsions to the core PyTorch team for how to improve these aspects.
+
+## Roadmap
+
+* unify API with Captum API: a single class that's callable per "technique"(? check for details before implementing)
+* Consider if we need an abstraction around "an optimization process" (in terms of stopping criteria, reporting losses, etc) or if there are sufficiently strong conventions in PyTorch land for such tasks
+* integrate Eli's FFT param changes (mostly for simplification) 
+* make a table of PyTorch interpretability tools for readme?
+* do we need image viewing helpers and io helpers or throw those out?
+* can we integrate paper references closer with the code?

captum/optim/__init__.py

@@ -0,0 +1,27 @@
+from typing import Dict, Optional, Union, Callable, Iterable
+from typing_extensions import Protocol
+
+import torch
+import torch.nn as nn
+
+ParametersForOptimizers = Iterable[Union[torch.Tensor, Dict[str, torch.tensor]]]
+
+
+class HasLoss(Protocol):
+    def loss(self) -> torch.Tensor:
+        ...
+
+
+class Parameterized(Protocol):
+    parameters: ParametersForOptimizers
+
+
+class Objective(Parameterized, HasLoss):
+    def cleanup(self):
+        pass
+
+
+ModuleOutputMapping = Dict[nn.Module, Optional[torch.Tensor]]
+
+StopCriteria = Callable[[int, Objective, torch.optim.Optimizer], bool]
+

captum/optim/_scrap_and_testing.py

@@ -0,0 +1,139 @@
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision
+import requests
+from PIL import Image
+from IPython.display import display
+
+from clarity.pytorch.inception_v1 import googlenet
+from lucid.misc.io import show, load, save
+from lucid.modelzoo.other_models import InceptionV1
+
+# get a test image
+img_url = (
+    "https://lucid-static.storage.googleapis.com/building-blocks/examples/dog_cat.png"
+)
+img_tf = load(img_url)
+img_pt = torch.as_tensor(img_tf.transpose(2, 0, 1))[None, ...]
+img_pil = Image.open(requests.get(img_url, stream=True).raw)
+
+# instantiate ported model
+net = googlenet(pretrained=True)
+
+# get predictions
+out = net(img_pt)
+logits = out.detach().numpy()[0]
+top_k = np.argsort(-logits)[:5]
+
+# load labels
+labels = load(InceptionV1.labels_path, split=True)
+
+# show predictions
+for i, k in enumerate(top_k):
+    prediction = logits[k]
+    label = labels[k]
+    print(f"{i}: {label} ({prediction*100:.2f}%)")
+
+# transforms
+
+
+# def build_grid(source_size, target_size):
+#     k = float(target_size) / float(source_size)
+#     direct = (
+#         torch.linspace(0, k, target_size)
+#         .unsqueeze(0)
+#         .repeat(target_size, 1)
+#         .unsqueeze(-1)
+#     )
+#     full = torch.cat([direct, direct.transpose(1, 0)], dim=2).unsqueeze(0)
+#     return full.cuda()
+
+
+# def random_crop_grid(x, grid):
+#     d = x.size(2) - grid.size(1)
+#     grid = grid.repeat(x.size(0), 1, 1, 1).cuda()
+#     # Add random shifts by x
+#     grid[:, :, :, 0] += torch.FloatTensor(x.size(0)).cuda().random_(0, d).unsqueeze(
+#         -1
+#     ).unsqueeze(-1).expand(-1, grid.size(1), grid.size(2)) / x.size(2)
+#     # Add random shifts by y
+#     grid[:, :, :, 1] += torch.FloatTensor(x.size(0)).cuda().random_(0, d).unsqueeze(
+#         -1
+#     ).unsqueeze(-1).expand(-1, grid.size(1), grid.size(2)) / x.size(2)
+#     return grid
+
+
+# # We want to crop a 80x80 image randomly for our batch
+# # Building central crop of 80 pixel size
+# grid_source = build_grid(224, 80)
+# # Make radom shift for each batch
+# grid_shifted = random_crop_grid(batch, grid_source)
+# # Sample using grid sample
+# sampled_batch = F.grid_sample(batch, grid_shifted)
+
+
+from clarity.pytorch.transform import RandomSpatialJitter, RandomUpsample
+
+# crop = torchvision.transforms.RandomCrop(
+#     224, padding=34, pad_if_needed=True, padding_mode="reflect"
+# )
+jitter = RandomSpatialJitter(16)
+ups = RandomUpsample()
+for i in range(10):
+    cropped = ups(img_pt)
+    show(cropped.numpy()[0].transpose(1, 2, 0))
+    # display(cropped)
+
+
+# result = param().cpu().detach().numpy()[0].transpose(1, 2, 0)
+# loss_curve = objective.history
+
+# 2019-11-21 notes from Pytorch team
+# Set up model
+# net = googlenet(pretrained=True)
+# parameterization = Image()  # TODO: make size adjustable, currently hardcoded
+# input_image = parameterization()
+
+# writer = SummaryWriter()
+# writer.add_graph(net, (input_image,))
+# writer.close()
+
+# Specify target module / "objective"
+# target_module = net.mixed3b._pool_reduce[1]
+# target_channel = 54
+# hook = OutputHook(target_module) # TODO: investigate detach on rerun
+# parameterization = Image()  # TODO: make size adjustable, currently hardcoded
+# optimizer = optim.Adam(parameterization.parameters, lr=0.025)
+
+# device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+# net = net.to(device)
+# parameterization = parameterization.to(device)
+# for i in range(1000):
+#     optimizer.zero_grad()
+
+#     # forward pass through entire net
+#     input_image = parameterization()
+#     with suppress(AbortForwardException):
+#         _ = net(input_image.to(device))
+
+#     # activations were stored during forward pass
+#     assert hook.saved_output is not None
+#     loss = -hook.saved_output[:, target_channel, :, :].sum()  # channel 13
+
+#     loss.backward()
+#     optimizer.step()
+
+#     if i % 100 == 0:
+#         print("Loss: ", -loss.cpu().detach().numpy())
+#         url = show(
+#             parameterization.raw_image.cpu()
+#             .detach()
+#             .numpy()[0]
+#             .transpose(1, 2, 0)
+#         )
+
+# traced_net = torch.jit.trace(net, example_inputs=(input_image,))
+# print(traced_net.graph)

captum/optim/io/__init__.py

@@ -0,0 +1 @@
+from .io import show

captum/optim/io/fixtures.py

@@ -0,0 +1,13 @@
+import torch
+
+# TODO: use imageio to redo load and avoid TF dependency
+from lucid.misc.io import load
+
+DOG_CAT_URL = (
+    "https://lucid-static.storage.googleapis.com/building-blocks/examples/dog_cat.png"
+)
+
+
+def image(url: str = DOG_CAT_URL):
+    img_np = load(url)
+    return torch.as_tensor(img_np.transpose(2, 0, 1))

captum/optim/io/formatters.py

@@ -0,0 +1,22 @@
+from io import BytesIO
+
+import torch
+from torchvision import transforms
+
+from IPython import display, get_ipython
+
+
+def tensor_jpeg(tensor: torch.Tensor):
+    if tensor.dim() == 3:
+        pil_image = transforms.ToPILImage()(tensor.cpu().detach()).convert("RGB")
+        buffer = BytesIO()
+        pil_image.save(buffer, format="jpeg")
+        data = buffer.getvalue()
+        return data
+    else:
+        return tensor
+
+
+def register_formatters():
+    jpeg_formatter = get_ipython().display_formatter.formatters["image/jpeg"]
+    jpeg_formatter.for_type(torch.Tensor, tensor_jpeg)

captum/optim/io/io.py

@@ -0,0 +1,11 @@
+# TODO: redo show using display or register handler for jupyter display directly
+# maybe we could even have subtypes of tensors that are "ImageTensors" or "ActivationTensors" etc
+from lucid.misc.io import show as lucid_show
+
+
+def show(thing):
+    if len(thing.shape) == 3:
+        numpy_thing = thing.cpu().detach().numpy().transpose(1, 2, 0)
+    elif len(thing.shape) == 4:
+        numpy_thing = thing.cpu().detach().numpy()[0].transpose(1, 2, 0)
+    lucid_show(numpy_thing)

captum/optim/models/__init__.py

@@ -0,0 +1 @@
+from .inception_v1 import googlenet

captum/optim/models/conv2d.py

@@ -0,0 +1,119 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+
+
+def _is_static_pad(kernel_size, stride=1, dilation=1, **_):
+    return stride == 1 and (dilation * (kernel_size - 1)) % 2 == 0
+
+
+def _get_padding(kernel_size, stride=1, dilation=1, **_):
+    padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2
+    return padding
+
+
+def _calc_same_pad(i, k, s, d):
+    return max((math.ceil(i / s) - 1) * s + (k - 1) * d + 1 - i, 0)
+
+
+def _split_channels(num_chan, num_groups):
+    split = [num_chan // num_groups for _ in range(num_groups)]
+    split[0] += num_chan - sum(split)
+    return split
+
+
+class Conv2dSame(nn.Conv2d):
+    """ Tensorflow like 'SAME' convolution wrapper for 2D convolutions
+    """
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
+                 padding=0, dilation=1, groups=1, bias=True):
+        super(Conv2dSame, self).__init__(
+            in_channels, out_channels, kernel_size, stride, 0, dilation,
+            groups, bias)
+
+    def forward(self, x):
+        ih, iw = x.size()[-2:]
+        kh, kw = self.weight.size()[-2:]
+        pad_h = _calc_same_pad(ih, kh, self.stride[0], self.dilation[0])
+        pad_w = _calc_same_pad(iw, kw, self.stride[1], self.dilation[1])
+        if pad_h > 0 or pad_w > 0:
+            x = F.pad(x, [pad_w//2, pad_w - pad_w//2, pad_h//2, pad_h - pad_h//2])
+        return F.conv2d(x, self.weight, self.bias, self.stride,
+                        self.padding, self.dilation, self.groups)
+
+
+# def conv2d_pad(in_chs, out_chs, kernel_size, **kwargs):
+#     padding = kwargs.pop('padding', '')
+#     kwargs.setdefault('bias', False)
+#     if isinstance(padding, str):
+#         # for any string padding, the padding will be calculated for you, one of three ways
+#         padding = padding.lower()
+#         if padding == 'same':
+#             # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact
+#             if _is_static_pad(kernel_size, **kwargs):
+#                 # static case, no extra overhead
+#                 padding = _get_padding(kernel_size, **kwargs)
+#                 return nn.Conv2d(in_chs, out_chs, kernel_size, padding=padding, **kwargs)
+#             else:
+#                 # dynamic padding
+#                 return Conv2dSame(in_chs, out_chs, kernel_size, **kwargs)
+#         elif padding == 'valid':
+#             # 'VALID' padding, same as padding=0
+#             return nn.Conv2d(in_chs, out_chs, kernel_size, padding=0, **kwargs)
+#         else:
+#             # Default to PyTorch style 'same'-ish symmetric padding
+#             padding = _get_padding(kernel_size, **kwargs)
+#             return nn.Conv2d(in_chs, out_chs, kernel_size, padding=padding, **kwargs)
+#     else:
+#         # padding was specified as a number or pair
+#         return nn.Conv2d(in_chs, out_chs, kernel_size, padding=padding, **kwargs)
+
+
+# class MixedConv2d(nn.Module):
+#     """ Mixed Grouped Convolution
+#     Based on MDConv and GroupedConv in MixNet impl:
+#       https://github.com/tensorflow/tpu/blob/master/models/official/mnasnet/mixnet/custom_layers.py
+#     """
+
+#     def __init__(self, in_channels, out_channels, kernel_size=3,
+#                  stride=1, padding='', dilated=False, depthwise=False, **kwargs):
+#         super(MixedConv2d, self).__init__()
+
+#         kernel_size = kernel_size if isinstance(kernel_size, list) else [kernel_size]
+#         num_groups = len(kernel_size)
+#         in_splits = _split_channels(in_channels, num_groups)
+#         out_splits = _split_channels(out_channels, num_groups)
+#         for idx, (k, in_ch, out_ch) in enumerate(zip(kernel_size, in_splits, out_splits)):
+#             d = 1
+#             # FIXME make compat with non-square kernel/dilations/strides
+#             if stride == 1 and dilated:
+#                 d, k = (k - 1) // 2, 3
+#             conv_groups = out_ch if depthwise else 1
+#             # use add_module to keep key space clean
+#             self.add_module(
+#                 str(idx),
+#                 conv2d_pad(
+#                     in_ch, out_ch, k, stride=stride,
+#                     padding=padding, dilation=d, groups=conv_groups, **kwargs)
+#             )
+#         self.splits = in_splits
+
+#     def forward(self, x):
+#         x_split = torch.split(x, self.splits, 1)
+#         x_out = [c(x) for x, c in zip(x_split, self._modules.values())]
+#         x = torch.cat(x_out, 1)
+#         return x
+
+
+# # helper method
+# def select_conv2d(in_chs, out_chs, kernel_size, **kwargs):
+#     assert 'groups' not in kwargs  # only use 'depthwise' bool arg
+#     if isinstance(kernel_size, list):
+#         # We're going to use only lists for defining the MixedConv2d kernel groups,
+#         # ints, tuples, other iterables will continue to pass to normal conv and specify h, w.
+#         return MixedConv2d(in_chs, out_chs, kernel_size, **kwargs)
+#     else:
+#         depthwise = kwargs.pop('depthwise', False)
+#         groups = out_chs if depthwise else 1
+#         return conv2d_pad(in_chs, out_chs, kernel_size, groups=groups, **kwargs)