Datasets:

---

vikp

/

code_with_explanations

like 2

ERROR: type should be string, got "https://github.com/utkuozbulak/pytorch-cnn-visualizations\nhttps://towardsdatascience.com/how-to-visualize-convolutional-features-in-40-lines-of-code-70b7d87b0030\n\n```\nimport h5py\nimport os\nfrom os.path import join\nfrom glob import glob\nfrom scipy.io import loadmat\nimport cv2\nimport torch\nfrom torch.autograd import Variable\nfrom torch.autograd import grad\nimport numpy as np\nimport matplotlib.pylab as plt\nimport csv\nfrom time import time\ntorch.cuda.get_device_name(device=None)\nimport torchvision\nalexnet=torchvision.models.alexnet(pretrained=True)\nnet = torchvision.models.vgg16(pretrained=True)\nnet\nfrom torchvision import transforms\nnormalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],\n std=[0.229, 0.224, 0.225]) # Note without normalization, the\ndenormalize = transforms.Normalize(\n mean=[-0.485/0.229, -0.456/0.224, -0.406/0.225],\n std=[1/0.229, 1/0.224, 1/0.225])\nclass SaveFeatures():\n def __init__(self, module):\n self.hook = module.register_forward_hook(self.hook_fn)\n def hook_fn(self, module, input, output):\n self.features = output#torch.tensor(output,requires_grad=True).cuda()\n def close(self):\n self.hook.remove()\n\ndef val_tfms(img_np):\n img = torch.from_numpy(img_np.astype(np.float32)).permute(2, 0, 1)\n nimg = normalize(img).unsqueeze(0).cuda()\n return nimg\n\ndef val_detfms(img_tsr):\n img = denormalize(img_tsr.squeeze()).permute(1,2,0)\n return img.detach().cpu().numpy()\nMAX_TRIAL = 100\nclass FilterVisualizer():\n def __init__(self, model, size=56, upscaling_steps=12, upscaling_factor=1.2):\n self.size, self.upscaling_steps, self.upscaling_factor = size, upscaling_steps, upscaling_factor\n self.model = model #alexnet.features.cuda().eval()\n # set_trainable(self.model, False)\n for param in self.model.parameters():\n param.requires_grad = False\n\n def visualize(self, layer, filter, lr=0.05, opt_steps=20, blur=None):\n sz = self.size\n img = np.uint8(np.random.uniform(50, 250, (sz, sz, 3))) / 255 # generate random image\n activations = SaveFeatures(list(self.model.children())[layer]) # register hook\n\n for _ in range(self.upscaling_steps): # scale the image up upscaling_steps times\n # train_tfms, val_tfms = tfms_from_model(vgg16, sz)\n img_var = Variable(val_tfms(img), requires_grad=True) # convert image to Variable that requires grad\n optimizer = torch.optim.Adam([img_var], lr=lr, weight_decay=1e-6)\n for n in range(MAX_TRIAL):\n optimizer.zero_grad()\n self.model(img_var)\n loss = -activations.features[0, filter].mean()\n loss.backward()\n if img_var.grad.norm()<1E-6:\n img_var = Variable(val_tfms(img + np.random.randn(*img.shape)), requires_grad=True) # convert image to Variable that requires grad\n optimizer = torch.optim.Adam([img_var], lr=lr, weight_decay=1e-6)\n print(\"Optimizer restart\")\n else:\n break\n for n in range(opt_steps): # optimize pixel values for opt_steps times\n optimizer.zero_grad()\n self.model(img_var)\n loss = -activations.features[0, filter].mean()\n loss.backward()\n optimizer.step()\n print(loss.data.cpu())\n img = val_detfms(img_var.data.cpu())\n self.output = img\n plt.figure(figsize=[8,8])\n plt.imshow(FVis.output)\n plt.show()\n sz = int(self.upscaling_factor * sz) # calculate new image size\n img = cv2.resize(img, (sz, sz), interpolation=cv2.INTER_CUBIC) # scale image up\n if blur is not None: img = cv2.blur(img, (blur, blur)) # blur image to reduce high frequency patterns\n self.save(layer, filter)\n activations.close()\n\n def save(self, layer, filter):\n plt.imsave(\"layer_\" + str(layer) + \"_filter_\" + str(filter) + \".jpg\", np.clip(self.output, 0, 1))\n\n#%%\nfeat = alexnet.features.cuda().eval()\nfeat = net.features.cuda().eval()\nFVis = FilterVisualizer(feat, size=227, upscaling_steps=2, upscaling_factor=1.2)\nFVis.visualize(14, 20, blur=10, opt_steps=20)\nplt.figure(figsize=[8,8])\nplt.imshow(FVis.output)\nplt.show()\nnet = alexnet.features[:]\nsz = 224\nlr = 0.1\nopt_steps = 100\nupscaling_factor = 1.2\nfilter = 10\nloss_arr = []\nactivations = SaveFeatures(list(alexnet.features.children())[8]) # register hook\n\nimg = np.uint8(np.random.uniform(150, 180, (sz, sz, 3))) / 255\nimg_var = Variable(val_tfms(img), requires_grad=True) # convert image to Variable that requires grad\noptimizer = torch.optim.Adam([img_var], lr=lr, weight_decay=1e-6)\nfor n in range(opt_steps): # optimize pixel values for opt_steps times\n optimizer.zero_grad()\n alexnet.features(img_var)\n loss = -activations.features[0, filter].mean()\n loss.backward()\n optimizer.step()\n loss_arr.append(loss.data.cpu())\nimg = val_detfms(img_var.data.cpu()).numpy()\noutput = img\nsz = int(upscaling_factor * sz) # calculate new image size\nimg = cv2.resize(img, (sz, sz), interpolation=cv2.INTER_CUBIC)\nalexnet.features(img_var)\nloss = -activations.features[0, filter].mean()\nFVis.output[:,:,0].mean()\nimg_np = np.random.rand(5,5,3)\nimg_tsr = val_tfms(img_np)\nimg_out = val_detfms(img_tsr)\n```\n\n## Printing Version of FilterVisualization\n\n```\n# def np2tensor(image,dtype):\n# \"Convert np.array (sz,sz,3) to tensor (1,3,sz,sz), imagenet normalized\"\n\n# a = np.asarray(image)\n# if a.ndim==2 : a = np.expand_dims(a,2)\n# a = np.transpose(a, (1, 0, 2))\n# a = np.transpose(a, (2, 1, 0))\n \n# #Imagenet norm\n# mean=np.array([0.485, 0.456, 0.406])[...,np.newaxis,np.newaxis]\n# std = np.array([0.229, 0.224, 0.225])[...,np.newaxis,np.newaxis]\n# a = (a-mean)/std\n# a = np.expand_dims(a,0)\n# return torch.from_numpy(a.astype(dtype, copy=False) )\n\n# def tensor2np(img_tensor):\n# \"Convert tensor (1,3,sz,sz) back to np.array (sz,sz,3), imagenet DEnormalized\"\n# a = np.squeeze(to_np(img_tensor))\n \n# mean=np.array([0.485, 0.456, 0.406])[...,np.newaxis,np.newaxis]\n# std = np.array([0.229, 0.224, 0.225])[...,np.newaxis,np.newaxis]\n# a = a*std + mean\n# return np.transpose(a, (1,2,0))\n\nclass FilterVisualizer():\n def __init__(self,model):\n self.model = model\n self.weights = None\n\n def visualize(self, sz, layer, filter, weights=None, \n upscaling_steps=12, upscaling_factor=1.2, lr=0.1, opt_steps=20, blur=None, print_losses=False):\n '''Add weights to support visualize combination of channels'''\n if weights is not None:\n assert len(weights) == len(filter)\n self.weights = torch.tensor(weights,dtype=torch.float,device='cuda')\n img = (np.random.random((sz,sz, 3)) * 20 + 128.)/255 # value b/t 0 and 1 \n activations = SaveFeatures(layer) # register hook\n\n for i in range(upscaling_steps): \n # convert np to tensor + channel first + new axis, and apply imagenet norm\n img_tensor = val_tfms(img)#,np.float32)\n img_tensor = img_tensor.cuda()\n img_tensor.requires_grad_();\n if not img_tensor.grad is None:\n img_tensor.grad.zero_(); \n \n \n optimizer = torch.optim.Adam([img_tensor], lr=0.1, weight_decay=1e-6)\n if i > upscaling_steps/2:\n opt_steps_ = int(opt_steps*1.3)\n else:\n opt_steps_ = opt_steps\n for n in range(opt_steps_): # optimize pixel values for opt_steps times\n optimizer.zero_grad()\n _=self.model(img_tensor)\n if weights is None:\n loss = -1*activations.features[0, filter].mean()\n else: \n loss = -1*torch.einsum(\"ijk,i->jk\", activations.features[0, filter], self.weights).mean()\n if print_losses:\n if i%3==0 and n%5==0:\n print(f'{i} - {n} - {float(-loss)}')\n loss.backward()\n optimizer.step()\n \n # convert tensor back to np\n img = val_detfms(img_tensor)\n self.output = img\n sz = int(upscaling_factor * sz) # calculate new image size\n# print(f'Upscale img to: {sz}')\n img = cv2.resize(img, (sz, sz), interpolation = cv2.INTER_CUBIC) # scale image up\n if blur is not None: img = cv2.blur(img,(blur,blur)) # blur image to reduce high frequency patterns\n \n activations.close()\n return np.clip(self.output, 0, 1)\n \n def get_transformed_img(self,img,sz):\n '''\n Scale up/down img to sz. Channel last (same as input)\n image: np.array [sz,sz,3], already divided by 255\" \n '''\n return cv2.resize(img, (sz, sz), interpolation = cv2.INTER_CUBIC)\n \n def most_activated(self, img, layer):\n '''\n image: np.array [sz,sz,3], already divided by 255\" \n '''\n img = cv2.resize(img, (224,224), interpolation = cv2.INTER_CUBIC)\n activations = SaveFeatures(layer)\n img_tensor = val_tfms(img)#,np.float32)\n img_tensor = img_tensor.cuda()\n \n _=self.model(img_tensor)\n mean_act = [np.squeeze(to_np(activations.features[0,i].mean())) for i in range(activations.features.shape[1])]\n activations.close()\n return mean_act\nfeat = alexnet.features.cuda().eval()\nFVis = FilterVisualizer(feat)\nimg = FVis.visualize(sz=227, layer=feat[8], filter=[1,5,3,10], weights=[1,3,1,7], blur=10, opt_steps=20, upscaling_steps=3, upscaling_factor=1.2, print_losses=True)\nplt.figure(figsize=[8,8])\nplt.imshow(FVis.output)\nplt.show()\n```\n\n## Using GAN to visualize Convolutiona Layers\n\n```\nBGR_mean = torch.tensor([104.0, 117.0, 123.0])\nBGR_mean = torch.reshape(BGR_mean, (1, 3, 1, 1))\ndef visualize(G, code):\n \"\"\"Do the De-caffe transform (Validated)\"\"\"\n code = code.reshape(-1, 4096).astype(np.float32)\n blobs = G(torch.from_numpy(code))\n out_img = blobs['deconv0'] # get raw output image from GAN\n clamp_out_img = torch.clamp(out_img + BGR_mean, 0, 255)\n vis_img = clamp_out_img[:, [2, 1, 0], :, :].permute([2, 3, 1, 0]).squeeze() / 255\n return vis_img\n\ndef visualize_for_torchnet(G, code):\n \"\"\"Do the De-caffe transform (Validated)\"\"\"\n blobs = G(code)\n out_img = blobs['deconv0'] # get raw output image from GAN\n clamp_out_img = torch.clamp(out_img + BGR_mean, 0, 255) / 255\n vis_img = clamp_out_img[:, [2, 1, 0], :, :] # still use BCHW sequence \n return vis_img\nclass FilterVisualizerGAN():\n def __init__(self,model):\n self.model = model\n self.G = load_generator()\n self.weights = None\n\n def visualize(self, sz, layer, filter, weights=None, \n lr=0.1, opt_steps=20, blur=None, print_losses=False): #upscaling_steps=12, upscaling_factor=1.2, \n '''Add weights to support visualize combination of channels'''\n if weights is not None:\n assert len(weights) == len(filter)\n self.weights = torch.tensor(weights,dtype=torch.float,device='cuda')\n \n activations = SaveFeatures(layer) # register hook\n feat = 0.01 * np.random.rand(1, 4096)\n feat = torch.from_numpy(np.float32(feat))\n feat = Variable(feat, requires_grad = True).cuda()\n img = visualize_for_torchnet(self.G, feat) \n resz_img = F.interpolate(img, (sz, sz), mode='bilinear', align_corners=True)\n # img = (np.random.random((sz,sz, 3)) * 20 + 128.)/255 # value b/t 0 and 1 \n # img_tensor = val_tfms(resz_img) \n img_tensor = normalize(resz_img.squeeze()).unsqueeze(0)\n# img_tensor = img_tensor.cuda()\n# img_tensor.requires_grad_();\n optimizer = optim.SGD([feat], lr=0.05,momentum=0.3,dampening=0.1)\n if not img_tensor.grad is None:\n img_tensor.grad.zero_(); \n\n for n in range(opt_steps_): # optimize pixel values for opt_steps times\n optimizer.zero_grad()\n img = visualize_for_torchnet(self.G, feat) \n resz_img = F.interpolate(img, (sz, sz), mode='bilinear', align_corners=True)\n img_tensor = normalize(resz_img.squeeze()).unsqueeze(0)\n _=self.model(img_tensor)\n if weights is None:\n loss = -1*activations.features[0, filter].mean()\n else: \n loss = -1*torch.einsum(\"ijk,i->jk\", activations.features[0, filter], self.weights).mean()\n if print_losses:\n if n%5==0:\n print(f'{n} - {float(-loss)}')\n loss.backward()\n optimizer.step()\n\n # convert tensor back to np\n img = val_detfms(img_tensor)\n self.output = img\n # sz = int(upscaling_factor * sz) # calculate new image size\n# print(f'Upscale img to: {sz}')\n img = cv2.resize(img, (sz, sz), interpolation = cv2.INTER_CUBIC) # scale image up\n if blur is not None: img = cv2.blur(img,(blur,blur)) # blur image to reduce high frequency patterns\n \n activations.close()\n return np.clip(self.output, 0, 1)\n \n def get_transformed_img(self,img,sz):\n '''\n Scale up/down img to sz. Channel last (same as input)\n image: np.array [sz,sz,3], already divided by 255\" \n '''\n return cv2.resize(img, (sz, sz), interpolation = cv2.INTER_CUBIC)\n \n# def most_activated(self, img, layer):\n# '''\n# image: np.array [sz,sz,3], already divided by 255\" \n# '''\n# img = cv2.resize(img, (224,224), interpolation = cv2.INTER_CUBIC)\n# activations = SaveFeatures(layer)\n# img_tensor = val_tfms(img)#,np.float32)\n# img_tensor = img_tensor.cuda()\n \n# _=self.model(img_tensor)\n# mean_act = [np.squeeze(to_np(activations.features[0,i].mean())) for i in range(activations.features.shape[1])]\n# activations.close()\n# return mean_act\nfrom torch_net_utils import visualize, load_generator\nfeat = alexnet.features.cuda().eval()\nFVisG = FilterVisualizerGAN(feat)\nimg = FVisG.visualize(sz=227, layer=feat[8], filter=[1,5,3,10], weights=[1,3,1,7], blur=10, opt_steps=20, upscaling_steps=3, upscaling_factor=1.2, print_losses=True)\nplt.figure(figsize=[8,8])\nplt.imshow(FVis.output)\nplt.show()\n```\n\n"