app.py

from fastai.vision.all import *
import cv2
import gradio as gr
import glob

class Hook():
  def hook_func(self, m, i, o): self.stored = o.detach().clone()

learn = load_learner("resnet152_fit_one_cycle_freeze_91acc.pkl", cpu=True)

#@title DataLoader
path = "train_val_cropped"
dblock = DataBlock(blocks = (ImageBlock, CategoryBlock),
                   get_items = get_image_files,
                   splitter = GrandparentSplitter(train_name="train",
                                                  valid_name="valid"),
                   get_y=parent_label,
                   item_tfms=RandomResizedCrop(128, min_scale=0.7),
                   batch_tfms=[*aug_transforms(),
                   Normalize.from_stats(*imagenet_stats)])
dls_augmented = dblock.dataloaders(path, shuffle=True)
dls_augmented.train.show_batch(max_n=8, nrows=2, figsize=(28,10))

class Hook():
  def hook_func(self, m, i, o): self.stored = o.detach().clone()

def gradcam(img_create):

  pred,idx,probs=learn.predict(img_create)
    
  x,= first(dls_augmented.test_dl([img_create]))
  hook_output = Hook()
  hook = learn.model[0].register_forward_hook(hook_output.hook_func)
  with torch.no_grad(): output = learn.model.eval()(x)
  act = hook_output.stored[0]
  hook.remove()

  input_size=act.shape[0]
  out_size=learn.model[1][-1].in_features
  kernel_size=act.shape[1]
  a=act
  new_act=tensor(np.zeros((out_size,kernel_size,kernel_size)))
  sum=tensor(np.zeros((1,kernel_size,kernel_size)))
  for i in range(0,input_size,4):
    sum=tensor(np.zeros((1,kernel_size,kernel_size)))
    for j in range(i,i+4):
      sum=sum+act[j,:,:]
    new_act[int(i/4),:,:]=sum/4
  cam_map = torch.einsum('ck,kij->cij', learn.model[1][-1].weight, new_act)

  x_dec = TensorImage(dls_augmented.train.decode((x,))[0][0])
  fig,ax = plt.subplots()
  x_dec.show(ctx=ax)
  ax.imshow(cam_map[1].detach().cpu(), alpha=0.6, extent=(0,128,128,0), interpolation='bilinear', cmap='magma');
  fig.savefig("gcam.jpg")
  im=cv2.imread("gcam.jpg",1)
  im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
  #im=np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')

  return dict(zip(categories, map(float, probs))), im

categories = ('arbanasi', 'filibe', 'gjirokoster', 'iskodra', 'kula', 'kuzguncuk', 'larissa_ampelakia', 'mardin', 'ohrid', 'pristina', 'safranbolu', 'selanik', 'sozopol_suzebolu', 'tiran', 'varna')
def classify_img(img):
  pred,idx,probs=learn.predict(img)
  return dict(zip(categories, map(float, probs)))

image=gr.inputs.Image(shape=(128,128))
label=gr.outputs.Label()
examples_=[]
for i in glob.glob("valid/**/*.jpg", recursive=True):
    examples_.append(i)

examples=["filibe-1-1.jpg",
          "ohrid-3-1.jpg",
          "varna-1-1.jpg"]


demo = gr.Interface(fn=gradcam, inputs=image, outputs=[label, gr.inputs.Image(shape=(128,128))], examples=examples)

demo.launch(inline=False)