!pip install fastai --upgrade

import torch
import fastai as fastai
import fastai.vision.all as vision

import pandas as pd
import matplotlib.pyplot as plt
from functools import partial

path = vision.untar_data(vision.URLs.PASCAL_2007)
path.ls()

train_img = '/root/.fastai/data/pascal_2007/train/'

csv = pd.read_csv(path.ls()[2])

csv.info()

def get_x(df): return train_img+df['fname'] #return df['fname'] 

def get_y(df): return df['labels'].split()#return csv[csv['fname'] == fname]['labels'].iloc[0].split()

def train_test(df): 
  train = df.index[~df['is_valid']].tolist()
  valid = df.index[df['is_valid']].tolist()

  return train, valid

db = vision.DataBlock(
    blocks=(vision.ImageBlock, vision.MultiCategoryBlock),
    get_x=get_x,
    get_y=get_y,
    splitter=train_test,
    item_tfms=vision.RandomResizedCrop(128, min_scale=0.35)
)

db.summary(csv)

dataset = db.datasets(csv)

print(dataset.vocab[torch.where(dataset[171][1]==1)[0]])
dataset[171][0]

dls = db.dataloaders(csv, bs=32)

dls.show_batch(nrows=1, ncols=3)

model = vision.cnn_learner(dls, vision.resnet18)

for data, label in dls.train:
  print(data.shape, label.shape)
  break

model.model

x, y = vision.to_cpu(dls.train.one_batch())

out = model.model(x)
print(out.shape, model.loss_func)

out[0]

def idx2vocab(tens):
  print(dataset.vocab[torch.where(tens.round()==1)[0]])

idx2vocab(out[10])

def binary_cross_entropy(pred, target):
  pred = pred.sigmoid()

  return -torch.where(target==1, pred, 1-pred).log().mean()

def acc_multi(pred, target, thresh=0.5, sigmoid=True):
  if sigmoid: pred = pred.sigmoid()

  return ((pred>thresh) == target.bool()).float().mean()

acc_multi(out[0], out[0].sigmoid().round())

model = vision.cnn_learner(dls, vision.resnet18, metrics=partial(acc_multi, thresh=0.57))
model.lr_find()

model = vision.cnn_learner(dls, vision.resnet18, metrics=partial(acc_multi, thresh=0.57))
model.fine_tune(3, base_lr=0.003, max_lr=slice(0.002, 0.1), freeze_epochs=4)

pred, targ = model.get_preds()

pred[0], targ[1]

model.metrics = partial(acc_multi, thresh=0.5)
loss, acc = model.validate()
print("val_acc: {}\tval_loss: {}".format(loss, acc))

acc_multi(pred, targ, thresh=0.55, sigmoid=False)

x = torch.linspace(0.05, 0.95)
y = [acc_multi(pred, targ, thresh=i, sigmoid=False) for i in x]

plt.plot(x, y, 'b-')

model.metrics = partial(acc_multi, thresh=0.57)
model.validate()

output = model.model(dls.one_batch()[0])

dls.one_batch()[0][0].shape

model.metrics = partial(acc_multi, thresh=0.55)
model.show_results()