Trained 170 classes, but cannot run inference

[shubhank008]

So I used almost 98% of your code as it is to train 170 classes as I wanted to try quick draw with image classification while most other repos are through using stroke data.

Here is how my model looks:

Model: "sequential_17"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_20 (Conv2D)           (None, 1, 28, 30)         21030     
_________________________________________________________________
max_pooling2d_8 (MaxPooling2 (None, 1, 14, 15)         0         
_________________________________________________________________
conv2d_21 (Conv2D)           (None, 1, 14, 15)         2040      
_________________________________________________________________
max_pooling2d_9 (MaxPooling2 (None, 1, 7, 7)           0         
_________________________________________________________________
dropout_3 (Dropout)          (None, 1, 7, 7)           0         
_________________________________________________________________
flatten_3 (Flatten)          (None, 49)                0         
_________________________________________________________________
dense_7 (Dense)              (None, 128)               6400      
_________________________________________________________________
dense_8 (Dense)              (None, 50)                6450      
_________________________________________________________________
dense_9 (Dense)              (None, 170)               8670      
=================================================================
Total params: 44,590
Trainable params: 44,590
Non-trainable params: 0
_________________________________________________________________

I was able to train and save my model but not able to run inference properly. Here are the some problems I am facing:

1. The test plots/graphs doesn't work, I am using loops to fill my labels for 170 classes

2. OpenCV gives me bad and blurry image, I use a different way for better image result but then my inference is very bad and random, even though on test data my inference was very accurate.
   Open CV result

Using alternate way

Inference result

My notebook is also attached, please ignore all the extra code that **import operator block of yours
Quick Draw.zip

**